1
00:00:00,000 --> 00:00:03,803
Audrow Nash: Welcome to the
Audrow Nash Podcast. I'm your

2
00:00:03,803 --> 00:00:07,805
host, Audrow Nash, and today
we're diving into a fascinating

3
00:00:07,805 --> 00:00:11,541
world of AI powered food
automation with Rajat Bhageria,

4
00:00:11,541 --> 00:00:15,344
the founder and CEO of Chef
Robotics. In this episode, we

5
00:00:15,344 --> 00:00:18,880
explore how Chef Robotics is
revolutionizing the food

6
00:00:18,880 --> 00:00:22,149
industry by developing AI
enabled robots for Food

7
00:00:22,149 --> 00:00:26,085
Assembly. Rajat shares insights
on why the food industry is

8
00:00:26,085 --> 00:00:30,154
arguably the best market for AI
today, discussing the massive

9
00:00:30,154 --> 00:00:33,757
labor shortages, challenging
work environment, and the

10
00:00:33,757 --> 00:00:37,826
potential of AI to transform
food production at scale. You'll

11
00:00:37,826 --> 00:00:41,095
learn about Chef Robotics
journey from concept to

12
00:00:41,095 --> 00:00:44,831
production, including their
strategic approach to market

13
00:00:44,831 --> 00:00:48,767
entry, data collection, and
scaling their technology. Rajat

14
00:00:48,767 --> 00:00:52,236
offers valuable lessons for
aspiring roboticists and

15
00:00:52,236 --> 00:00:56,172
investors covering topics like
the importance of real world

16
00:00:56,172 --> 00:01:00,242
training data in robotics, how
to bootstrap a robotic startup

17
00:01:00,242 --> 00:01:04,245
and secure customer commitments,
the challenges of deploying

18
00:01:04,245 --> 00:01:07,714
robots in production
environments, and the future of

19
00:01:07,714 --> 00:01:10,849
food automation and Its
potential impact on the

20
00:01:10,849 --> 00:01:14,452
industry. Whether you're a
roboticist, an investor, or

21
00:01:14,452 --> 00:01:18,721
simply curious about the future
of food production, this episode

22
00:01:18,721 --> 00:01:22,591
offers a wealth of insights into
one of the most promising

23
00:01:22,591 --> 00:01:26,660
applications of AI and robotics
today, So grab a snack, maybe

24
00:01:26,660 --> 00:01:30,796
one assembled by a robot in the
near future, and join us for a

25
00:01:30,796 --> 00:01:34,666
fascinating exploration of AI,
robotics, and the future of

26
00:01:34,666 --> 00:01:36,000
food. Let's dive in.

27
00:01:36,000 --> 00:01:38,000
Audrow Nash: Would you introduce
yourself?

28
00:01:38,000 --> 00:01:40,824
Rajat Bhageria: Hey Audrow!
Yeah, I'm Rajat. I'm the founder

29
00:01:40,824 --> 00:01:42,000
and CEO of Chef Robotics.

30
00:01:42,000 --> 00:01:45,522
Audrow Nash: Awesome. And would
you tell me some of the details

31
00:01:45,522 --> 00:01:48,764
about Chef Robotics? Like how
many employees you are, what

32
00:01:48,764 --> 00:01:51,000
stage of funding, these kinds of
things?

33
00:01:51,000 --> 00:01:54,379
Rajat Bhageria: Yeah, Chef
Robotics, we're like 32, 33

34
00:01:54,379 --> 00:01:57,884
employees. We're based out of
San Francisco, California.

35
00:01:57,884 --> 00:02:01,826
Funding wise, I think, we're at
22 and a half million in equity

36
00:02:01,826 --> 00:02:05,643
and debt. I think 18 or so of
that is equity. And the rest we

37
00:02:05,643 --> 00:02:09,586
use for equipment financing, to
fund the robotics as a service.

38
00:02:09,586 --> 00:02:12,840
we have robots actually now
deployed, all over North

39
00:02:12,840 --> 00:02:16,282
America, including, Canada.
We're actually now actively

40
00:02:16,282 --> 00:02:20,036
thinking about Europe as the
next kind of market. Yeah. And,

41
00:02:20,036 --> 00:02:23,791
things are going relatively
well. So we've now actually made

42
00:02:23,791 --> 00:02:27,233
around 17. 7, I believe is the
latest, million meals in

43
00:02:27,233 --> 00:02:31,175
production, which is awesome. So
that's broadly what we do. And

44
00:02:31,175 --> 00:02:34,930
of course we, make, AI enabled
robots for the food industry.

45
00:02:34,930 --> 00:02:38,685
And of course there's a lot more
depth there, but broadly AI

46
00:02:38,685 --> 00:02:41,000
enabled robots for the food
industry.

47
00:02:41,000 --> 00:02:44,746
Audrow Nash: Hell yeah. Now,
when we talked in a quick call

48
00:02:44,746 --> 00:02:48,683
before this interview, you gave
me a, an overview how, did you

49
00:02:48,683 --> 00:02:52,429
pick this market and then how
you started Chef Robotics and

50
00:02:52,429 --> 00:02:56,175
how I don't know, the whole
problem space that you're going

51
00:02:56,175 --> 00:03:00,111
into. Would you tell me some of
that? Start at the motivation.

52
00:03:00,111 --> 00:03:01,000
Or the market.

53
00:03:01,000 --> 00:03:03,000
Rajat Bhageria: Yeah, sure.

54
00:03:03,000 --> 00:03:06,462
Rajat Bhageria: in, in college
actually, so I went to Penn for

55
00:03:06,462 --> 00:03:09,812
undergrad and grad school, I was
working on a company called

56
00:03:09,812 --> 00:03:12,882
Third Eye. We were using
computer vision to essentially

57
00:03:12,882 --> 00:03:16,400
recognize what's in front of you
for the visually impaired. And

58
00:03:16,400 --> 00:03:19,806
worked on this company for three
and a half years, and we can

59
00:03:19,806 --> 00:03:23,379
talk about Third Eye as well if
you're curious, but I think this

60
00:03:23,379 --> 00:03:26,785
company really introduced me to
the power of computer vision,

61
00:03:26,785 --> 00:03:30,246
more than anything else. before
then, I had done a lot of, web

62
00:03:30,246 --> 00:03:33,150
programming and, desktop
programs, but for the first

63
00:03:33,150 --> 00:03:36,276
time, a computer could make
sense of the physical world,

64
00:03:36,276 --> 00:03:39,403
which was just insane to me,
right? And it really got me

65
00:03:39,403 --> 00:03:42,809
thinking about, the power of AI
and, computer vision. And, my

66
00:03:42,809 --> 00:03:46,047
next kind of question is, okay,
Third Eye was completely a

67
00:03:46,047 --> 00:03:49,397
software, right? Product, right?
we had, we were using smart

68
00:03:49,397 --> 00:03:52,914
glasses, but of course those are
off the shelf or mobile phones

69
00:03:52,914 --> 00:03:56,376
also off the shelf. So it was
purely a software product. And I

70
00:03:56,376 --> 00:03:59,782
was like, okay, is the biggest
application of ai? Is it gonna

71
00:03:59,782 --> 00:04:03,299
be purely in the digital world
or is it really gonna be. In the

72
00:04:03,299 --> 00:04:06,593
physical world. I, was just
curious. And, I just started to

73
00:04:06,593 --> 00:04:09,999
do some brainstorming. It's and
it became abundantly obvious,

74
00:04:09,999 --> 00:04:13,516
like the, like 90 plus percent
of global GDP is in the physical

75
00:04:13,516 --> 00:04:16,420
world, right? It's like things
like construction and

76
00:04:16,420 --> 00:04:19,602
agriculture and food and
beverage and healthcare and, all

77
00:04:19,602 --> 00:04:22,785
these other industries, right?
these are really where the

78
00:04:22,785 --> 00:04:26,135
majority of the wealth of the
world is. So I was like, okay,

79
00:04:26,135 --> 00:04:29,205
of course, then if you're
interacting with the physical

80
00:04:29,205 --> 00:04:32,723
world, then, you can take sensor
data in using a camera, but of

81
00:04:32,723 --> 00:04:36,129
course you also probably need to
actuate. So I think I became

82
00:04:36,129 --> 00:04:39,311
pretty excited about the
prospect of kind of combining AI

83
00:04:39,311 --> 00:04:42,549
plus robotics to have embodied
AI, right? And this kind of

84
00:04:42,549 --> 00:04:46,011
became like extremely, exciting
for me because, I, I feel like

85
00:04:46,011 --> 00:04:49,640
I'm very lucky to enjoy my work.
I really enjoy my work, but I do

86
00:04:49,640 --> 00:04:52,767
know there's hundreds of
millions of humans, billions of

87
00:04:52,767 --> 00:04:56,061
humans who. Don't have that
luxury and it seems wouldn't it

88
00:04:56,061 --> 00:04:59,522
be nice if we lived in a world
where there was abundance where

89
00:04:59,522 --> 00:05:02,705
Robots can do some of the dull
dangers and dirty jobs and

90
00:05:02,705 --> 00:05:06,222
robots can do And humans can do
something hopefully more humane

91
00:05:06,222 --> 00:05:09,628
and better. So that vision was
exciting and you know I wanted

92
00:05:09,628 --> 00:05:12,978
to really focus on embodied AI
And, that's decided to get my

93
00:05:12,978 --> 00:05:16,384
master's in that and focus in on
that. but I didn't know what

94
00:05:16,384 --> 00:05:19,790
product to build yet or what
industry yet to focus on, right?

95
00:05:19,790 --> 00:05:23,363
and I took a bit of a detour to
figure that out. a friend and I,

96
00:05:23,363 --> 00:05:26,657
started a small VC fund called
Prototype Capital. basically

97
00:05:26,657 --> 00:05:29,951
investing in founders who are
using new technologies in old

98
00:05:29,951 --> 00:05:33,078
industries. Old industries like
oil and gas and food and

99
00:05:33,078 --> 00:05:36,539
beverage and all the industries
I just alluded to, right? and,

100
00:05:36,539 --> 00:05:40,001
again, we can also talk about
the prototype experience, but at

101
00:05:40,001 --> 00:05:43,407
least what's relevant for the,
for Chef is that, that allowed

102
00:05:43,407 --> 00:05:46,980
me some time to just understand
the industry, the physical world

103
00:05:46,980 --> 00:05:50,386
a little bit more, talk to a lot
of people in these different

104
00:05:50,386 --> 00:05:51,000
industries.

105
00:05:51,000 --> 00:05:54,336
Rajat Bhageria: What, I, learned
is actually that, the food

106
00:05:54,336 --> 00:05:57,784
industry is probably one of the
best places to apply AI, just

107
00:05:57,784 --> 00:06:01,177
from a market size perspective.
Period. And specifically the

108
00:06:01,177 --> 00:06:04,682
reason I say this is that, if
you look at the Bureau for Labor

109
00:06:04,682 --> 00:06:08,074
Statistics, the, you know, the
domestic population in the U.

110
00:06:08,074 --> 00:06:11,353
S., the most common job are
things like the number one and

111
00:06:11,353 --> 00:06:14,859
number two most common jobs are
retail salespeople and nursing

112
00:06:14,859 --> 00:06:18,364
aides and personal aides. those
two jobs I don't think will be

113
00:06:18,364 --> 00:06:21,700
automated by AI robots anytime
soon. It's a very human job.

114
00:06:21,700 --> 00:06:25,148
okay. It was like, okay, what's
number three? Number three is

115
00:06:25,148 --> 00:06:28,145
food service and food
preparation. So that felt a lot

116
00:06:28,145 --> 00:06:31,594
more tractable to me, right? And
by the way, I'm not from the

117
00:06:31,594 --> 00:06:34,760
food industry, right? So I
didn't totally understand it,

118
00:06:34,760 --> 00:06:38,152
but I can at least, I go to
Chipotle and Sweetgreen and I've

119
00:06:38,152 --> 00:06:41,770
seen those people assemble those
meals. It seems like you take a

120
00:06:41,770 --> 00:06:44,993
camera and a manipulator and
some different utensils and,

121
00:06:44,993 --> 00:06:48,442
make a robot that does that too.
So it just felt a little bit

122
00:06:48,442 --> 00:06:51,777
more tractable to me. that idea
of, okay, it seems like the

123
00:06:51,777 --> 00:06:55,283
market size is pretty big, was
exciting. And specifically what

124
00:06:55,283 --> 00:06:58,505
specifically, there's an
interesting insight, which is if

125
00:06:58,505 --> 00:07:01,898
you, agree that the biggest
impact for AI is in the physical

126
00:07:01,898 --> 00:07:05,177
world, because 90 percent of
global GDP is in the physical

127
00:07:05,177 --> 00:07:08,569
world, then it's okay, what part
of the physical world do we

128
00:07:08,569 --> 00:07:11,905
focus on? And then we said,
okay, what's the biggest market

129
00:07:11,905 --> 00:07:15,466
in the physical world? And it's
the labor market, which is like

130
00:07:15,466 --> 00:07:18,972
45 trillion of GDP. And then,
okay, what's the biggest part of

131
00:07:18,972 --> 00:07:22,251
the labor market? it's retail
salespeople and nursing, but

132
00:07:22,251 --> 00:07:25,473
that's not tractable. the
biggest tractable market for AI

133
00:07:25,473 --> 00:07:28,866
today, From this, analysis I
did, at least, was food service

134
00:07:28,866 --> 00:07:31,975
and food production, food
preparation. it's okay, there

135
00:07:31,975 --> 00:07:35,367
seems just from a market size
perspective, this is something

136
00:07:35,367 --> 00:07:38,364
that I want to, really
investigate and look into. But

137
00:07:38,364 --> 00:07:41,587
of course, like I said, I didn't
know much about the food

138
00:07:41,587 --> 00:07:45,035
industry, so the next step was
just talking to as many people

139
00:07:45,035 --> 00:07:48,258
as I could possibly talk to.
And, I think this was a very

140
00:07:48,258 --> 00:07:51,481
broad search, right? This was
everything from fast casual

141
00:07:51,481 --> 00:07:54,929
operators to ghost kitchens. at
the time, Travis Kalanick had

142
00:07:54,929 --> 00:07:58,491
just raised a bunch of money for
cloud kitchens. and, companies

143
00:07:58,491 --> 00:08:01,827
like Reef and Kitchen United
were blossoming. This is still

144
00:08:01,827 --> 00:08:05,445
pre COVID by the way. we talked
to like food truck operators, we

145
00:08:05,445 --> 00:08:08,611
talked to airline catering
companies, food manufacturing

146
00:08:08,611 --> 00:08:11,834
companies, what's the biggest
pain point you face, right?

147
00:08:11,834 --> 00:08:15,000
What's going on, right? And what
I, learned is that the.

148
00:08:15,000 --> 00:08:18,331
Rajat Bhageria: The biggest
thing that everyone said over

149
00:08:18,331 --> 00:08:22,130
and over, without exception, was
there's a really, crushing labor

150
00:08:22,130 --> 00:08:25,344
shortage. And, of course in
robotics you hear this very

151
00:08:25,344 --> 00:08:28,675
frequently. It's not a new piece
of information. What was

152
00:08:28,675 --> 00:08:31,831
interesting with the food
industry, they said, is that

153
00:08:31,831 --> 00:08:35,513
oftentimes food environments are
extremely cold, 34, 35 degrees

154
00:08:35,513 --> 00:08:39,078
Fahrenheit, or extremely hot,
right? 80 plus, 90 plus degrees

155
00:08:39,078 --> 00:08:42,526
Fahrenheit. hiring is actually
Oftentimes harder than other

156
00:08:42,526 --> 00:08:46,149
industries like warehousing and
logistics and supply chain. So

157
00:08:46,149 --> 00:08:49,539
it seems like there was a very
pressing need where AI plus

158
00:08:49,539 --> 00:08:51,000
robots could really help.

159
00:08:51,000 --> 00:08:54,429
Audrow Nash: And you're saying
that reason that, or one of the

160
00:08:54,429 --> 00:08:57,416
reasons that the hiring is
difficult is because of the

161
00:08:57,416 --> 00:09:00,458
environmental conditions that
it's hot or cold. So it's

162
00:09:00,458 --> 00:09:03,611
unpleasant to work there and
thus people don't work there

163
00:09:03,611 --> 00:09:06,708
often. And I saw like the
turnover is super high for the

164
00:09:06,708 --> 00:09:09,805
people that are in it. So you
think it's it's miserable,

165
00:09:09,805 --> 00:09:13,235
possibly partially because of
the environmental conditions and

166
00:09:13,235 --> 00:09:16,000
that just exacerbates these
labor shortages. Okay.

167
00:09:16,000 --> 00:09:19,027
Rajat Bhageria: Yes, and not
only that, you have these

168
00:09:19,027 --> 00:09:22,221
redundant motions, right? you're
like, you're often scoop

169
00:09:22,221 --> 00:09:25,808
chopping and scooping and you're
doing a lot of redundant motion

170
00:09:25,808 --> 00:09:29,339
So that like back injuries and
there's just a lot it's a lot of

171
00:09:29,339 --> 00:09:32,646
physical exertion in a very
uncomfortable environment Which

172
00:09:32,646 --> 00:09:35,000
comes together to not be the
greatest job,

173
00:09:35,000 --> 00:09:36,000
Audrow Nash: for sure.

174
00:09:36,000 --> 00:09:39,347
Rajat Bhageria: So You know, I
think this obviously has a lot

175
00:09:39,347 --> 00:09:42,419
of implications. for the
business itself, I learned, and

176
00:09:42,419 --> 00:09:45,602
this is again true from the big
kind of food manufacturing

177
00:09:45,602 --> 00:09:48,235
companies to the small food
trucks, the business

178
00:09:48,235 --> 00:09:51,418
implications are also severe,
which is that they often are

179
00:09:51,418 --> 00:09:54,600
leaving revenue on the table.
Specifically, what I mean by

180
00:09:54,600 --> 00:09:58,057
this is if they could have made,
x meals or volume, now because

181
00:09:58,057 --> 00:10:01,458
they don't have enough labor,
they can only make 70 percent of

182
00:10:01,458 --> 00:10:05,025
x. So they're leaving revenue on
the table. their supply of labor

183
00:10:05,025 --> 00:10:07,000
is not able to meet demand of
labor.

184
00:10:07,000 --> 00:10:10,403
Audrow Nash: That's wild. Okay.
Yes, so, they literally are just

185
00:10:10,403 --> 00:10:13,592
not making as much money as they
could if the production was

186
00:10:13,592 --> 00:10:16,782
higher for this kind of thing.
And that probably also drives

187
00:10:16,782 --> 00:10:19,813
inflation in their costs and
things like this, because it

188
00:10:19,813 --> 00:10:23,215
costs more to get out less, and
then they're making less profit,

189
00:10:23,215 --> 00:10:26,352
and then they have to charge
more for what they do make. So

190
00:10:26,352 --> 00:10:28,000
that's a vicious cycle. I think

191
00:10:28,000 --> 00:10:31,039
Rajat Bhageria: yes, and there's
another vicious cycle in there

192
00:10:31,039 --> 00:10:34,079
as well, which is that because
you don't have enough labor, you

193
00:10:34,079 --> 00:10:36,829
have to ask your current folks
to work overtime, which of

194
00:10:36,829 --> 00:10:39,000
course, it's not because there's
turnover, so

195
00:10:39,000 --> 00:10:42,784
Audrow Nash: or but even
increases it further because of

196
00:10:42,784 --> 00:10:44,000
this they burn out

197
00:10:44,000 --> 00:10:47,147
Rajat Bhageria: yes. They burn
out. They burn out. So it's a

198
00:10:47,147 --> 00:10:50,031
pretty rough situation. And by
the way, that's from the

199
00:10:50,031 --> 00:10:53,336
business side. There's another
like more like socioeconomic and

200
00:10:53,336 --> 00:10:56,587
nationalistic reason. Like some
of these companies, especially

201
00:10:56,587 --> 00:10:59,262
the bigger ones, the
manufacturing ones, right? The

202
00:10:59,262 --> 00:11:02,252
food manufacturing companies,
they were actively at least

203
00:11:02,252 --> 00:11:05,608
considering, hey, how do we how
do we make this work? Obviously,

204
00:11:05,608 --> 00:11:08,808
robotics is on the table, but
also Can we take at least parts

205
00:11:08,808 --> 00:11:12,007
of our supply chain? Maybe not
everything, because it's food,

206
00:11:12,007 --> 00:11:15,311
there's a freshness factor, but
can we take parts of our supply

207
00:11:15,311 --> 00:11:18,510
chain offshore? So these are,
both these things combined, and

208
00:11:18,510 --> 00:11:21,395
it felt big market, like
arguably number one market for

209
00:11:21,395 --> 00:11:24,594
AI, market size, from the market
size perspective. Seems like

210
00:11:24,594 --> 00:11:27,741
there's a hair on fire problem,
right? And okay, I was like,

211
00:11:27,741 --> 00:11:29,000
let's dig deeper, right?

212
00:11:29,000 --> 00:11:31,440
Audrow Nash: do you mean hair on
fire like it's a big problem

213
00:11:31,440 --> 00:11:33,000
right now is that what you mean
by that

214
00:11:33,000 --> 00:11:36,970
Rajat Bhageria: What I mean by
that is I think like in startup

215
00:11:36,970 --> 00:11:40,877
world, one of the lessons, at
least I've learned is that, you

216
00:11:40,877 --> 00:11:45,039
can have products that are like,
the classic expression that many

217
00:11:45,039 --> 00:11:46,000
VCs use is like

218
00:11:46,000 --> 00:11:49,000
Audrow Nash: pain or vitamin
yeah

219
00:11:49,000 --> 00:11:51,992
Rajat Bhageria: So a vitamin,
like you can sell it, but it's

220
00:11:51,992 --> 00:11:54,984
it's never priority number one.
It's let's take care of that

221
00:11:54,984 --> 00:11:58,076
next quarter. It's sales cycles
are long and It doesn't really

222
00:11:58,076 --> 00:12:00,669
have that, exponential growth,
whereas if you have a

223
00:12:00,669 --> 00:12:03,612
painkiller, it's like sales
cycles are quick, it's like the

224
00:12:03,612 --> 00:12:06,753
Warren Buffett quote, I want a
business that even a dumb person

225
00:12:06,753 --> 00:12:08,000
can run, right? It's that

226
00:12:08,000 --> 00:12:09,000
Audrow Nash: yeah

227
00:12:09,000 --> 00:12:10,000
Rajat Bhageria: right?

228
00:12:09,000 --> 00:12:12,000
Audrow Nash: it's a great quote
okay

229
00:12:12,000 --> 00:12:13,000
Rajat Bhageria: yes.

230
00:12:12,000 --> 00:12:14,000
Audrow Nash: all right so hair
on fire

231
00:12:14,000 --> 00:12:17,324
Rajat Bhageria: It's a hair and
fire problem, and they really

232
00:12:17,324 --> 00:12:20,321
desperately need a solution.
it's, another way, another

233
00:12:20,321 --> 00:12:23,699
analogy that's coming to my mind
right now is when, SpaceX was

234
00:12:23,699 --> 00:12:27,131
founded, there wasn't really a
great launch solution, right? Of

235
00:12:27,131 --> 00:12:30,183
course you had governments,
they're expensive, extremely

236
00:12:30,183 --> 00:12:31,000
expensive, very

237
00:12:31,000 --> 00:12:33,000
Audrow Nash: oh yeah

238
00:12:33,000 --> 00:12:35,637
Rajat Bhageria: And SpaceX was
able to actually come in and

239
00:12:35,637 --> 00:12:38,453
people were willing to work with
them even though they actually

240
00:12:38,453 --> 00:12:40,955
hadn't launched any successful
rockets in the early days

241
00:12:40,955 --> 00:12:43,547
because they were the only game
in town, right? So because

242
00:12:43,547 --> 00:12:46,363
there's a hair on fire problem
for people to take satellites to

243
00:12:46,363 --> 00:12:49,134
space, there wasn't really an
alternative. People were willing

244
00:12:49,134 --> 00:12:51,592
to like, work with a small,
scrappy startup and make it

245
00:12:51,592 --> 00:12:53,961
work, right? If that makes
sense. So there's a lot of

246
00:12:53,961 --> 00:12:56,732
reasons why that kind of problem
is a good problem to solve, I

247
00:12:56,732 --> 00:12:57,000
think.

248
00:12:57,000 --> 00:13:00,000
Audrow Nash: gotcha okay

249
00:13:00,000 --> 00:13:03,235
Rajat Bhageria: the next
question is, okay, great, it

250
00:13:03,235 --> 00:13:07,019
seems like market's big, there's
a big problem that we can, AI

251
00:13:07,019 --> 00:13:10,742
and robots can help with. The
next kind of question I had is,

252
00:13:10,742 --> 00:13:13,000
okay, what kind of task do I
work on?

253
00:13:13,000 --> 00:13:15,377
Rajat Bhageria: and again, I'm
not from the food universe, so

254
00:13:15,377 --> 00:13:17,909
what I did is just interviewed a
lot of people and what I learned

255
00:13:17,909 --> 00:13:20,130
is that broad strokes, and
everything except fine dining,

256
00:13:20,130 --> 00:13:22,000
because again, we're not trying
to focus on fine

257
00:13:22,000 --> 00:13:27,000
Audrow Nash: yeah

258
00:13:27,000 --> 00:13:30,453
Rajat Bhageria: the world of a
kitchen into three broad strokes

259
00:13:30,453 --> 00:13:33,852
tasks. You have to prep food,
which is cutting things, washing

260
00:13:33,852 --> 00:13:37,251
things, getting the ingredients
ready. Then you have to do the

261
00:13:37,251 --> 00:13:40,649
cooking itself. and then finally
you have to do the plating or

262
00:13:40,649 --> 00:13:43,664
the assembly, just to put the
plating or assembly into,

263
00:13:43,664 --> 00:13:47,172
context and help, help, help you
visualize it. Essentially, it's

264
00:13:47,172 --> 00:13:50,681
what you would see at Chipotle,
right? So the person who's kinda

265
00:13:50,681 --> 00:13:54,079
scooping the chicken, the rice
into your burrito or your. Your

266
00:13:54,079 --> 00:13:57,094
bowl. but, really the, thinking
there is you have to Be

267
00:13:57,094 --> 00:13:59,670
flexible, do lots of
ingredients, deal with the

268
00:13:59,670 --> 00:14:02,904
different portion sizes, place
in different compartments or

269
00:14:02,904 --> 00:14:06,193
parts of the tray or burrito or
the wrap or sandwich, things

270
00:14:06,193 --> 00:14:09,482
like that, right? Not spill it,
not damage ingredients, etc.

271
00:14:09,482 --> 00:14:12,771
What was really interesting for
me is that my impression was

272
00:14:12,771 --> 00:14:16,005
that cooking is going to be the
most expensive part of this

273
00:14:16,005 --> 00:14:18,801
process from a labor
perspective. What I learned is

274
00:14:18,801 --> 00:14:22,200
that actually that's not true.
the most expensive part of this

275
00:14:22,200 --> 00:14:25,653
process is food assembly, which
is interesting. And, the reason

276
00:14:25,653 --> 00:14:29,052
it was interesting, is that what
I learned is that from a prep

277
00:14:29,052 --> 00:14:32,450
perspective, there's actually a
lot of traditional automation,

278
00:14:32,450 --> 00:14:35,904
essentially food processors, is
the way you can think about it.

279
00:14:35,904 --> 00:14:37,000
Now, there might be,

280
00:14:37,000 --> 00:14:39,000
Audrow Nash: industrial scale

281
00:14:39,000 --> 00:14:42,053
Rajat Bhageria: yes, no,
exactly, but, even like food

282
00:14:42,053 --> 00:14:45,452
trucks use it, even like fast
casuals use it, and of course

283
00:14:45,452 --> 00:14:48,793
food manufacturers use it.
different sizes, of course, but

284
00:14:48,793 --> 00:14:52,018
yeah, I think you get the
analogy. food prep is, it does

285
00:14:52,018 --> 00:14:55,129
require labor, but it's not
necessarily something that

286
00:14:55,129 --> 00:14:58,528
scales linearly. In other words,
if you need more volume of

287
00:14:58,528 --> 00:15:02,099
material, you don't need more
human beings, right? Because you

288
00:15:02,099 --> 00:15:04,000
have this traditional
automation,

289
00:15:04,000 --> 00:15:05,000
Audrow Nash: it's very

290
00:15:04,000 --> 00:15:06,696
Rajat Bhageria: scales more sub
linearly. It's very efficient.

291
00:15:06,696 --> 00:15:08,000
It scales sub linearly, right?

292
00:15:08,000 --> 00:15:10,678
Audrow Nash: It's one person
driving a huge tractor or

293
00:15:10,678 --> 00:15:13,752
something like that, as opposed
to a bunch of little tractors.

294
00:15:13,752 --> 00:15:14,000
Okay.

295
00:15:14,000 --> 00:15:17,731
Rajat Bhageria: That's right.
That's right. That same analogy

296
00:15:17,731 --> 00:15:21,523
is true in cooking, but for a
different reason. So in cooking,

297
00:15:21,523 --> 00:15:25,437
you can cook in batches, right?
one person can cook a soup for a

298
00:15:25,437 --> 00:15:29,352
thousand people. One person can
be at the grill flipping burgers

299
00:15:29,352 --> 00:15:32,532
for 60 people, right? again,
same analogy. It scales

300
00:15:32,532 --> 00:15:36,385
relatively well. It scales sub
linearly, right? you can imagine

301
00:15:36,385 --> 00:15:39,994
that if you need to make, if
you're like, a Chipotle during

302
00:15:39,994 --> 00:15:43,725
dinner rush in Manhattan, you
can have a few cooks, like two,

303
00:15:43,725 --> 00:15:47,150
three. They can cook for
everyone because you can do big

304
00:15:47,150 --> 00:15:50,820
dishes. Portion sizes, very big
batches. So that leads us to

305
00:15:50,820 --> 00:15:54,306
assembly. Assembly is the only
part of this equation that

306
00:15:54,306 --> 00:15:58,159
actually scales more linearly.
In other words, if you need more

307
00:15:58,159 --> 00:16:01,523
output or you need more
throughput, you need more human

308
00:16:01,523 --> 00:16:05,193
beings. Because it's an assembly
line, right? Each person is

309
00:16:05,193 --> 00:16:08,740
like, if you need, if you
literally need more bowls coming

310
00:16:08,740 --> 00:16:12,532
out, you need more humans making
those bowls in parallel. Does

311
00:16:12,532 --> 00:16:14,000
that kind of make sense?

312
00:16:14,000 --> 00:16:17,303
Audrow Nash: Yeah. 'cause
you're, basically, I'm thinking

313
00:16:17,303 --> 00:16:20,896
of, assembly is the good word
for it, but if you had, I dunno,

314
00:16:20,896 --> 00:16:24,257
you're just, you're adding a
single and each person may be

315
00:16:24,257 --> 00:16:27,791
adding one ingredient in one
spot and it's that's pretty, you

316
00:16:27,791 --> 00:16:31,326
have to do that manually. You
can't do it. There's not like a

317
00:16:31,326 --> 00:16:34,861
big scooper with a hundred ends
that you can just, especially

318
00:16:34,861 --> 00:16:38,280
for custom dishes, especially
for like restaurants, I would

319
00:16:38,280 --> 00:16:41,525
assume at like end point
locations, like right where the

320
00:16:41,525 --> 00:16:45,002
customer is eating, but it's
probably true elsewhere in like

321
00:16:45,002 --> 00:16:48,652
those grocery store, buy a meal
in a bowl kits and this kind of

322
00:16:48,652 --> 00:16:49,000
thing.

323
00:16:49,000 --> 00:16:50,000
Rajat Bhageria: that's right.

324
00:16:50,000 --> 00:16:55,526
Audrow Nash: linear. Almost.
It's, probably basically linear

325
00:16:55,526 --> 00:16:57,000
in terms of huh.

326
00:16:57,000 --> 00:17:00,575
Rajat Bhageria: basically what I
learned is that like 60 percent

327
00:17:00,575 --> 00:17:03,983
of the labor in a commercial or
industrial kitchen that makes

328
00:17:03,983 --> 00:17:07,000
meals is in food assembly.
Otherwise known as plating,

329
00:17:07,000 --> 00:17:09,835
Audrow Nash: bonkers. I would
not have thought, but it makes

330
00:17:09,835 --> 00:17:12,669
sense now that you're explaining
it, but I really would have

331
00:17:12,669 --> 00:17:15,409
thought prep would be larger,
but I see what you're saying

332
00:17:15,409 --> 00:17:18,244
about you get the efficiencies
of it. And same with cooking,

333
00:17:18,244 --> 00:17:21,173
you can batch, whereas assembly
you can't batch, and therefore

334
00:17:21,173 --> 00:17:24,055
when you're doing scale, that's
the one. Because it's linear,

335
00:17:24,055 --> 00:17:25,000
it's such a problem.

336
00:17:25,000 --> 00:17:28,887
Rajat Bhageria: that's right, it
became pretty clear that if we

337
00:17:28,887 --> 00:17:32,465
were to help with this labor
shortage, if we were to solve

338
00:17:32,465 --> 00:17:36,105
one problem, we should really
focus on food assembly. Right

339
00:17:36,105 --> 00:17:40,115
now it was like, okay, great. we
have a big market, we have, it's

340
00:17:40,115 --> 00:17:43,878
a big pain point. We think that
the biggest thing we can help

341
00:17:43,878 --> 00:17:47,703
across the market, not just in
like fast casuals, but not just

342
00:17:47,703 --> 00:17:50,788
in food trucks, but really
across the market, food

343
00:17:50,788 --> 00:17:53,996
manufacturing, it's food
assembly. Then it was like,

344
00:17:53,996 --> 00:17:57,821
okay, like who are going to be
our customers, right? Who do we

345
00:17:57,821 --> 00:18:01,337
sell to? And of course, the
first question is, this a B2C

346
00:18:01,337 --> 00:18:05,162
company or is it a B2B company?
we considered B2C actually, at

347
00:18:05,162 --> 00:18:08,000
the very, beginning of Chef, I
was like, okay,

348
00:18:08,000 --> 00:18:11,012
Audrow Nash: Would that mean
like I buy a product for my home

349
00:18:11,012 --> 00:18:12,000
kind of thing? Yeah,

350
00:18:12,000 --> 00:18:15,888
Rajat Bhageria: So I thought
about, both you buy a product

351
00:18:15,888 --> 00:18:19,911
for your home and also you buy
the end sandwich or burger or

352
00:18:19,911 --> 00:18:24,000
coffee, right? So the product
for your home idea, I didn't, I

353
00:18:24,000 --> 00:18:27,000
Audrow Nash: That one's hard and
expensive. Yeah.

354
00:18:27,000 --> 00:18:30,600
Rajat Bhageria: One's hard and
expensive and actually like I, I

355
00:18:30,600 --> 00:18:33,971
actually did a bit of a thought
experiment with my friends.

356
00:18:33,971 --> 00:18:37,343
basically I posted them the
question, okay, look, let's say

357
00:18:37,343 --> 00:18:40,829
you have two potential options.
Option one is This like black

358
00:18:40,829 --> 00:18:44,486
box in your house and it's gonna
make anything you want. I don't

359
00:18:44,486 --> 00:18:48,029
know how it's gonna work yet,
but just assume it exists. would

360
00:18:48,029 --> 00:18:51,571
you rather have that or let's
say option B is imagine a future

361
00:18:51,571 --> 00:18:54,771
where you know, we have the
proliferation of robots that

362
00:18:54,771 --> 00:18:58,086
actually make the food, which
means your food is like And,

363
00:18:58,086 --> 00:19:01,514
customizable than, essentially
as customizable, as you would

364
00:19:01,514 --> 00:19:04,943
have in your house. it's fast.
It's probably cheaper, right?

365
00:19:04,943 --> 00:19:08,429
imagine that, these robots are
in ghost kitchens. And so what

366
00:19:08,429 --> 00:19:09,000
that means

367
00:19:09,000 --> 00:19:11,826
Audrow Nash: they just output an
order and it's delivered to you.

368
00:19:11,826 --> 00:19:13,000
They handle stocking. Okay,

369
00:19:13,000 --> 00:19:16,000
Rajat Bhageria: And it's
probably cheaper as well. And

370
00:19:16,000 --> 00:19:18,000
Audrow Nash: Scale. Okay. That
sounds great.

371
00:19:18,000 --> 00:19:20,240
Rajat Bhageria: of the scale,
and then finally it's also

372
00:19:20,240 --> 00:19:22,000
delivered to you by a robot. So
it's cheaper

373
00:19:22,000 --> 00:19:23,000
Audrow Nash: Ha Sounds hard.

374
00:19:23,000 --> 00:19:26,311
Rajat Bhageria: but what was
interesting is like you can

375
00:19:26,311 --> 00:19:29,681
visualize that future, right?
And what was interesting is

376
00:19:29,681 --> 00:19:33,229
every single person said option
B, which is to say like they

377
00:19:33,229 --> 00:19:36,776
would rather, and what I asked
about why, and they said look

378
00:19:36,776 --> 00:19:40,383
like it's because first of all I
don't have to take care of a

379
00:19:40,383 --> 00:19:43,990
robot, I don't have to refill
it, I don't have to clean it. I

380
00:19:43,990 --> 00:19:47,419
don't have to pay for it. I
would rather rent the food, if

381
00:19:47,419 --> 00:19:50,730
you will, than buy the food
equipment. and, plus, I, get

382
00:19:50,730 --> 00:19:54,455
access to my favorite chefs. I,
like my, the brands, I like, I,

383
00:19:54,455 --> 00:19:58,003
don't wanna have to give up
those brands. And, anyways, like

384
00:19:58,003 --> 00:20:01,550
this thought experiment was a
very quick thought experiment,

385
00:20:01,550 --> 00:20:05,334
by the way. But it, it basically
put a big dent in my head about

386
00:20:05,334 --> 00:20:08,468
at home robots. I became more
and more convinced that

387
00:20:08,468 --> 00:20:09,000
actually,

388
00:20:09,000 --> 00:20:12,306
Audrow Nash: think that
generalizes to, other, I, I

389
00:20:12,306 --> 00:20:16,000
suspect how you say it, that it
probably does generalize.

390
00:20:16,000 --> 00:20:18,431
Rajat Bhageria: I think it does.
a good example of this is

391
00:20:18,431 --> 00:20:20,862
clothing, by the way. back in
the day, people used to make

392
00:20:20,862 --> 00:20:23,000
textiles at their house. A
cottage industry, right?

393
00:20:23,000 --> 00:20:24,000
Audrow Nash: Yeah.

394
00:20:24,000 --> 00:20:27,472
Rajat Bhageria: but then over
time, of course, we learned that

395
00:20:27,472 --> 00:20:30,776
there's efficiencies of scale,
right? it's better to have a

396
00:20:30,776 --> 00:20:34,024
central factory near the water
where you can make all your

397
00:20:34,024 --> 00:20:37,216
clothing and then you can buy
the clothing. so I think, I

398
00:20:37,216 --> 00:20:40,744
think it generally scales. you
can imagine a future where a lot

399
00:20:40,744 --> 00:20:44,272
of the day to day stuff you do
at home, you ship out, just like

400
00:20:44,272 --> 00:20:47,464
a ghost kitchen's kind of a
central factory, if you will,

401
00:20:47,464 --> 00:20:50,656
you ship it out of your home.
Because you have autonomous

402
00:20:50,656 --> 00:20:53,960
delivery, right? That's the key
assumption here. Autonomous

403
00:20:53,960 --> 00:20:57,096
delivery, ship it out to a
central location, they do the

404
00:20:57,096 --> 00:21:00,568
processing and ship it back to
you. even clothing, so like at,

405
00:21:00,568 --> 00:21:03,984
in big factories and industrial
settings, like companies like

406
00:21:03,984 --> 00:21:07,120
Aramark and Cisco, like they
basically, like the Smocks,

407
00:21:07,120 --> 00:21:10,536
those are actually cleaned in a
separate facility and Aramark

408
00:21:10,536 --> 00:21:13,728
delivers them to like your
restaurant or your plant every

409
00:21:13,728 --> 00:21:16,976
day. It's, already a thing,
right? So I think that analogy

410
00:21:16,976 --> 00:21:20,000
actually could work. There's
certain things that need,

411
00:21:20,000 --> 00:21:23,069
Audrow Nash: that's, probably up
and coming. I'm thinking even

412
00:21:23,069 --> 00:21:25,000
like delivery, laundry services
kind of

413
00:21:25,000 --> 00:21:27,822
Rajat Bhageria: that's right.
That's right. I think there's

414
00:21:27,822 --> 00:21:30,501
certain things that just need to
stay in the house, like

415
00:21:30,501 --> 00:21:33,515
cleaning, cleaning, just, it's a
fundamental, which is why like

416
00:21:33,515 --> 00:21:36,289
companies like Roomba have
actually done well in robotics,

417
00:21:36,289 --> 00:21:39,399
But anyways, I, like all of that
to say that I was less convinced

418
00:21:39,399 --> 00:21:42,269
about like selling robots to
consumers. I'm still, less than

419
00:21:42,269 --> 00:21:44,948
I think robotics is more
fundamentally a B2B business. I

420
00:21:44,948 --> 00:21:46,000
just think that's what

421
00:21:46,000 --> 00:21:48,000
Audrow Nash: I agree.

422
00:21:48,000 --> 00:21:50,833
Rajat Bhageria: So unless you're
the end consumer, right? So

423
00:21:50,833 --> 00:21:53,618
unless you're Tesla and you're
building your own robots or,

424
00:21:53,618 --> 00:21:56,309
like Tesla makes their own
robots for their own facility,

425
00:21:56,309 --> 00:21:59,000
right? But they're also selling
to a factory if you will.

426
00:21:59,000 --> 00:22:00,000
Audrow Nash: Still B2B. Ah

427
00:22:00,000 --> 00:22:03,362
Rajat Bhageria: Still B2B. okay,
so no to selling robots to

428
00:22:03,362 --> 00:22:06,781
consumers. Then the other idea
is, okay, what if we actually

429
00:22:06,781 --> 00:22:10,257
made our own restaurant? Like
what, Creator and everyone else

430
00:22:10,257 --> 00:22:13,676
has done? That also, it didn't
really hold water for me. And

431
00:22:13,676 --> 00:22:17,096
there's a few reasons for this.
just on the surface, it felt

432
00:22:17,096 --> 00:22:20,287
very much and we've all read
Zero to One, right? Like it

433
00:22:20,287 --> 00:22:23,535
very, much felt The perfect
competitive market that Peter

434
00:22:23,535 --> 00:22:27,182
Thiel talked about. Like it's, a
little bit better, but it's not

435
00:22:27,182 --> 00:22:30,658
that much better. And in another
restaurant opens up down the

436
00:22:30,658 --> 00:22:31,000
street

437
00:22:31,000 --> 00:22:33,000
Audrow Nash: Yeah, it's just a

438
00:22:33,000 --> 00:22:34,000
Rajat Bhageria: out of luck.

439
00:22:34,000 --> 00:22:37,241
Audrow Nash: basically. Like
it's not, you're right, maybe

440
00:22:37,241 --> 00:22:40,091
it's two times better or
something, maybe two times

441
00:22:40,091 --> 00:22:43,500
faster or something, but it's
not like that much and you have

442
00:22:43,500 --> 00:22:47,076
to wait, probably like lines and
all the things to get to it, so

443
00:22:47,076 --> 00:22:50,150
it gets half, I don't know,
twice as good, half as good

444
00:22:50,150 --> 00:22:53,000
because of the lines, and then
it's about the same.

445
00:22:53,000 --> 00:22:55,836
Rajat Bhageria: And by the way,
this, idea of making your own

446
00:22:55,836 --> 00:22:58,718
restaurant, it's it's built on
an interesting foundation. So I

447
00:22:58,718 --> 00:23:01,692
actually, again, I interviewed a
lot of robotics founders when I

448
00:23:01,692 --> 00:23:04,435
was trying to figure out what to
do. it's built on a really

449
00:23:04,435 --> 00:23:07,456
interesting foundation, which is
look, trying to deploy robots in

450
00:23:07,456 --> 00:23:10,246
other customers site is really
freaking hard. And the reason

451
00:23:10,246 --> 00:23:12,895
it's hard is that every customer
does things a little bit

452
00:23:12,895 --> 00:23:15,777
differently. There's different,
Ingredients for us in the food

453
00:23:15,777 --> 00:23:18,287
industry. There's different
palettes in the palatizing

454
00:23:18,287 --> 00:23:20,704
industry, etc. So what if you
could just control the

455
00:23:20,704 --> 00:23:23,447
environment? You have your own
environment. You're your own

456
00:23:23,447 --> 00:23:26,236
customer. it's Nolan Bushnell
like selling, Atari to cheese,

457
00:23:26,236 --> 00:23:28,979
Chuck E. Cheese's, which he
founded both of them. He, built

458
00:23:28,979 --> 00:23:31,582
Chuck E. Cheese's and he built
Atari to sell to Chuck E.

459
00:23:31,582 --> 00:23:32,000
Cheese's.

460
00:23:32,000 --> 00:23:33,000
Audrow Nash: I didn't know
that's so

461
00:23:33,000 --> 00:23:34,000
Rajat Bhageria: that idea.

462
00:23:34,000 --> 00:23:35,000
Audrow Nash: Yeah,

463
00:23:35,000 --> 00:23:37,430
Rajat Bhageria: So it's, and you
constrain the environment. It's

464
00:23:37,430 --> 00:23:38,000
like you really

465
00:23:38,000 --> 00:23:40,000
Audrow Nash: there is the
appeal. Yeah,

466
00:23:40,000 --> 00:23:43,205
Rajat Bhageria: Exactly. The
technology problem is a lot

467
00:23:43,205 --> 00:23:46,182
easier, right? Which is,
notable. But the issue I've

468
00:23:46,182 --> 00:23:49,902
learned from these same founders
is what you called out, which is

469
00:23:49,902 --> 00:23:53,394
fair, and there's two others.
Number one, even if you get the

470
00:23:53,394 --> 00:23:56,657
robots right, oftentimes it's
really hard to get the food

471
00:23:56,657 --> 00:23:57,000
right.

472
00:23:57,000 --> 00:23:58,000
Audrow Nash: ah, yeah.

473
00:23:58,000 --> 00:24:00,871
Rajat Bhageria: just for
example, like many, I won't call

474
00:24:00,871 --> 00:24:03,741
out names, but many companies
face this, right? They made

475
00:24:03,741 --> 00:24:06,863
really cool robots, but then,
consumers would buy the food and

476
00:24:06,863 --> 00:24:09,734
it was like, eh, fine. But I
actually prefer, Domino's. I

477
00:24:09,734 --> 00:24:12,000
prefer Domino's. I prefer
Starbucks. I prefer

478
00:24:12,000 --> 00:24:13,000
Audrow Nash: Whatever else.

479
00:24:13,000 --> 00:24:17,000
Rajat Bhageria: whatever else.
And then I, also learned that

480
00:24:17,000 --> 00:24:20,000
you're essentially stacking two
businesses on

481
00:24:20,000 --> 00:24:22,000
Audrow Nash: exactly. That's
what I was going to say.

482
00:24:22,000 --> 00:24:25,041
Rajat Bhageria: is very hard,
and you have a food business,

483
00:24:25,041 --> 00:24:27,000
and you're stacking the
probabilities.

484
00:24:27,000 --> 00:24:31,110
Audrow Nash: Yes, 100%. It's
much harder. You have to be two

485
00:24:31,110 --> 00:24:32,000
CEOs at once.

486
00:24:32,000 --> 00:24:33,000
Rajat Bhageria: Yes.

487
00:24:32,000 --> 00:24:35,750
Audrow Nash: And I think one of
the things also with this is

488
00:24:35,750 --> 00:24:39,438
it's a huge upfront investment
because you have to get both

489
00:24:39,438 --> 00:24:41,000
businesses going at once.

490
00:24:41,000 --> 00:24:42,000
Rajat Bhageria: Yes.

491
00:24:42,000 --> 00:24:46,030
Audrow Nash: But to me, okay,
we're going to talk more about

492
00:24:46,030 --> 00:24:50,261
what, you guys are doing. But if
you guys want to scale to that

493
00:24:50,261 --> 00:24:54,493
in the future, you can gradually
build up capabilities and you,

494
00:24:54,493 --> 00:24:58,119
so from that, you'll be in a
much stronger position to

495
00:24:58,119 --> 00:25:02,149
automate the rest of it. Cause
you're basically doing one of

496
00:25:02,149 --> 00:25:06,515
these problems at a time, versus
both at one time and requiring a

497
00:25:06,515 --> 00:25:09,000
huge capital investment, all at
once.

498
00:25:09,000 --> 00:25:10,000
Rajat Bhageria: Yes.

499
00:25:10,000 --> 00:25:11,000
Audrow Nash: Uhhuh

500
00:25:10,000 --> 00:25:13,397
Rajat Bhageria: That's right.
anyways, okay, the question is,

501
00:25:13,397 --> 00:25:16,682
okay, who do we sell to? So
selling the robots to consumers

502
00:25:16,682 --> 00:25:19,633
is out. Selling burgers and
sandwiches and, coffee to

503
00:25:19,633 --> 00:25:23,253
consumers is also out. Which, by
the way, you've seen this in the

504
00:25:23,253 --> 00:25:26,594
industry, right? the people who
try to build at home robots,

505
00:25:26,594 --> 00:25:29,935
they've all not done well. The
people who have tried to sell

506
00:25:29,935 --> 00:25:33,443
food, they've also not all done
well, but most of them have not

507
00:25:33,443 --> 00:25:34,000
done well.

508
00:25:34,000 --> 00:25:36,876
Audrow Nash: Uhhuh . Yeah. I
literally have not had any on

509
00:25:36,876 --> 00:25:39,752
this podcast 'cause I didn't
like any of them before this.

510
00:25:39,752 --> 00:25:40,000
Yeah.

511
00:25:40,000 --> 00:25:41,000
Rajat Bhageria: Yes

512
00:25:41,000 --> 00:25:44,000
Audrow Nash: But

513
00:25:44,000 --> 00:25:48,000
Rajat Bhageria: So now it's okay
B2B,

514
00:25:48,000 --> 00:25:51,085
Audrow Nash: Great. Okay, so
your B2B, who do you sell? Who

515
00:25:51,085 --> 00:25:54,379
do you sell to in B2B? Or which,
market? Segment of the B2B set

516
00:25:54,379 --> 00:25:56,000
of companies you could sell to.

517
00:25:56,000 --> 00:26:00,588
Rajat Bhageria: yes, so there's
some interesting story there,

518
00:26:00,588 --> 00:26:05,027
too so when we first started our
natural inclination was to

519
00:26:05,027 --> 00:26:09,765
really get on the bandwagon of
delivery so Uber Eats was really

520
00:26:09,765 --> 00:26:13,000
picking up steam. This is,
again, still pre

521
00:26:13,000 --> 00:26:17,030
Audrow Nash: Yeah. Actually, can
you tell me the timeline of

522
00:26:17,030 --> 00:26:20,724
where we are when you started
this kind of thing? Just.

523
00:26:20,724 --> 00:26:22,000
Broader context for

524
00:26:22,000 --> 00:26:25,053
Rajat Bhageria: basically, I
started the company mid 2019,

525
00:26:25,053 --> 00:26:27,000
and this is when I was doing all
this

526
00:26:27,000 --> 00:26:29,000
Audrow Nash: Just before COVID.
Okay.

527
00:26:29,000 --> 00:26:33,041
Rajat Bhageria: Just before
COVID, yes. And, yeah, so I, I,

528
00:26:33,041 --> 00:26:36,945
what I, what we were thinking
about is, okay, delivery is

529
00:26:36,945 --> 00:26:41,329
really starting to take off. And
this is, again, from, empirical

530
00:26:41,329 --> 00:26:45,438
data from the customers. We,
really tried to talk to, try to

531
00:26:45,438 --> 00:26:49,000
talk to the customers early on,
before we even built

532
00:26:49,000 --> 00:26:52,000
Audrow Nash: And customers being
businesses, right?

533
00:26:52,000 --> 00:26:54,000
Rajat Bhageria: at the time it
was just food businesses, broad,

534
00:26:54,000 --> 00:26:55,841
super broad strokes, like any
food business, let's talk to

535
00:26:55,841 --> 00:26:56,000
them,

536
00:26:56,000 --> 00:26:59,000
Audrow Nash: Hell yeah.

537
00:26:59,000 --> 00:27:02,441
Rajat Bhageria: Fast Casual was
something that was very obvious,

538
00:27:02,441 --> 00:27:04,000
it seemed very obvious to us,

539
00:27:04,000 --> 00:27:08,000
Audrow Nash: Can you define fast
casual? is that Chipotle? Or,

540
00:27:08,000 --> 00:27:11,659
Rajat Bhageria: it's Chipotle,
basically. Chipotle, Sweetgreen,

541
00:27:11,659 --> 00:27:15,317
Dig In, Kava, Mixed, Salad Go,
any kind of place where you have

542
00:27:15,317 --> 00:27:18,918
the count It's the original guys
here were like Subway, right?

543
00:27:18,918 --> 00:27:22,518
You have the original counter,
and you're assembling the meal.

544
00:27:22,518 --> 00:27:26,235
the ingredients are pre cut and
pre cooked. And you have a human

545
00:27:26,235 --> 00:27:29,777
who's assembling the meal into A
sandwich, a burrito, a wrap,

546
00:27:29,777 --> 00:27:33,261
whatever, right? And, so what
was really interesting here is

547
00:27:33,261 --> 00:27:36,920
that delivery was really taking
off. Cloud kitchens were really

548
00:27:36,920 --> 00:27:40,288
starting to gain steam. And
again, this is even pre COVID,

549
00:27:40,288 --> 00:27:43,830
but even pre COVID, 40%, 30 40
percent of volume was delivery

550
00:27:43,830 --> 00:27:46,792
orders, which was very
fascinating. in other words,

551
00:27:46,792 --> 00:27:50,102
people were not going into the
store and picking up their

552
00:27:50,102 --> 00:27:53,470
Chipotle or their Subway or
their Dig In. They were rather

553
00:27:53,470 --> 00:27:56,955
getting it delivered to them
using, one of these aggregators

554
00:27:56,955 --> 00:27:58,000
like Uber Eats and

555
00:27:58,000 --> 00:28:01,810
Audrow Nash: Was that across the
U. S. or in specific locations?

556
00:28:01,810 --> 00:28:03,000
I'm wondering, where

557
00:28:03,000 --> 00:28:04,000
Rajat Bhageria: guessing it was
more of a coastal.

558
00:28:04,000 --> 00:28:05,000
Audrow Nash: Yeah, that's what I
would

559
00:28:05,000 --> 00:28:06,000
Rajat Bhageria: it was one of
the coast

560
00:28:06,000 --> 00:28:09,389
Audrow Nash: it seems like, in
San Francisco maybe that's the

561
00:28:09,389 --> 00:28:12,833
case. but I can, I don't know,
middle of nowhere farm country,

562
00:28:12,833 --> 00:28:15,000
I doubt this is the case kind of
thing.

563
00:28:15,000 --> 00:28:16,589
Rajat Bhageria: That's right.
and by the way, that changed

564
00:28:16,589 --> 00:28:17,000
after COVID, of

565
00:28:17,000 --> 00:28:18,000
Audrow Nash: Yeah.

566
00:28:18,000 --> 00:28:21,981
Rajat Bhageria: but before
COVID, you are totally right. so

567
00:28:21,981 --> 00:28:25,895
anyways, we, like when we first
started off, like we like,

568
00:28:25,895 --> 00:28:30,146
really thought about, okay, fast
casual seems like a good go to

569
00:28:30,146 --> 00:28:34,195
market, right? And we actually
and, I'm Cutting the chase to

570
00:28:34,195 --> 00:28:38,041
really answer your question
about who we sell to now, we,

571
00:28:38,041 --> 00:28:41,550
actually got some early
customers. We got some early

572
00:28:41,550 --> 00:28:45,397
prototypes. we, that's how we
raised some of the earliest

573
00:28:45,397 --> 00:28:48,501
funding we had. And, we,
actually shipped some

574
00:28:48,501 --> 00:28:52,685
prototypes. At the time, they
were just prototypes to our, one

575
00:28:52,685 --> 00:28:56,868
of our early customers in San
Francisco. And, we're like, hey,

576
00:28:56,868 --> 00:29:00,985
customer, I'm not going to say
the customer's name right now,

577
00:29:00,985 --> 00:29:05,303
but hey, customer, built a robot
that can like flexibly assemble

578
00:29:05,303 --> 00:29:09,285
leafy greens for you. Is that
useful for you? and, what we,

579
00:29:09,285 --> 00:29:13,131
realize is that, what we realize
is that in a fast casual

580
00:29:13,131 --> 00:29:17,382
environment, that is not useful.
And the reason it's not useful

581
00:29:17,382 --> 00:29:21,363
is Let's talk about an employee
who works at, a Sweetgreen.

582
00:29:21,363 --> 00:29:25,075
Let's just say arbitrarily,
right? It could be any fast

583
00:29:25,075 --> 00:29:28,381
casual chain, or a ghost
kitchen. The volumes are

584
00:29:28,381 --> 00:29:32,430
actually relatively low. It's
not like you're making tens of

585
00:29:32,430 --> 00:29:36,141
thousands of meals. You're
making a few hundred meals a

586
00:29:36,141 --> 00:29:40,190
day. does that mean? What it
means is that the jobs are very

587
00:29:40,190 --> 00:29:41,000
generalized.

588
00:29:41,000 --> 00:29:44,894
Audrow Nash: You go down the
whole line giving the items, not

589
00:29:44,894 --> 00:29:47,000
just one specific thing. Not
just

590
00:29:47,000 --> 00:29:50,670
Rajat Bhageria: That's right. in
other words, not just the leafy

591
00:29:50,670 --> 00:29:54,111
greens. in other words, if we
wanted to say, Hey, Sally. The

592
00:29:54,111 --> 00:29:57,552
store manager wants to say, hey
Sally, I want you to do this

593
00:29:57,552 --> 00:30:00,993
other thing today. We have this
robot that's going to do the

594
00:30:00,993 --> 00:30:04,376
assembly. for Sally to actually
be able to fully leave, our

595
00:30:04,376 --> 00:30:07,416
robot needs to do 100
ingredients. Everything. And if

596
00:30:07,416 --> 00:30:10,857
we cannot do everything, then
Sally cannot do, go do another

597
00:30:10,857 --> 00:30:14,412
task. By the way, there's not a
labor shortage. So there's, no

598
00:30:14,412 --> 00:30:17,968
risk of Sally getting laid off.
It's just, Sally's going to go

599
00:30:17,968 --> 00:30:19,000
do something else.

600
00:30:19,000 --> 00:30:21,348
Audrow Nash: you mean there is a
labor shortage, so sadly he's

601
00:30:21,348 --> 00:30:23,432
not going to get laid off. I
might have heard the exact

602
00:30:23,432 --> 00:30:24,000
opposite. Okay.

603
00:30:24,000 --> 00:30:27,692
Rajat Bhageria: There is a labor
shortage. My bad. that's what I

604
00:30:27,692 --> 00:30:30,923
meant. so, this was really
interesting, right? So what I

605
00:30:30,923 --> 00:30:34,154
learned there is that in
robotics to actually be able to

606
00:30:34,154 --> 00:30:37,385
add value to a customer, you
essentially need to have at

607
00:30:37,385 --> 00:30:40,731
least a human equivalent. At
least if you have less than a

608
00:30:40,731 --> 00:30:44,423
human equivalent. Like it's one
or zero, right? If you have less

609
00:30:44,423 --> 00:30:45,000
than human

610
00:30:45,000 --> 00:30:48,901
Audrow Nash: In that specific
domain, you're saying. In the

611
00:30:48,901 --> 00:30:53,000
domain of fast casual, this is
what you're In other areas too?

612
00:30:53,000 --> 00:30:57,053
Rajat Bhageria: I think it's at
other areas too, right? So the,

613
00:30:57,053 --> 00:31:01,041
place where this changes is if
you have extremely high volume,

614
00:31:01,041 --> 00:31:04,000
right? If you have extremely
high volume, then

615
00:31:04,000 --> 00:31:05,000
Audrow Nash: it.

616
00:31:05,000 --> 00:31:08,296
Rajat Bhageria: exactly, each of
the roles are more specialized,

617
00:31:08,296 --> 00:31:11,077
but, even in that case, you
still have to have a human

618
00:31:11,077 --> 00:31:14,373
equivalent. It can be any of the
humans, but they have to have a

619
00:31:14,373 --> 00:31:17,000
human equivalent of the tasks
they're doing, right?

620
00:31:17,000 --> 00:31:19,080
Audrow Nash: What do you mean
human equivalent? I don't quite

621
00:31:19,080 --> 00:31:20,000
understand what that means.

622
00:31:20,000 --> 00:31:24,267
Rajat Bhageria: Okay, so so let
me let me paint it this way like

623
00:31:24,267 --> 00:31:28,467
in fast casual Sally's got to do
100 gradients. If we cannot do

624
00:31:28,467 --> 00:31:32,400
the 100 gradients we have we
don't have a Sally equivalent.

625
00:31:32,400 --> 00:31:36,000
In other words, I cannot go to
this the I cannot go to

626
00:31:36,000 --> 00:31:39,692
Audrow Nash: Sally equivalent is
a very funny way to put it.

627
00:31:39,692 --> 00:31:40,000
Okay.

628
00:31:40,000 --> 00:31:43,794
Rajat Bhageria: Yes, if I don't
have that, then I can't go to

629
00:31:43,794 --> 00:31:47,402
this customer and say, Hey, you
should pay me like you pay

630
00:31:47,402 --> 00:31:51,072
Sally, which is a salary. A
robotics as a service salary. I

631
00:31:51,072 --> 00:31:53,000
can't do that because I haven't

632
00:31:53,000 --> 00:31:54,000
Audrow Nash: You haven't done
the

633
00:31:54,000 --> 00:31:57,236
Rajat Bhageria: offset Sally so
she can do something else.

634
00:31:57,236 --> 00:32:00,472
Exactly. on the other hand, you
have like more high volume

635
00:32:00,472 --> 00:32:03,764
production. In high volume
production, if you have somebody

636
00:32:03,764 --> 00:32:07,112
on assembly line, they're doing
one task. Now, oftentimes in

637
00:32:07,112 --> 00:32:10,348
manufacturing for example,
there's still changeovers. It's

638
00:32:10,348 --> 00:32:13,751
not like you're doing that, that
one task 24 hours, or sorry,

639
00:32:13,751 --> 00:32:17,155
eight hours a day. Maybe you
might be doing it for five hours

640
00:32:17,155 --> 00:32:20,614
and then doing another task with
three hours. There might be a

641
00:32:20,614 --> 00:32:23,571
changeover, but you have a
couple tasks. A few tasks,

642
00:32:23,571 --> 00:32:27,142
right? So if I can have a system
that does a few of those tasks,

643
00:32:27,142 --> 00:32:30,210
let's say five tasks, or five
ingredients, in this case

644
00:32:30,210 --> 00:32:33,000
because it's food, then I have a
Sally equivalent.

645
00:32:33,000 --> 00:32:37,322
Audrow Nash: In those tasks
because it's high volume. So the

646
00:32:37,322 --> 00:32:41,644
distinction here, it's not the,
fast casual dining. It's the

647
00:32:41,644 --> 00:32:45,822
volume because with higher
volume, you have the ability to

648
00:32:45,822 --> 00:32:50,000
do specialization and this is
where robots are beneficial.

649
00:32:50,000 --> 00:32:53,593
Rajat Bhageria: that's right.
what we basically realized is,

650
00:32:53,593 --> 00:32:57,305
look, at the time when we were
really focusing on fast casual,

651
00:32:57,305 --> 00:33:00,000
this is I'm telling you the
sequential story,

652
00:33:00,000 --> 00:33:03,000
Audrow Nash: yeah,

653
00:33:03,000 --> 00:33:05,440
Rajat Bhageria: it became
pretty, it was like an

654
00:33:05,440 --> 00:33:08,642
interesting time. It's okay, we
know this is a big market. It's

655
00:33:08,642 --> 00:33:11,540
arguably the biggest market for
AI. We know there's a big

656
00:33:11,540 --> 00:33:14,437
problem. We know that food
assembly is the right thing to

657
00:33:14,437 --> 00:33:17,385
solve. we believe we have the
right go to market, which is

658
00:33:17,385 --> 00:33:20,079
selling to other food
businesses, as opposed to being

659
00:33:20,079 --> 00:33:22,926
the food business itself, or
even selling the robots. We

660
00:33:22,926 --> 00:33:26,179
believe that, this technology is
possible. we, believe that this

661
00:33:26,179 --> 00:33:29,381
all is possible. And further, we
even believe we have the right

662
00:33:29,381 --> 00:33:32,533
technology architecture, which
is more of a pick and place, an

663
00:33:32,533 --> 00:33:35,786
AI based system, which is a lot
more flexible than a dispensing,

664
00:33:35,786 --> 00:33:39,039
hard automation system. we're a
little bit like, ah, we, we feel

665
00:33:39,039 --> 00:33:42,038
like we got all these things
right. And yet, it's still not

666
00:33:42,038 --> 00:33:44,987
there yet, because we can't do
100 ingredients. So then we

667
00:33:44,987 --> 00:33:48,087
started to ask ourselves, okay,
how do we do 100 ingredients?

668
00:33:48,087 --> 00:33:50,731
What's the bottleneck, right?
What we learned is the

669
00:33:50,731 --> 00:33:53,730
bottleneck is very simply just
AI trading data, which is to

670
00:33:53,730 --> 00:33:57,034
say, the thing with food is that
food is very highly dimensional,

671
00:33:57,034 --> 00:33:59,931
and it's very long tail.
obviously, different ingredients

672
00:33:59,931 --> 00:34:03,083
are different than each other. a
potato mash is very different

673
00:34:03,083 --> 00:34:06,133
than diced chicken, and diced
chicken is very different than

674
00:34:06,133 --> 00:34:09,335
shredded chicken, and shredded
chicken is very different than a

675
00:34:09,335 --> 00:34:12,029
potato mash. Thin sauce. So
obviously ingredients are

676
00:34:12,029 --> 00:34:15,181
different from each other. But
the thing is there's also a lot

677
00:34:15,181 --> 00:34:18,027
of differences within the
ingredient. So that same diced

678
00:34:18,027 --> 00:34:21,128
chicken, it's being cut by a
human being. So every single day

679
00:34:21,128 --> 00:34:24,279
it's going to be a little bit
differently cut. And by the way,

680
00:34:24,279 --> 00:34:27,228
it's also being cooked by a
human being. So sometimes it's

681
00:34:27,228 --> 00:34:30,430
going to be cooked a little bit
longer, sometimes it's cooked a

682
00:34:30,430 --> 00:34:33,632
little bit less. Sometimes the
cook's going to add a little bit

683
00:34:33,632 --> 00:34:36,377
extra oil, which means the
stickiness is different. In

684
00:34:36,377 --> 00:34:39,325
other words, it's a very highly
dimensional It's extremely

685
00:34:39,325 --> 00:34:42,274
variable. And by the way, this
is not even to say anything

686
00:34:42,274 --> 00:34:44,968
organic, about the organic
structure. It's a, organic

687
00:34:44,968 --> 00:34:47,967
material. So even growing it is
different. to, to deal with

688
00:34:47,967 --> 00:34:50,000
this, AI, we learned, is just
necessary,

689
00:34:50,000 --> 00:34:53,527
Audrow Nash: Because it's an
edgy problem, like you spend a

690
00:34:53,527 --> 00:34:57,054
lot of time on edge cases, and
because of this AI is better

691
00:34:57,054 --> 00:35:00,103
than like really hard coded
rules, a more classical

692
00:35:00,103 --> 00:35:01,000
approach. Okay.

693
00:35:01,000 --> 00:35:03,962
Rajat Bhageria: that's right,
you, I, that's exactly right.

694
00:35:03,962 --> 00:35:06,723
let, maybe another way, to
express this is, let's say I

695
00:35:06,723 --> 00:35:09,886
wanted to make a robot that just
does rice, right? It just, and

696
00:35:09,886 --> 00:35:12,697
it's not even just rice, it
does, one particular kind of

697
00:35:12,697 --> 00:35:15,759
rice that a machine cooks the
rice, and even, it, the machine

698
00:35:15,759 --> 00:35:18,972
does everything, so you know the
input is exactly the same every

699
00:35:18,972 --> 00:35:21,934
time. yeah, like I could hard
code a system, maybe not hard

700
00:35:21,934 --> 00:35:24,845
coded, but I could have a rules
based system, right? That,

701
00:35:24,845 --> 00:35:26,000
probably would do fine.

702
00:35:26,000 --> 00:35:27,000
Audrow Nash: totally.

703
00:35:27,000 --> 00:35:28,975
Rajat Bhageria: but now if I
want to have a flexible system

704
00:35:28,975 --> 00:35:30,615
where like the inputs are
variable every time and

705
00:35:30,615 --> 00:35:32,590
different customers make their
rice differently, and by the

706
00:35:32,590 --> 00:35:34,598
way, they don't just have rice,
they have trillions of other

707
00:35:34,598 --> 00:35:35,000
ingredients,

708
00:35:35,000 --> 00:35:37,000
Audrow Nash: And also,

709
00:35:37,000 --> 00:35:38,000
Rajat Bhageria: it's a little
bit harder.

710
00:35:38,000 --> 00:35:41,853
Audrow Nash: sure, and but so
with the way where you have the

711
00:35:41,853 --> 00:35:45,707
exact same rice every single
time and it's made by a machine,

712
00:35:45,707 --> 00:35:49,245
that requires a lot more
investment in infrastructure to

713
00:35:49,245 --> 00:35:53,161
have that machine to transport
it to arrange it the exact same

714
00:35:53,161 --> 00:35:56,825
way every time. and so that's a
bigger expenditure for the

715
00:35:56,825 --> 00:36:00,616
customers. And so that would
probably limit. The market that

716
00:36:00,616 --> 00:36:04,280
you could go after for this kind
of thing. So having it be

717
00:36:04,280 --> 00:36:08,259
flexible makes it so more people
can do it. You want to make it

718
00:36:08,259 --> 00:36:11,797
so you can try your robot, or
deploy just one of them or

719
00:36:11,797 --> 00:36:15,777
something, have it at relatively
low cost to the customer. They

720
00:36:15,777 --> 00:36:19,820
don't have to change their whole
like floor plan or anything. so

721
00:36:19,820 --> 00:36:23,231
flexible is way better for lots
of reasons. I imagine,

722
00:36:23,231 --> 00:36:25,000
especially for the customer.

723
00:36:25,000 --> 00:36:28,451
Rajat Bhageria: that's a hundred
percent true. And, actually the

724
00:36:28,451 --> 00:36:31,147
industry has a lot of what's
called low mix Right?

725
00:36:31,147 --> 00:36:34,382
Automation. low mix is what
you're saying, right? it, you're

726
00:36:34,382 --> 00:36:37,672
trying to make a system that
does one thing. anoth an example

727
00:36:37,672 --> 00:36:40,853
from like electronics might be
you're trying to make a line

728
00:36:40,853 --> 00:36:44,196
that is iPhone 15s. That's all
it's gonna do. If you're trying

729
00:36:44,196 --> 00:36:47,000
to do Macs on that line, it's
not gonna work, right?

730
00:36:47,000 --> 00:36:49,891
Audrow Nash: Yeah, I think of
molding generally for this kind

731
00:36:49,891 --> 00:36:52,640
of thing where it's like you
have one mold That's all it's

732
00:36:52,640 --> 00:36:55,341
going to make this kind of
thing. You can make a bunch of

733
00:36:55,341 --> 00:36:57,000
them, but you can't change
anything

734
00:36:57,000 --> 00:37:00,218
Rajat Bhageria: That's right.
And the same thing is true in

735
00:37:00,218 --> 00:37:03,382
the food industry, by the way.
So if I'm like Kraft Heinz,

736
00:37:03,382 --> 00:37:06,436
Kraft Heinz, for example, makes
ketchup. No, if you make

737
00:37:06,436 --> 00:37:09,818
ketchup, you don't have 50 kinds
of ketchup, right? You have a

738
00:37:09,818 --> 00:37:12,000
few SKUs, right? So you're gonna
go to a

739
00:37:12,000 --> 00:37:17,000
Audrow Nash: super low mix and
high volume. Yeah

740
00:37:17,000 --> 00:37:20,624
Rajat Bhageria: integrator,
you're gonna pay them 10 million

741
00:37:20,624 --> 00:37:24,127
bucks, but they'll make you a
system that, perfectly makes

742
00:37:24,127 --> 00:37:27,509
those bottles of ketchup for
you, given the same inputs.

743
00:37:27,509 --> 00:37:30,831
Which is, which you can
basically guarantee because you

744
00:37:30,831 --> 00:37:34,454
have a custom line, right? we
were like, okay, look, back to

745
00:37:34,454 --> 00:37:37,897
the chef world, right? we had
just done this fast, casual

746
00:37:37,897 --> 00:37:41,279
experiment, we're like, 100
ingredients. To add value to

747
00:37:41,279 --> 00:37:44,903
FastCasual. we can't do that
right now because we don't have

748
00:37:44,903 --> 00:37:48,104
training data for 100
Ingredients. And by the way, we

749
00:37:48,104 --> 00:37:51,728
can talk about that too. Why
don't we have training data for

750
00:37:51,728 --> 00:37:55,351
100 Ingredients? It's because,
unlike something like a large

751
00:37:55,351 --> 00:37:58,613
language model, I can't just
download the internet for

752
00:37:58,613 --> 00:38:02,478
training data, right? That's not
possible. unlike a software and

753
00:38:02,478 --> 00:38:06,042
car company, or even like a
company that's doing rigid body

754
00:38:06,042 --> 00:38:09,726
manipulation, palletizing, case
packing, I actually can't use

755
00:38:09,726 --> 00:38:13,470
simulation because simulation
really assumes Newtonian physics

756
00:38:13,470 --> 00:38:17,396
and non deformable stuff Whereas
food is deformable, sticky, wet,

757
00:38:17,396 --> 00:38:18,000
malleable,

758
00:38:18,000 --> 00:38:20,980
Audrow Nash: all the things
physics doesn't do well, huh or

759
00:38:20,980 --> 00:38:23,000
our current physics models in
simulation

760
00:38:23,000 --> 00:38:26,293
Rajat Bhageria: physics. Yeah,
exactly. So simulation also

761
00:38:26,293 --> 00:38:29,587
doesn't work. So then you're
like, okay, you can bootstrap

762
00:38:29,587 --> 00:38:32,937
your data, right? You can like
in a lab bootstrap the data,

763
00:38:32,937 --> 00:38:36,117
which is fine to get started,
right? But of course as we

764
00:38:36,117 --> 00:38:39,580
talked about, given the highly
dimensional nature of food and

765
00:38:39,580 --> 00:38:42,931
the number of ingredients And
the fact that every customer,

766
00:38:42,931 --> 00:38:46,224
every, kind of, they're all
different, like that only gets

767
00:38:46,224 --> 00:38:49,744
you so far. I can bootstrap one
ingredient or two ingredients,

768
00:38:49,744 --> 00:38:53,322
but I can't bootstrap a hundred
ingredients. So we went back to

769
00:38:53,322 --> 00:38:56,729
the drawing board and we're
like, okay, who else did we talk

770
00:38:56,729 --> 00:38:59,965
to in this search when we
decided what kind of, customers

771
00:38:59,965 --> 00:39:02,975
to go after? and we stumbled
upon, of course, Google.

772
00:39:02,975 --> 00:39:06,382
Production Manufacturing, right?
So initially, my impression

773
00:39:06,382 --> 00:39:08,539
actually was like,
manufacturing, food

774
00:39:08,539 --> 00:39:12,060
manufacturing, that's already
solved, right? Like it's, these,

775
00:39:12,060 --> 00:39:15,410
machines that we've, all seen
growing up with how it works,

776
00:39:15,410 --> 00:39:18,817
right? These are traditional
automation systems that kind of

777
00:39:18,817 --> 00:39:21,940
just do the same thing over and
over again. That was my

778
00:39:21,940 --> 00:39:25,404
impression. what I learned is
that there's another really big

779
00:39:25,404 --> 00:39:28,697
market within manufacturing
called high mix manufacturing.

780
00:39:28,697 --> 00:39:32,104
High mix is essentially, the
opposite of low mix. Low mix is

781
00:39:32,104 --> 00:39:35,397
you can get these traditional
pieces of automation, it's a

782
00:39:35,397 --> 00:39:38,804
Kraft Heinz example, and it's
solved, low mix manufacturing.

783
00:39:38,804 --> 00:39:41,984
High mix is a little bit
different. So high mix might be

784
00:39:41,984 --> 00:39:45,618
like, I want to make, a prepared
meal, and to make this prepared

785
00:39:45,618 --> 00:39:48,912
meal, I know that every human
wants a little bit. Slightly

786
00:39:48,912 --> 00:39:52,205
different, stuff. Some people
want meat lovers meals, some

787
00:39:52,205 --> 00:39:55,555
people are vegetarians, some
people are vegans, some people

788
00:39:55,555 --> 00:39:58,849
are gluten free. And within
these, there's 30 kind of SKUs

789
00:39:58,849 --> 00:40:02,483
within each of these categories.
what we learned is that there's

790
00:40:02,483 --> 00:40:06,003
this really incredible market,
from a market size perspective,

791
00:40:06,003 --> 00:40:09,467
where, there's these companies
that'll have hundreds of SKUs.

792
00:40:09,467 --> 00:40:11,000
And if you have hundreds of

793
00:40:11,000 --> 00:40:13,000
Audrow Nash: is an individual
item or what's an SKU?

794
00:40:13,000 --> 00:40:15,000
Rajat Bhageria: like an
individual meal, let's say.

795
00:40:15,000 --> 00:40:16,000
Audrow Nash: Oh, okay. What does
it stand

796
00:40:16,000 --> 00:40:20,000
Rajat Bhageria: So each,

797
00:40:20,000 --> 00:40:21,000
Audrow Nash: Is it an acronym?

798
00:40:21,000 --> 00:40:23,000
Rajat Bhageria: it probably is
an acronym.

799
00:40:23,000 --> 00:40:26,169
Audrow Nash: But it's a meal.
It's like an instance of a meal

800
00:40:26,169 --> 00:40:27,000
in a permutation

801
00:40:27,000 --> 00:40:30,313
Rajat Bhageria: yeah, like
different cans of LaCroix or

802
00:40:30,313 --> 00:40:32,000
like SKUs, right? Different,

803
00:40:32,000 --> 00:40:33,000
Audrow Nash: different flavors
you're

804
00:40:33,000 --> 00:40:35,000
Rajat Bhageria: basically like a
identifiable product

805
00:40:35,000 --> 00:40:43,000
Audrow Nash: Ah, okay.

806
00:40:43,000 --> 00:40:46,164
Rajat Bhageria: okay, so it's
okay, interesting. So what was

807
00:40:46,164 --> 00:40:49,117
interesting about these though
is that because they have

808
00:40:49,117 --> 00:40:52,439
hundreds of SKUs, they are not
gonna have hundreds of dedicated

809
00:40:52,439 --> 00:40:55,550
lines like the Lomix companies
do. Instead, you're going to

810
00:40:55,550 --> 00:40:58,503
have way less lines, like 10
lines arbitrarily. And then

811
00:40:58,503 --> 00:41:01,826
you're going to do changeovers.
Which is to say you're going to

812
00:41:01,826 --> 00:41:05,042
go from doing pad thai in the
morning, to thai green curry in

813
00:41:05,042 --> 00:41:08,312
the afternoon, to maybe doing a
Mexican casserole at night. So

814
00:41:08,312 --> 00:41:11,739
the line is flexible, right? And
so of course, as you alluded to,

815
00:41:11,739 --> 00:41:15,009
traditional automation doesn't
work. And the reason it doesn't

816
00:41:15,009 --> 00:41:18,278
work is traditional automation
is really designed, you bolt to

817
00:41:18,278 --> 00:41:21,547
the ground, it just does that
rice. And the rice robot is not,

818
00:41:21,547 --> 00:41:24,870
the rice machine I should say,
is not then going to go do diced

819
00:41:24,870 --> 00:41:28,034
chicken. But of course, for high
mix manufacturing, you need

820
00:41:28,034 --> 00:41:30,987
that, you need something
flexible, you need something to

821
00:41:30,987 --> 00:41:33,781
go from doing You know, Paptide
and Indian Korma too.

822
00:41:33,781 --> 00:41:36,893
TigerBeanCurry. traditional
animation doesn't work. So what

823
00:41:36,893 --> 00:41:40,004
I learned is that they use
humans, and specifically the way

824
00:41:40,004 --> 00:41:43,062
that production works is that
you have these long assembly

825
00:41:43,062 --> 00:41:46,279
lines, and these assembly lines
are often in very cold rooms,

826
00:41:46,279 --> 00:41:49,654
34, 36 degree Fahrenheit rooms,
and essentially each person's on

827
00:41:49,654 --> 00:41:52,765
a station on the line, and with
their station, they have an

828
00:41:52,765 --> 00:41:56,035
ingredient. okay, I'm making the
pad type. Person 1 is putting

829
00:41:56,035 --> 00:41:59,251
the trays onto the line. Person
2 is doing the noodles person

830
00:41:59,251 --> 00:42:02,626
and there's two lanes let's say
so the person three on the other

831
00:42:02,626 --> 00:42:05,421
side of the light conveyor is
also doing noodles then

832
00:42:05,421 --> 00:42:08,216
somebody's doing mixed
vegetables etc what was really

833
00:42:08,216 --> 00:42:11,169
interesting from a technology
perspective for us is that

834
00:42:11,169 --> 00:42:14,438
because it's much higher volume
the jobs are much special much

835
00:42:14,438 --> 00:42:17,708
more specialized now sally if
she were to work at one of these

836
00:42:17,708 --> 00:42:20,766
companies sally is on station
three and sally is doing the

837
00:42:20,766 --> 00:42:23,877
broccoli Now, Sally isn't just
doing broccoli for her eight

838
00:42:23,877 --> 00:42:27,199
hour shift, she might be doing,
in three hours there's going to

839
00:42:27,199 --> 00:42:30,522
be a changeover, she's going to
do something else, but probably

840
00:42:30,522 --> 00:42:33,000
Sally is doing five ingredients
per day, right?

841
00:42:33,000 --> 00:42:36,033
Rajat Bhageria: ingredients I
can bootstrap in a lab, I can

842
00:42:36,033 --> 00:42:39,375
bootstrap the training data, and
I can get something good enough.

843
00:42:39,375 --> 00:42:42,254
To deploy something in the
world, right? And this became

844
00:42:42,254 --> 00:42:45,339
our kind of like thesis, which
is okay look, we believe that

845
00:42:45,339 --> 00:42:48,167
for robotics, like really
embodied AI to work, you need

846
00:42:48,167 --> 00:42:51,201
real world training data. Sim
doesn't work, right? It could

847
00:42:51,201 --> 00:42:54,183
work for some industries, but
food it doesn't really work.

848
00:42:54,183 --> 00:42:56,702
can't download the training
data, there's no food

849
00:42:56,702 --> 00:42:59,838
manipulation data on the web.
Bootstrapping only takes you so

850
00:42:59,838 --> 00:43:02,923
far, you really need real world
training data. So it's okay,

851
00:43:02,923 --> 00:43:05,802
we're going to bootstrap
something just to get something

852
00:43:05,802 --> 00:43:08,578
good enough to add value to
these customers to do five

853
00:43:08,578 --> 00:43:11,406
ingredients. If we can do five
ingredients, we can ship

854
00:43:11,406 --> 00:43:14,285
something into production. If we
can ship something into

855
00:43:14,285 --> 00:43:17,113
production, over time, we'll
just see a lot more. Yeah,

856
00:43:17,113 --> 00:43:20,301
exactly. And by the way, if we
get more data, then our AI gets

857
00:43:20,301 --> 00:43:22,820
better. We do better
manipulation. What I mean by

858
00:43:22,820 --> 00:43:25,956
better manipulation is where,
we're more consistent, we spill

859
00:43:25,956 --> 00:43:28,887
less, we don't damage the
ingredient more, we are, we can

860
00:43:28,887 --> 00:43:31,458
do more portion size, we can
place it in different

861
00:43:31,458 --> 00:43:34,440
compartments. that's better
manipulation. If we can do all

862
00:43:34,440 --> 00:43:37,730
of that better, then our robots
Are more flexible. If our robots

863
00:43:37,730 --> 00:43:40,506
are more flexible than our
current customers, like our

864
00:43:40,506 --> 00:43:43,643
current customer batch is gonna
use those robots more. If our

865
00:43:43,643 --> 00:43:46,625
current customers have high
utilization with their current

866
00:43:46,625 --> 00:43:49,812
robots, what are they gonna do?
They're gonna get more robots.

867
00:43:49,812 --> 00:43:53,000
So now we have even more robots.
So we have even more training

868
00:43:53,000 --> 00:43:55,000
Audrow Nash: virtuous cycle.
Yeah,

869
00:43:55,000 --> 00:43:58,279
Rajat Bhageria: it's a virtuous
cycle. And by the way, now when

870
00:43:58,279 --> 00:44:01,037
I go to my next customer,
there's a bunch of benefits

871
00:44:01,037 --> 00:44:03,900
here. First of all, I can
probably already have a good.

872
00:44:03,900 --> 00:44:06,814
It's even better for them
because I have a good starting

873
00:44:06,814 --> 00:44:09,781
policy of how to manipulate
Maybe not all the ingredients

874
00:44:09,781 --> 00:44:12,487
are even because every
ingredients their ingredients

875
00:44:12,487 --> 00:44:15,610
are slightly different, but I
have a starting policy, right?

876
00:44:15,610 --> 00:44:18,732
Probably I have a good case
study from my existing customer,

877
00:44:18,732 --> 00:44:21,491
right? I might even have
operation operationally some

878
00:44:21,491 --> 00:44:24,353
local people who can actually
help service the thing So

879
00:44:24,353 --> 00:44:27,320
there's all these things
actually make the whole flywheel

880
00:44:27,320 --> 00:44:30,494
just get better But of course
the core of the flywheel is the

881
00:44:30,494 --> 00:44:33,825
data flywheel, right? So, we're
like, ah Aha, this is what, this

882
00:44:33,825 --> 00:44:36,896
is the go to market, right? And,
I just love to learn about

883
00:44:36,896 --> 00:44:40,175
startups and business. So I was
like, it felt a little bit like

884
00:44:40,175 --> 00:44:43,349
we were trying to build a Model
S before, which is like going

885
00:44:43,349 --> 00:44:46,576
directly to fast casual, right?
I'm sorry, Model 3 right away.

886
00:44:46,576 --> 00:44:49,803
And, now it's okay, really right
now we just need to build the

887
00:44:49,803 --> 00:44:51,000
Roadster. And this is a

888
00:44:51,000 --> 00:44:52,000
Audrow Nash: Love it. Uhhuh.

889
00:44:51,000 --> 00:44:53,921
Rajat Bhageria: something that
can like, get to market, add

890
00:44:53,921 --> 00:44:57,139
value, and we learn how to like,
really build the technology. And

891
00:44:57,139 --> 00:45:00,208
we create this flywheel. And,
but of course the vision was the

892
00:45:00,208 --> 00:45:03,129
same. The vision of Chef has
always been, we want to deploy

893
00:45:03,129 --> 00:45:05,901
AI enabled robots in the food
industry, period. We wanna

894
00:45:05,901 --> 00:45:08,277
deploy AI enabled robot
everywhere and, to every

895
00:45:08,277 --> 00:45:11,000
commercial kitchen, right?
Because by the way, like the

896
00:45:11,000 --> 00:45:14,119
market needs that and the, the
demand is there. and if somebody

897
00:45:14,119 --> 00:45:16,000
were to solve this, like the
demand is

898
00:45:16,000 --> 00:45:19,333
Audrow Nash: Yeah. It would
skyrocket inval. Like you could,

899
00:45:19,333 --> 00:45:22,000
you have a huge market you could
just gobble up.

900
00:45:22,000 --> 00:45:23,000
Rajat Bhageria: That's right.

901
00:45:23,000 --> 00:45:25,796
Rajat Bhageria: So it was like,
okay, how do we then think,

902
00:45:25,796 --> 00:45:28,592
okay, we, we started here, we
started with the model three.

903
00:45:28,592 --> 00:45:31,293
We'd like to go back there at
some point. We're right now

904
00:45:31,293 --> 00:45:33,994
where the road are. And then
we're like, okay, we have to

905
00:45:33,994 --> 00:45:36,885
connect the two dots. And, what
was really interesting is, of

906
00:45:36,885 --> 00:45:39,728
course, if I can manipulate
diced chicken in a food factory,

907
00:45:39,728 --> 00:45:42,619
why can't I do it in a ghost
kitchen? And if I can manipulate

908
00:45:42,619 --> 00:45:45,273
diced chicken in a ghost
kitchen, why can't I do it in a

909
00:45:45,273 --> 00:45:48,258
fast casual, right? Manipulation
is manipulation. The hard part

910
00:45:48,258 --> 00:45:50,722
is manipulation. The grasping.
That's the hard part.

911
00:45:50,722 --> 00:45:53,424
everything, there's other
issues, by the way, right? Like

912
00:45:53,424 --> 00:45:55,698
form factor, there's other
issues. but those are

913
00:45:55,698 --> 00:45:58,542
engineering issues that can be
solved. The thing that's like

914
00:45:58,542 --> 00:46:00,958
the hard thing is the
manipulation, right? So we're

915
00:46:00,958 --> 00:46:03,897
like, huh. So not only are we,
do we have this flywheel within

916
00:46:03,897 --> 00:46:06,692
manufacturing, which is like the
more robots we deploy, the

917
00:46:06,692 --> 00:46:09,773
better the system gets, but that
same training data and that same

918
00:46:09,773 --> 00:46:12,000
AI can then be used to go to
step two. Step two

919
00:46:12,000 --> 00:46:13,000
Audrow Nash: Yeah, exactly.

920
00:46:12,000 --> 00:46:14,000
Rajat Bhageria: Go's Kitchens
and Cloud Kitchens.

921
00:46:14,000 --> 00:46:18,099
Audrow Nash: Hell yeah. Yeah.
That's what I was meaning where

922
00:46:18,099 --> 00:46:22,063
it's you guys can just do, you
start with five products and

923
00:46:22,063 --> 00:46:26,162
then eventually once you have
scaled up to a hundred products

924
00:46:26,162 --> 00:46:30,328
or 500 products, then you can
maybe then open your own kitchen

925
00:46:30,328 --> 00:46:34,427
and you have the robot worker
part solved. It's but and then.

926
00:46:34,427 --> 00:46:38,391
On the way to going from robot
serving five things to robot

927
00:46:38,391 --> 00:46:42,490
restaurant, you can serve all
the things that are in between,

928
00:46:42,490 --> 00:46:46,387
which is the rest of the food
industry. So it seems like a

929
00:46:46,387 --> 00:46:48,000
good growing plan to me.

930
00:46:48,000 --> 00:46:50,929
Rajat Bhageria: Yes. Yes,
exactly. So like right now we're

931
00:46:50,929 --> 00:46:53,000
focused on food factories, big
giant food

932
00:46:53,000 --> 00:46:54,000
Audrow Nash: Oh yeah. Might as
well.

933
00:46:54,000 --> 00:46:57,549
Rajat Bhageria: step is exactly,
next step is regional food

934
00:46:57,549 --> 00:47:00,496
factories. Regional food
factories would be cloud

935
00:47:00,496 --> 00:47:03,624
kitchens, ghost kitchens. These
are smaller, they're

936
00:47:03,624 --> 00:47:07,293
aggregations of restaurants. By
the way, they are essentially

937
00:47:07,293 --> 00:47:10,541
factories, they have a lot of
real estate, and they're

938
00:47:10,541 --> 00:47:14,331
delivering delivery orders, so
it is a food facility, it's just

939
00:47:14,331 --> 00:47:17,699
smaller than the ones we're
going after right now. Lower

940
00:47:17,699 --> 00:47:21,429
volume, right? And then finally,
we get to, if you think about

941
00:47:21,429 --> 00:47:25,218
right now we're like, we're You
know, relatively, we're low mix

942
00:47:25,218 --> 00:47:28,286
on the relative scale to
restaurants, but very high

943
00:47:28,286 --> 00:47:31,955
volume. And then, as we go to
step two, the mix gets a little

944
00:47:31,955 --> 00:47:35,323
bit higher, the volume gets a
little bit lower. And then

945
00:47:35,323 --> 00:47:38,632
finally, when we get to
commercial kitchens, the mix is

946
00:47:38,632 --> 00:47:42,000
back way at the top, but the
volume is way, lower again.

947
00:47:42,000 --> 00:47:43,000
Audrow Nash: Oh yeah. That's
super

948
00:47:43,000 --> 00:47:45,000
Rajat Bhageria: So that's that's
what we're thinking about

949
00:47:45,000 --> 00:47:48,745
Audrow Nash: Yes. And I would
think if we're sticking with the

950
00:47:48,745 --> 00:47:52,369
like model S to three, then you
opening your own restaurant,

951
00:47:52,369 --> 00:47:54,000
that can be your Cybertruck

952
00:47:54,000 --> 00:47:55,000
Rajat Bhageria: RoboTaxi.

953
00:47:55,000 --> 00:47:59,104
Audrow Nash: or the Robotaxi,
yeah, okay, hell yeah. So that

954
00:47:59,104 --> 00:48:03,414
sounds really cool. getting the
data is very clever and you can

955
00:48:03,414 --> 00:48:07,244
train on that. And I think just
to make it a little more

956
00:48:07,244 --> 00:48:10,870
explicit or make sure I
understand. So you're getting

957
00:48:10,870 --> 00:48:14,700
training data on these five
items that now are you're in

958
00:48:14,700 --> 00:48:19,010
production and you're actually
adding value for a customer. now

959
00:48:19,010 --> 00:48:23,182
you're recording all of that
data and you are then using that

960
00:48:23,182 --> 00:48:26,808
to train. better general
policies that you can use to

961
00:48:26,808 --> 00:48:31,049
bootstrap additional things. is
that, kind of how it's working

962
00:48:31,049 --> 00:48:35,358
where it's, and then you can do,
you can add another five items

963
00:48:35,358 --> 00:48:37,000
or whatever it might be.

964
00:48:37,000 --> 00:48:41,286
Rajat Bhageria: yes, that's,
right, exactly, right? I think,

965
00:48:41,286 --> 00:48:45,571
like, where we really started
is, if you think about a meal,

966
00:48:45,571 --> 00:48:46,000
right?

967
00:48:46,000 --> 00:48:49,196
Rajat Bhageria: The majority of
the meal is what you might

968
00:48:49,196 --> 00:48:52,558
consider like scoopable right in
other words that you can use

969
00:48:52,558 --> 00:48:55,919
your hands or you can use like
something like a scoop like an

970
00:48:55,919 --> 00:48:59,115
ice cream scoop or something to
pick it up and that's like

971
00:48:59,115 --> 00:49:02,587
things like starches like rices
and pastas and things like that

972
00:49:02,587 --> 00:49:05,728
then you have things like
different meats right it can be

973
00:49:05,728 --> 00:49:09,089
like diced meat or shredded meat
or Pulled meats, things like

974
00:49:09,089 --> 00:49:12,451
that. Then you have different
kinds of vegetables. Of course,

975
00:49:12,451 --> 00:49:15,536
there's lots of different kinds
of vegetables, different

976
00:49:15,536 --> 00:49:18,898
combinations of vegetables, but
they're always usually cut in

977
00:49:18,898 --> 00:49:22,314
some way. Then you might have a
garnish, right? You might have

978
00:49:22,314 --> 00:49:25,676
something like a spice, or you
might have like more vegetable

979
00:49:25,676 --> 00:49:28,982
toppings, something like that.
Maybe you have a sauce, maybe

980
00:49:28,982 --> 00:49:32,013
you have something like a mash,
right? All of these are

981
00:49:32,013 --> 00:49:35,044
scoopable. They're very
different, by the way, shredded

982
00:49:35,044 --> 00:49:38,130
meat is very different than a
cheese, and cheese is very

983
00:49:38,130 --> 00:49:41,381
different than a sauce, and a
sauce is very different than,

984
00:49:41,381 --> 00:49:44,853
than broccoli. They're all very
different, but they're all like

985
00:49:44,853 --> 00:49:48,104
this, you can imagine an idea
where, a camera to understand

986
00:49:48,104 --> 00:49:51,520
where do I pick from and where
do I place, A 6DoF arm with the

987
00:49:51,520 --> 00:49:54,000
right, manipulator at the, the
right, utensil

988
00:49:54,000 --> 00:49:57,000
Audrow Nash: An appropriate
manipulator. Yeah.

989
00:49:57,000 --> 00:50:00,060
Rajat Bhageria: the combination
of computer vision, motion

990
00:50:00,060 --> 00:50:02,910
planning control, the right
robot policy, motion pilot

991
00:50:02,910 --> 00:50:06,181
policies, and the right utensils
can hopefully build a general

992
00:50:06,181 --> 00:50:09,084
policy on how to manipulate
food. And the more of these

993
00:50:09,084 --> 00:50:12,091
ingredients it sees, and the
more, it's basically breath.

994
00:50:12,091 --> 00:50:15,363
different ingredients and also
depth. Like, how many of those,

995
00:50:15,363 --> 00:50:18,634
the same ingredients used over
and over. you can make a system

996
00:50:18,634 --> 00:50:21,853
that's very flexible, but of
course it's still got to be very

997
00:50:21,853 --> 00:50:25,072
consistent. I think consistency
is the hard part here, right?

998
00:50:25,072 --> 00:50:28,396
It's gotta, it can't be picking
up, if the, if it's a one ounce

999
00:50:28,396 --> 00:50:31,562
portion, you can't be picking up
three ounces one time, zero

1000
00:50:31,562 --> 00:50:34,781
ounces the next time, half an
ounce the next time. It's gotta

1001
00:50:34,781 --> 00:50:38,000
be, like, basically very close
to three ounces over and over.

1002
00:50:38,000 --> 00:50:43,188
Audrow Nash: Okay, so this, to
me, this sounds like a very hard

1003
00:50:43,188 --> 00:50:48,376
problem and it also sounds hard
to get your, so you are getting

1004
00:50:48,376 --> 00:50:52,824
data and I understand in the
abstract how more data is

1005
00:50:52,824 --> 00:50:56,776
improvement, and you've
mentioned so robotic arm

1006
00:50:56,776 --> 00:51:01,965
appropriate gripper and you're
using mostly computer vision and

1007
00:51:01,965 --> 00:51:06,412
that lends itself from your
background which is really

1008
00:51:06,412 --> 00:51:11,435
great, how is that the only, is
vision the only sensor you're

1009
00:51:11,435 --> 00:51:16,294
using other than maybe like
joint torques or something like

1010
00:51:16,294 --> 00:51:21,482
this? and then like, how do you,
annotate your data? How do you

1011
00:51:21,482 --> 00:51:26,424
establish what was good or what
was not good? how do you use

1012
00:51:26,424 --> 00:51:31,282
production data in actually
training the system to, do what

1013
00:51:31,282 --> 00:51:34,000
you want it to do? Like, how,
do,

1014
00:51:34,000 --> 00:51:35,000
Rajat Bhageria: yeah. It's a,

1015
00:51:35,000 --> 00:51:36,000
Audrow Nash: go ahead.

1016
00:51:36,000 --> 00:51:39,105
Rajat Bhageria: there's a, lot
there, and all very good

1017
00:51:39,105 --> 00:51:42,437
questions. one thing that's
really interesting is that most

1018
00:51:42,437 --> 00:51:45,937
of our team is from the software
and car industry. So we have,

1019
00:51:45,937 --> 00:51:49,269
the vast majority of our team is
from, Cruise, Neuro, Zoox,

1020
00:51:49,269 --> 00:51:52,544
Waymo, Embark, and all these
different AV companies and of

1021
00:51:52,544 --> 00:51:55,988
course like what you essentially
ask is what AV companies are

1022
00:51:55,988 --> 00:51:59,601
very good at. They're very good
at like Annotation and pipelines

1023
00:51:59,601 --> 00:52:02,989
and, testing and test infra,
right? So essentially, okay, so

1024
00:52:02,989 --> 00:52:06,377
let's, start at like where we
started. So the first step, of

1025
00:52:06,377 --> 00:52:09,256
course is having the right
metrics. So what are the

1026
00:52:09,256 --> 00:52:12,814
relevant metrics? So consistency
is one of the main metrics. So

1027
00:52:12,814 --> 00:52:16,258
how do you measure, and that's
consistent with weight, right?

1028
00:52:16,258 --> 00:52:19,363
The way, by the way, And this
also gets to your sensors

1029
00:52:19,363 --> 00:52:22,695
question. of course, we have
RGBD. We have a couple of RGBD

1030
00:52:22,695 --> 00:52:25,969
sensors. we also use force
torque, also have weigh scales,

1031
00:52:25,969 --> 00:52:29,301
essentially like scales, right?
and so the way this kind of

1032
00:52:29,301 --> 00:52:32,406
works is that what we can do is
we say, okay, we have a

1033
00:52:32,406 --> 00:52:35,512
topography. Based on this
topography and the ingredient

1034
00:52:35,512 --> 00:52:38,561
I'm picking, And the other
sensor data, inputs I have,

1035
00:52:38,561 --> 00:52:42,061
where, what pose do I pick at in
the pan to be consistent? And

1036
00:52:42,061 --> 00:52:45,223
again, that'll take into
consideration a bunch of stuff.

1037
00:52:45,223 --> 00:52:48,329
Like all, like the entire robot
policy, it'll take into

1038
00:52:48,329 --> 00:52:51,378
consideration any fine tuning
the robot did before the

1039
00:52:51,378 --> 00:52:54,878
production run started, which
the robots fine tune themselves.

1040
00:52:54,878 --> 00:52:57,984
Like it'll pick and dump to fine
tune itself before the

1041
00:52:57,984 --> 00:52:59,000
production starts.

1042
00:52:59,000 --> 00:53:00,000
Audrow Nash: cool.

1043
00:53:00,000 --> 00:53:02,813
Rajat Bhageria: account. Yeah.
and that's because there's so

1044
00:53:02,813 --> 00:53:03,000
much

1045
00:53:03,000 --> 00:53:05,892
Audrow Nash: You're calibrating.
So you have, like your base

1046
00:53:05,892 --> 00:53:08,783
layer of understanding, but then
it's going to fine tune the

1047
00:53:08,783 --> 00:53:11,819
network a little bit by picking
and dumping just a few times to

1048
00:53:11,819 --> 00:53:14,711
be like, okay, let me get the
consistence of this. That's so

1049
00:53:14,711 --> 00:53:15,000
funny.

1050
00:53:15,000 --> 00:53:17,256
Rajat Bhageria: it's a very
practical, it's a very practical

1051
00:53:17,256 --> 00:53:19,323
thing, right? Because every
batch of food is a tiny bit

1052
00:53:19,323 --> 00:53:20,000
different. It just

1053
00:53:20,000 --> 00:53:23,834
Audrow Nash: Oh, definitely.
What a funny thing. Okay. I've

1054
00:53:23,834 --> 00:53:27,863
seen people do that too. if you
see like people waiting, while

1055
00:53:27,863 --> 00:53:31,697
they're whatever, they'll pick
and scoop whatever. And it's

1056
00:53:31,697 --> 00:53:35,596
like, they are calibrating in
some sense and that's so funny

1057
00:53:35,596 --> 00:53:38,000
the robots doing that. Okay, so
cool.

1058
00:53:38,000 --> 00:53:40,707
Rajat Bhageria: so then
basically now, we have, we've

1059
00:53:40,707 --> 00:53:43,976
figured out, okay, what pose do
I pick at? We'll execute a pick,

1060
00:53:43,976 --> 00:53:46,938
and then of course what our
scales do is they're take away

1061
00:53:46,938 --> 00:53:49,849
scales. So we can measure the
weight lost from the pan of

1062
00:53:49,849 --> 00:53:52,863
food. They're underneath the pan
of food. So we measure the

1063
00:53:52,863 --> 00:53:56,081
weight lost. What that allows us
to do is close the loop. Okay,

1064
00:53:56,081 --> 00:53:57,000
we, based on based

1065
00:53:57,000 --> 00:54:00,529
Audrow Nash: you're making your
own supervised training data

1066
00:54:00,529 --> 00:54:04,294
basically because you have that
scale and it's all weight driven

1067
00:54:04,294 --> 00:54:05,000
Interesting.

1068
00:54:05,000 --> 00:54:07,000
Rajat Bhageria: so that's one,
metric, right?

1069
00:54:07,000 --> 00:54:08,000
Audrow Nash: Yeah, okay.

1070
00:54:08,000 --> 00:54:11,185
Rajat Bhageria: That's one. So
it probably arguably one of the

1071
00:54:11,185 --> 00:54:14,421
most important ones, but that is
one of them. so weight is one.

1072
00:54:14,421 --> 00:54:17,605
another thing we think about is
like placement, right? So like

1073
00:54:17,605 --> 00:54:20,944
what, what we do is we have a, a
neur network that basically does

1074
00:54:20,944 --> 00:54:24,180
like the before and after. It's
kinda qa. So we like, it's part

1075
00:54:24,180 --> 00:54:27,159
of our product. So what we can
do is we, we. We say, okay,

1076
00:54:27,159 --> 00:54:30,189
look, we like the food to be
distributed like this, whether

1077
00:54:30,189 --> 00:54:33,014
it's in a clump, or maybe it
needs to be spread out, or

1078
00:54:33,014 --> 00:54:36,199
whatever the customer wants. The
customer can actually we give

1079
00:54:36,199 --> 00:54:39,332
them a nice web app where they
can say how they want the food

1080
00:54:39,332 --> 00:54:40,000
to be placed,

1081
00:54:40,000 --> 00:54:43,611
Audrow Nash: That's so cool. And
they're like I want it dispersed

1082
00:54:43,611 --> 00:54:47,000
like this or mushed or like a
Perfect All that looks like the

1083
00:54:47,000 --> 00:54:50,000
Rajat Bhageria: yeah, or in this
compartment, or whatever else,

1084
00:54:50,000 --> 00:54:51,000
Audrow Nash: Okay, that's really
cool.

1085
00:54:51,000 --> 00:54:53,899
Rajat Bhageria: Yeah, and then
of course what we can do is use

1086
00:54:53,899 --> 00:54:56,612
it, we can have a before and
after, right? And then we can

1087
00:54:56,612 --> 00:54:59,558
give a scoring to that. So like
how good was the deposit from a

1088
00:54:59,558 --> 00:55:02,363
placement perspective? So that's
another metric. We have the

1089
00:55:02,363 --> 00:55:04,000
placement quality basically,
right?

1090
00:55:04,000 --> 00:55:07,733
Audrow Nash: you just look at
like the difference in the

1091
00:55:07,733 --> 00:55:11,600
before and after image and then
you assume that's what was

1092
00:55:11,600 --> 00:55:15,333
placed and then you evaluate
that based on some intended

1093
00:55:15,333 --> 00:55:16,000
placement,

1094
00:55:16,000 --> 00:55:18,915
Rajat Bhageria: There's a bunch
of different ways to score it

1095
00:55:18,915 --> 00:55:21,735
and different customers care
about different things, right?

1096
00:55:21,735 --> 00:55:24,697
But we have depth as well. So we
can see, okay, like what does

1097
00:55:24,697 --> 00:55:27,612
the mound look like? Or may,
maybe they care, may, maybe they

1098
00:55:27,612 --> 00:55:30,527
care just about look, we want
the entire thing to be covered.

1099
00:55:30,527 --> 00:55:33,442
there's different ways to score
it, but there's some score we

1100
00:55:33,442 --> 00:55:36,310
assign for the placement for
that customer, right? so that's

1101
00:55:36,310 --> 00:55:39,320
metric too. The third metric we
look at is, throughput, because

1102
00:55:39,320 --> 00:55:42,283
of course it needs to be fast.
if it's not fast, then honestly

1103
00:55:42,283 --> 00:55:43,000
it's not great.

1104
00:55:43,000 --> 00:55:44,000
Audrow Nash: Yeah.

1105
00:55:44,000 --> 00:55:47,017
Rajat Bhageria: so that's
another metric we look at.

1106
00:55:47,017 --> 00:55:50,497
another metric we look at is.
This is just core reliability.

1107
00:55:50,497 --> 00:55:54,036
And reliability, there's many
dimensions to reliability, but,

1108
00:55:54,036 --> 00:55:57,459
some to call out would be things
like missed bowls. Did we,

1109
00:55:57,459 --> 00:56:00,823
detect a bowl, but we couldn't
place into it? for whatever

1110
00:56:00,823 --> 00:56:04,188
reason. did we do a double
deposit where we tried to place

1111
00:56:04,188 --> 00:56:05,000
twice? did we,

1112
00:56:05,000 --> 00:56:07,000
Audrow Nash: These are bugs,
basically.

1113
00:56:07,000 --> 00:56:10,320
Rajat Bhageria: yeah, basically
bugs and reliability. Did we try

1114
00:56:10,320 --> 00:56:13,381
to accelerate so fast that we
use collaborative robots? Did

1115
00:56:13,381 --> 00:56:16,494
the robot protect us? Stop. and
so that also gets taken into

1116
00:56:16,494 --> 00:56:19,347
consideration. So anyways, we
have metrics for the most

1117
00:56:19,347 --> 00:56:22,460
important things our customers
care about, and all of that's

1118
00:56:22,460 --> 00:56:25,469
considered when you're making
this policy to, To make sure

1119
00:56:25,469 --> 00:56:28,737
that like you don't get a system
that's flexible, but then like

1120
00:56:28,737 --> 00:56:32,057
it's super slow or it's flexible
But it's spilling food all over

1121
00:56:32,057 --> 00:56:35,170
the place or it's flexible But
like it's super ugly when you

1122
00:56:35,170 --> 00:56:36,000
place it, right?

1123
00:56:36,000 --> 00:56:37,000
Audrow Nash: huh.

1124
00:56:37,000 --> 00:56:40,657
Rajat Bhageria: All of these
things need to be taken into

1125
00:56:40,657 --> 00:56:44,700
consideration and then you know
the final part of your question

1126
00:56:44,700 --> 00:56:48,100
was like, okay the data
annotation and stuff a lot of

1127
00:56:48,100 --> 00:56:52,271
it's automated, so we've built a
lot of tooling to do this. So if

1128
00:56:52,271 --> 00:56:56,249
there's a high weight event, we
wanted x, and after some sigma

1129
00:56:56,249 --> 00:56:59,778
standard deviation, we'll
automatically annotate, we'll

1130
00:56:59,778 --> 00:57:03,820
automatically annotate that and
say, oh, this was a high weight

1131
00:57:03,820 --> 00:57:07,798
event, let's collect a, we use
ROS, so let's collect a ROS bag

1132
00:57:07,798 --> 00:57:11,648
of all the information, like a
little bit of the information

1133
00:57:11,648 --> 00:57:15,754
before and after. along, and now
we can debug it, right? And all

1134
00:57:15,754 --> 00:57:19,604
of that, we have a data pipeline
that's all, as soon as that

1135
00:57:19,604 --> 00:57:23,390
event happens The robot, will
send that rolling recorder to

1136
00:57:23,390 --> 00:57:27,175
the cloud. And then we have a
nice interface where, we will

1137
00:57:27,175 --> 00:57:31,282
get, an alert on, our on Slack,
and if it's a P zero O or L zero

1138
00:57:31,282 --> 00:57:35,067
alert PagerDuty. then the, the
person on call, the customer

1139
00:57:35,067 --> 00:57:38,981
support person on call can, if
it's really kind of production

1140
00:57:38,981 --> 00:57:42,638
downtime impacting, like they
can, They can actually like

1141
00:57:42,638 --> 00:57:46,424
triage that so so that's I
talked about the support process

1142
00:57:46,424 --> 00:57:50,146
But that same thing is used with
ML too, right? Because if

1143
00:57:50,146 --> 00:57:53,290
there's issues like we've
automatically annotated

1144
00:57:53,290 --> 00:57:57,268
everything It's one in the same
process like it's just it goes

1145
00:57:57,268 --> 00:57:59,000
different places like one's

1146
00:57:59,000 --> 00:58:03,393
Audrow Nash: Yeah, for sure.
Makes sense. Yeah, you take the

1147
00:58:03,393 --> 00:58:08,079
data, you route it both places.
That's pretty cool. Yeah, I like

1148
00:58:08,079 --> 00:58:12,326
that quite a bit. That's, so
you're getting your data, and

1149
00:58:12,326 --> 00:58:16,499
then You do the scooping to fine
tune. is there any trick

1150
00:58:16,499 --> 00:58:21,185
involved with going from, you're
going from, you try to probably

1151
00:58:21,185 --> 00:58:25,358
sequence what you, the food that
you work on based on its

1152
00:58:25,358 --> 00:58:29,825
similarity with the past foods
that you've trained on, so you

1153
00:58:29,825 --> 00:58:34,364
don't go from, like, all rice to
shredded chicken or something

1154
00:58:34,364 --> 00:58:37,000
that might be a big difference.
how,

1155
00:58:37,000 --> 00:58:39,000
Rajat Bhageria: In terms of like
our own training

1156
00:58:39,000 --> 00:58:43,304
Audrow Nash: yeah, so in terms
of, so you support five items,

1157
00:58:43,304 --> 00:58:47,678
say. if you're going to add item
six, you probably, if you did

1158
00:58:47,678 --> 00:58:52,052
like rice and a bunch of other
small grains, then you probably

1159
00:58:52,052 --> 00:58:56,356
don't want to go to shredded
chicken or broccoli or something

1160
00:58:56,356 --> 00:59:00,095
like this. That's a big
departure from it. how do you

1161
00:59:00,095 --> 00:59:02,000
sequence the items you add?

1162
00:59:02,000 --> 00:59:04,538
Rajat Bhageria: yeah, so it's
actually an interesting

1163
00:59:04,538 --> 00:59:07,507
question. this is actually if it
was a, if it was like perfect

1164
00:59:07,507 --> 00:59:10,285
world we'd do exactly what
you're saying. But the thing is

1165
00:59:10,285 --> 00:59:12,870
like now that we're in
production, like again, this is

1166
00:59:12,870 --> 00:59:15,696
a sequential story. Now that
we're in production, we have a

1167
00:59:15,696 --> 00:59:18,713
customer who's paying us money.
We have to do what the customer

1168
00:59:18,713 --> 00:59:19,000
wants.

1169
00:59:19,000 --> 00:59:20,000
Audrow Nash: yeah.

1170
00:59:20,000 --> 00:59:22,255
Rajat Bhageria: So what the
customer wants is the

1171
00:59:22,255 --> 00:59:24,603
ingredients that are the
toughest for their people,

1172
00:59:24,603 --> 00:59:27,042
right? So and this was
interesting, right? we learned

1173
00:59:27,042 --> 00:59:29,895
that like risotto, which is a
very thick viscous material It's

1174
00:59:29,895 --> 00:59:31,000
like really hard because

1175
00:59:31,000 --> 00:59:35,231
Audrow Nash: is that like
cheese? Risotto means cheese,

1176
00:59:35,231 --> 00:59:38,000
right? Or is it a cheesy rice?
Okay.

1177
00:59:38,000 --> 00:59:40,254
Rajat Bhageria: It's very dense
basically, right? So it's like a

1178
00:59:40,254 --> 00:59:42,225
very, it's like a hard, like
your arms are hurting, your

1179
00:59:42,225 --> 00:59:43,000
back's hurting, right?

1180
00:59:43,000 --> 00:59:45,000
Audrow Nash: Do a lot of work to
do to pull through it. Okay.

1181
00:59:45,000 --> 00:59:48,367
Rajat Bhageria: For example,
right? My point is that like

1182
00:59:48,367 --> 00:59:52,030
Sometimes we didn't have, we,
it's like a, it's a negotiation.

1183
00:59:52,030 --> 00:59:55,515
Hey customer, we've done your
basmati rice. How about we do

1184
00:59:55,515 --> 00:59:59,000
your brown rice? The customer is
no, I, you, I know you can

1185
00:59:59,000 --> 01:00:00,000
Audrow Nash: the brown

1186
00:59:59,000 --> 01:00:01,805
Rajat Bhageria: I want you to do
this, and, so it's, a

1187
01:00:01,805 --> 01:00:04,610
negotiation, which of course
isn't great from a purely

1188
01:00:04,610 --> 01:00:07,000
technology perspective, because
like you said,

1189
01:00:07,000 --> 01:00:09,000
Audrow Nash: you would want a
sequence. Yeah.

1190
01:00:09,000 --> 01:00:12,088
Rajat Bhageria: you would want
to sequence it, but sometimes

1191
01:00:12,088 --> 01:00:15,125
you're not able to. You as an
early stage startup, you take

1192
01:00:15,125 --> 01:00:16,000
what you can get.

1193
01:00:16,000 --> 01:00:19,750
Audrow Nash: Makes sense. So
then you, what, you probably do

1194
01:00:19,750 --> 01:00:23,500
then is you take whatever the
material they want and you can

1195
01:00:23,500 --> 01:00:27,125
bootstrap something with it and
that gets you working good

1196
01:00:27,125 --> 01:00:30,813
enough. and eventually maybe
your learned scooping policies

1197
01:00:30,813 --> 01:00:34,750
will be general enough to work
with a lot of things, but you're

1198
01:00:34,750 --> 01:00:38,000
bootstrapping specific things
that you need to work.

1199
01:00:38,000 --> 01:00:40,747
Rajat Bhageria: the way we
thought about it is and again,

1200
01:00:40,747 --> 01:00:43,735
this is it's changed through the
years, right? Like per class,

1201
01:00:43,735 --> 01:00:46,723
we'd have like policies, right?
right? If you will, a class of

1202
01:00:46,723 --> 01:00:49,759
food, like a mash is a class of
food, and it's a very different

1203
01:00:49,759 --> 01:00:50,000
class

1204
01:00:50,000 --> 01:00:53,563
Audrow Nash: how technical, I
just, talking to you about this,

1205
01:00:53,563 --> 01:00:57,184
I don't know very much about the
food industry, but having you,

1206
01:00:57,184 --> 01:01:00,805
and you not being a food person,
but you've come in and you are

1207
01:01:00,805 --> 01:01:04,023
dissecting it, and you have
this, I don't know, semantic

1208
01:01:04,023 --> 01:01:05,000
understanding of,

1209
01:01:05,000 --> 01:01:06,000
Rajat Bhageria: It's like a

1210
01:01:06,000 --> 01:01:09,959
Audrow Nash: yeah, exactly, It's
so funny. But okay, so you have

1211
01:01:09,959 --> 01:01:12,000
your mashes, and, all these
other

1212
01:01:12,000 --> 01:01:15,176
Rajat Bhageria: And they're very
different than. Exactly. we

1213
01:01:15,176 --> 01:01:18,247
basically had like different
policies per class and it was

1214
01:01:18,247 --> 01:01:21,318
an, it was like an inheritance
tree almost, right? So like

1215
01:01:21,318 --> 01:01:24,547
within a mash then you could
have like, a cauliflower mash is

1216
01:01:24,547 --> 01:01:27,618
probably pretty similar to a
potato mash. I just made that

1217
01:01:27,618 --> 01:01:30,635
up, but you get the point. So
that was good enough to get

1218
01:01:30,635 --> 01:01:33,918
started, right? like you said.
And then, of course, as you run

1219
01:01:33,918 --> 01:01:37,041
it more, more matches, you see
more pulled chicken than, or

1220
01:01:37,041 --> 01:01:40,165
pulled meats than, your entire
policy gets better. And then

1221
01:01:40,165 --> 01:01:43,447
over time, what we're trying to
do, of course, is to have more

1222
01:01:43,447 --> 01:01:46,676
of a generic, a unified policy,
right? And, of course, that's

1223
01:01:46,676 --> 01:01:49,800
where things like diffusion
policies, and, exactly. And, by

1224
01:01:49,800 --> 01:01:52,871
the way, the new stuff, right?
The diffusion policies, and

1225
01:01:52,871 --> 01:01:55,994
learning from demonstration, and
all of that work is really

1226
01:01:55,994 --> 01:01:57,000
important for that.

1227
01:01:57,000 --> 01:02:00,382
Audrow Nash: Ah, okay. I would
like to talk about that more

1228
01:02:00,382 --> 01:02:03,134
later. Let me write it down.
Yeah, learning from

1229
01:02:03,134 --> 01:02:06,000
demonstration and all that. It's
very interesting.

1230
01:02:06,000 --> 01:02:09,387
Audrow Nash: Are you guys using
learning from demonstration? Is

1231
01:02:09,387 --> 01:02:11,000
that how you're bootstrapping?

1232
01:02:11,000 --> 01:02:12,587
Rajat Bhageria: at the
beginning, no, because that

1233
01:02:12,587 --> 01:02:13,000
didn't exist.

1234
01:02:13,000 --> 01:02:14,000
Audrow Nash: because

1235
01:02:14,000 --> 01:02:15,000
Rajat Bhageria: but more,

1236
01:02:15,000 --> 01:02:18,123
Audrow Nash: pretty early,
although there are some pretty

1237
01:02:18,123 --> 01:02:21,247
compelling research results
lately, although I don't know

1238
01:02:21,247 --> 01:02:24,479
all the caveats to the demos
that I've seen and things like

1239
01:02:24,479 --> 01:02:27,000
this, and I keep finding more
when I look, but

1240
01:02:27,000 --> 01:02:30,420
Rajat Bhageria: In the early
days of Chef, like in 2019,

1241
01:02:30,420 --> 01:02:33,902
2020, this wasn't really,
obviously it existed, it wasn't

1242
01:02:33,902 --> 01:02:37,383
really good, right? Imitation
learning wasn't really, if,

1243
01:02:37,383 --> 01:02:41,109
imitation learning felt like
reinforcement learning, which is

1244
01:02:41,109 --> 01:02:44,957
yeah, we're going to do it, but
I didn't, I, honestly, I hadn't

1245
01:02:44,957 --> 01:02:48,744
really heard of companies who
had productionized, right? RL or

1246
01:02:48,744 --> 01:02:52,592
IL. Yeah. as of late though, I
think I think there's a lot more

1247
01:02:52,592 --> 01:02:56,317
exciting work. So what we are
doing now is if we're trying to

1248
01:02:56,317 --> 01:02:59,860
onboard a new, a completely new
class. So I'll give you an

1249
01:02:59,860 --> 01:03:03,586
example. we are now, we've now
productionized. Yeah. Actually

1250
01:03:03,586 --> 01:03:07,067
even more interesting, like
we're doing, picking like one

1251
01:03:07,067 --> 01:03:11,037
chicken breast, that, remember I
talked about like a meal, It has

1252
01:03:11,037 --> 01:03:14,335
all these like different
ingredients. Sometimes at the

1253
01:03:14,335 --> 01:03:17,878
top there's like a piece of
meat, right? A piece of salmon

1254
01:03:17,878 --> 01:03:18,000
or

1255
01:03:18,000 --> 01:03:21,000
Audrow Nash: delicious. Yes.

1256
01:03:21,000 --> 01:03:24,391
Rajat Bhageria: yeah, so,
before, we'd have to like,

1257
01:03:24,391 --> 01:03:27,978
bootstrap and then get
production data and there's this

1258
01:03:27,978 --> 01:03:32,087
whole thing you gotta do to like
onboard a new class that takes

1259
01:03:32,087 --> 01:03:36,000
some time. with LFD and IL,
that, that time is a lot shorter

1260
01:03:36,000 --> 01:03:39,000
Audrow Nash: demonstration and
what was that? What's IO?

1261
01:03:39,000 --> 01:03:40,000
Rajat Bhageria: imitation

1262
01:03:40,000 --> 01:03:43,000
Audrow Nash: Imitation learning.
IL?

1263
01:03:43,000 --> 01:03:48,000
Rajat Bhageria: you can have,
Yeah, you can have 60 or

1264
01:03:48,000 --> 01:03:49,000
Audrow Nash: You can

1265
01:03:48,000 --> 01:03:50,635
Rajat Bhageria: 70
demonstrations, you can speed it

1266
01:03:50,635 --> 01:03:53,683
up much faster. like you said,
it's definitely newer, but I

1267
01:03:53,683 --> 01:03:56,731
think we're finding some good
results there as well. we're,

1268
01:03:56,731 --> 01:03:59,934
thinking that's actually going
to be a really powerful way to,

1269
01:03:59,934 --> 01:04:02,000
really, really accelerate that
flywheel.

1270
01:04:02,000 --> 01:04:03,000
Audrow Nash: That's

1271
01:04:03,000 --> 01:04:05,458
Rajat Bhageria: Because we
already, and by the way, the, to

1272
01:04:05,458 --> 01:04:08,042
do learning from demonstration
well, it's a combination of the

1273
01:04:08,042 --> 01:04:10,458
demonstrations and production
data now. It's not just the,

1274
01:04:10,458 --> 01:04:11,000
it's not just

1275
01:04:11,000 --> 01:04:14,369
Audrow Nash: Oh, you have your
own demonstrations. That's

1276
01:04:14,369 --> 01:04:17,680
super, super cool. Yeah. So you
can choose a more stable

1277
01:04:17,680 --> 01:04:21,286
algorithm or something because
you have all of your data from

1278
01:04:21,286 --> 01:04:23,000
your robot actually doing it.

1279
01:04:23,000 --> 01:04:25,000
Rajat Bhageria: And it's
labeled, it's

1280
01:04:25,000 --> 01:04:27,943
Audrow Nash: and it's
supervised. Yeah. Oh, that's a

1281
01:04:27,943 --> 01:04:31,453
super cool thing. What do you
think? So in the recent learning

1282
01:04:31,453 --> 01:04:34,792
from demonstration work, I
haven't looked into it too much.

1283
01:04:34,792 --> 01:04:37,849
What's changed? Why is it
better? Is it, like all this

1284
01:04:37,849 --> 01:04:40,962
transformer? Are we using
transformers? And is that why

1285
01:04:40,962 --> 01:04:43,000
it's great now? Or what's,
happened?

1286
01:04:43,000 --> 01:04:46,081
Rajat Bhageria: I think,
transformers are quite powerful.

1287
01:04:46,081 --> 01:04:47,000
so I think, yeah,

1288
01:04:47,000 --> 01:04:50,388
Audrow Nash: that is it. Okay. I
didn't actually research into

1289
01:04:50,388 --> 01:04:53,776
it, but transformers, as far as
I understanding, give a better

1290
01:04:53,776 --> 01:04:57,000
weighted look back at context
from the past. is that a fair

1291
01:04:57,000 --> 01:05:00,377
Rajat Bhageria: yeah, exactly.
Exactly. And I think it, so I

1292
01:05:00,377 --> 01:05:03,416
think there's, I think
transformers are really helpful

1293
01:05:03,416 --> 01:05:06,623
because you have better, it's
using, it's, called, it's a

1294
01:05:06,623 --> 01:05:09,887
diffusion policy, right? So it's
similar to the same, it's

1295
01:05:09,887 --> 01:05:13,489
similar to the technology that,
something like a Sora might use,

1296
01:05:13,489 --> 01:05:16,866
or Obviously, DALI might use.
It's like you're generating an

1297
01:05:16,866 --> 01:05:19,567
image right here, you're
generating a policy and

1298
01:05:19,567 --> 01:05:22,944
predicting like the in joint
space like where what the joint

1299
01:05:22,944 --> 01:05:26,264
show of the robot should be,
right? I think the transformer

1300
01:05:26,264 --> 01:05:29,472
certainly helped. but I think
it's it's really a bunch of

1301
01:05:29,472 --> 01:05:33,130
things coming together, right? I
honestly it's like GPUs are also

1302
01:05:33,130 --> 01:05:36,675
getting better and processing. I
think the VR technology, to be

1303
01:05:36,675 --> 01:05:40,221
actually able to train, do the
demonstrations getting better. I

1304
01:05:40,221 --> 01:05:43,597
think there's a bunch of things,
but I think, yeah, like the

1305
01:05:43,597 --> 01:05:47,199
pivotal thing would probably be
transformers and diffusion as an

1306
01:05:47,199 --> 01:05:49,000
idea of really becoming a thing.

1307
01:05:49,000 --> 01:05:52,897
Audrow Nash: What a thing. Okay.
Yeah. It's amazing. I do feel

1308
01:05:52,897 --> 01:05:56,919
like, robotics is changing. Like
the technologies that we use as

1309
01:05:56,919 --> 01:06:00,816
an industry. are really gonna
the next 10 years is going to be

1310
01:06:00,816 --> 01:06:04,712
different than the last 10 years
and i know that kind of seems

1311
01:06:04,712 --> 01:06:08,358
obvious but it's just it's like
a lot of the things i even

1312
01:06:08,358 --> 01:06:12,128
learned in school which is not
that long ago i was in school

1313
01:06:12,128 --> 01:06:16,214
five years ago or something like
this are not used anymore and we

1314
01:06:16,214 --> 01:06:20,173
have different methods for doing
things It's crazy. some of the

1315
01:06:20,173 --> 01:06:24,133
low level stuff, say, ASTAR or
something, we still use that for

1316
01:06:24,133 --> 01:06:27,652
navigation. more optimal
planning in some state space or

1317
01:06:27,652 --> 01:06:31,486
something, if it's small, or
RRTs or something. But, a lot of

1318
01:06:31,486 --> 01:06:34,000
these perception and
manipulation tasks,

1319
01:06:34,000 --> 01:06:35,000
Rajat Bhageria: Yeah.

1320
01:06:35,000 --> 01:06:39,671
Audrow Nash: they just seem, it
seems like these, imitation

1321
01:06:39,671 --> 01:06:44,025
learning and all the machine
learning and AI, it's like

1322
01:06:44,025 --> 01:06:48,854
they're fundamentally different
approaches, and they enable a

1323
01:06:48,854 --> 01:06:52,971
lot more performance and
handling of Less Structured

1324
01:06:52,971 --> 01:06:54,000
Environments.

1325
01:06:54,000 --> 01:06:57,211
Rajat Bhageria: Yes. And, it's,
really interesting. So like I

1326
01:06:57,211 --> 01:07:00,421
alluded to, most of our team is
from autonomous vehicles. I'm

1327
01:07:00,421 --> 01:07:03,316
like, honestly, I'm not the
expert as much as they are,

1328
01:07:03,316 --> 01:07:06,316
right? the AV industry went
through this, right? Which is

1329
01:07:06,316 --> 01:07:09,684
like the AV industry for a while
is very much rule based, right?

1330
01:07:09,684 --> 01:07:12,789
Like you detect red light,
right? And then you stop, right?

1331
01:07:12,789 --> 01:07:16,053
Or you, detect the car and then
you track the car and then you

1332
01:07:16,053 --> 01:07:19,316
make sure you don't like, when
you're doing a lane change, you

1333
01:07:19,316 --> 01:07:22,474
don't collide with the car. I'm
simplifying, the point. it's

1334
01:07:22,474 --> 01:07:25,684
very much rules based. But then
what happens is that like the

1335
01:07:25,684 --> 01:07:28,947
world is, highly dimensional.
It's confusing and then you have

1336
01:07:28,947 --> 01:07:32,000
all these like It's just an
extremely kind of wrangly code

1337
01:07:32,000 --> 01:07:35,316
where like it's like all these
rules, right? And like of course

1338
01:07:35,316 --> 01:07:38,684
you can't get to every edge. And
by the way, there's still a lot

1339
01:07:38,684 --> 01:07:41,842
of machine learning It's a lot
of machine learning, but it's

1340
01:07:41,842 --> 01:07:45,000
like rules like it's you know,
detect the red light and then

1341
01:07:45,000 --> 01:07:48,706
Audrow Nash: Yeah, it's the
machine learning is a lot of

1342
01:07:48,706 --> 01:07:52,743
time for waiting things not for
changing policy like changing

1343
01:07:52,743 --> 01:07:54,000
what it does per se

1344
01:07:54,000 --> 01:07:57,179
Rajat Bhageria: And that same
thing was true for us, right? So

1345
01:07:57,179 --> 01:08:00,103
we started off that way, right?
Because that was what the

1346
01:08:00,103 --> 01:08:03,077
industry did, right? And if
you're doing robotics in 2020,

1347
01:08:03,077 --> 01:08:05,795
of course that's what you're
going to do. That's what

1348
01:08:05,795 --> 01:08:08,821
everyone does. and, I think that
was fine because it got us

1349
01:08:08,821 --> 01:08:12,051
started. But I think, what you
find is that you have these like

1350
01:08:12,051 --> 01:08:15,333
very gnarly policies. Like our,
policy at that point is like 200

1351
01:08:15,333 --> 01:08:18,000
parameters, right? Just to
manipulate an ingredient.

1352
01:08:18,000 --> 01:08:19,000
Audrow Nash: Yeah.

1353
01:08:18,000 --> 01:08:22,940
Rajat Bhageria: that works fine.
but then the. The, the customer,

1354
01:08:22,940 --> 01:08:27,728
like you go from like customer
one to two, and customer two has

1355
01:08:27,728 --> 01:08:31,832
the same kind of, let's say,
cucumbers. Yeah, but, and

1356
01:08:31,832 --> 01:08:36,240
exactly. And you can have this
inheritance idea, but it's,

1357
01:08:36,240 --> 01:08:37,000
different.

1358
01:08:37,000 --> 01:08:40,768
Audrow Nash: Yeah I think a lot
of times with this You make a

1359
01:08:40,768 --> 01:08:44,784
design architecture and then you
find out by some awful exception

1360
01:08:44,784 --> 01:08:48,552
that you need data routing from
one spot to another where you

1361
01:08:48,552 --> 01:08:51,950
didn't intend and it's
incredibly painful to get it and

1362
01:08:51,950 --> 01:08:55,718
then the whole thing becomes a
hairball if you would do it or

1363
01:08:55,718 --> 01:08:59,425
you have to redesign which is
painful especially if you have

1364
01:08:59,425 --> 01:09:03,193
your 200 conditions. So yeah,
whereas this you just learn and

1365
01:09:03,193 --> 01:09:07,023
it learns all the connections
for things and you don't have to

1366
01:09:07,023 --> 01:09:09,000
worry about that hairball of and

1367
01:09:09,000 --> 01:09:12,000
Rajat Bhageria: And I think one
thing, I think I think I, at

1368
01:09:12,000 --> 01:09:15,100
least from what I've understood,
a lot of the AV companies are

1369
01:09:15,100 --> 01:09:17,000
also like now going in this
direction,

1370
01:09:17,000 --> 01:09:18,000
Audrow Nash: Yeah,

1371
01:09:18,000 --> 01:09:21,172
Rajat Bhageria: I think Wave,
for example, Wave, I think, in,

1372
01:09:21,172 --> 01:09:24,448
they, just raised a big round,
and for a while, this end to end

1373
01:09:24,448 --> 01:09:27,724
learning idea was like, I don't
know, people are very skeptical

1374
01:09:27,724 --> 01:09:31,000
about it, and people still are
very skeptical about it, I think

1375
01:09:31,000 --> 01:09:34,587
Audrow Nash: I still have plenty
of skepticism about it, but I

1376
01:09:34,587 --> 01:09:38,000
think there's a lot of potential
too for specific problems.

1377
01:09:38,000 --> 01:09:40,584
Rajat Bhageria: And I think
Tesla FSD 12 is like one

1378
01:09:40,584 --> 01:09:43,466
interesting example of like many
people, many friends have

1379
01:09:43,466 --> 01:09:46,000
mentioned that they like, really
FSD 12, and that's

1380
01:09:46,000 --> 01:09:48,000
Audrow Nash: Is it the current
operating

1381
01:09:48,000 --> 01:09:49,627
Rajat Bhageria: that's the new
one, and apparently that's the

1382
01:09:49,627 --> 01:09:50,000
one where they

1383
01:09:50,000 --> 01:09:54,000
Audrow Nash: It is self driving.
It's at some level.

1384
01:09:54,000 --> 01:09:55,908
Rajat Bhageria: but apparently
they actually they cut 300, 000

1385
01:09:55,908 --> 01:09:57,754
lines of code and they're really
using imitation learning to

1386
01:09:57,754 --> 01:09:58,000
actually

1387
01:09:58,000 --> 01:10:02,357
Audrow Nash: it's crazy. Yeah,
what a cool thing. I can't wait

1388
01:10:02,357 --> 01:10:06,573
to, to try it because I see
great things about it, but okay.

1389
01:10:06,573 --> 01:10:10,578
So yeah, and that's an end to
end approach pretty much, I

1390
01:10:10,578 --> 01:10:11,000
think,

1391
01:10:11,000 --> 01:10:12,000
Rajat Bhageria: That's right.
That's right.

1392
01:10:12,000 --> 01:10:15,875
Audrow Nash: You have code to
hook up stuff, and then the rest

1393
01:10:15,875 --> 01:10:19,000
of it, like the policies
probably, mostly learned.

1394
01:10:19,000 --> 01:10:22,559
Rajat Bhageria: that's right.
And what's, nice, by the way, is

1395
01:10:22,559 --> 01:10:25,946
that we can do, I, think the
winner, I, really am convinced

1396
01:10:25,946 --> 01:10:29,505
about this now more than ever
before, is the winner in AI plus

1397
01:10:29,505 --> 01:10:32,604
robots is going to be the
companies that have the most

1398
01:10:32,604 --> 01:10:35,991
production data. you can and
should use Imitation learning,

1399
01:10:35,991 --> 01:10:38,000
like that is useful to get
started,

1400
01:10:38,000 --> 01:10:39,000
Audrow Nash: Yeah, but it's a

1401
01:10:38,000 --> 01:10:40,911
Rajat Bhageria: But at the end
of the day, it's really good for

1402
01:10:40,911 --> 01:10:43,914
bootstrapping, but at the end of
the day, every, in our industry,

1403
01:10:43,914 --> 01:10:46,917
like food, it applies for really
any robotics problem, but in the

1404
01:10:46,917 --> 01:10:49,827
food industry, every customer is
gonna have different broccoli.

1405
01:10:49,827 --> 01:10:52,230
They're gonna cut it
differently, they're gonna cook

1406
01:10:52,230 --> 01:10:55,001
it differently. It, it gets you
started, but you just simply

1407
01:10:55,001 --> 01:10:57,912
need to deploy lots of robots at
lots of customer sites. And by

1408
01:10:57,912 --> 01:11:00,453
the way, different, within the
same customer, different

1409
01:11:00,453 --> 01:11:03,364
customer sites are going to do
it differently. And, so you just

1410
01:11:03,364 --> 01:11:06,367
need a boatload of robots in the
world to really have enough data

1411
01:11:06,367 --> 01:11:09,000
to actually make this all,
really work too well together.

1412
01:11:09,000 --> 01:11:13,241
Audrow Nash: Yeah, it's a really
interesting approach. And it's,

1413
01:11:13,241 --> 01:11:17,349
the, challenge that I think it
becomes is how do you grow your

1414
01:11:17,349 --> 01:11:21,127
company through the stages as
you are collecting data and

1415
01:11:21,127 --> 01:11:25,102
automating more and more of the
process? Because you have to

1416
01:11:25,102 --> 01:11:29,277
find the clever, low impedance
way to grow your company because

1417
01:11:29,277 --> 01:11:33,187
you can't go let me go nothing
to solve all the way and I'm

1418
01:11:33,187 --> 01:11:37,229
going to deploy all this money
and lose it all to just try to

1419
01:11:37,229 --> 01:11:40,807
see if I have market fit or
something like you have to

1420
01:11:40,807 --> 01:11:42,000
gradually grow and

1421
01:11:42,000 --> 01:11:43,000
Rajat Bhageria: By the way,

1422
01:11:43,000 --> 01:11:44,000
Audrow Nash: what

1423
01:11:44,000 --> 01:11:46,927
Rajat Bhageria: you said,
something really interesting,

1424
01:11:46,927 --> 01:11:49,907
right? Which is basically You
essentially called out the

1425
01:11:49,907 --> 01:11:53,313
Cruise and Waymo approach, which
is let's raise tens of billions

1426
01:11:53,313 --> 01:11:56,612
of dollars And let's like just
jump to the end goal versus the

1427
01:11:56,612 --> 01:11:59,805
Tesla approach, which is hey,
let's just ship a vehicle It's

1428
01:11:59,805 --> 01:12:02,945
not gonna have any autopilot and
over time we're gonna have

1429
01:12:02,945 --> 01:12:06,137
billions of miles of FSD and
like Hopefully one day, like we

1430
01:12:06,137 --> 01:12:08,000
have a really good autonomy
system.

1431
01:12:08,000 --> 01:12:11,741
Audrow Nash: you're right yeah I
didn't mean it intentionally but

1432
01:12:11,741 --> 01:12:15,194
yeah I think that is a good it
is the different stories that

1433
01:12:15,194 --> 01:12:16,000
you see there.

1434
01:12:16,000 --> 01:12:17,000
Rajat Bhageria: exactly.

1435
01:12:17,000 --> 01:12:20,568
Audrow Nash: huh. And it's also
you guys, it's Electric Sheep.

1436
01:12:20,568 --> 01:12:23,791
you and Electric Sheep, to me,
are taking a very similar

1437
01:12:23,791 --> 01:12:27,302
approach, which I think makes a
lot of sense. of just getting

1438
01:12:27,302 --> 01:12:30,525
lots of data, deploying out
there, starting as simple as

1439
01:12:30,525 --> 01:12:33,000
possible, but then growing into
complexity.

1440
01:12:33,000 --> 01:12:36,706
Rajat Bhageria: That's right.
That's right. And, yeah, I'm,

1441
01:12:36,706 --> 01:12:40,412
getting, I think for startups,
look, if you're like there's

1442
01:12:40,412 --> 01:12:44,307
certain companies who can do the
Waymo approach. But, I really

1443
01:12:44,307 --> 01:12:48,075
think for startups, like you,
you like you said, you have to

1444
01:12:48,075 --> 01:12:51,844
find what's the Uber black,
right? what's the thing that you

1445
01:12:51,844 --> 01:12:55,299
can get to market or what's the
roadster or what's the.

1446
01:12:55,299 --> 01:12:58,000
Microsoft, MS DOS, like with
some, like the

1447
01:12:58,000 --> 01:13:01,000
Audrow Nash: start up that
flywheel.

1448
01:13:01,000 --> 01:13:03,937
Rajat Bhageria: start up that
flywheel. what's the Amazon,

1449
01:13:03,937 --> 01:13:05,000
selling books on the,

1450
01:13:05,000 --> 01:13:07,634
Audrow Nash: Yeah, your entry
point. Your beachhead or

1451
01:13:07,634 --> 01:13:09,000
whatever, to the rest of it.

1452
01:13:09,000 --> 01:13:10,000
Rajat Bhageria: That's right.

1453
01:13:09,000 --> 01:13:11,000
Audrow Nash: The thing,

1454
01:13:11,000 --> 01:13:13,850
Rajat Bhageria: and then you
have to grow that, but while

1455
01:13:13,850 --> 01:13:17,050
also telling the story, because
from, a fundraising perspective,

1456
01:13:17,050 --> 01:13:20,250
like Uber Block honestly is not
that exciting. Like another taxi

1457
01:13:20,250 --> 01:13:23,000
service for like rich people is
not that exciting. Yes,

1458
01:13:23,000 --> 01:13:25,400
Audrow Nash: Is that where they
started? Okay, I didn't know. I

1459
01:13:25,400 --> 01:13:27,000
don't know much of this history,
but okay.

1460
01:13:27,000 --> 01:13:30,835
Rajat Bhageria: basically had it
was literally a limousine. It

1461
01:13:30,835 --> 01:13:34,485
had people wearing like a nice
like suit. It was like super

1462
01:13:34,485 --> 01:13:38,443
high end, right? But it allowed
them to get started, and wealthy

1463
01:13:38,443 --> 01:13:42,278
folks in San Francisco used it.
But over time, of course, they

1464
01:13:42,278 --> 01:13:45,000
built UberX. They built
UberPool. UberX and,

1465
01:13:45,000 --> 01:13:46,000
Audrow Nash: and that's Tesla.

1466
01:13:46,000 --> 01:13:47,000
Rajat Bhageria: a hundred
billion and

1467
01:13:47,000 --> 01:13:50,625
Audrow Nash: It's interesting
because in my mind Like, okay,

1468
01:13:50,625 --> 01:13:54,432
there, I bet there, so I guess
there is nuance in this and that

1469
01:13:54,432 --> 01:13:58,178
there are very different ways of
starting this. You can do the

1470
01:13:58,178 --> 01:14:01,804
Uber Black or you can do the
Tesla approach where you go for

1471
01:14:01,804 --> 01:14:05,550
the luxury markets and that's a
way of spinning up production.

1472
01:14:05,550 --> 01:14:08,994
you can go, it's like the
project you're doing, as far as

1473
01:14:08,994 --> 01:14:12,498
I am aware, it's not like the
luxury and food part. It's a

1474
01:14:12,498 --> 01:14:16,124
very pragmatic, Useful part of
the food preparation one. you

1475
01:14:16,124 --> 01:14:19,628
have Amazon with their sending
books by mail. That's very,

1476
01:14:19,628 --> 01:14:23,375
that's not very, I guess maybe
that's rich, richer people that

1477
01:14:23,375 --> 01:14:27,000
are reading and want access to
their books. But I guess it's

1478
01:14:27,000 --> 01:14:28,000
Rajat Bhageria: I guess what,
it's not,

1479
01:14:28,000 --> 01:14:32,179
Audrow Nash: finding your market
initially that is a low

1480
01:14:32,179 --> 01:14:33,000
investment.

1481
01:14:33,000 --> 01:14:36,229
Rajat Bhageria: I think it's
finding the people. It's not

1482
01:14:36,229 --> 01:14:39,685
even like wealth per se, but
it's like finding the people who

1483
01:14:39,685 --> 01:14:43,197
have like big pain points. So in
the Amazon use case, It like,

1484
01:14:43,197 --> 01:14:46,653
what I mean by that is like in
the Amazon use case, like the,

1485
01:14:46,653 --> 01:14:49,995
pain point they were solving is
If I want a long tail book,

1486
01:14:49,995 --> 01:14:53,507
right? if I want like the, if I
were to rank, rank books, if I

1487
01:14:53,507 --> 01:14:56,000
want the 10, 000th most common
book, In the,

1488
01:14:56,000 --> 01:14:58,000
Audrow Nash: not in any
bookstores. Yeah.

1489
01:14:58,000 --> 01:15:01,048
Rajat Bhageria: yeah, exactly.
and I live in like Omaha,

1490
01:15:01,048 --> 01:15:04,367
Nebraska. Like I probably can't
go to the bookstore next door

1491
01:15:04,367 --> 01:15:07,469
and find that 10, 000th ranked
book. But Amazon's kind of

1492
01:15:07,469 --> 01:15:10,626
clever insight was if I can
aggregate all the books into a

1493
01:15:10,626 --> 01:15:13,946
central warehouse, then I can
have 20 copies of the 10, 000th

1494
01:15:13,946 --> 01:15:17,320
ranked book. Oh, you can have
all the longtail ones. So it was

1495
01:15:17,320 --> 01:15:20,422
like solving like a big pain
point for like book readers,

1496
01:15:20,422 --> 01:15:23,633
which is like I want this long
tailed book and now with the

1497
01:15:23,633 --> 01:15:26,898
internet I can take advantage of
it, right? I think for Uber

1498
01:15:26,898 --> 01:15:30,000
Black, it's it's a, different
analogy, but it's okay it's

1499
01:15:30,000 --> 01:15:32,000
Audrow Nash: would pay for this?
Rich people who

1500
01:15:32,000 --> 01:15:35,471
Rajat Bhageria: would pay for
it? That, yes. And I think for

1501
01:15:35,471 --> 01:15:39,058
Chef, it's given the technology
available today, which, what's

1502
01:15:39,058 --> 01:15:42,529
the market that has the biggest,
again, hair on fire problem

1503
01:15:42,529 --> 01:15:46,000
where I can actually add value,
which is food manufacturers?

1504
01:15:46,000 --> 01:15:49,875
Audrow Nash: Hell yeah. That
seems great. Let's see. And so

1505
01:15:49,875 --> 01:15:53,421
just, the last part of the
picture that I don't really

1506
01:15:53,421 --> 01:15:57,493
understand, or I guess, has have
some ambiguity around. So the

1507
01:15:57,493 --> 01:16:01,630
end effector or end effector to
the robot, it's mostly a scoop,

1508
01:16:01,630 --> 01:16:05,570
but you're looking into like
pinch grip or something. So you

1509
01:16:05,570 --> 01:16:08,000
can pick up chicken or different
meat

1510
01:16:08,000 --> 01:16:11,105
Rajat Bhageria: Yeah, it's a
good, it's a good question. So

1511
01:16:11,105 --> 01:16:14,421
the way you can think about it
is let me talk to you about what

1512
01:16:14,421 --> 01:16:17,789
it is now and like how we think
about it. so each robot, so each

1513
01:16:17,789 --> 01:16:21,158
of our systems, we make modules.
These modules are basically the

1514
01:16:21,158 --> 01:16:23,842
exact same hardware. Every
single customer, without

1515
01:16:23,842 --> 01:16:26,789
exception, has the same
hardware. That's like a hard and

1516
01:16:26,789 --> 01:16:27,000
fast

1517
01:16:27,000 --> 01:16:28,000
Audrow Nash: Love it. That's so
much

1518
01:16:28,000 --> 01:16:31,824
Rajat Bhageria: we don't want to
make maintenance, that's how you

1519
01:16:31,824 --> 01:16:34,000
have, that's how you build a
scalable

1520
01:16:34,000 --> 01:16:37,765
Audrow Nash: And also your data,
like it makes it easier for you

1521
01:16:37,765 --> 01:16:41,000
to train your data because you
all have the same setup.

1522
01:16:41,000 --> 01:16:43,000
Rajat Bhageria: It's, more
homogenous, like you said,

1523
01:16:43,000 --> 01:16:44,000
Audrow Nash: You need a good

1524
01:16:43,000 --> 01:16:45,000
Rajat Bhageria: a geometry
perspective.

1525
01:16:45,000 --> 01:16:48,437
Audrow Nash: probably for oh,
you want to swap your sensor or

1526
01:16:48,437 --> 01:16:51,479
something like that. I'm
actually, I'm very curious to

1527
01:16:51,479 --> 01:16:54,915
how you solve that problem, but
I'll ask that in a bit. okay.

1528
01:16:54,915 --> 01:16:57,000
You have this homogenous setup.
Yeah.

1529
01:16:57,000 --> 01:17:00,342
Rajat Bhageria: the same, it's
the same footprint as a person,

1530
01:17:00,342 --> 01:17:03,522
and then, from an end effector
perspective, we, each of our

1531
01:17:03,522 --> 01:17:06,809
robots, without exception, has
the same actuator. Actuator as

1532
01:17:06,809 --> 01:17:09,828
in the pneumatic actuator
itself. And then that actuator

1533
01:17:09,828 --> 01:17:13,331
has different, has an interface,
and on that interface, we have a

1534
01:17:13,331 --> 01:17:16,672
set of different utensils, and
we have a library, essentially,

1535
01:17:16,672 --> 01:17:19,852
of different utensils, and this
library, what's interesting

1536
01:17:19,852 --> 01:17:23,194
about this library is that, It
scales very logarithmically. In

1537
01:17:23,194 --> 01:17:26,158
other words, there's, like I
said, there's trillions of

1538
01:17:26,158 --> 01:17:29,554
different ingredients, and the
way that a particular ingredient

1539
01:17:29,554 --> 01:17:32,734
was cut and prepped yesterday,
it will never exist in human

1540
01:17:32,734 --> 01:17:35,968
history probably ever again.
That particular way, right? But

1541
01:17:35,968 --> 01:17:39,040
that doesn't mean you need
thousands of end effectors and

1542
01:17:39,040 --> 01:17:42,543
thousands of utensils. We have a
small class of utensils, we mass

1543
01:17:42,543 --> 01:17:45,400
manufacture them, we have
patents around all of them,

1544
01:17:45,400 --> 01:17:48,741
that's it, right? we send the
entire library to our customers,

1545
01:17:48,741 --> 01:17:52,083
so each customer has an entire
library, each customer gets the

1546
01:17:52,083 --> 01:17:55,425
same hardware, each customer has
the same code base, software,

1547
01:17:55,425 --> 01:17:58,604
and then of course there's
different policies based on what

1548
01:17:58,604 --> 01:18:01,299
you need for your ingredients
that you're actually

1549
01:18:01,299 --> 01:18:02,000
manipulating.

1550
01:18:02,000 --> 01:18:06,037
Audrow Nash: Yeah. Okay. So
almost everything is the same,

1551
01:18:06,037 --> 01:18:10,282
except the end effector may vary
with the food and the policy

1552
01:18:10,282 --> 01:18:14,040
varies with the food most
likely. And so that lets you

1553
01:18:14,040 --> 01:18:18,286
control a lot and you get to
reuse a lot of the same code and

1554
01:18:18,286 --> 01:18:21,000
everything like this, and then
you just

1555
01:18:21,000 --> 01:18:22,000
Rajat Bhageria: It also means
that it can evolve,

1556
01:18:22,000 --> 01:18:23,000
Audrow Nash: it

1557
01:18:23,000 --> 01:18:26,132
Rajat Bhageria: brain gets
better. The brain gets better.

1558
01:18:26,132 --> 01:18:29,429
if you have too much bespoke
stuff, it seems like, sometimes

1559
01:18:29,429 --> 01:18:32,780
we'll talk to our customers and
we're like, ah, you're, like,

1560
01:18:32,780 --> 01:18:36,077
They're like, oh, you're not
making me custom stuff. But no,

1561
01:18:36,077 --> 01:18:39,264
that's not, that's the whole
point. The fact that it's not

1562
01:18:39,264 --> 01:18:42,176
custom means that it's, your
thing's gonna get better

1563
01:18:42,176 --> 01:18:43,000
literally every

1564
01:18:43,000 --> 01:18:47,309
Audrow Nash: Yeah, that's so
cool. What a funny thing where

1565
01:18:47,309 --> 01:18:51,545
it's no, Because we do this
generalization, we are able to

1566
01:18:51,545 --> 01:18:56,000
benefit from everything that
occurs. And, that's really nice.

1567
01:18:56,000 --> 01:18:57,000
Rajat Bhageria: Yes.

1568
01:18:57,000 --> 01:19:01,072
Audrow Nash: How do you, so
what's your, do you have a plan

1569
01:19:01,072 --> 01:19:04,730
for how to upgrade your
hardware? So if say, I'm sure

1570
01:19:04,730 --> 01:19:08,870
you basically, you did a ton of
research before. So that you

1571
01:19:08,870 --> 01:19:12,459
could lock in and feel
confident, look at the supply

1572
01:19:12,459 --> 01:19:16,255
chains, look at the cost, look
at the ecosystem and the

1573
01:19:16,255 --> 01:19:20,189
support. So lock that in a
really good one. But a painful

1574
01:19:20,189 --> 01:19:24,330
thing that I have not yet seen a
very good answer to, is the

1575
01:19:24,330 --> 01:19:28,401
eventual process of upgrading.
So you have, yeah, it's five

1576
01:19:28,401 --> 01:19:32,611
years later and a new camera
comes out. And I guess Maybe you

1577
01:19:32,611 --> 01:19:36,476
can mix the fleet or something
like that or something or

1578
01:19:36,476 --> 01:19:40,548
provide actually probably just
put both on and then you can

1579
01:19:40,548 --> 01:19:44,689
relate it but do you have a do
you have a good upgrade story

1580
01:19:44,689 --> 01:19:46,000
yet or what are you

1581
01:19:46,000 --> 01:19:49,444
Rajat Bhageria: Yeah. Yeah,
totally. upgrades are, updates

1582
01:19:49,444 --> 01:19:53,125
and upgrades are both part of
the robotics as a service story.

1583
01:19:53,125 --> 01:19:56,688
So maybe we talk, about both of
them because I think they're

1584
01:19:56,688 --> 01:20:00,428
both relevant. here's the way we
think about it, right? So like

1585
01:20:00,428 --> 01:20:04,109
the hardware itself is mostly
off the shelf for us. Almost all

1586
01:20:04,109 --> 01:20:05,000
of it's off the

1587
01:20:05,000 --> 01:20:07,000
Audrow Nash: Love it.

1588
01:20:07,000 --> 01:20:10,605
Rajat Bhageria: Why do I, call
that out? there's a bunch of

1589
01:20:10,605 --> 01:20:14,149
benefits you get from that.
Number one, each component has

1590
01:20:14,149 --> 01:20:17,815
been productized by a different
company and these are really

1591
01:20:17,815 --> 01:20:21,420
great manufacturers, right? So
Our, CPU, or our GPU, or our

1592
01:20:21,420 --> 01:20:25,086
HMI, the touchscreen, or our
light tower, each, there's some

1593
01:20:25,086 --> 01:20:28,874
manufacturer who's world class
at making these components, and

1594
01:20:28,874 --> 01:20:32,418
so what that means is our mean
time before failure for the

1595
01:20:32,418 --> 01:20:36,022
hardware itself is quite high.
In other words, the hardware

1596
01:20:36,022 --> 01:20:37,000
actually doesn't

1597
01:20:37,000 --> 01:20:39,000
Audrow Nash: it's robust for
sure.

1598
01:20:39,000 --> 01:20:41,980
Rajat Bhageria: It's robust.
Number two, another benefit of

1599
01:20:41,980 --> 01:20:44,000
that is that we have built our
software,

1600
01:20:44,000 --> 01:20:45,000
Audrow Nash: on it.

1601
01:20:45,000 --> 01:20:48,863
Rajat Bhageria: yes, and the
software is built to be hardware

1602
01:20:48,863 --> 01:20:52,790
agnostic. So what we can do is
we have three or four different

1603
01:20:52,790 --> 01:20:56,780
vendors per component. So if any
particular component vendor is

1604
01:20:56,780 --> 01:21:00,833
having some supply chain issues
or whatever, great, we can go to

1605
01:21:00,833 --> 01:21:04,000
them. So that's really nice.
that also by the way,

1606
01:21:04,000 --> 01:21:07,289
Audrow Nash: Like I'm imagining,
did you, is it like you have

1607
01:21:07,289 --> 01:21:10,578
three different vendors where
you can buy depth cameras from?

1608
01:21:10,578 --> 01:21:13,436
And it's a, it's all so one,
you're using an Oakty, a

1609
01:21:13,436 --> 01:21:15,000
RealSense and something else.

1610
01:21:15,000 --> 01:21:18,964
Rajat Bhageria: the code base is
built to be hardware agnostic.

1611
01:21:18,964 --> 01:21:22,865
So like basically there's like a
capture node, right? And like

1612
01:21:22,865 --> 01:21:26,514
the camera capture node can,
can, have the, core code, the

1613
01:21:26,514 --> 01:21:30,352
core, like logic is agnostic to
the hardware, but there's one

1614
01:21:30,352 --> 01:21:33,938
file, which is okay, like the
drivers for that particular

1615
01:21:33,938 --> 01:21:37,839
camera, if that makes sense. And
then you can say, okay, this,

1616
01:21:37,839 --> 01:21:41,740
instance is using this. Camera
node, basically. Camera capture

1617
01:21:41,740 --> 01:21:45,767
node. Or that same thing is true
for scales. we can very quickly

1618
01:21:45,767 --> 01:21:49,416
bring up new hardware in our
stack. Or new, robot arms. Or

1619
01:21:49,416 --> 01:21:53,003
new sensors. Or maybe we want
another camera. we can very

1620
01:21:53,003 --> 01:21:56,337
quickly bring up new sensors,
trivially. Not, totally

1621
01:21:56,337 --> 01:21:59,420
trivially, but basically
trivially. Because we've

1622
01:21:59,420 --> 01:22:02,000
designed the system to be like,
extremely

1623
01:22:02,000 --> 01:22:05,292
Audrow Nash: yeah, so you can
just sub out parts of it. How

1624
01:22:05,292 --> 01:22:08,472
does that work with your
training data? Or does it matter

1625
01:22:08,472 --> 01:22:11,541
because the things you're
labeling are, you can control

1626
01:22:11,541 --> 01:22:15,000
everything anyways. Like you're
just getting an RGBD image and

1627
01:22:15,000 --> 01:22:17,683
Rajat Bhageria: Yes. I think
it's a, Very good question.

1628
01:22:17,683 --> 01:22:20,557
Right now, like I'm saying we
could do this. Honestly, we've

1629
01:22:20,557 --> 01:22:23,000
been pretty happy with our
components so far. There

1630
01:22:23,000 --> 01:22:24,000
Audrow Nash: Yeah.

1631
01:22:23,000 --> 01:22:25,677
Rajat Bhageria: need, but, you'd
be right. that would be a

1632
01:22:25,677 --> 01:22:26,000
concern

1633
01:22:26,000 --> 01:22:27,000
Audrow Nash: Okay, yeah.

1634
01:22:27,000 --> 01:22:28,000
Rajat Bhageria: about.

1635
01:22:28,000 --> 01:22:31,179
Audrow Nash: But I get what
you're saying where you've built

1636
01:22:31,179 --> 01:22:34,358
it in a way that everything is
modular and that's definitely

1637
01:22:34,358 --> 01:22:37,589
like as far as I'm aware that's
the best practice for sure to

1638
01:22:37,589 --> 01:22:40,768
make it so that you're not like
super coupled to you're just

1639
01:22:40,768 --> 01:22:43,682
using like general ROS
architecture principles which is

1640
01:22:43,682 --> 01:22:44,000
great.

1641
01:22:44,000 --> 01:22:47,839
Rajat Bhageria: yes, exactly. by
the way, this very well leads to

1642
01:22:47,839 --> 01:22:51,677
service and upgrades. okay, from
a service perspective, let's say

1643
01:22:51,677 --> 01:22:55,339
some part of the system breaks.
And of course, what we have, a

1644
01:22:55,339 --> 01:22:59,000
pareto of, hardware failures,
that we try our customer support

1645
01:22:59,000 --> 01:23:02,199
Audrow Nash: You mean a Pareto
distribution? Is that what

1646
01:23:02,199 --> 01:23:05,735
you're referring to? Nice. So 80
20 rule kind of thing? Okay so

1647
01:23:05,735 --> 01:23:09,327
20 percent will fail 80 percent
of the time. Is that what you're

1648
01:23:09,327 --> 01:23:10,000
saying? Okay

1649
01:23:10,000 --> 01:23:12,000
Rajat Bhageria: Yes. Yes.

1650
01:23:12,000 --> 01:23:15,379
Rajat Bhageria: and so then what
we can do is we can have FRUZE,

1651
01:23:15,379 --> 01:23:18,018
field replaceable units for
things that might fail

1652
01:23:18,018 --> 01:23:21,133
frequently. And oftentimes those
FRUZE will, just ship them

1653
01:23:21,133 --> 01:23:24,459
initially with the deploy, like
the, a new robot. like we only,

1654
01:23:24,459 --> 01:23:27,626
we, we usually, ship starter
packs, six robots at a time for

1655
01:23:27,626 --> 01:23:30,952
a new customer. So like you get
a starter pack, we'll, ship you

1656
01:23:30,952 --> 01:23:34,172
some spare parts, from the get
go, right? And then of course,

1657
01:23:34,172 --> 01:23:37,392
like. we'll also have them at
our HQ so we can overnight them

1658
01:23:37,392 --> 01:23:40,507
if, something actually needs a
break. By the way, this is a

1659
01:23:40,507 --> 01:23:43,885
very rare occurrence because our
meantime before failure is like

1660
01:23:43,885 --> 01:23:47,000
180 days. So it's it's not, like
hardware's failing all the

1661
01:23:47,000 --> 01:23:49,000
Audrow Nash: awesome. Especially
for

1662
01:23:49,000 --> 01:23:50,000
Rajat Bhageria: then if
something

1663
01:23:50,000 --> 01:23:53,587
Audrow Nash: how early, like you
guys are clearly put together

1664
01:23:53,587 --> 01:23:57,289
and everything, but it's you're
still beginning, like you're not

1665
01:23:57,289 --> 01:24:00,760
at, you're not at scaling just
yet, I think. And so the fact

1666
01:24:00,760 --> 01:24:04,000
that you're failing this
infrequently is just wonderful.

1667
01:24:04,000 --> 01:24:05,000
Rajat Bhageria: yes. Yeah.

1668
01:24:05,000 --> 01:24:07,564
Rajat Bhageria: I, it's funny,
like we've stayed under the

1669
01:24:07,564 --> 01:24:10,171
radar quite a bit, and now is
really when we're like really

1670
01:24:10,171 --> 01:24:12,823
starting to scale, which is
cool. And we can talk more about

1671
01:24:12,823 --> 01:24:13,000
that

1672
01:24:13,000 --> 01:24:14,000
Audrow Nash: Yeah, sure.

1673
01:24:13,000 --> 01:24:16,252
Rajat Bhageria: yeah. but yeah,
back to, the service and

1674
01:24:16,252 --> 01:24:19,910
upgrades, so then what we do is
we actually have partnered with

1675
01:24:19,910 --> 01:24:23,510
a third party service org. these
guys have 10, 000 technicians

1676
01:24:23,510 --> 01:24:26,645
all over the world. from a
scaling perspective, we can

1677
01:24:26,645 --> 01:24:30,419
deploy, and that's why we talked
about Europe a little bit, We're

1678
01:24:30,419 --> 01:24:31,000
not super,

1679
01:24:31,000 --> 01:24:32,000
Audrow Nash: Geographically

1680
01:24:32,000 --> 01:24:34,344
Rajat Bhageria: there's certain
things, yeah, like certainly

1681
01:24:34,344 --> 01:24:36,688
there's things to be careful
about because of course we want

1682
01:24:36,688 --> 01:24:37,000
the best

1683
01:24:37,000 --> 01:24:39,392
Audrow Nash: I would think their
data policy and stuff might be

1684
01:24:39,392 --> 01:24:40,000
things you might

1685
01:24:40,000 --> 01:24:42,476
Rajat Bhageria: exactly. so
there's, legal financial

1686
01:24:42,476 --> 01:24:44,000
ramifications that do need to be

1687
01:24:44,000 --> 01:24:46,400
Audrow Nash: but you could
literally do it. Yeah. From an

1688
01:24:46,400 --> 01:24:48,000
operations point, you're fine
with it.

1689
01:24:48,000 --> 01:24:51,535
Rajat Bhageria: yes, that's
right. and so what'll happen, is

1690
01:24:51,535 --> 01:24:54,834
we have a really nice repair
manual, like a giant repair

1691
01:24:54,834 --> 01:24:58,369
manual, And we'll train the
trainer, if you will, like we'll

1692
01:24:58,369 --> 01:25:01,846
train their, I'm a person who
trains other people how to do

1693
01:25:01,846 --> 01:25:05,439
these repairs with these crews
and Anytime something actually

1694
01:25:05,439 --> 01:25:09,151
fails. We'll say hey, we got an
alert that something failed the

1695
01:25:09,151 --> 01:25:12,627
customer told us something
failed They'll drive out an hour

1696
01:25:12,627 --> 01:25:16,280
because everything is an hour
away from wherever they might be

1697
01:25:16,280 --> 01:25:19,639
in the country and or North
America And they'll change it

1698
01:25:19,639 --> 01:25:23,056
apart. And if they need to,
they'll FaceTime us. So that's

1699
01:25:23,056 --> 01:25:25,000
like servicing, for any
hardware.

1700
01:25:25,000 --> 01:25:28,580
Rajat Bhageria: And then from an
upgrades perspective, I think

1701
01:25:28,580 --> 01:25:32,217
it's a very similar thing. So we
do these like ECOs, basically,

1702
01:25:32,217 --> 01:25:35,739
engineering change orders, where
it's okay, we, we'd like, so

1703
01:25:35,739 --> 01:25:39,261
I'll give you an interesting
example, We, we had these scales

1704
01:25:39,261 --> 01:25:42,783
that we had shipped and they
were like IP67, which is already

1705
01:25:42,783 --> 01:25:46,074
fairly high water and grass
rating. but we're in the food

1706
01:25:46,074 --> 01:25:49,076
industry and people like love
water. They just spray

1707
01:25:49,076 --> 01:25:50,000
everything down.

1708
01:25:50,000 --> 01:25:52,000
Audrow Nash: Yeah. Yeah.

1709
01:25:52,000 --> 01:25:55,993
Rajat Bhageria: was getting
water damage and, that was right

1710
01:25:55,993 --> 01:25:59,852
when we deployed, like early on
when we deployed, this was

1711
01:25:59,852 --> 01:26:04,044
happening. So we did an ECO and
we changed out to IP69K scales.

1712
01:26:04,044 --> 01:26:08,303
And, it's pretty, pretty simple.
We just ship them out, and, you

1713
01:26:08,303 --> 01:26:12,030
just install them and, again,
it's very, we already have

1714
01:26:12,030 --> 01:26:15,823
another, node for, the new,
driver for the new scales and

1715
01:26:15,823 --> 01:26:19,549
you're up and running. I think
upgrades, because it's so

1716
01:26:19,549 --> 01:26:23,608
modular, it's not for example,
there's no PCBs at Chef. it's,

1717
01:26:23,608 --> 01:26:27,401
electronics are all off the
shelf, right? It's, there's a

1718
01:26:27,401 --> 01:26:28,000
DIN rail.

1719
01:26:28,000 --> 01:26:29,000
Audrow Nash: you

1720
01:26:29,000 --> 01:26:30,000
Rajat Bhageria: like a very
simple,

1721
01:26:30,000 --> 01:26:31,000
Audrow Nash: That's it.

1722
01:26:30,000 --> 01:26:33,179
Rajat Bhageria: it's extremely,
like if you were to look at it,

1723
01:26:33,179 --> 01:26:35,904
it's an extremely simple
hardware system. Which by the

1724
01:26:35,904 --> 01:26:38,881
way, it seems Simplicity, I
think, is really hard to do. So

1725
01:26:38,881 --> 01:26:41,656
I think there's a lot of
complexity in the software, by

1726
01:26:41,656 --> 01:26:44,683
the way. There's, because we
don't have hardware complexity,

1727
01:26:44,683 --> 01:26:47,560
we have a lot of software
complexity, but from a hardware

1728
01:26:47,560 --> 01:26:50,739
perspective, then if we need to
change out the tech, like we're

1729
01:26:50,739 --> 01:26:53,716
going to, there's a new PCAP
display that we're going to do

1730
01:26:53,716 --> 01:26:56,743
an ECO for. It's like much, much
more reactive. There's like

1731
01:26:56,743 --> 01:26:59,619
multi touch and it's nice
because in food facilities, you

1732
01:26:59,619 --> 01:27:02,193
have like gloves. So it's so
anyways, there's some,

1733
01:27:02,193 --> 01:27:05,119
oftentimes these gloves have
food in them. So there's some

1734
01:27:05,119 --> 01:27:07,894
practical reasons why we're
doing this. But you want to

1735
01:27:07,894 --> 01:27:10,922
change the touch screen, no big
deal. You want to change the

1736
01:27:10,922 --> 01:27:14,000
camera, no big deal. new GPU,
touch it onto the dendril. it's

1737
01:27:14,000 --> 01:27:17,454
Audrow Nash: that's really nice.
Yeah, I would. The only concern

1738
01:27:17,454 --> 01:27:20,961
that I would have with new parts
would be that it would hurt your

1739
01:27:20,961 --> 01:27:24,199
ability to reuse past training
data. But I guess that you're

1740
01:27:24,199 --> 01:27:27,059
working on a number of
platforms, that is going to be

1741
01:27:27,059 --> 01:27:30,189
in your favor because they're
all going to be a little bit

1742
01:27:30,189 --> 01:27:33,265
different. and then I would
imagine if there's maybe it's

1743
01:27:33,265 --> 01:27:36,610
just good enough you just pass
the image and it figures it out

1744
01:27:36,610 --> 01:27:39,524
but I would imagine say it's
like a slightly different

1745
01:27:39,524 --> 01:27:42,438
resolution or the colors are a
little bit different or

1746
01:27:42,438 --> 01:27:45,514
something like that I could
imagine that the camera could

1747
01:27:45,514 --> 01:27:46,000
mess with

1748
01:27:46,000 --> 01:27:49,263
Rajat Bhageria: itself is quite
important, yes. No, you're 100

1749
01:27:49,263 --> 01:27:51,000
percent you're 100 percent
right.

1750
01:27:51,000 --> 01:27:53,854
Audrow Nash: the rest of it yeah
it makes perfect sense that you

1751
01:27:53,854 --> 01:27:56,662
could swap the touch screen and
that wouldn't have any negative

1752
01:27:56,662 --> 01:27:58,000
impacts for this kind of thing

1753
01:27:58,000 --> 01:28:00,970
Rajat Bhageria: By the way, one
thing that's also pretty cool

1754
01:28:00,970 --> 01:28:03,940
that's happening in the world,
which I'm sure you've seen, is

1755
01:28:03,940 --> 01:28:06,861
like there's like exponentially
decreasing prices for stuff.

1756
01:28:06,861 --> 01:28:09,734
Like back in the day, like it
used to cost 60, 000 to buy a

1757
01:28:09,734 --> 01:28:12,412
robot arm. Then of course
Universal Robots had this big

1758
01:28:12,412 --> 01:28:15,431
breakthrough with collaborative
robots and like they got it on

1759
01:28:15,431 --> 01:28:18,109
to 30 or so, right? And now
there's a bunch of these, a

1760
01:28:18,109 --> 01:28:21,030
bunch of new companies like
Doosan Robotics is one that just

1761
01:28:21,030 --> 01:28:24,000
went public in Korea and like
they have this really, nice arm

1762
01:28:24,000 --> 01:28:26,727
that's like substantially
cheaper, right? And There's a,

1763
01:28:26,727 --> 01:28:29,794
and by the way, that's just like
I'm talking about, robot arms,

1764
01:28:29,794 --> 01:28:32,715
but even the Intel RealSense
camera is 200, which is insane.

1765
01:28:32,715 --> 01:28:35,637
the Jetson suite at NVIDIA,
there's all these components and

1766
01:28:35,637 --> 01:28:37,000
the cost is so much cheaper,

1767
01:28:37,000 --> 01:28:38,000
Audrow Nash: It's

1768
01:28:37,000 --> 01:28:39,577
Rajat Bhageria: right? And I
think that's really cool, too. I

1769
01:28:39,577 --> 01:28:42,239
think the hardware, people, talk
all about, we're talking about

1770
01:28:42,239 --> 01:28:44,690
transformers and, all this
stuff. the software's obviously

1771
01:28:44,690 --> 01:28:46,972
getting really good, but, I
think the hardware is also

1772
01:28:46,972 --> 01:28:49,000
getting much better and cheaper,
which is really

1773
01:28:49,000 --> 01:28:52,733
Audrow Nash: Totally. And you
need it to do things in the

1774
01:28:52,733 --> 01:28:56,729
physical world. Like you need an
expensive sensors and things

1775
01:28:56,729 --> 01:29:00,920
like this. I don't know. I just,
one of the things that blows me

1776
01:29:00,920 --> 01:29:04,785
away is I was, an undergrad
maybe 10 years ago or something

1777
01:29:04,785 --> 01:29:08,322
like this. And I did an
internship at a company and we

1778
01:29:08,322 --> 01:29:12,382
were using an RTK GPS. and it
was, Like a 10k GPS or something

1779
01:29:12,382 --> 01:29:16,247
like that like it was absurdly
expensive and don't quote my

1780
01:29:16,247 --> 01:29:20,242
numbers But ballpark of what I
remember and now companies are

1781
01:29:20,242 --> 01:29:23,713
getting them for 100 or
something like really low and

1782
01:29:23,713 --> 01:29:27,774
it's really amazing. So you're
seeing that fall, especially, I

1783
01:29:27,774 --> 01:29:31,835
think part of it is maybe more
efficient processes for things.

1784
01:29:31,835 --> 01:29:36,027
And then another part is larger
markets like robotics is growing

1785
01:29:36,027 --> 01:29:39,826
up and all the cars are using
sensors and because of that,

1786
01:29:39,826 --> 01:29:43,166
there's more demands for
sensors. So then it's more

1787
01:29:43,166 --> 01:29:45,000
efficient is what I imagine.

1788
01:29:45,000 --> 01:29:49,293
Rajat Bhageria: And, by the way,
I really hope, I really hope

1789
01:29:49,293 --> 01:29:53,585
that this like recent advance,
or I think ChatGPT was really,

1790
01:29:53,585 --> 01:29:57,737
awesome actually, because it,
like inspired the world to do

1791
01:29:57,737 --> 01:30:01,819
AI. Like, before then, if you
remember, like during COVID,

1792
01:30:01,819 --> 01:30:04,000
crypto was a cool thing, right?

1793
01:30:04,000 --> 01:30:06,000
Audrow Nash: yeah,

1794
01:30:06,000 --> 01:30:08,000
Rajat Bhageria: and

1795
01:30:08,000 --> 01:30:10,000
Audrow Nash: I love the point.
Yeah, definitely.

1796
01:30:10,000 --> 01:30:13,483
Rajat Bhageria: I think yeah,
like I, it's a broader point,

1797
01:30:13,483 --> 01:30:17,262
which is I think the world needs
a, the world needs like a, like

1798
01:30:17,262 --> 01:30:20,863
a figurehead, right? It's like
the four minute mile. Somebody

1799
01:30:20,863 --> 01:30:24,170
needs to do the four minute
mile, and then everyone else

1800
01:30:24,170 --> 01:30:27,712
comes after. In EVs, that was
Tesla. Tesla did. EVs, and now

1801
01:30:27,712 --> 01:30:30,959
everyone's doing EVs. In
semiconductors, there's Intel.

1802
01:30:30,959 --> 01:30:34,325
Intel, really, productionized
the, the transistor into an

1803
01:30:34,325 --> 01:30:37,808
integrated circuit and then sold
them, and then, of course,

1804
01:30:37,808 --> 01:30:41,173
there's lots of people who are
doing integrated circuits.

1805
01:30:41,173 --> 01:30:44,716
Facebook did this for social
networks, and Uber did this for

1806
01:30:44,716 --> 01:30:48,317
marketplaces. I think OpenAI did
this for AI. Now, AI, the AI

1807
01:30:48,317 --> 01:30:51,506
they're doing is more large
language models, but they,

1808
01:30:51,506 --> 01:30:55,166
they've woke the world up when
it comes to, AI. AI, which is I

1809
01:30:55,166 --> 01:30:58,531
think really good because now
there's a lot of engineers,

1810
01:30:58,531 --> 01:31:01,956
entrepreneurs, operators,
investors, writers, journalists,

1811
01:31:01,956 --> 01:31:05,439
everyone else who's thinking
about AI, which is really good

1812
01:31:05,439 --> 01:31:08,627
for the world. And you know
what's interesting is that

1813
01:31:08,627 --> 01:31:12,347
what's the natural progression
of this like Hype in, or yeah, I

1814
01:31:12,347 --> 01:31:15,889
guess it is hype. what's an
actual progression? I think it's

1815
01:31:15,889 --> 01:31:19,432
two things, right? one is like
people are more interested in

1816
01:31:19,432 --> 01:31:22,797
energy, right? Like how do you
actually get the energy to

1817
01:31:22,797 --> 01:31:26,399
actually power these models,
right? they're really interested

1818
01:31:26,399 --> 01:31:29,646
in compute. So GPUs is a very
natural extension. And of

1819
01:31:29,646 --> 01:31:33,129
course, what's the third one?
Robotics. It's you, take the,

1820
01:31:33,129 --> 01:31:36,317
you take these models and you
apply them in a physical

1821
01:31:36,317 --> 01:31:39,565
environment. So I'm hopeful
that. I'm hopeful that this

1822
01:31:39,565 --> 01:31:43,048
energy that we have around AI
right now will translate to a

1823
01:31:43,048 --> 01:31:46,295
lot more breakthroughs in
robotics, but also just a lot

1824
01:31:46,295 --> 01:31:49,661
more people doing robotics and
deploying successfully and

1825
01:31:49,661 --> 01:31:53,498
scaling robotics companies. And,
hopefully if, there's at least a

1826
01:31:53,498 --> 01:31:56,923
couple winners, that really
scale hundreds of thousands of

1827
01:31:56,923 --> 01:32:00,229
robots, everyone else, it'll,
open the floodgates. It'll

1828
01:32:00,229 --> 01:32:02,000
inspire other people to do it.

1829
01:32:02,000 --> 01:32:06,145
Audrow Nash: Yeah, I think it's
very true. I also think there's

1830
01:32:06,145 --> 01:32:10,092
additional market forces that
are making it so that robotics

1831
01:32:10,092 --> 01:32:14,237
is heavily invested in, like the
labor shortages you mentioned.

1832
01:32:14,237 --> 01:32:17,855
Because there is the point
where, what was the You said

1833
01:32:17,855 --> 01:32:21,539
something like there's 30
percent labor shortage or what

1834
01:32:21,539 --> 01:32:22,000
was the

1835
01:32:22,000 --> 01:32:26,000
Rajat Bhageria: Yeah, exactly.

1836
01:32:26,000 --> 01:32:29,994
Audrow Nash: that might get
worse I think it's likely that

1837
01:32:29,994 --> 01:32:34,332
gets worse. a lot of people are
retiring The biggest cohort the

1838
01:32:34,332 --> 01:32:38,602
baby boomers are retiring and so
they were largely blue collar

1839
01:32:38,602 --> 01:32:42,802
workers or a lot of them were
blue collar workers and so then

1840
01:32:42,802 --> 01:32:47,072
We're gonna need robots and so
this is going to drive robotics

1841
01:32:47,072 --> 01:32:50,928
investment because they are a
feasible solution. there's

1842
01:32:50,928 --> 01:32:55,335
technical risk of course, but if
you set up a company in a smart

1843
01:32:55,335 --> 01:32:59,674
way, like in my opinion you guys
are doing, like you can really

1844
01:32:59,674 --> 01:33:04,150
dig into a lot of these problems
and make a lot of value, and you

1845
01:33:04,150 --> 01:33:08,213
don't require like a massive
overhaul of all the facilities

1846
01:33:08,213 --> 01:33:12,000
that they do the work in. you
can just drop a robot in.

1847
01:33:12,000 --> 01:33:15,097
Rajat Bhageria: that's right.
And I think, yes, all of that's

1848
01:33:15,097 --> 01:33:18,245
really true. And I think what's
also true is, I think, we were

1849
01:33:18,245 --> 01:33:21,495
talking about this a little bit
before, but my, like our company

1850
01:33:21,495 --> 01:33:24,542
and our cohort of companies, our
peers in the space, I think

1851
01:33:24,542 --> 01:33:27,537
they've also learned from the
past a little bit, like about

1852
01:33:27,537 --> 01:33:30,635
what's good and what's not good.
I'll give you some examples,

1853
01:33:30,635 --> 01:33:33,681
right? and these are like more
practical, like robot, how do

1854
01:33:33,681 --> 01:33:36,677
you build robotics? I think
there's some practical lessons,

1855
01:33:36,677 --> 01:33:39,520
right? one is, off the shelf
hardware. It's really good.

1856
01:33:39,520 --> 01:33:42,719
there's many reasons we talked
about it. Robotics as a Service.

1857
01:33:42,719 --> 01:33:45,817
I think what's really nice about
Robotics as a Service is you

1858
01:33:45,817 --> 01:33:48,762
don't have to convince somebody
to put down half a million

1859
01:33:48,762 --> 01:33:51,910
bucks. that's really powerful.
Now, by the way, this caveat is

1860
01:33:51,910 --> 01:33:54,956
that some customers, especially
in industrial settings, they

1861
01:33:54,956 --> 01:33:57,901
don't, they just want to buy the
thing. it's not perfectly

1862
01:33:57,901 --> 01:34:00,948
perfect. Perfect there. But
generally it's nicer not to have

1863
01:34:00,948 --> 01:34:04,197
to put down this big expenditure
for a startup that you've never

1864
01:34:04,197 --> 01:34:07,092
heard of. Which is the case in
selling like us. Like most

1865
01:34:07,092 --> 01:34:10,240
people haven't heard of Chef yet
in our customer base. Another

1866
01:34:10,240 --> 01:34:12,982
idea that I think people are
realizing is don't try to

1867
01:34:12,982 --> 01:34:16,028
automate the full process. And
what I mean by that is if you

1868
01:34:16,028 --> 01:34:18,821
try to automate the full
process, you have to, it's too

1869
01:34:18,821 --> 01:34:21,969
much. You have to deal with all
these things. Ideally you have

1870
01:34:21,969 --> 01:34:24,000
this little cell that can just
slide on.

1871
01:34:24,000 --> 01:34:25,000
Audrow Nash: Totally agree.

1872
01:34:25,000 --> 01:34:28,125
Rajat Bhageria: No retrofitting,
there's, all these lessons.

1873
01:34:28,125 --> 01:34:31,354
Another one that's interesting
that's like less technical, but

1874
01:34:31,354 --> 01:34:34,271
for a while people used to like
fund inventory for their

1875
01:34:34,271 --> 01:34:35,000
hardware using

1876
01:34:35,000 --> 01:34:38,000
Audrow Nash: Yeah.

1877
01:34:38,000 --> 01:34:40,954
Rajat Bhageria: and now I think
the world is realizing that, Oh,

1878
01:34:40,954 --> 01:34:43,677
we have a, it's a real hard
asset. Like I don't need to get

1879
01:34:43,677 --> 01:34:46,585
equity. I can just get debt. And
that's really nice to actually

1880
01:34:46,585 --> 01:34:49,077
finance the inventory to
actually do these robotics as

1881
01:34:49,077 --> 01:34:50,000
service deployments.

1882
01:34:50,000 --> 01:34:52,635
Audrow Nash: Interesting. I
don't fully understand that.

1883
01:34:52,635 --> 01:34:54,000
Would you tell me a bit more?

1884
01:34:54,000 --> 01:34:57,715
Rajat Bhageria: yeah. okay. So
here's the way you can think

1885
01:34:57,715 --> 01:35:01,304
about it. It's okay, like I
signed a customer, right? Our

1886
01:35:01,304 --> 01:35:05,081
business model is robotics as
service, right? we have to buy

1887
01:35:05,081 --> 01:35:09,111
them, bill materials upfront. So
we have to outlay all this cash

1888
01:35:09,111 --> 01:35:11,000
upfront. But we don't get paid

1889
01:35:11,000 --> 01:35:12,000
Audrow Nash: For a while.

1890
01:35:12,000 --> 01:35:15,271
Rajat Bhageria: all at once,
right? Exactly. we get paid in

1891
01:35:15,271 --> 01:35:18,376
bits and pieces as a service.
now there's a big cashflow

1892
01:35:18,376 --> 01:35:21,979
issue, right? So if I have let's
say I raised 10 million bucks as

1893
01:35:21,979 --> 01:35:25,195
a seed round and I want to
deploy 10 robots, I don't know,

1894
01:35:25,195 --> 01:35:28,577
I'm just saying arbitrary BOM,
but you, get the point. Like I

1895
01:35:28,577 --> 01:35:32,125
have to out, I've already spent
a bunch of my cash on inventory.

1896
01:35:32,125 --> 01:35:35,341
wouldn't it be nice if that 10
million bucks is really for

1897
01:35:35,341 --> 01:35:38,667
keeping the lights on, hiring
engineers, hiring deployments,

1898
01:35:38,667 --> 01:35:39,000
folks,

1899
01:35:39,000 --> 01:35:45,000
Audrow Nash: Actual operating
expenses. Yeah.

1900
01:35:45,000 --> 01:35:48,763
Rajat Bhageria: and the way that
might work is like. I sign a

1901
01:35:48,763 --> 01:35:52,649
customer, I take that contract
to the bank, advances me cash, I

1902
01:35:52,649 --> 01:35:55,857
use that cash to buy the
inventory, I, provision the

1903
01:35:55,857 --> 01:35:59,497
hardware with my software, I
configure it and then I deploy

1904
01:35:59,497 --> 01:36:03,383
it, I pass site acceptance test,
the customer starts paying me,

1905
01:36:03,383 --> 01:36:04,000
and then I

1906
01:36:04,000 --> 01:36:06,000
Audrow Nash: pay the loan. Yep.
I

1907
01:36:06,000 --> 01:36:09,176
Rajat Bhageria: what this means
is that the cash flow, the cash

1908
01:36:09,176 --> 01:36:12,050
flow cycle is much better
because my, I have, instead of,

1909
01:36:12,050 --> 01:36:15,176
waiting X months to return the
cash from my bill of materials,

1910
01:36:15,176 --> 01:36:18,000
I can be net cash flow positive
on a unit basis in a few

1911
01:36:18,000 --> 01:36:21,348
Audrow Nash: Yeah. And you don't
have to trade equity, which is

1912
01:36:21,348 --> 01:36:24,377
way better because you can
incentivize employees. You can

1913
01:36:24,377 --> 01:36:27,512
keep more for yourself. You
don't have to raise more rounds

1914
01:36:27,512 --> 01:36:29,000
of investment than you need.

1915
01:36:29,000 --> 01:36:31,935
Rajat Bhageria: and, you can
grow the company. and all of

1916
01:36:31,935 --> 01:36:35,230
that, and remember what you said
earlier, which is robotics, you

1917
01:36:35,230 --> 01:36:38,011
have to continue growing
gradually, right? You have to

1918
01:36:38,011 --> 01:36:41,203
show like the test analogy. if
you have more, if you have more

1919
01:36:41,203 --> 01:36:44,241
equity, you have more runway,
which means you can ship more

1920
01:36:44,241 --> 01:36:47,382
robots and show more progress,
which then allows you to raise

1921
01:36:47,382 --> 01:36:48,000
more capital

1922
01:36:48,000 --> 01:36:51,774
Audrow Nash: Even better. Yeah,
you're right. That is great.

1923
01:36:51,774 --> 01:36:55,673
Let's see. So I know you wanted
to talk about a few things and

1924
01:36:55,673 --> 01:36:58,000
we are coming to the end of our
time.

1925
01:36:58,000 --> 01:37:04,359
Audrow Nash: but why, are many
robotic companies failing? what

1926
01:37:04,359 --> 01:37:06,000
have we learned?

1927
01:37:06,000 --> 01:37:09,000
Rajat Bhageria: Yeah, I think,
like I said, I tried to really

1928
01:37:09,000 --> 01:37:12,049
study the industry early on when
thinking about Chef because I

1929
01:37:12,049 --> 01:37:15,000
was just like, it was like,
okay, what's the biggest thing I

1930
01:37:15,000 --> 01:37:18,148
could possibly, I was like, what
are the big things I could work

1931
01:37:18,148 --> 01:37:21,246
on? And I was like, energy seems
really big because, obviously,

1932
01:37:21,246 --> 01:37:24,000
there's a big climate change
issue and, energy is really

1933
01:37:24,000 --> 01:37:27,362
Audrow Nash: And we're building
our industrial plant. Like we

1934
01:37:27,362 --> 01:37:30,724
need way more. Probably three
times or something crazy. Okay.

1935
01:37:30,724 --> 01:37:31,000
Yeah,

1936
01:37:31,000 --> 01:37:34,269
Rajat Bhageria: and then it was
like AI, and within AI and the

1937
01:37:34,269 --> 01:37:37,643
physical world, and I was like,
so convinced that this is such a

1938
01:37:37,643 --> 01:37:41,018
powerful thing, like I think AI
and the physical world is really

1939
01:37:41,018 --> 01:37:44,445
going to be like electricity, so
I was like, I was super excited.

1940
01:37:44,445 --> 01:37:47,555
But then I look at the industry,
And it's a graveyard! it's

1941
01:37:47,555 --> 01:37:50,561
robotics has been tough,
honestly, that's just the truth,

1942
01:37:50,561 --> 01:37:53,935
right? There has, been, like, a
lot of failures. the AV industry

1943
01:37:53,935 --> 01:37:57,204
has sucked in tens of billions
of dollars, and yet, the output

1944
01:37:57,204 --> 01:38:00,525
has been millions of dollars in
revenue. single digit millions,

1945
01:38:00,525 --> 01:38:01,000
probably.

1946
01:38:01,000 --> 01:38:05,000
Audrow Nash: it's

1947
01:38:05,000 --> 01:38:08,433
Rajat Bhageria: anyways, I,
really tried to, learn from the

1948
01:38:08,433 --> 01:38:12,157
industry. I, think one, one big
thing I learned, there's a bunch

1949
01:38:12,157 --> 01:38:15,532
of lessons which we can talk
about, dozens, probably, but,

1950
01:38:15,532 --> 01:38:18,907
one, one, thing that I think,
puts the story to life is, I

1951
01:38:18,907 --> 01:38:22,399
think a lot of people who do
robotics, the founders, they're

1952
01:38:22,399 --> 01:38:25,657
extremely technically minded
folks, right? They're, very

1953
01:38:25,657 --> 01:38:29,207
smart folks because he, and,
very, usually generalist because

1954
01:38:29,207 --> 01:38:32,407
it's such a, it's such a
breathful field. Like you need

1955
01:38:32,407 --> 01:38:36,073
somebody who can, like at least
understand everything from like

1956
01:38:36,073 --> 01:38:39,681
computer science and AI to then
of course the electrical stuff

1957
01:38:39,681 --> 01:38:43,114
to the mechanical stuff to,
Everything in between firmware,

1958
01:38:43,114 --> 01:38:46,838
manufacturing, etc. Even service
and customer success, there's a

1959
01:38:46,838 --> 01:38:50,330
lot of facets to it. what does
that mean? what that means is

1960
01:38:50,330 --> 01:38:53,938
that there's a lot of folks who
just like to build cool stuff,

1961
01:38:53,938 --> 01:38:57,662
right? it's like robots are cool
to build. And I think you saw a

1962
01:38:57,662 --> 01:39:00,629
lot of robotic companies who
were like, in previous

1963
01:39:00,629 --> 01:39:03,946
generations, who were like, you
could tell it's just hey,

1964
01:39:03,946 --> 01:39:06,972
somebody's let me build
something cool. and I think,

1965
01:39:06,972 --> 01:39:10,463
that can work, right? And it has
worked in the past in other

1966
01:39:10,463 --> 01:39:13,722
industries. But the issue is
that the iteration loops in

1967
01:39:13,722 --> 01:39:17,039
robotics are much slower than
something like software. So

1968
01:39:17,039 --> 01:39:20,472
here's what I mean by that. if I
am a software as a service

1969
01:39:20,472 --> 01:39:23,789
company, right? Or I'm a
consumer software company, I can

1970
01:39:23,789 --> 01:39:26,000
build a prototype in a weekend,
right?

1971
01:39:26,000 --> 01:39:27,000
Audrow Nash: Yeah,

1972
01:39:26,000 --> 01:39:29,250
Rajat Bhageria: can get
something work. It's a hack,

1973
01:39:29,250 --> 01:39:32,813
hackathon style, right? MVP I
can get, up in 24 hours, 48

1974
01:39:32,813 --> 01:39:36,688
hours, whatever. it might not be
per perfect, but you, get the

1975
01:39:36,688 --> 01:39:37,000
idea.

1976
01:39:37,000 --> 01:39:38,000
Audrow Nash: You're Parerito.

1977
01:39:38,000 --> 01:39:39,000
Rajat Bhageria: very quickly.

1978
01:39:39,000 --> 01:39:41,844
Audrow Nash: you get 80 percent
with 20 percent of the effort or

1979
01:39:41,844 --> 01:39:43,000
something like this. Okay.

1980
01:39:43,000 --> 01:39:46,579
Rajat Bhageria: That's right.
And then you can iterate, right?

1981
01:39:46,579 --> 01:39:50,215
And this is the entire Eric Lee,
Eric, Lee startup, Lee startup

1982
01:39:50,215 --> 01:39:53,505
model, right? Which is you have
a hypothesis, you build a

1983
01:39:53,505 --> 01:39:57,141
prototype, and you iterate until
you get to product market fit.

1984
01:39:57,141 --> 01:40:00,605
and, I think Silicon Valley, a
lot of Silicon Valley, modern

1985
01:40:00,605 --> 01:40:03,895
Silicon Valley, like software,
not the old semi conductor

1986
01:40:03,895 --> 01:40:07,589
Silicon Valley, is built on that
idea. if I'm doing robotics, or

1987
01:40:07,589 --> 01:40:11,052
autonomy, or AI more generally,
the iteration loops are just

1988
01:40:11,052 --> 01:40:14,342
fundamentally slower. Even if
it's like autonomy, is slow

1989
01:40:14,342 --> 01:40:17,632
because it's like hard. Like
autonomy is much harder than

1990
01:40:17,632 --> 01:40:21,268
making a web app, right? So and
that's just the truth, it's not

1991
01:40:21,268 --> 01:40:24,732
you need both, right? You need
both, but it's just, but, the

1992
01:40:24,732 --> 01:40:28,195
iteration loops are slower. So
instead of having 10 shots on

1993
01:40:28,195 --> 01:40:31,543
goal with a pre seed round or
the seed round, you have One

1994
01:40:31,543 --> 01:40:35,237
shot on goal, before you run out
of money, or maybe you have two

1995
01:40:35,237 --> 01:40:38,873
shots on goal. So this idea of,
let me just build a cool thing,

1996
01:40:38,873 --> 01:40:42,163
and then I'll iterate. It
doesn't really work because you

1997
01:40:42,163 --> 01:40:45,626
end up building the wrong thing
and then you fail. I've seen

1998
01:40:45,626 --> 01:40:48,916
this over and over, like
founders build a cool thing, and

1999
01:40:48,916 --> 01:40:52,264
then you just, you end up not
being able to raise them the

2000
01:40:52,264 --> 01:40:55,958
next round, think now what a lot
more companies are doing in our

2001
01:40:55,958 --> 01:40:59,537
cohort, the more, more kind of
contemporary robotics companies

2002
01:40:59,537 --> 01:41:03,000
is, you still need the same
technically minded founders, but

2003
01:41:03,000 --> 01:41:04,000
Audrow Nash: Gotta be market

2004
01:41:04,000 --> 01:41:06,707
Rajat Bhageria: people are a
little bit more business,

2005
01:41:06,707 --> 01:41:09,614
Business focused. so it's,
things like, okay. for example,

2006
01:41:09,614 --> 01:41:12,471
we sold the product before we
built it. Like when we were

2007
01:41:12,471 --> 01:41:15,729
doing fast casuals, even like we
got contracts really making sure

2008
01:41:15,729 --> 01:41:18,686
that people really want this
thing. Like we know for a fact

2009
01:41:18,686 --> 01:41:21,693
people want this thing. And of
course we reduce the risk for

2010
01:41:21,693 --> 01:41:24,700
the customer too. We added a
site acceptance test that said,

2011
01:41:24,700 --> 01:41:27,808
look, you're not actually going
to pay the recurring fee until

2012
01:41:27,808 --> 01:41:30,815
the thing works. So we reduce
risk for the customers, but we

2013
01:41:30,815 --> 01:41:33,823
knew that they wanted it. for
example, and, but what selling

2014
01:41:33,823 --> 01:41:36,680
before building got us is We
know they want it. We have a

2015
01:41:36,680 --> 01:41:39,587
good product requirements
document where we've synthesized

2016
01:41:39,587 --> 01:41:42,394
all the requirements from
multiple customers. So we know

2017
01:41:42,394 --> 01:41:45,000
we're not overfitting on any
particular customer. So

2018
01:41:45,000 --> 01:41:46,000
Audrow Nash: Definitely.

2019
01:41:46,000 --> 01:41:48,672
Rajat Bhageria: I think, what
that allowed us to do is we

2020
01:41:48,672 --> 01:41:51,438
didn't have to iterate towards
product market fit. We had a

2021
01:41:51,438 --> 01:41:54,438
product market fit, if you will,
from the get go. Now, obviously

2022
01:41:54,438 --> 01:41:57,344
there's still lessons. There's
still unknown unknowns and even

2023
01:41:57,344 --> 01:42:00,391
known unknowns. So there's still
lessons, but you're less likely,

2024
01:42:00,391 --> 01:42:03,109
like we were able to really get
product market fit and One

2025
01:42:03,109 --> 01:42:06,016
iteration loop, which is fast
casual, and the second iteration

2026
01:42:06,016 --> 01:42:08,688
loop was food manufacturing, as
opposed to seven or eight

2027
01:42:08,688 --> 01:42:11,688
iteration loops, which is really
where you end up not working, I

2028
01:42:11,688 --> 01:42:14,266
think. That's one lesson I would
say that's really been

2029
01:42:14,266 --> 01:42:17,078
interesting, which is selling
before building. And coming at

2030
01:42:17,078 --> 01:42:19,797
it from oh, we're really solving
a customer pain point, as

2031
01:42:19,797 --> 01:42:22,000
opposed to just trying to build
something cool.

2032
01:42:22,000 --> 01:42:26,269
Audrow Nash: totally. selling
before building. I hear the idea

2033
01:42:26,269 --> 01:42:30,675
a lot for SaaS projects now, at
least on X and things like this.

2034
01:42:30,675 --> 01:42:34,875
how do you get a customer to
sign off that they, is it like a

2035
01:42:34,875 --> 01:42:39,351
contract that they will pay once
you deliver something? Or how do

2036
01:42:39,351 --> 01:42:43,000
you, sell before you build? What
does that look like?

2037
01:42:43,000 --> 01:42:46,072
Rajat Bhageria: Yeah, I think,
yeah, so I, I think where this

2038
01:42:46,072 --> 01:42:48,993
starts is you have to get them
to realize that they have a

2039
01:42:48,993 --> 01:42:52,065
really big pain point and tell
you. So this isn't like you go

2040
01:42:52,065 --> 01:42:55,187
to them and say, I'm going to
build, not really a sale, right?

2041
01:42:55,187 --> 01:42:58,259
It's like really them selling
you almost, right? Like you go,

2042
01:42:58,259 --> 01:43:01,381
no, I guess what I'm saying is
you go to them and you're like,

2043
01:43:01,381 --> 01:43:04,353
okay, customer. What's your
biggest pain point? And this is

2044
01:43:04,353 --> 01:43:07,475
what we, this is literally what
we did. It's just interviewing

2045
01:43:07,475 --> 01:43:10,597
them almost. what's your biggest
pain point? And they say food

2046
01:43:10,597 --> 01:43:13,669
assembly. Then they're like,
okay, can robots help? Yeah. And

2047
01:43:13,669 --> 01:43:16,791
then it's okay, if we were to do
this, would this be something

2048
01:43:16,791 --> 01:43:19,914
you pay for? Yeah. And then, so,
in other words, they're like,

2049
01:43:19,914 --> 01:43:22,835
they're selling you, like
they've, it's, like how. How big

2050
01:43:22,835 --> 01:43:26,007
of a problem is it for them, I
guess is what I'm saying, right?

2051
01:43:26,007 --> 01:43:29,180
if it's a really big problem for
them, and there's no solutions

2052
01:43:29,180 --> 01:43:32,151
in the market, which is what was
the case in our case, then

2053
01:43:32,151 --> 01:43:35,122
they're going to be a lot more
willing to work with you. So

2054
01:43:35,122 --> 01:43:38,094
what we did is we went to them
and said, look, customer, we

2055
01:43:38,094 --> 01:43:41,367
know what your requirements are,
and here they are. We've written

2056
01:43:41,367 --> 01:43:44,590
them into this Word document. do
you agree that these are indeed

2057
01:43:44,590 --> 01:43:46,000
your technical requirements?

2058
01:43:46,000 --> 01:43:49,977
Audrow Nash: the way, we're,
we're at our ending time, are

2059
01:43:49,977 --> 01:43:54,160
you okay to go for another maybe
like 5 10 minutes, or do you

2060
01:43:54,160 --> 01:43:58,000
have a hard cutoff? Okay, hell
yeah. Sorry to interrupt.

2061
01:43:58,000 --> 01:44:01,216
Rajat Bhageria: No, good. here's
your technical requirements, and

2062
01:44:01,216 --> 01:44:04,136
do you agree with those? Yeah.
Yeah. And then we say, okay,

2063
01:44:04,136 --> 01:44:07,204
look, we've done this financial
analysis, and we think that if

2064
01:44:07,204 --> 01:44:10,025
we were to charge you X dollars
per year, this would be a

2065
01:44:10,025 --> 01:44:12,944
financial win for you, and
here's why we think that. Do you

2066
01:44:12,944 --> 01:44:15,913
agree with that? Assuming these
assumptions are indeed true,

2067
01:44:15,913 --> 01:44:18,784
yeah, we agree with that. Okay,
how about this? Let's work

2068
01:44:18,784 --> 01:44:21,703
together and sign this contract
that says, if we build this

2069
01:44:21,703 --> 01:44:22,000
thing,

2070
01:44:22,000 --> 01:44:24,000
Audrow Nash: You'll buy for
this.

2071
01:44:24,000 --> 01:44:27,120
Rajat Bhageria: and it meets
this site acceptance test. This

2072
01:44:27,120 --> 01:44:30,031
is the key part, the site
acceptance test. It meets this

2073
01:44:30,031 --> 01:44:33,047
set of requirements, and these
set of requirements will be

2074
01:44:33,047 --> 01:44:36,374
things like Is this consistent?
Is this fast? Is this footprint?

2075
01:44:36,374 --> 01:44:39,754
Is this food safety rating? Etc,
etc. It can do this washdown for

2076
01:44:39,754 --> 01:44:42,925
sanitation. if it meets these
requirements, Then you will pay

2077
01:44:42,925 --> 01:44:46,045
us contractually this amount of
money. And just to make sure

2078
01:44:46,045 --> 01:44:49,425
you're really serious, how about
you put down XK as a deposit and

2079
01:44:49,425 --> 01:44:52,232
it's, not a non-recurring
engineering, but it's just a

2080
01:44:52,232 --> 01:44:54,000
deposit to really make sure
you're

2081
01:44:54,000 --> 01:44:55,000
Audrow Nash: Ah. Okay.

2082
01:44:55,000 --> 01:44:58,728
Rajat Bhageria: this is, it is
a, true contractual agreement,

2083
01:44:58,728 --> 01:45:02,700
but the customer hasn't outlayed
a ton of cash, but they have put

2084
01:45:02,700 --> 01:45:06,000
their commitment down that
they're serious about this.

2085
01:45:06,000 --> 01:45:09,610
Audrow Nash: I love it. That's
great. What a cool idea to,

2086
01:45:09,610 --> 01:45:13,344
because I've been seeing this in
SaaS where people are like,

2087
01:45:13,344 --> 01:45:17,079
don't build it first. but seeing
it in robotics where it's a

2088
01:45:17,079 --> 01:45:20,689
hardware project is a new idea
to me and it's really cool.

2089
01:45:20,689 --> 01:45:21,000
Yeah.

2090
01:45:21,000 --> 01:45:23,891
Rajat Bhageria: I actually,
don't think it's as new as, I

2091
01:45:23,891 --> 01:45:27,036
thought this was new too, but
then I like, I learned that like

2092
01:45:27,036 --> 01:45:29,674
a lot of industries work this
way. So like a systems

2093
01:45:29,674 --> 01:45:32,971
integrator, a systems integrator
doesn't, I was talking about the

2094
01:45:32,971 --> 01:45:36,167
Heinz line, for making ketchup.
some, like JR Automation or JMP

2095
01:45:36,167 --> 01:45:39,109
Automation is not putting up 10
million bucks of their own

2096
01:45:39,109 --> 01:45:42,152
capital and, going to Heinz and
being like, Hey, do you like

2097
01:45:42,152 --> 01:45:45,297
this thing? Okay. We iterate a
little bit. Do you like it now?

2098
01:45:45,297 --> 01:45:48,290
Oh, we changed a little bit.
Okay. Do you like it now? They

2099
01:45:48,290 --> 01:45:49,000
don't do that.

2100
01:45:49,000 --> 01:45:50,000
Audrow Nash: Yeah.

2101
01:45:50,000 --> 01:45:54,947
Rajat Bhageria: Instead, Heinz
goes to JR Automation and Heinz

2102
01:45:54,947 --> 01:45:59,495
said, Hey, here's what we need
to build. And JR will put,

2103
01:45:59,495 --> 01:46:04,681
together a. A quote, right? And,
then Hez can sign off and that's

2104
01:46:04,681 --> 01:46:05,000
when

2105
01:46:05,000 --> 01:46:06,000
Audrow Nash: the ticket.

2106
01:46:06,000 --> 01:46:10,058
Rajat Bhageria: automation
starts to work. There's a lot of

2107
01:46:10,058 --> 01:46:14,116
examples also about this. I
generally, I, think if I'm like

2108
01:46:14,116 --> 01:46:18,243
Boeing, right? and I am like
Delta, I just made that up, but

2109
01:46:18,243 --> 01:46:22,302
Boeing is not gonna make Delta
some planes without. Without

2110
01:46:22,302 --> 01:46:26,360
having a PO. And by the way,
that also goes for new models.

2111
01:46:26,360 --> 01:46:30,624
So there's, an interesting like
story with Pan Am Airlines and

2112
01:46:30,624 --> 01:46:34,751
Juan Tripp. Juan Tripp is like
the guy who like made Pan Am.

2113
01:46:34,751 --> 01:46:38,672
and he was negotiating, I think
with the CEO of Boeing or

2114
01:46:38,672 --> 01:46:42,730
something at the time for, he
wanted like a really big jet,

2115
01:46:42,730 --> 01:46:47,063
and basically the Boeing CEO was
like, look, if you Pan Am give

2116
01:46:47,063 --> 01:46:51,534
me an order for X planes, then I
will build a 747 for you. And so

2117
01:46:51,534 --> 01:46:55,661
Pan Am put down a commitment.
Okay. Okay. Boeing. So long as

2118
01:46:55,661 --> 01:46:59,169
you meet these commitments.
These agreements, these

2119
01:46:59,169 --> 01:47:03,365
requirements, iPanAm Airlines
will buy X of your 747s. So now

2120
01:47:03,365 --> 01:47:07,561
the Boeing CEO has commitment
that people want his plane, and

2121
01:47:07,561 --> 01:47:11,481
now he's willing to spend
billions of dollars of R&amp; D

2122
01:47:11,481 --> 01:47:15,540
into making this 747, which is
how the 747 started. I think

2123
01:47:15,540 --> 01:47:19,392
there's a lot of examples
actually where the selling for

2124
01:47:19,392 --> 01:47:23,794
before building, Bill Gates did
this with MS DOS. He sold MS DOS

2125
01:47:23,794 --> 01:47:27,921
to IBM before he spent time and
money. he didn't drop out of

2126
01:47:27,921 --> 01:47:32,185
Harvard. he built, he sold MS
DOS to IBM and then he, built MS

2127
01:47:32,185 --> 01:47:36,587
DOS. Or Wozniak, or sorry, Paul
Allen did, but yeah, you get the

2128
01:47:36,587 --> 01:47:37,000
point.

2129
01:47:37,000 --> 01:47:41,724
Audrow Nash: Gotcha. Yeah. I bet
you're right that it's more

2130
01:47:41,724 --> 01:47:46,606
common. I just haven't seen it
in a robotics context. at least

2131
01:47:46,606 --> 01:47:51,331
that I've noticed, which is
quite cool. Okay. Hell yeah. And

2132
01:47:51,331 --> 01:47:55,740
then, a last thing that I wanted
to talk about is why is

2133
01:47:55,740 --> 01:47:57,000
production hard?

2134
01:47:57,000 --> 01:48:00,175
Rajat Bhageria: Yeah, this was
an interesting lesson for me,

2135
01:48:00,175 --> 01:48:03,190
right? Like I, I think like
robotics is already extremely

2136
01:48:03,190 --> 01:48:06,630
hard as it is, right? it just so
hard to just get zero to one. it

2137
01:48:06,630 --> 01:48:09,910
takes it can take a year or two
years to get even zero to one.

2138
01:48:09,910 --> 01:48:13,138
And autonomous vehicles took 10,
20 years to get from zero to

2139
01:48:13,138 --> 01:48:16,259
one, but the production is just
At least 10 to 100x harder,

2140
01:48:16,259 --> 01:48:19,540
right? And what I mean by that
is the amount of edge cases are

2141
01:48:19,540 --> 01:48:22,873
just insane. So for, I'll give
you an interesting example. Like

2142
01:48:22,873 --> 01:48:26,206
food facilities are sloped and
the reason they're sloped is for

2143
01:48:26,206 --> 01:48:27,000
water drainage.

2144
01:48:27,000 --> 01:48:28,000
Audrow Nash: Yeah. It makes

2145
01:48:27,000 --> 01:48:30,205
Rajat Bhageria: Now, by the way,
I, we just talked about selling

2146
01:48:30,205 --> 01:48:33,260
before building. We talked about
getting product requirements

2147
01:48:33,260 --> 01:48:36,115
early. Like I'm all about the
product part early, just to

2148
01:48:36,115 --> 01:48:38,820
really understand the
requirements, but even I did not

2149
01:48:38,820 --> 01:48:41,524
notice that the floor is sloped,
right? Like it's just

2150
01:48:41,524 --> 01:48:44,379
imperceptible. It's not, and
it's not even a question, to

2151
01:48:44,379 --> 01:48:47,484
ask. You just don't even think
about asking it. we shipped our

2152
01:48:47,484 --> 01:48:50,489
robots to production, and the
very first thing we noticed is

2153
01:48:50,489 --> 01:48:53,694
the robots like, shaking around.
This is like day one, right? We

2154
01:48:53,694 --> 01:48:56,750
shipped a robot to production.
It's oh damn, what's happening

2155
01:48:56,750 --> 01:48:57,000
here?

2156
01:48:57,000 --> 01:48:58,000
Audrow Nash: is

2157
01:48:57,000 --> 01:48:59,652
Rajat Bhageria: this wasn't
showing up. Why is this

2158
01:48:59,652 --> 01:49:02,928
happening? And then, of course
we asked them, and they're like,

2159
01:49:02,928 --> 01:49:05,945
oh, that's what's happening. so
anyways, that's a hardware

2160
01:49:05,945 --> 01:49:08,753
thing. Okay, let's add
stabilizing feet. But then it's

2161
01:49:08,753 --> 01:49:11,873
also then the next thing we
notice is, okay, the robots like

2162
01:49:11,873 --> 01:49:14,993
this, the conveyor might be
slanted. Okay. so now if I like,

2163
01:49:14,993 --> 01:49:18,166
I have to detect the slope of
the conveyor and place lower or

2164
01:49:18,166 --> 01:49:21,546
higher. Okay. okay. Now it's all
great. We solved that issue. But

2165
01:49:21,546 --> 01:49:24,562
then like now instead of day
one, now we're day, now we've

2166
01:49:24,562 --> 01:49:27,578
run for 16 hours a day. And
after 16 hours of the conveyor

2167
01:49:27,578 --> 01:49:30,438
running, the conveyor is
shaking. It's like vibrating a

2168
01:49:30,438 --> 01:49:33,611
lot too. So now the conveyor,
it's already like this, but now

2169
01:49:33,611 --> 01:49:36,731
it's like it walked. It's a
little bit, it's like this, it's

2170
01:49:36,731 --> 01:49:40,111
skewed. So now you have to place
lower and closer and farther and

2171
01:49:40,111 --> 01:49:43,075
higher. Again, this is just a
very simple example. now we

2172
01:49:43,075 --> 01:49:45,988
deploy with our next customer
and that next customer has

2173
01:49:45,988 --> 01:49:48,900
cheese grits and every single
day those cheese grits are

2174
01:49:48,900 --> 01:49:51,864
totally different. That's just
what it is. and then, at a

2175
01:49:51,864 --> 01:49:54,932
different customer you have
humans who are like eager to go

2176
01:49:54,932 --> 01:49:58,052
home and they're like, they're
randomly taking bowls out of,

2177
01:49:58,052 --> 01:50:01,329
they, they pull the ball from in
front of the robot, and so the

2178
01:50:01,329 --> 01:50:04,449
robot has detected the ball,
it's tracking the ball, and now

2179
01:50:04,449 --> 01:50:07,413
the ball disappears, so that
track is like, what, I don't

2180
01:50:07,413 --> 01:50:10,065
know what to do, what's
happened. There's all these

2181
01:50:10,065 --> 01:50:13,030
weird things that happen with
different ingredients. Like

2182
01:50:13,030 --> 01:50:15,890
other humans that you're
interacting with, detect, like

2183
01:50:15,890 --> 01:50:18,906
detecting the, the, velocity of
a conveyor. You think it's

2184
01:50:18,906 --> 01:50:22,026
simple, monotonic, right? It's
like probably simple to use a

2185
01:50:22,026 --> 01:50:25,406
camera to detect the velocity of
conveyor. It's super noisy. It's

2186
01:50:25,406 --> 01:50:28,423
it's all over the place and
every conveyor is a little bit

2187
01:50:28,423 --> 01:50:31,595
different. So how do you have
really good speed estimation of

2188
01:50:31,595 --> 01:50:34,351
the conveyor? So you can
actually place onto a moving

2189
01:50:34,351 --> 01:50:37,315
conveyor and get into these very
small compartments. It's

2190
01:50:37,315 --> 01:50:40,644
actually not trivial. So I think
there's just a lot of And I can

2191
01:50:40,644 --> 01:50:43,660
name hundreds of examples like
this. there's just so many,

2192
01:50:43,660 --> 01:50:46,000
weird things you just never even
think about.

2193
01:50:46,000 --> 01:50:47,000
Audrow Nash: To totally.

2194
01:50:47,000 --> 01:50:50,077
Rajat Bhageria: And, that's,
honestly, when you think about,

2195
01:50:50,077 --> 01:50:53,154
a mode, I think the data mode's
really good, and I think the

2196
01:50:53,154 --> 01:50:56,333
service mode, actually having
this, network of people all over

2197
01:50:56,333 --> 01:50:59,564
the country. But, from a product
perspective, just dealing with

2198
01:50:59,564 --> 01:51:02,590
all these edge cases, even if
you copy my hardware, even if

2199
01:51:02,590 --> 01:51:03,000
you copy

2200
01:51:03,000 --> 01:51:04,000
Audrow Nash: Ah,

2201
01:51:04,000 --> 01:51:07,587
Rajat Bhageria: dealing with all
these edge cases and making a

2202
01:51:07,587 --> 01:51:11,175
solution is pretty hard. I'll
give you one more example, final

2203
01:51:11,175 --> 01:51:14,415
example. Usability, right? Many
of our users, the people

2204
01:51:14,415 --> 01:51:17,771
actually on the line who are
like telling Chef what to do,

2205
01:51:17,771 --> 01:51:21,069
they're often not English
speaking, they're often Spanish

2206
01:51:21,069 --> 01:51:24,772
speaking or Farsi or they speak
different languages and, they're

2207
01:51:24,772 --> 01:51:28,128
often not necessarily college
educated or even high school

2208
01:51:28,128 --> 01:51:31,484
educated. we need to make a
system that's like hundreds of

2209
01:51:31,484 --> 01:51:35,072
thousands of lines of Python and
C code and all this like cool

2210
01:51:35,072 --> 01:51:38,486
hardware, be used by somebody
who, hasn't really dealt with

2211
01:51:38,486 --> 01:51:42,015
machinery or robots or software
really ever. the usability is

2212
01:51:42,015 --> 01:51:45,487
actually a lot more important
than you might think, actually

2213
01:51:45,487 --> 01:51:49,190
having everything in Spanish or
being, whatever language that is

2214
01:51:49,190 --> 01:51:52,314
most prevalent at that site,
having the system so it's

2215
01:51:52,314 --> 01:51:55,555
extraordinarily obvious that
this is the thing to do, if

2216
01:51:55,555 --> 01:51:59,258
something like if a bowl is too
far on the conveyor so the robot

2217
01:51:59,258 --> 01:52:02,845
can't reach, throwing a really
nice visual 3D warning, oh, the

2218
01:52:02,845 --> 01:52:06,201
bowl is too far. Or maybe
somebody comes up in the morning

2219
01:52:06,201 --> 01:52:09,499
and they're like, I want to go
home, so they crank up the

2220
01:52:09,499 --> 01:52:13,029
conveyor speed. So the conveyor
speed is way faster than the,

2221
01:52:13,029 --> 01:52:16,616
the ODD, or like the expected
speed, yeah. And now the robot's

2222
01:52:16,616 --> 01:52:20,204
I can't keep up with all this.
And now the user's really upset

2223
01:52:20,204 --> 01:52:23,560
because they're missing bowls.
Missing containers, but the

2224
01:52:23,560 --> 01:52:26,800
robot can't do anything. So
throwing a nice warning. Ah,

2225
01:52:26,800 --> 01:52:30,387
you're on the dashboard Oh, your
conveyor is moving a lot fast

2226
01:52:30,387 --> 01:52:33,975
and a lot faster than it should
but doing that in a way that's

2227
01:52:33,975 --> 01:52:36,000
not in words, but rather in
Images.

2228
01:52:36,000 --> 01:52:40,000
Audrow Nash: something easy to
explain. Yeah, images,

2229
01:52:40,000 --> 01:52:43,375
Rajat Bhageria: that really
separates like a prototype

2230
01:52:43,375 --> 01:52:46,938
company from a production
company And then going from one

2231
01:52:46,938 --> 01:52:50,500
robots of two robots One robot,
zero robots, one robot in

2232
01:52:50,500 --> 01:52:54,188
production is a big deal, and
then ten, and then, it's just

2233
01:52:54,188 --> 01:52:58,063
every layer it's just like the
amount of volume of stuff going

2234
01:52:58,063 --> 01:53:01,313
on is just absurd. Like now
we've made 17. 7 million

2235
01:53:01,313 --> 01:53:05,000
servings, meals in production,
like it's just the amount of

2236
01:53:05,000 --> 01:53:08,938
stuff going, happening, is just
a lot. And you have to have the

2237
01:53:08,938 --> 01:53:12,938
right data infrastructure to be
able to process all this, alert,

2238
01:53:12,938 --> 01:53:16,813
alert the right people, annotate
all of this, you can actually

2239
01:53:16,813 --> 01:53:18,000
triage it, all this

2240
01:53:18,000 --> 01:53:20,000
Audrow Nash: you got to keep up
with it. Do you have any,

2241
01:53:20,000 --> 01:53:21,000
Rajat Bhageria: up with that.

2242
01:53:21,000 --> 01:53:24,824
Audrow Nash: do you have any
heuristics or like lessons or

2243
01:53:24,824 --> 01:53:28,912
any things that would have any
mindsets or anything that would

2244
01:53:28,912 --> 01:53:33,132
have made this whole process of
dealing with all these difficult

2245
01:53:33,132 --> 01:53:36,758
edge cases that pop up in
production a little easier or

2246
01:53:36,758 --> 01:53:40,516
just be ready for them and
steele yourself for it or what

2247
01:53:40,516 --> 01:53:44,802
do you think like anything would
have helped you in preparing for

2248
01:53:44,802 --> 01:53:45,000
all

2249
01:53:45,000 --> 01:53:48,788
Rajat Bhageria: I think, there's
a few like strategic things and

2250
01:53:48,788 --> 01:53:52,340
a few practical kind of tactical
things that come to mind. I

2251
01:53:52,340 --> 01:53:55,773
think like the thought that's
coming to my mind is just we

2252
01:53:55,773 --> 01:53:59,442
really think a lot about hiring
and hiring really good people,

2253
01:53:59,442 --> 01:54:03,231
but just honestly just continue
hiring the absolute best people.

2254
01:54:03,231 --> 01:54:07,019
If you have a really good team,
they can, knock this out, right?

2255
01:54:07,019 --> 01:54:10,570
I think it just starts with
that, your team, right? and then

2256
01:54:10,570 --> 01:54:13,944
also having, like I, yeah, like
it, it seems like, On the

2257
01:54:13,944 --> 01:54:17,495
surface, you might be looking
for a very tactical thing, but

2258
01:54:17,495 --> 01:54:20,988
like having a culture where
everyone really cares about the

2259
01:54:20,988 --> 01:54:24,835
customer, right? I think there's
a lot of companies who like They

2260
01:54:24,835 --> 01:54:28,564
talk about customer centricity,
but they really don't care that

2261
01:54:28,564 --> 01:54:32,411
much about the customer. Whereas
if you can build a culture where

2262
01:54:32,411 --> 01:54:35,726
everyone across the company just
knows everything that's

2263
01:54:35,726 --> 01:54:39,159
happening with the customer.
They've been to the customer,

2264
01:54:39,159 --> 01:54:42,474
they know the customer, they
know why the customer cares

2265
01:54:42,474 --> 01:54:45,966
about something. Then if there's
a little bug that happens,

2266
01:54:45,966 --> 01:54:49,695
people can empathize. It's ah,
my code is doing that. I need to

2267
01:54:49,695 --> 01:54:53,069
go fix it. And there's a
personal, I, if I, if that means

2268
01:54:53,069 --> 01:54:56,798
I'm going to stay till 8 fixing
this bug, great, because I care

2269
01:54:56,798 --> 01:55:00,349
about this customer, and I know
what it, what, how impactful

2270
01:55:00,349 --> 01:55:03,841
this is for the customer. So I
think that's some, strategic

2271
01:55:03,841 --> 01:55:07,333
stuff. On a practical level, I
think, would say a couple of

2272
01:55:07,333 --> 01:55:10,174
things, really, we invested a
decent bit in test

2273
01:55:10,174 --> 01:55:13,548
infrastructure, and Data
pipelines. I think we could have

2274
01:55:13,548 --> 01:55:17,159
done even more, right? Just like
really good internal tooling

2275
01:55:17,159 --> 01:55:20,651
around data infra. So like even
better tagging, even better

2276
01:55:20,651 --> 01:55:23,788
annotation, even better
processes so that if there is

2277
01:55:23,788 --> 01:55:27,340
like creating federations so
there's 20 weight events. Let's

2278
01:55:27,340 --> 01:55:30,950
act, let's aggregate all those
into one bug ticket so that we

2279
01:55:30,950 --> 01:55:34,146
have 20 bagged files so that
when an engineer goes and

2280
01:55:34,146 --> 01:55:37,579
triages them They have all the
data they need right there,

2281
01:55:37,579 --> 01:55:41,249
right? or another example might
be like, okay, if I patch a PR

2282
01:55:41,249 --> 01:55:45,037
to fix some thing, let me patch
it to, automatically patch it to

2283
01:55:45,037 --> 01:55:48,648
every branch that we have in
production. And these, these are

2284
01:55:48,648 --> 01:55:52,199
like not, these are commonly,
these are software tools that,

2285
01:55:52,199 --> 01:55:55,632
in SaaS all the time and like
production companies all the

2286
01:55:55,632 --> 01:55:58,000
time, but robotics doesn't have
a ton of

2287
01:55:58,000 --> 01:56:02,632
Audrow Nash: I agree. Yeah, so
really investing in that tooling

2288
01:56:02,632 --> 01:56:06,897
because that improves process
and every time you, when you

2289
01:56:06,897 --> 01:56:11,456
have a crappy process, it eats
future time because you have to

2290
01:56:11,456 --> 01:56:13,000
go do whatever it is.

2291
01:56:13,000 --> 01:56:14,844
Rajat Bhageria: the way, it's
tough though, like I'm saying

2292
01:56:14,844 --> 01:56:15,000
this,

2293
01:56:15,000 --> 01:56:18,089
Audrow Nash: you don't want to
optimize too much is the other

2294
01:56:18,089 --> 01:56:19,000
side of that evil.

2295
01:56:19,000 --> 01:56:22,567
Rajat Bhageria: well, and, once
you, like, when you're trying to

2296
01:56:22,567 --> 01:56:25,855
do zero to one, you're not
thinking about tooling, you just

2297
01:56:25,855 --> 01:56:29,255
want to ship the thing, right?
It's whatever it takes to ship

2298
01:56:29,255 --> 01:56:32,766
the thing out the door. And then
as soon as you ship it, you're

2299
01:56:32,766 --> 01:56:36,166
delus you're like, underwater
with bugs. You're diluted with,

2300
01:56:36,166 --> 01:56:39,677
er, yeah, and, then then you're
like, there's a lot of pressure

2301
01:56:39,677 --> 01:56:43,021
to grow, and, so anyways, it
feels, it's you want to do this

2302
01:56:43,021 --> 01:56:46,643
internal tooling stuff, but it's
like, how do you create the time

2303
01:56:46,643 --> 01:56:50,210
to do it, right? there's always
this, hard battle, but you, need

2304
01:56:50,210 --> 01:56:53,331
to do it. It's necessary, or
else you're never gonna get

2305
01:56:53,331 --> 01:56:54,000
ahead of it.

2306
01:56:54,000 --> 01:56:58,524
Audrow Nash: Yeah, I feel like,
I don't know, for maybe, I guess

2307
01:56:58,524 --> 01:57:02,695
it probably comes back to the
people you hire to some level

2308
01:57:02,695 --> 01:57:07,007
for this kind of thing. Because
one of the joys that I've had

2309
01:57:07,007 --> 01:57:11,532
working on the ROS2 project is I
have wrote a lot of tools and I

2310
01:57:11,532 --> 01:57:15,561
wrote a lot of scripts to
automate really dull things for

2311
01:57:15,561 --> 01:57:19,661
us. because one, I like to,
Doing that, and two, they make

2312
01:57:19,661 --> 01:57:24,256
everything else easier. And when
you have more consistency across

2313
01:57:24,256 --> 01:57:28,003
projects, it's much easier to
write additional future

2314
01:57:28,003 --> 01:57:32,032
automations. And so having
people that want to go out and

2315
01:57:32,032 --> 01:57:36,274
do these kinds of things, and
know about them, that there is

2316
01:57:36,274 --> 01:57:40,728
the option to do this thing and
improve the process, that might

2317
01:57:40,728 --> 01:57:42,000
be a good way too.

2318
01:57:42,000 --> 01:57:44,000
Rajat Bhageria: Yes. No, I
totally agree.

2319
01:57:44,000 --> 01:57:46,000
Audrow Nash: Oh, yeah. Yeah.
Hiring, as you said.

2320
01:57:46,000 --> 01:57:48,206
Rajat Bhageria: Hiring is
everything. some of our

2321
01:57:48,206 --> 01:57:50,908
engineers will just they're
like, okay, I need to solve this

2322
01:57:50,908 --> 01:57:53,294
thing. And if I can spend
another half an hour, I can

2323
01:57:53,294 --> 01:57:56,085
build this quick little internal
tool. And it's not even, it's

2324
01:57:56,085 --> 01:57:58,922
not, like a part of the sprint.
They're just like, I need to do

2325
01:57:58,922 --> 01:58:01,758
this because it's going to help
me. And I really care about the

2326
01:58:01,758 --> 01:58:04,370
company and I have a lot of
equity in this company. So let

2327
01:58:04,370 --> 01:58:05,000
me just do it.

2328
01:58:05,000 --> 01:58:08,000
Audrow Nash: Yeah, totally.

2329
01:58:08,000 --> 01:58:11,269
Rajat Bhageria: And by the way,
I say that by the way, because

2330
01:58:11,269 --> 01:58:14,590
it's an important point, which
is if you hire people where it's

2331
01:58:14,590 --> 01:58:17,226
not just a job, but like
they're, it's really they

2332
01:58:17,226 --> 01:58:20,231
really, believe in it. if, in
other words, you treat your

2333
01:58:20,231 --> 01:58:23,184
people like investors. They're
like investing into their

2334
01:58:23,184 --> 01:58:26,241
company with their time. They
really believe in it. And if

2335
01:58:26,241 --> 01:58:29,405
they really believe in it and
they have a lot of skin in the

2336
01:58:29,405 --> 01:58:32,568
game, then it's less, it's like,
it's not okay, I just do my

2337
01:58:32,568 --> 01:58:35,573
sprint. I'm gonna go home,
right? It's I want to make the

2338
01:58:35,573 --> 01:58:38,631
product and the company and the
machine, being the company

2339
01:58:38,631 --> 01:58:39,000
better.

2340
01:58:39,000 --> 01:58:43,398
Audrow Nash: Yeah, it's mission
oriented trying to help the

2341
01:58:43,398 --> 01:58:47,795
whole thing. Yeah. Do you have
any thoughts on how to hire?

2342
01:58:47,795 --> 01:58:51,000
Good people or how do you find
good people?

2343
01:58:51,000 --> 01:58:55,111
Rajat Bhageria: Yeah, that's
very tough. Hiring is actually a

2344
01:58:55,111 --> 01:58:59,154
lot. like people think like
fundraising is hard. I Hiring is

2345
01:58:59,154 --> 01:59:03,062
extremely tough, I think,
honestly. I have a few thoughts.

2346
01:59:03,062 --> 01:59:07,307
one thought I have is that, I,
watch this, famous clip. I don't

2347
01:59:07,307 --> 01:59:11,485
know if it's famous or not, but
this clip by, Derek Sivers. He

2348
01:59:11,485 --> 01:59:15,798
talks about, basically, he's the
guy who made Seedy Baby back in

2349
01:59:15,798 --> 01:59:20,111
the day. And, he had this famous
clip, this, clip, which is, the

2350
01:59:20,111 --> 01:59:24,491
first follower is the person who
turns a lone nut into leader. So

2351
01:59:24,491 --> 01:59:28,668
he shows this like little clip
of somebody like dancing around

2352
01:59:28,668 --> 01:59:32,712
the middle of a field, right?
Just randomly and he was like,

2353
01:59:32,712 --> 01:59:37,024
oh, who's this weird person? And
then one other person joins and

2354
01:59:37,024 --> 01:59:41,000
as soon as that one other person
joins everybody else comes

2355
01:59:41,000 --> 01:59:44,130
Audrow Nash: you're like an
eccentric doing your thing until

2356
01:59:44,130 --> 01:59:47,000
others follow you and then
you're the leader Hilarious.

2357
01:59:47,000 --> 01:59:49,950
Rajat Bhageria: And I think
that's how hiring works. when I

2358
01:59:49,950 --> 01:59:52,700
try to find, my approach was
always to get, very senior

2359
01:59:52,700 --> 01:59:55,650
folks. I want senior staff level
engineers from the get go.

2360
01:59:55,650 --> 01:59:58,650
Which, by the way, was hard
because, we're this tiny company

2361
01:59:58,650 --> 02:00:01,750
who can't pay a ton of money in
terms of salary, right? at the

2362
02:00:01,750 --> 02:00:04,950
time, right? This was like back
in 2019, 2020. so it's like, how

2363
02:00:04,950 --> 02:00:07,850
do we convince somebody? So
initially it was just a lot of

2364
02:00:07,850 --> 02:00:11,000
rejection. We talked to a lot of
people and a lot of people are

2365
02:00:11,000 --> 02:00:14,000
like, look, this is great, but I
have a mortgage or I have a

2366
02:00:14,000 --> 02:00:16,600
partner and I have kids and
sorry. but we ultimately

2367
02:00:16,600 --> 02:00:19,150
convinced one person who had a
really, great set of

2368
02:00:19,150 --> 02:00:21,800
experiences, quite senior, had
been at a bunch of top

2369
02:00:21,800 --> 02:00:24,800
companies, had been a lead, and
like brand names that people

2370
02:00:24,800 --> 02:00:27,000
know of, right? Like
respectable, basically.

2371
02:00:27,000 --> 02:00:28,000
Audrow Nash: hell yeah.

2372
02:00:28,000 --> 02:00:31,054
Rajat Bhageria: Getting him was
huge because as soon as we got

2373
02:00:31,054 --> 02:00:33,862
him. When I talked to the next
engineer, I wasn't the one

2374
02:00:33,862 --> 02:00:36,719
selling that next engineer. This
person was the one that's

2375
02:00:36,719 --> 02:00:38,000
selling the next engineer,

2376
02:00:38,000 --> 02:00:41,497
Audrow Nash: hell yeah. That's
so funny. It's a lot like that

2377
02:00:41,497 --> 02:00:45,108
podcasting, by the way, which is
very funny. Getting guests and

2378
02:00:45,108 --> 02:00:47,000
this kind of thing. Same
dynamic.

2379
02:00:47,000 --> 02:00:48,000
Rajat Bhageria: Cool.

2380
02:00:47,000 --> 02:00:51,000
Audrow Nash: Okay. And

2381
02:00:51,000 --> 02:00:53,973
Rajat Bhageria: having a high
bar early on. I think it's in

2382
02:00:53,973 --> 02:00:57,197
the early days it's easier just
to take whoever because like You

2383
02:00:57,197 --> 02:01:00,271
don't have much to offer so it's
come anyone who's willing to

2384
02:01:00,271 --> 02:01:03,294
come but I think if you have a
really high bar early on then

2385
02:01:03,294 --> 02:01:06,266
you set yourself up in many ways
because this thing we just

2386
02:01:06,266 --> 02:01:09,188
talked about from the hiring
perspective but a really good

2387
02:01:09,188 --> 02:01:12,363
engineer has good architecture
decisions because like it's like

2388
02:01:12,363 --> 02:01:15,335
the famous thing like Yeah,
like, Facebook still uses a lot

2389
02:01:15,335 --> 02:01:18,509
of PHP. It's like, why do they
use PHP? it's because originally

2390
02:01:18,509 --> 02:01:21,532
Facebook was written in PHP, so
it persists. There's inertia

2391
02:01:21,532 --> 02:01:24,706
there. and I think that's true
of like technical decisions. for

2392
02:01:24,706 --> 02:01:27,729
example, if you start off with
ROS 1, it's really hard to do

2393
02:01:27,729 --> 02:01:30,803
ROS 2. But if you have somebody
who is, who's been around the

2394
02:01:30,803 --> 02:01:34,027
block, they might recommend ROS
2 from the get go. That's a good

2395
02:01:34,027 --> 02:01:36,798
architecture decision and
probably gonna pay off in the

2396
02:01:36,798 --> 02:01:37,000
long

2397
02:01:37,000 --> 02:01:39,000
Audrow Nash: Hell yeah.

2398
02:01:39,000 --> 02:01:42,592
Rajat Bhageria: But right?
There's many reasons to get a

2399
02:01:42,592 --> 02:01:46,377
more senior person from the
early days. It's harder, but if

2400
02:01:46,377 --> 02:01:50,355
you can do that, then it makes
hiring other people easier. And

2401
02:01:50,355 --> 02:01:54,525
then, after that, just keeping a
really high bar, having a really

2402
02:01:54,525 --> 02:01:56,000
good interview process,

2403
02:01:56,000 --> 02:01:57,000
Audrow Nash: Hell yeah.

2404
02:01:56,000 --> 02:01:58,486
Rajat Bhageria: It's not easy,
honestly. Like I think it's

2405
02:01:58,486 --> 02:01:59,000
really hard.

2406
02:01:59,000 --> 02:02:03,728
Audrow Nash: Yeah, I think so
too. And I love that lone nut

2407
02:02:03,728 --> 02:02:08,857
idea. hilarious. But so we have
gone quite a bit over time. I am

2408
02:02:08,857 --> 02:02:13,906
wondering if you have any last
thoughts or things you'd like to

2409
02:02:13,906 --> 02:02:18,875
share with our audience. what,
do you hope they take away from

2410
02:02:18,875 --> 02:02:22,000
this interview and about Chef
Robotics?

2411
02:02:22,000 --> 02:02:25,337
Rajat Bhageria: Yeah. I think
I'll say on a more broad scale,

2412
02:02:25,337 --> 02:02:28,675
a scope, like broad scope thing,
which is I really, think the

2413
02:02:28,675 --> 02:02:32,012
biggest impact for AI is going
to be on the physical world in

2414
02:02:32,012 --> 02:02:35,404
the labor market. And I think by
the way, it's not this is not

2415
02:02:35,404 --> 02:02:38,687
like a dystopian, take your job
thing. I really believe in a

2416
02:02:38,687 --> 02:02:42,079
future of abundance where, you
know, 50 percent of the economy

2417
02:02:42,079 --> 02:02:45,471
is manual labor right now. what
if robots are doing that? And,

2418
02:02:45,471 --> 02:02:48,480
by the way, there's been a lot
of instances where like.

2419
02:02:48,480 --> 02:02:51,325
Automation has created a lot,
like every instance of

2420
02:02:51,325 --> 02:02:54,827
automation has created more jobs
than it's taken. So the G Throw

2421
02:02:54,827 --> 02:02:58,109
Toll, or the printing press, the
computer, the steam engine.

2422
02:02:58,109 --> 02:03:01,392
These created a lot more jobs
than it took, but the jobs are

2423
02:03:01,392 --> 02:03:04,456
better. And I really am a big
fan of that, so I'm really

2424
02:03:04,456 --> 02:03:07,739
optimistic and hope that a lot
more people are going to take

2425
02:03:07,739 --> 02:03:10,693
their AI skills, right? Let's
take those skills you've

2426
02:03:10,693 --> 02:03:13,647
developed in machine learning
and AI, or even software

2427
02:03:13,647 --> 02:03:16,875
engineering, and, like really
apply them to do embodied AI.

2428
02:03:16,875 --> 02:03:20,377
Because I think like the really
important problems of the world,

2429
02:03:20,377 --> 02:03:23,824
the really important problems of
the world, they're not in bits

2430
02:03:23,824 --> 02:03:27,161
that you can solve it. You can
do a lot in bits. You can do a

2431
02:03:27,161 --> 02:03:30,389
lot in bits, right? Palantir is
a cool example of a company

2432
02:03:30,389 --> 02:03:33,781
doing a lot in bits. Uber is a
cool company. It's moving atoms

2433
02:03:33,781 --> 02:03:34,000
with

2434
02:03:34,000 --> 02:03:35,000
Audrow Nash: Yeah.

2435
02:03:34,000 --> 02:03:37,186
Rajat Bhageria: But I think like
really the most important

2436
02:03:37,186 --> 02:03:40,372
problems in the world are
physical world problems that you

2437
02:03:40,372 --> 02:03:43,614
can't solve. The Global Hunger
Crisis or Water Pollution or

2438
02:03:43,614 --> 02:03:46,690
Climate Change using just
software. You just cannot. And

2439
02:03:46,690 --> 02:03:50,206
I, really hope that more people
take those skills and apply them

2440
02:03:50,206 --> 02:03:53,337
to the physical world. I think
that'll be really good for

2441
02:03:53,337 --> 02:03:56,743
humanity in the world. and then
when it comes to Chef, I think

2442
02:03:56,743 --> 02:04:00,204
I, think like we have tried to
stay very grounded, let's have a

2443
02:04:00,204 --> 02:04:03,555
big vision and mission, but stay
really grounded and actually

2444
02:04:03,555 --> 02:04:07,071
shipping real robots that really
help solve an actual pain point

2445
02:04:07,071 --> 02:04:10,092
for the customer. I think. I
think that's really key in

2446
02:04:10,092 --> 02:04:13,498
robotics. I think it's so hard
to do robotics as it is, but if

2447
02:04:13,498 --> 02:04:16,629
you have a customer, or ideally
a set of customers, who's

2448
02:04:16,629 --> 02:04:19,871
telling you what to do, and of
course you be careful not to

2449
02:04:19,871 --> 02:04:23,112
overfit, and you can really
build something that's actually

2450
02:04:23,112 --> 02:04:26,243
useful, and hopefully, if you
have the right vision, then

2451
02:04:26,243 --> 02:04:29,759
something that can become quite
scalable. And I think, if that's

2452
02:04:29,759 --> 02:04:33,000
exciting for you, then of course
we'd love to chat as well.

2453
02:04:33,000 --> 02:04:40,329
Audrow Nash: Hell yeah. All
right. It was a pleasure. And

2454
02:04:40,329 --> 02:04:42,000
bye everyone.

2455
02:04:42,000 --> 02:04:43,000
Rajat Bhageria: Thanks for
having me, Audrow.

2456
02:04:43,000 --> 02:04:46,715
Audrow Nash: As we wrap up this
fascinating conversation with

2457
02:04:46,715 --> 02:04:50,309
Rajat Bhageria, I hope you are
as excited as I am about the

2458
02:04:50,309 --> 02:04:53,842
potential of AI and robotics in
the food industry. Rajat's

2459
02:04:53,842 --> 02:04:57,740
insights into the challenges and
opportunities in this space are

2460
02:04:57,740 --> 02:05:01,455
truly eye opening. Before we go,
I'd like to leave you with a

2461
02:05:01,455 --> 02:05:05,353
question to ponder: how might AI
powered food automation reshape

2462
02:05:05,353 --> 02:05:09,008
not just the food industry, but
our entire relationship with

2463
02:05:09,008 --> 02:05:12,601
food production and consumption?
Will we see a future where

2464
02:05:12,601 --> 02:05:16,073
customized restaurant quality
meals are as accessible and

2465
02:05:16,073 --> 02:05:19,423
affordable as fast food is
today? If you found value in

2466
02:05:19,423 --> 02:05:23,017
today's discussion, I'd be
incredibly grateful if you leave

2467
02:05:23,017 --> 02:05:26,367
a rating or review. Your
feedback not only helps others

2468
02:05:26,367 --> 02:05:29,838
discover the show, but also
helps me to improve and bring

2469
02:05:29,838 --> 02:05:33,371
you the content you want to
hear. Thank you for listening.

2470
02:05:33,371 --> 02:05:36,843
And until next time, keep
exploring the fascinating world

2471
02:05:36,843 --> 02:05:38,000
of robotics and AI.
