1
00:00:00,000 --> 00:00:03,415
Audrow Nash: I've talked to a
lot of people about this

2
00:00:03,415 --> 00:00:07,335
interview, and I'm excited that
I get to share it with you. In

3
00:00:07,335 --> 00:00:10,939
it, I talk with Nag and Mike
from Electric Sheep. They're

4
00:00:10,939 --> 00:00:14,607
doing many things differently
than most robotics companies

5
00:00:14,607 --> 00:00:18,464
that I've talked to, and they're
making big bets that I think

6
00:00:18,464 --> 00:00:22,005
will pay off. Here are three
examples to show you what I

7
00:00:22,005 --> 00:00:25,357
mean. First, they're throwing
away classical robotics

8
00:00:25,357 --> 00:00:29,340
approaches, and instead... using
machine learning. I'm not just

9
00:00:29,340 --> 00:00:33,387
talking about for perception or
parameter optimization, but even

10
00:00:33,387 --> 00:00:37,118
for things like localization or
high level control. Second,

11
00:00:37,118 --> 00:00:40,786
they've turned the lawn mowing
problem on its head to make

12
00:00:40,786 --> 00:00:44,453
robots that are intrinsically
safer. And third, instead of

13
00:00:44,453 --> 00:00:48,247
selling their robots or doing a
subscription model, they buy

14
00:00:48,247 --> 00:00:51,282
profitable landscaping
businesses and give those

15
00:00:51,282 --> 00:00:55,076
companies robots. There are a
lot of advantages to this last

16
00:00:55,076 --> 00:00:59,123
point and you'll see the details
of each during the interview. I

17
00:00:59,123 --> 00:01:03,170
hope it surprises you as much as
it did me. I think you'll enjoy

18
00:01:03,170 --> 00:01:07,091
this interview if you're curious
about how AI and robotics can

19
00:01:07,091 --> 00:01:10,442
fit together in a real
application and if you want to

20
00:01:10,442 --> 00:01:14,173
see a new robotics business
model that I think will be very

21
00:01:14,173 --> 00:01:18,156
popular in the near future. I'm
Audrow Nash, this is the Audrow

22
00:01:18,156 --> 00:01:22,014
Nash Podcast, so let's see why
Electric Sheep is doing things

23
00:01:22,014 --> 00:01:25,871
so differently, and I'd love to
know in the comments or on X.

24
00:01:25,871 --> 00:01:29,665
What do you think about their
bets? Would you take them? Are

25
00:01:29,665 --> 00:01:33,269
you convinced? Or do you see
some weak points? I hope you

26
00:01:33,269 --> 00:01:37,000
enjoy. Hi everyone. Nag, can I
have you introduce yourself?

27
00:01:37,000 --> 00:01:41,340
Nag Murthy: Yeah. Hi everyone.
I'm Nag Murthy. I'm the CEO and

28
00:01:41,340 --> 00:01:44,000
co founder of Electric Sheep
Robotics.

29
00:01:44,000 --> 00:01:46,000
Audrow Nash: And Mike, would you
introduce yourself?

30
00:01:46,000 --> 00:01:49,278
Michael Laskey: Hi everyone. I'm
Mike Laskey. I am the Vice

31
00:01:49,278 --> 00:01:52,000
President of Autonomy at
Electric Sheep Robotics.

32
00:01:52,000 --> 00:01:55,773
Audrow Nash: Awesome. Now, Nag,
would you tell me a bit about

33
00:01:55,773 --> 00:01:58,000
Electric Sheep? What do you guys
do?

34
00:01:58,000 --> 00:02:01,985
Nag Murthy: Sure, yeah. We are
an outdoor maintenance company

35
00:02:01,985 --> 00:02:05,318
powered by artificial
intelligence and robotics. So

36
00:02:05,318 --> 00:02:09,042
the way we work is we acquire
traditional outdoor service

37
00:02:09,042 --> 00:02:12,439
providers. And we then
progressively transform their

38
00:02:12,439 --> 00:02:16,294
operations by deploying our own
proprietary AI software and

39
00:02:16,294 --> 00:02:20,410
robots. and basically what we're
doing at the back end is we're

40
00:02:20,410 --> 00:02:24,461
building this large world model
that can effectively drive all

41
00:02:24,461 --> 00:02:28,381
outdoor autonomous equipment.
And then over time, we plan to

42
00:02:28,381 --> 00:02:32,236
use those, automated robots to
improve our own margins over

43
00:02:32,236 --> 00:02:34,000
time. So that's what we do.

44
00:02:34,000 --> 00:02:37,000
Audrow Nash: And you're starting
with lawnmowing at the moment.

45
00:02:37,000 --> 00:02:39,000
So that's the first embodiment,
I suppose?

46
00:02:39,000 --> 00:02:43,290
Nag Murthy: That's right. So
that's the first instance of

47
00:02:43,290 --> 00:02:46,000
what the model, you know,
automates.

48
00:02:46,000 --> 00:02:49,979
Audrow Nash: Gotcha. And so
there's so much that's

49
00:02:49,979 --> 00:02:54,675
interesting to unpack there.
let's start with why, like the

50
00:02:54,675 --> 00:02:59,609
approach you guys are taking.
Yeah. I suppose. Like, so why is

51
00:02:59,609 --> 00:03:04,225
it these world models? Like,
what is that? And why is that

52
00:03:04,225 --> 00:03:08,682
interesting? maybe Mike, is
that, is that a good one for

53
00:03:08,682 --> 00:03:09,000
you?

54
00:03:09,000 --> 00:03:12,435
Michael Laskey: Yeah, I can take
that one. So I think, I think

55
00:03:12,435 --> 00:03:15,871
there's like a, there's kind of
two fronts of it. One is like,

56
00:03:15,871 --> 00:03:19,473
what's, what's the architecture?
What's the approach of how we go

57
00:03:19,473 --> 00:03:22,908
about designing these robots?
And then the other side is like,

58
00:03:22,908 --> 00:03:26,288
how do we get data? How do we
actually train these models? So

59
00:03:26,288 --> 00:03:29,557
at Electric Sheep, we're very
much like, we believe that...

60
00:03:29,557 --> 00:03:32,771
The future of robots is a
learning robot. A robot that can

61
00:03:32,771 --> 00:03:35,985
understand the world in an
intuitive way, can relate to it

62
00:03:35,985 --> 00:03:39,420
with more of a natural instinct
that progresses over time from

63
00:03:39,420 --> 00:03:42,468
experiences, and can be
interacted with in sort of the,

64
00:03:42,468 --> 00:03:45,571
the UX that you might interact
with like an outdoor work

65
00:03:45,571 --> 00:03:48,896
animal. So natural language,
visual, like there should be no

66
00:03:48,896 --> 00:03:52,165
app. So none of our robots have
apps. They should just have

67
00:03:52,165 --> 00:03:55,600
simple like on off buttons, and
that's about it from a sort of

68
00:03:55,600 --> 00:03:58,537
like button interface
standpoint. now, to create this

69
00:03:58,537 --> 00:04:01,973
sort of intelligence and this
experience, there's sort of like

70
00:04:01,973 --> 00:04:05,186
two ways you can do it in
robotics, right? There's, direct

71
00:04:05,186 --> 00:04:08,566
policy learning, where you say,
okay, well, I'm going to like

72
00:04:08,566 --> 00:04:12,002
push mow and just take all my
push mowing experience and train

73
00:04:12,002 --> 00:04:15,493
the robot to do push mowing. And
then on the flip side of that,

74
00:04:15,493 --> 00:04:18,374
when you look at sort of like
different RL, like MDP

75
00:04:18,374 --> 00:04:21,920
celebrations, there's you learn
the model. So instead of saying,

76
00:04:21,920 --> 00:04:25,190
I'm going to teach this thing to
only push mow, it's you're

77
00:04:25,190 --> 00:04:28,625
saying, I'm going to teach this
thing to learn how the natural

78
00:04:28,625 --> 00:04:31,950
world works. And then you can
tell it to do different tasks.

79
00:04:31,950 --> 00:04:34,000
Then you can say, oh, oh yeah,
sorry?

80
00:04:34,000 --> 00:04:37,608
Audrow Nash: So you're saying
like in a reinforcement learning

81
00:04:37,608 --> 00:04:41,101
context, you are, it's almost
like you're generating like an

82
00:04:41,101 --> 00:04:44,767
embedding or something. You're
learning how to learn by getting

83
00:04:44,767 --> 00:04:48,317
like an efficient representation
of reality or something like

84
00:04:48,317 --> 00:04:51,693
this. And then you use that to
get more specialized into a

85
00:04:51,693 --> 00:04:55,418
specific task. Is that, like, is
it along those lines or what do

86
00:04:55,418 --> 00:04:56,000
you think?

87
00:04:56,000 --> 00:04:59,557
Michael Laskey: It's along those
lines. Like if you, you know, if

88
00:04:59,557 --> 00:05:02,293
you're familiar with kind of
like the old ideas of

89
00:05:02,293 --> 00:05:05,686
reinforcement learning, it was
the idea of the agent's policy.

90
00:05:05,686 --> 00:05:09,242
So it was like given a goal, how
it would try to like respond and

91
00:05:09,242 --> 00:05:12,416
achieve that goal. But then
there was the other concept of

92
00:05:12,416 --> 00:05:15,645
the environment model, or the
transition function. And that

93
00:05:15,645 --> 00:05:18,983
was how the agent would in its
head see, okay, well if I'm in

94
00:05:18,983 --> 00:05:22,102
this state and I take this
action, what happens next? And

95
00:05:22,102 --> 00:05:25,002
what's cool about that
representation is it's kind of

96
00:05:25,002 --> 00:05:28,395
like task agnostic. so it's more
like if we can give the robot

97
00:05:28,395 --> 00:05:31,842
intuitive physics and intuitive
sense of how the robot behaves,

98
00:05:31,842 --> 00:05:35,180
we can then have it synthesize
arbitrary tasks. because today

99
00:05:35,180 --> 00:05:38,409
we're push moving, but tomorrow
we might say, well, can you

100
00:05:38,409 --> 00:05:41,802
strengthen it? And it's like,
well, I know how movement works,

101
00:05:41,802 --> 00:05:45,194
I know how semantics works, I
know collision avoidance. Let me

102
00:05:45,194 --> 00:05:47,000
go and do those, tasks with
that.

103
00:05:47,000 --> 00:05:50,776
Audrow Nash: So you're learning
a lot of, like, fundamental

104
00:05:50,776 --> 00:05:54,616
things about how reality works.
That's the goal? Nag, do you

105
00:05:54,616 --> 00:05:55,000
wanna?

106
00:05:55,000 --> 00:05:58,633
Nag Murthy: Yeah, that's exactly
it, right? At, you know,

107
00:05:58,633 --> 00:06:02,393
basically, that's the world
model that, you know, the robot

108
00:06:02,393 --> 00:06:06,090
has in its head. And so, it
uses, it uses those intuitions

109
00:06:06,090 --> 00:06:09,532
to sort of perform these
different tasks. so all these

110
00:06:09,532 --> 00:06:13,547
tasks are like policies that you
can learn that are specific to

111
00:06:13,547 --> 00:06:17,435
that task being performed. So
mowing, sweeping, snow removal,

112
00:06:17,435 --> 00:06:21,451
but at its core, there is a task
agnostic world model, which is

113
00:06:21,451 --> 00:06:24,000
independent of the task being
performed.

114
00:06:24,000 --> 00:06:27,396
Audrow Nash: So that sounds
amazing to me. and I'm

115
00:06:27,396 --> 00:06:31,200
wondering, like, how do you, so
I would wonder, my first

116
00:06:31,200 --> 00:06:35,615
thinking would be physics. like,
our models of physics are pretty

117
00:06:35,615 --> 00:06:39,419
good for probably like lawn
mowing and things like that.

118
00:06:39,419 --> 00:06:43,155
They're not good for fine
manipulation yet, probably. I

119
00:06:43,155 --> 00:06:47,094
mean, I haven't seen work
anyways. how, how do you, do you

120
00:06:47,094 --> 00:06:51,170
just, do you kind of like, where
are you using things like a

121
00:06:51,170 --> 00:06:55,177
physics emulator, and where are
you learning stuff? Are you

122
00:06:55,177 --> 00:06:59,592
learning everything, or how does
it, how does it work? Maybe Mike

123
00:06:59,592 --> 00:07:00,000
again?

124
00:07:00,000 --> 00:07:03,655
Michael Laskey: Yeah, no, great
questions. So I think when you

125
00:07:03,655 --> 00:07:07,251
look at You know, TAS in outdoor
landscape, you're right. You

126
00:07:07,251 --> 00:07:10,670
kind of have like two camps
right now of TAS. There's your

127
00:07:10,670 --> 00:07:14,089
mechanized surface coverages,
your snow plowing, your push

128
00:07:14,089 --> 00:07:17,508
mowing, your street sweeping.
And then there's like really

129
00:07:17,508 --> 00:07:21,045
fine grain, like tree trimming,
where pick an apple, prune a

130
00:07:21,045 --> 00:07:24,464
bush. so what we want to do is
definitely take this like a

131
00:07:24,464 --> 00:07:27,883
roadmap. And we first started
off saying like, okay, well.

132
00:07:27,883 --> 00:07:31,243
There's a huge amount of bulk
labor in mechanized surface

133
00:07:31,243 --> 00:07:34,603
coverage. So let's first just
think, what would the world

134
00:07:34,603 --> 00:07:38,316
model need to do to solve these
tasks? And then build in a way,

135
00:07:38,316 --> 00:07:41,500
though, where it can be
scaffolded to Dexterous later.

136
00:07:41,500 --> 00:07:44,742
so, like, it's, it's pretty
generic, like... AI kind of

137
00:07:44,742 --> 00:07:48,043
tech, so it's definitely
scalable, but now when you look

138
00:07:48,043 --> 00:07:51,403
at like something like a snow
plowing or push mowing, you

139
00:07:51,403 --> 00:07:54,822
really have like, four key
challenging things that need to

140
00:07:54,822 --> 00:07:58,477
be done in order for the robot
to affect the report, like four

141
00:07:58,477 --> 00:08:02,131
like kind of key AI concepts. So
one of them is, so think just

142
00:08:02,131 --> 00:08:05,079
for mowing for a cleaning
sample, you need to have

143
00:08:05,079 --> 00:08:08,616
collision avoidance. But it's
subtle because some things you

144
00:08:08,616 --> 00:08:12,153
actually have to get very close
to and touch. So you have to

145
00:08:12,153 --> 00:08:15,395
have this concept we would call
like semantic collision

146
00:08:15,395 --> 00:08:18,873
avoidance. Like understanding
like, oh, I can push my mower

147
00:08:18,873 --> 00:08:22,645
right next to a bush and mow an
area, but when I'm in like, when

148
00:08:22,645 --> 00:08:26,300
I'm next to like a human or a
tree, I can't. and then you need

149
00:08:26,300 --> 00:08:29,955
obviously mapping. You have to
have some sort of global map of

150
00:08:29,955 --> 00:08:33,492
the world to know what you've
mowed before and what you have

151
00:08:33,492 --> 00:08:37,206
mowed. And then for commercial
mowing, you need actually, a lot

152
00:08:37,206 --> 00:08:40,448
of people don't, this is
actually a really hard nuanced

153
00:08:40,448 --> 00:08:43,926
problem of mowing. You need to
lay down these very precise,

154
00:08:43,926 --> 00:08:47,522
straight lines, so you need high
quality localization. better

155
00:08:47,522 --> 00:08:50,941
than like most people would try
and get for in like a self

156
00:08:50,941 --> 00:08:54,772
driving car application. cause a
professional landscaper looks at

157
00:08:54,772 --> 00:08:58,309
these lines and they basically
say if this robot was good or

158
00:08:58,309 --> 00:09:01,669
not. And to do that you need to
have this very precise...

159
00:09:01,669 --> 00:09:05,383
Ability to follow lines and lay
them down over like hundreds of

160
00:09:05,383 --> 00:09:08,743
meters of stripes. so when you
combine all that, you need

161
00:09:08,743 --> 00:09:11,749
mapping, localization, 3D
awareness, the ability to

162
00:09:11,749 --> 00:09:15,286
understand what's traversable,
what's not. And these are the

163
00:09:15,286 --> 00:09:19,000
core concepts. So, like you're
saying, you don't really need...

164
00:09:19,000 --> 00:09:21,754
Audrow Nash: What was the fourth
one? I have collision avoidance,

165
00:09:21,754 --> 00:09:24,000
mapping, and straight lines.
What was the fourth one?

166
00:09:24,000 --> 00:09:26,990
Michael Laskey: Semantic
traversability, I would say.

167
00:09:26,990 --> 00:09:30,092
Like, really understanding,
like, what you can mow over

168
00:09:30,092 --> 00:09:33,646
versus not. which is almost like
a semantic problem of like, is

169
00:09:33,646 --> 00:09:35,000
this mowable, basically.

170
00:09:35,000 --> 00:09:39,063
Nag Murthy: So take leaves on
the ground in the fall, right?

171
00:09:39,063 --> 00:09:42,922
It's okay to go over them. Like
a toy bunny rabbit on the

172
00:09:42,922 --> 00:09:46,984
ground, not okay. You know, so,
it's a, sprinkler heads, not

173
00:09:46,984 --> 00:09:48,000
okay. Not okay.

174
00:09:48,000 --> 00:09:50,000
Audrow Nash: Big bummer. Mm hmm.

175
00:09:50,000 --> 00:09:53,765
Michael Laskey: Yeah, and like,
what most products in the market

176
00:09:53,765 --> 00:09:57,176
do today is you teach it,
through like, you know, precise,

177
00:09:57,176 --> 00:10:00,706
like, joysticking, how to, just
avoiding things. what's very

178
00:10:00,706 --> 00:10:04,176
different about what we're
building is there's no teaching.

179
00:10:04,176 --> 00:10:07,882
This robot must semantically see
the world and understand if it

180
00:10:07,882 --> 00:10:09,000
can go over or not.

181
00:10:09,000 --> 00:10:12,675
Audrow Nash: Okay, so how do you
do that? How do you semantically

182
00:10:12,675 --> 00:10:16,180
understand the world? and maybe
the best way to start would be

183
00:10:16,180 --> 00:10:19,233
like, are you using, what
sensors are you using? Is it

184
00:10:19,233 --> 00:10:22,625
your cameras primarily, is what
I would imagine? Maybe depth

185
00:10:22,625 --> 00:10:25,000
cameras, or what are you guys
doing there?

186
00:10:25,000 --> 00:10:29,371
Michael Laskey: Yeah, so our
robots are equipped, I'm a huge

187
00:10:29,371 --> 00:10:33,596
believer in stereo vision. So
the robots are equipped with

188
00:10:33,596 --> 00:10:36,000
just one stereo camera. just
one.

189
00:10:36,000 --> 00:10:39,686
Audrow Nash: That's super cool.
Just one, yeah. You don't have

190
00:10:39,686 --> 00:10:43,551
like 40 all posted around, which
is what it seems like many do. I

191
00:10:43,551 --> 00:10:47,000
mean like autonomous cars will
have them everywhere. Okay,

192
00:10:47,000 --> 00:10:50,155
Michael Laskey: yeah, I mean we
very reduced the like weight and

193
00:10:50,155 --> 00:10:52,718
size of our platform that we
felt like from a safety

194
00:10:52,718 --> 00:10:55,725
standpoint You could do one and
one's interesting because now

195
00:10:55,725 --> 00:10:58,585
you're in the domain of like a
human, right? You have like

196
00:10:58,585 --> 00:11:01,789
similar like human eyes what you
mostly see in other animals This

197
00:11:01,789 --> 00:11:04,894
also forces you to reason across
time and space so you can keep

198
00:11:04,894 --> 00:11:08,000
track of like what's around you
How do you look around and like

199
00:11:08,000 --> 00:11:11,235
Audrow Nash: which is super
cool? Yeah, that's why it's

200
00:11:11,235 --> 00:11:14,529
surprising that you have one
Okay, so then you take that

201
00:11:14,529 --> 00:11:18,118
stereo Image Data. So you're
getting 3d information about the

202
00:11:18,118 --> 00:11:21,706
world and image information.
You've kind of are superimposing

203
00:11:21,706 --> 00:11:25,000
them. And then you're running
that through some. Sort of

204
00:11:25,000 --> 00:11:28,294
neural network to establish
semantic understanding, like

205
00:11:28,294 --> 00:11:31,824
object recognition or what do
you do from there? Or it's, it

206
00:11:31,824 --> 00:11:35,294
goes into just some neural
network and, you'd say good, bad

207
00:11:35,294 --> 00:11:39,000
and it eventually does the good
thing, or what, how does it go?

208
00:11:39,000 --> 00:11:42,569
Michael Laskey: Yeah, so right
now how it works is we feed a

209
00:11:42,569 --> 00:11:46,257
stream of video stream stereo
video into what we're calling ES

210
00:11:46,257 --> 00:11:49,409
one. and then the, also the
other thing we feed in is

211
00:11:49,409 --> 00:11:52,680
uncorrected GPS, so non
R-T-K-G-P-S. This is relatively

212
00:11:52,680 --> 00:11:56,190
cheap, simple to get, and you
kind of want, like, Nog has a

213
00:11:56,190 --> 00:11:59,877
good analogy that kind of, like,
relates it to, like, pigeons,

214
00:11:59,877 --> 00:12:03,208
like, how they use magnetic
sensing to navigate. So, you

215
00:12:03,208 --> 00:12:06,717
know, it's kind of like, it's
something that is there, it's

216
00:12:06,717 --> 00:12:10,524
very easy to get when you don't
want non RTK, so we feed that in

217
00:12:10,524 --> 00:12:11,000
as well.

218
00:12:11,000 --> 00:12:14,046
Audrow Nash: I love that
approach, because, so you're

219
00:12:14,046 --> 00:12:17,724
learning, like, RTK GPS is very
accurate. It's like, what is it,

220
00:12:17,724 --> 00:12:21,000
like centimeter or millimeter
level accuracy? On the good

221
00:12:21,000 --> 00:12:23,000
Nag Murthy: days, yeah.

222
00:12:23,000 --> 00:12:26,100
Audrow Nash: Yeah. Well, yeah,
it depends on coverage, depends

223
00:12:26,100 --> 00:12:29,150
if you have beacons or something
set up for it to triangulate

224
00:12:29,150 --> 00:12:32,350
with, all sorts of things. But,
it's very expensive too. and all

225
00:12:32,350 --> 00:12:35,200
this thing. So you're just
using, or maybe it's come down

226
00:12:35,200 --> 00:12:36,000
in price, right?

227
00:12:36,000 --> 00:12:39,737
Nag Murthy: It's, it's not so
much the expense. I think, sure,

228
00:12:39,737 --> 00:12:43,293
there is, you, you get like
those Trimble GPSs, which costs

229
00:12:43,293 --> 00:12:46,909
like seven, 10, 000, maybe more,
but then you also have like

230
00:12:46,909 --> 00:12:50,646
these audio simple GPSs, right,
which are like in the, but the

231
00:12:50,646 --> 00:12:54,383
problem, the RTK corrections is
you can't rely on them. Right.

232
00:12:54,383 --> 00:12:57,698
So you have all these multi
pathing issues. When you go

233
00:12:57,698 --> 00:13:01,555
close to the walls or, you know,
when you're in an urban canyon,

234
00:13:01,555 --> 00:13:05,413
right, you really don't get like
the right, you know, signals to

235
00:13:05,413 --> 00:13:09,089
get to you, or there's a lot of
distortion of the signal that

236
00:13:09,089 --> 00:13:12,705
happens. Ionospheric activity
can like wreak havoc with like

237
00:13:12,705 --> 00:13:16,201
your GPS, right? All kinds of
crazy things like affect it.

238
00:13:16,201 --> 00:13:19,817
Cause I mean, think about it,
right? Like you're having like

239
00:13:19,817 --> 00:13:23,374
these little basketball size
satellites, like whizzing past

240
00:13:23,374 --> 00:13:26,628
it, like, I don't know, God
knows how, how fast you're

241
00:13:26,628 --> 00:13:30,124
trying to like get like signals
and synchronize them using

242
00:13:30,124 --> 00:13:33,379
atomic clocks or something,
right? Like, so it's, it's

243
00:13:33,379 --> 00:13:37,055
again, like very classic in its
sort of modality, right? Like

244
00:13:37,055 --> 00:13:40,551
precision GPS. That's kind of
where all of this is. And we

245
00:13:40,551 --> 00:13:44,167
don't like classic modalities
fundamentally because you have

246
00:13:44,167 --> 00:13:47,000
to get everything right and
everything precise.

247
00:13:47,000 --> 00:13:50,840
Audrow Nash: Oh, so you, I think
I want to talk about that quite

248
00:13:50,840 --> 00:13:54,740
a bit, but you, at a high level.
Think they're finicky because of

249
00:13:54,740 --> 00:13:58,580
the precision that you require.
And so you want robust solutions

250
00:13:58,580 --> 00:14:02,000
and that's why you rely on GPS
and these kinds of things.

251
00:14:02,000 --> 00:14:05,182
Michael Laskey: When you scale
the products with it, on good

252
00:14:05,182 --> 00:14:08,152
days when you get it, it's
really amazing. But there's a

253
00:14:08,152 --> 00:14:09,000
lot of bad days.

254
00:14:09,000 --> 00:14:12,642
Audrow Nash: But finicky and
lots of failure cases and, uh,

255
00:14:12,642 --> 00:14:14,000
sorry, Mike, go ahead.

256
00:14:14,000 --> 00:14:18,105
Michael Laskey: Yeah, exactly.
It's an unreliable service that

257
00:14:18,105 --> 00:14:21,813
in a lot of places in the
country you just can't rely on

258
00:14:21,813 --> 00:14:25,918
at all. Or on different weather
days. And you also are pushing

259
00:14:25,918 --> 00:14:29,361
your bill of materials to be
cheaper, which is a big

260
00:14:29,361 --> 00:14:33,599
advantage too. like having only
one stereo camera means you need

261
00:14:33,599 --> 00:14:37,505
only one input. One chip thing
to accept one stereo camera.

262
00:14:37,505 --> 00:14:41,743
Pair of information. Like, it's
not like you need a custom board

263
00:14:41,743 --> 00:14:45,981
that goes to 30 stereo cameras.
And it's like, it makes all the,

264
00:14:45,981 --> 00:14:50,086
and then you don't need a crazy
processor to process all that.

265
00:14:50,086 --> 00:14:54,058
You're not storing an insane
amount of data that you have to

266
00:14:54,058 --> 00:14:57,965
upload and figure out all that.
It makes all the downstream

267
00:14:57,965 --> 00:15:01,805
problems also easier. I would
imagine to do these kinds of

268
00:15:01,805 --> 00:15:05,977
make these decisions. It seems
like power moves to me, like you

269
00:15:05,977 --> 00:15:10,082
guys are making power moves in
how you're doing stuff. We said

270
00:15:10,082 --> 00:15:13,723
ML was going to drive every
major decision, and then we

271
00:15:13,723 --> 00:15:17,498
designed the stack to force ML
to do that. So we said, no

272
00:15:17,498 --> 00:15:21,272
external calibration, no
ability, no, yeah, no redundancy

273
00:15:21,272 --> 00:15:22,000
in sensing.

274
00:15:22,000 --> 00:15:25,729
Nag Murthy: Here's the, yeah,
this is the crazy part, right?

275
00:15:25,729 --> 00:15:29,645
You gotta, there's a history to
all of this, right? And we have

276
00:15:29,645 --> 00:15:33,126
flirted with like classical
approaches, scaled classical

277
00:15:33,126 --> 00:15:36,980
robots. In the first two years
of our existence, right? And we

278
00:15:36,980 --> 00:15:40,461
are deeply aware of like how
amazing it can be, but also

279
00:15:40,461 --> 00:15:44,004
deeply aware of how painful it
gets when you go down that

280
00:15:44,004 --> 00:15:47,857
route, right? Like we had like
these big massive ouster lidars

281
00:15:47,857 --> 00:15:51,835
and like, you know, all kinds of
crazy things. On our robot. And

282
00:15:51,835 --> 00:15:55,565
then we just said like, throw
away all that. Yeah. Hell with

283
00:15:55,565 --> 00:15:56,000
it. You

284
00:15:56,000 --> 00:15:59,031
Audrow Nash: know, I wonder if
it's like a pendulum and you

285
00:15:59,031 --> 00:16:02,216
guys have swung, you started one
way and you're going to swing

286
00:16:02,216 --> 00:16:05,041
the other, and then you're going
to find a happy medium

287
00:16:05,041 --> 00:16:08,123
eventually that's a bit more, in
the middle. I like, I don't

288
00:16:08,123 --> 00:16:11,000
know. And maybe the other
methods are working wonderful.

289
00:16:11,000 --> 00:16:14,906
Nag Murthy: I think, yeah, I
think the thing is, one way to

290
00:16:14,906 --> 00:16:18,878
think about this is, right, I
think that we won't swing back

291
00:16:18,878 --> 00:16:22,983
primarily because this is now a
data problem at its core. It's

292
00:16:22,983 --> 00:16:27,153
not like a sensor problem. It's
not a calibration problem. It's

293
00:16:27,153 --> 00:16:30,000
not a sensor quality problem.
None of that.

294
00:16:30,000 --> 00:16:33,124
Audrow Nash: Interesting, yeah,
so you don't have most of those

295
00:16:33,124 --> 00:16:36,000
problems now to solve with
classical methods. Interesting.

296
00:16:36,000 --> 00:16:38,685
Nag Murthy: Every time you have
a classical method, you always

297
00:16:38,685 --> 00:16:41,413
come down to like, how good is
your accelerometer, right? Like,

298
00:16:41,413 --> 00:16:43,925
do you have like high quality
accelerometer? Does it like,

299
00:16:43,925 --> 00:16:46,610
cause that's the thing that ties
in like your whole slam stack

300
00:16:46,610 --> 00:16:47,000
together.

301
00:16:47,000 --> 00:16:51,488
Audrow Nash: Yeah. So how does,
so with your normal GPS, you're

302
00:16:51,488 --> 00:16:55,334
basically saying we think this
is going to be somewhat

303
00:16:55,334 --> 00:16:59,679
unreliable and therefore your
system will behave accordingly.

304
00:16:59,679 --> 00:17:03,953
For this kind of thing, it's
just that you'll adjust for the

305
00:17:03,953 --> 00:17:08,584
information that isn't good, but
is something. is that, that kind

306
00:17:08,584 --> 00:17:13,000
of the thinking with not using
RTK and just using regular GPS?

307
00:17:13,000 --> 00:17:17,683
Michael Laskey: Yeah, cause, so
this all feeds into what we call

308
00:17:17,683 --> 00:17:21,634
AS1. One of the outputs of AS1
is essentially maps and

309
00:17:21,634 --> 00:17:25,659
localization. now... Non
corrected GPS is only so good,

310
00:17:25,659 --> 00:17:30,049
but when you combine that with,
let's say, a Learn VO that's

311
00:17:30,049 --> 00:17:34,000
happening interior to ES1, you
can... Visual odometry.

312
00:17:34,000 --> 00:17:36,766
Audrow Nash: Oh, okay. Acronyms.
I don't even know them in my own

313
00:17:36,766 --> 00:17:38,000
field. Okay, visual odometry.

314
00:17:38,000 --> 00:17:41,431
Michael Laskey: Perfect. Yeah,
so another way of doing

315
00:17:41,431 --> 00:17:45,054
localization is to use, of
course, camera data. It's like

316
00:17:45,054 --> 00:17:48,676
how humans do. Walk around, we
see movement with our head

317
00:17:48,676 --> 00:17:52,488
related to vision. So, internal
to our network, it's able to

318
00:17:52,488 --> 00:17:56,365
kind of fuse the GPS with visual
data and produce better pose

319
00:17:56,365 --> 00:17:57,000
estimates.

320
00:17:57,000 --> 00:18:02,060
Audrow Nash: Mm hmm. Okay, so
you get your pose estimate. So

321
00:18:02,060 --> 00:18:04,000
maybe Nag, what is ES1?

322
00:18:04,000 --> 00:18:06,919
Nag Murthy: I think that's a
question better answered by

323
00:18:06,919 --> 00:18:10,047
Mike, but I'll take a stab at
it. Like ES1 is essentially, I

324
00:18:10,047 --> 00:18:13,280
joke about it and call it like
ChatGPT, if you want to call it

325
00:18:13,280 --> 00:18:15,000
that, right? Which is
essentially

326
00:18:15,000 --> 00:18:18,000
Audrow Nash: like computer aided
design, like 3d parts.

327
00:18:18,000 --> 00:18:22,080
Nag Murthy: No, no, no. Like,
no, it's like, it was both. CAT,

328
00:18:22,080 --> 00:18:26,029
so it's, if you think, yeah, so
if you think of it this way,

329
00:18:26,029 --> 00:18:29,846
right, like ChatGPT was to
language, ChatGPT fundamentally

330
00:18:29,846 --> 00:18:34,124
groks language, right? Fine. All
the naysayers aside who question

331
00:18:34,124 --> 00:18:38,007
whether it has an internal
representation of what languages

332
00:18:38,007 --> 00:18:42,022
and all that, but it clearly has
some kind of, you know, it's

333
00:18:42,022 --> 00:18:45,510
some kind of magic going on in
there where it sort of

334
00:18:45,510 --> 00:18:49,130
understands the structure of
language, the grammar, the

335
00:18:49,130 --> 00:18:53,013
nuances, right? Not just one
language across languages. And

336
00:18:53,013 --> 00:18:57,093
we're saying there's got to be
something similar when it comes

337
00:18:57,093 --> 00:19:01,239
to embodied AI, where, and it's
non verbal intelligence, if you

338
00:19:01,239 --> 00:19:05,122
want to call it that. So ES1
really is something that, sort

339
00:19:05,122 --> 00:19:09,071
of captures sort of that non
verbal, intelligence. Like it's

340
00:19:09,071 --> 00:19:12,296
a way for... Like a robot to
sort of navigate its

341
00:19:12,296 --> 00:19:16,508
surroundings, or it's the means
by which the robot navigates its

342
00:19:16,508 --> 00:19:20,260
surroundings, manipulates
physical reality. And it's this

343
00:19:20,260 --> 00:19:24,274
one giant sort of world model
neural net, if you want to call

344
00:19:24,274 --> 00:19:28,355
it that. Maybe Mike can add to
this, but that's how I describe

345
00:19:28,355 --> 00:19:30,000
it as a product guy. When

346
00:19:30,000 --> 00:19:33,994
Audrow Nash: I love the, the cat
GPT, so cat, like the animal, it

347
00:19:33,994 --> 00:19:37,436
like gets around. It's a
physical representation. That's

348
00:19:37,436 --> 00:19:41,000
super cool. Okay. Mike, do you
have anything to add to it?

349
00:19:41,000 --> 00:19:44,677
Michael Laskey: No, I'm gonna
just completely go with what Nox

350
00:19:44,677 --> 00:19:48,176
said there. No, I think it is
that it's it's basically like

351
00:19:48,176 --> 00:19:51,853
right now it is like embodied AI
and it's supposed to be going

352
00:19:51,853 --> 00:19:55,234
after the predictions that a
physical agent would need to

353
00:19:55,234 --> 00:19:58,970
solve tasks and like look at the
key concepts that we're trying

354
00:19:58,970 --> 00:20:02,410
to do of like You know just
surface coverage right with no

355
00:20:02,410 --> 00:20:05,909
map, no teaching and right now
that's what ES1 is trying to

356
00:20:05,909 --> 00:20:09,468
predict is those fundamental
representations for an agent to

357
00:20:09,468 --> 00:20:13,145
do a task like push, mow, snow,
blow. and then we want to keep

358
00:20:13,145 --> 00:20:16,644
building on it to get more and
more of the embodied AI full

359
00:20:16,644 --> 00:20:17,000
stack.

360
00:20:17,000 --> 00:20:21,650
Audrow Nash: So, you, with ES1,
you take in GPS and video, and

361
00:20:21,650 --> 00:20:25,775
you output pose and maybe
actions, like motor commands,

362
00:20:25,775 --> 00:20:29,000
or, what is that, what is the
output again?

363
00:20:29,000 --> 00:20:31,864
Michael Laskey: So the output,
right now it kind of outputs

364
00:20:31,864 --> 00:20:34,000
like, kind of a plethora of
representations.

365
00:20:34,000 --> 00:20:36,000
Audrow Nash: Yeah, it's probably
a big tuple of a lot of things.

366
00:20:36,000 --> 00:20:38,845
Michael Laskey: It's a big
tuple. Yeah. It's like, it's

367
00:20:38,845 --> 00:20:41,897
like, which you kind of see in
like, you know, like, the, I

368
00:20:41,897 --> 00:20:44,948
think like perceiver.io like
papers where you have like now

369
00:20:44,948 --> 00:20:48,000
like many, like one to many type
of representation outputs.

370
00:20:48,000 --> 00:20:50,900
Audrow Nash: Yeah. 'cause you
need lots of information for

371
00:20:50,900 --> 00:20:52,000
your robots. For sure.

372
00:20:52,000 --> 00:20:55,479
Michael Laskey: Exactly. So it
some of the, so, you know, it's

373
00:20:55,479 --> 00:20:59,071
kind of interesting 'cause right
now we're sort of testing it by

374
00:20:59,071 --> 00:21:02,494
predicting like some of the more
classic representations. And

375
00:21:02,494 --> 00:21:05,973
then we have a planner that can
like synthesize the, basically

376
00:21:05,973 --> 00:21:09,508
policies can be synthesized with
a planner on top. but, I think

377
00:21:09,508 --> 00:21:12,988
as a whole, like, we'll probably
at one point start getting to

378
00:21:12,988 --> 00:21:16,411
learn planning. But, what it
does today, is it outputs, like,

379
00:21:16,411 --> 00:21:19,722
a point cloud of the world, a
semantic understanding of the

380
00:21:19,722 --> 00:21:23,033
world, so like a 3D semantic
point cloud. a bird's eye view

381
00:21:23,033 --> 00:21:26,231
map, so it says, here's where
you are, here's what you've

382
00:21:26,231 --> 00:21:29,262
covered before, the exact
position of the robot on the

383
00:21:29,262 --> 00:21:32,404
map. so... And then, also,
what's kind of interesting is

384
00:21:32,404 --> 00:21:35,547
like, low level obstacle
detection. So I would say like,

385
00:21:35,547 --> 00:21:38,745
oh, here's a manhole cover.
Like, oh, here's your child's

386
00:21:38,745 --> 00:21:42,000
bunny. This is a 3D cuboid.
Like, do not go near it. Yeah.

387
00:21:42,000 --> 00:21:46,413
Audrow Nash: How do you, I mean,
there's a lot of questions, I

388
00:21:46,413 --> 00:21:50,683
have a lot of questions about
that, just in general. So, how

389
00:21:50,683 --> 00:21:54,313
much is it like, So I
interviewed, like, many years

390
00:21:54,313 --> 00:21:58,868
ago, Sergey Levin from Berkeley.
so he's a professor who's like,

391
00:21:58,868 --> 00:22:03,281
End to end robotics stuff, so
you basically pass in raw sensor

392
00:22:03,281 --> 00:22:07,267
input and you output, like,
actions to do the task. this

393
00:22:07,267 --> 00:22:11,893
seems a bit like that, and maybe
you have a better description of

394
00:22:11,893 --> 00:22:16,377
his work for this kind of thing,
but how, it seems like this is

395
00:22:16,377 --> 00:22:20,719
very similar in spirit to me,
and it might be the same thing.

396
00:22:20,719 --> 00:22:22,000
what do you think?

397
00:22:22,000 --> 00:22:25,378
Michael Laskey: Yeah, I mean, so
he was, he was on my cause

398
00:22:25,378 --> 00:22:29,041
committee, so I'm pretty used to
debating with Sergey, actually.

399
00:22:29,041 --> 00:22:32,705
but, so I think, I'm, like, at a
very high level, like, I really

400
00:22:32,705 --> 00:22:35,854
agree with a lot of his
approaches. and if you look at,

401
00:22:35,854 --> 00:22:39,346
like, the body of his work, it
really says, like, yes, robots

402
00:22:39,346 --> 00:22:42,952
need to learn from actions and
experiences. You need to collect

403
00:22:42,952 --> 00:22:46,502
a lot of data, and they need to
have recitations that are good

404
00:22:46,502 --> 00:22:49,593
enough they can actually
efficiently learn. and search

405
00:22:49,593 --> 00:22:53,028
algorithms in RL. Where I think
ES1 is slightly different is

406
00:22:53,028 --> 00:22:56,062
instead of just predicting,
like, let's say like it's

407
00:22:56,062 --> 00:22:59,268
trained policy, for example,
well, okay, just a bit of a

408
00:22:59,268 --> 00:23:02,531
caveat. It's kind of hard to
argue with Sergei as a whole

409
00:23:02,531 --> 00:23:05,909
because he's done so many papers
that it's like, what's the

410
00:23:05,909 --> 00:23:09,000
specific one that, you know,
you're trying to predict?

411
00:23:09,000 --> 00:23:12,340
Audrow Nash: Yeah, for sure.
Because ES thinking is moving.

412
00:23:12,340 --> 00:23:15,000
It's like a big tree advancing.
Right, exactly.

413
00:23:15,000 --> 00:23:19,082
Michael Laskey: For sure. So,
but if you just look at maybe,

414
00:23:19,082 --> 00:23:23,367
like, the RL community, rather
than learning a specific policy,

415
00:23:23,367 --> 00:23:27,313
what we're trying to learn is
like a representation of the

416
00:23:27,313 --> 00:23:31,531
environment. And then this is
something that you can call like

417
00:23:31,531 --> 00:23:35,000
planners or learn planners that
can operate on top.

418
00:23:35,000 --> 00:23:38,672
Audrow Nash: Okay, the way that
I would think of it, and maybe

419
00:23:38,672 --> 00:23:42,523
this isn't correct, and I, maybe
just, when I was in grad school,

420
00:23:42,523 --> 00:23:45,425
all the things were like
embeddings and you learn

421
00:23:45,425 --> 00:23:48,920
efficient embeddings for things.
so you learn a good way of

422
00:23:48,920 --> 00:23:52,237
representing something and then
you can do like one shot

423
00:23:52,237 --> 00:23:55,376
learning or something if you
have a really efficient.

424
00:23:55,376 --> 00:23:58,812
embedding that lets you learn a
new task really quickly or

425
00:23:58,812 --> 00:24:02,484
something like this. Is it, is
it at all similar to this or is

426
00:24:02,484 --> 00:24:06,038
it, I don't know, how does it
fit with that? Or is that just

427
00:24:06,038 --> 00:24:09,000
like an outdated model? or just
a different model.

428
00:24:09,000 --> 00:24:12,672
Michael Laskey: I think that's
where you would actually want to

429
00:24:12,672 --> 00:24:16,110
go. So right now we actually
predict presentations that are

430
00:24:16,110 --> 00:24:19,781
more human interpretable, like a
point cloud or a semantic map.

431
00:24:19,781 --> 00:24:23,395
yeah. And then we call like, you
know, a planner that we might

432
00:24:23,395 --> 00:24:26,891
code on top of that. And that,
that works for today. It also

433
00:24:26,891 --> 00:24:29,747
works for scaling a product
because you want that

434
00:24:29,747 --> 00:24:33,302
interpretability as a, as an
engineer, especially when you're

435
00:24:33,302 --> 00:24:36,799
trying to. Product fit. There's
a lot of UX that you have to

436
00:24:36,799 --> 00:24:39,946
code. So there's a lot of
practical reasons that we do

437
00:24:39,946 --> 00:24:43,384
that. But as we evolve and
scale, our sole goal would be to

438
00:24:43,384 --> 00:24:47,056
replace everything, you know, to
you. Just get to a point where

439
00:24:47,056 --> 00:24:50,319
it actually predicts like,
here's a thousand dimensional

440
00:24:50,319 --> 00:24:53,000
float vector. And then the
planner takes that.

441
00:24:53,000 --> 00:24:57,000
Audrow Nash: but raw input from
the sensors. Yeah. All that.

442
00:24:57,000 --> 00:25:00,221
Michael Laskey: But like having
worked a lot in my PhD with

443
00:25:00,221 --> 00:25:03,496
systems like that, when you're
actually deploying something,

444
00:25:03,496 --> 00:25:06,608
you have an end customer and
they're like. Oh yeah, why'd

445
00:25:06,608 --> 00:25:09,829
your robot like hit a tree?
Yeah, then you, you really want

446
00:25:09,829 --> 00:25:12,886
somebody to fall back on of
saying, well this is what it

447
00:25:12,886 --> 00:25:16,216
thought, and you can as a human
interpret that because it's a

448
00:25:16,216 --> 00:25:19,000
representation that makes sense
to, people, really.

449
00:25:19,000 --> 00:25:23,242
Audrow Nash: Yep, so what's the,
what's the point of, so if

450
00:25:23,242 --> 00:25:27,124
you're outputting a point
cloud... How different is it

451
00:25:27,124 --> 00:25:31,294
than just a normal point cloud
that you'd get from a depth

452
00:25:31,294 --> 00:25:35,608
camera? I guess it has some
semantics included, or like, why

453
00:25:35,608 --> 00:25:39,993
not just go right to a point
cloud from your sensor? For this

454
00:25:39,993 --> 00:25:41,000
kind of thing.

455
00:25:41,000 --> 00:25:43,235
Michael Laskey: Yeah, I mean
well basically because we're

456
00:25:43,235 --> 00:25:45,471
just feeding in left and right
stereo You know one of the

457
00:25:45,471 --> 00:25:47,000
representations that will do is
create.

458
00:25:47,000 --> 00:25:50,396
Audrow Nash: Oh, you're not
doing any of the calibration and

459
00:25:50,396 --> 00:25:53,000
things like this You were saying
is that true?

460
00:25:53,000 --> 00:25:56,877
Michael Laskey: Yeah, so you can
just feed in like two images

461
00:25:56,877 --> 00:26:00,500
With minimal you can have a
pretty like rough estimate of

462
00:26:00,500 --> 00:26:04,250
your baseline, but you don't
need like machine tolerance on

463
00:26:04,250 --> 00:26:08,000
that Like you can be off with
like a couple of centimeters.

464
00:26:08,000 --> 00:26:12,739
Audrow Nash: Okay. So let's just
backing up a bit. So what we

465
00:26:12,739 --> 00:26:17,477
have so far is you guys have
this ES1 system. This ES1 system

466
00:26:17,477 --> 00:26:21,361
takes camera and GPS and it
outputs a big bunch of

467
00:26:21,361 --> 00:26:25,556
information, including like a
point cloud, maybe motor

468
00:26:25,556 --> 00:26:30,605
commands, all this stuff. and so
the goal is to have this be more

469
00:26:30,605 --> 00:26:35,422
of like a learning system that
you can make more flexible over

470
00:26:35,422 --> 00:26:39,850
time. And it's starting with
lawn mowing, but potentially

471
00:26:39,850 --> 00:26:44,355
getting into a bunch of other
applications. that all seems

472
00:26:44,355 --> 00:26:49,171
wonderful to me. It's, I wonder
how you, so how do, how do you

473
00:26:49,171 --> 00:26:53,988
get your data, I suppose, for
all this? Like, and like, do you

474
00:26:53,988 --> 00:26:58,804
use simulation or how did, how
did you get started? I suppose.

475
00:26:58,804 --> 00:27:03,465
cause getting started to me
sounds like it sounds wonderful,

476
00:27:03,465 --> 00:27:08,203
but it sounds like you need a
lot of data to even get started

477
00:27:08,203 --> 00:27:11,000
and maybe I misunderstand
something.

478
00:27:11,000 --> 00:27:14,701
Nag Murthy: Yeah. I think Mike
can go into the Simtorial part,

479
00:27:14,701 --> 00:27:18,463
but like there is a precursor to
this, which is, you know, when

480
00:27:18,463 --> 00:27:21,687
we were in operation for the
first couple of years, we

481
00:27:21,687 --> 00:27:25,328
generated a ton of data. Right.
That helps seed some of these

482
00:27:25,328 --> 00:27:28,970
efforts. And then Mike brought
in sort of the Simtorial angle

483
00:27:28,970 --> 00:27:32,433
to this, which he could talk
about, and then sort of using

484
00:27:32,433 --> 00:27:35,836
both of those got us to like a
80, 90 percent good enough

485
00:27:35,836 --> 00:27:39,597
product, and then from there on,
yeah, it's, it's like our data

486
00:27:39,597 --> 00:27:43,358
engine that really drives those.
Like, corner case improvements

487
00:27:43,358 --> 00:27:47,000
over time, but Mike can delve
more into the sim to real part.

488
00:27:47,000 --> 00:27:50,529
Audrow Nash: Or just before we
get into the sim to real part,

489
00:27:50,529 --> 00:27:54,000
so, more background on Electric
Sheep. So, how long have you

490
00:27:54,000 --> 00:27:56,950
guys been around, and Nag,
you've been, you're a co

491
00:27:56,950 --> 00:28:00,306
founder, right? That's right,
yeah. How long have you guys

492
00:28:00,306 --> 00:28:01,000
been around?

493
00:28:01,000 --> 00:28:05,624
Nag Murthy: three and a half
years. we got our first, we

494
00:28:05,624 --> 00:28:10,000
raised, like, a pre seed round
in, like, Jan of 2020.

495
00:28:10,000 --> 00:28:13,000
Audrow Nash: Wow, right as the
world shut down.

496
00:28:13,000 --> 00:28:16,663
Nag Murthy: I know, like just
snuck it right in, you know,

497
00:28:16,663 --> 00:28:19,000
just like before the world shut
down.

498
00:28:19,000 --> 00:28:22,644
Audrow Nash: Okay. Yeah. So then
for the first two years or

499
00:28:22,644 --> 00:28:26,288
something, you said you were
trying classical methods, like

500
00:28:26,288 --> 00:28:30,240
it was the same idea in a sense,
because I mean we talked before

501
00:28:30,240 --> 00:28:34,131
this conversation, Nag, you and
I, and Like you guys have a lot

502
00:28:34,131 --> 00:28:37,714
of interesting ideas about how
to do things and it's, it's

503
00:28:37,714 --> 00:28:41,543
surprising to hear, it's like
revolut I don't know, awesome on

504
00:28:41,543 --> 00:28:45,248
the technical side, awesome on
the business side, awesome on

505
00:28:45,248 --> 00:28:49,078
the, everything. So it's, it's
cool to see it coming together.

506
00:28:49,078 --> 00:28:52,907
But it makes sense that you guys
have a bit more history where

507
00:28:52,907 --> 00:28:56,427
you've been trying things and
you probably had those like

508
00:28:56,427 --> 00:29:00,318
different business model ideas
back then and then you're coming

509
00:29:00,318 --> 00:29:03,777
around to the technical
decisions on things and bringing

510
00:29:03,777 --> 00:29:06,000
on great people like Mike. For
this.

511
00:29:06,000 --> 00:29:10,063
Nag Murthy: Exactly. I think
it's been, we've always wanted

512
00:29:10,063 --> 00:29:14,194
to sort of build this full stack
business. You know, when we

513
00:29:14,194 --> 00:29:18,326
first sketched out the idea on a
paper napkin between me and

514
00:29:18,326 --> 00:29:22,526
Jared, we felt like Jared is my
co founder. He comes from the

515
00:29:22,526 --> 00:29:26,795
commercial landscaping industry
and he's like, yeah, he's like

516
00:29:26,795 --> 00:29:30,858
the man when it comes to all
things landscaping. Right. So,

517
00:29:30,858 --> 00:29:34,645
and it's, it's, it's important
to sort of bring that up

518
00:29:34,645 --> 00:29:38,570
because. What shapes our
company, right, is this sense of

519
00:29:38,570 --> 00:29:43,046
intense pragmatism about what we
want to solve, right, and how we

520
00:29:43,046 --> 00:29:47,384
want to solve it. and it's, it's
sort of the confluence of, you

521
00:29:47,384 --> 00:29:51,723
know, three things really. One
is this, sort of mindset on like

522
00:29:51,723 --> 00:29:55,647
the, the business model is the
ultimate goal. That's that

523
00:29:55,647 --> 00:29:59,641
mindset combined with sort of
the operations know how that

524
00:29:59,641 --> 00:30:03,842
Jared brings combined with the
machine learning and autonomy,

525
00:30:03,842 --> 00:30:07,698
smart that Mike brings. Those
three things all shape our

526
00:30:07,698 --> 00:30:12,174
company and shape our vision and
they reinforce each other really

527
00:30:12,174 --> 00:30:13,000
at the core.

528
00:30:13,000 --> 00:30:16,041
Audrow Nash: They're the pillars
that everything builds on. You

529
00:30:16,041 --> 00:30:19,034
have the strong business model,
the great operations, and then

530
00:30:19,034 --> 00:30:20,000
the great technical.

531
00:30:20,000 --> 00:30:23,782
Nag Murthy: And you need all
three to sort of come together,

532
00:30:23,782 --> 00:30:27,185
for this to actually happen,
because this is a complex

533
00:30:27,185 --> 00:30:30,966
business to build. It isn't
like, you know, and, but, but we

534
00:30:30,966 --> 00:30:34,811
feel like this is the only way
to sort of to build. To do it.

535
00:30:34,811 --> 00:30:38,529
Yeah. Like if, if your end goal
is this moonshot, you know,

536
00:30:38,529 --> 00:30:42,374
large world model, which is a
chat GPT equivalent for outdoor

537
00:30:42,374 --> 00:30:46,345
robotics. This is the only way
to do it. We cannot think of any

538
00:30:46,345 --> 00:30:50,000
other way and we've tried a
bunch of ways. We don't, yeah.

539
00:30:50,000 --> 00:30:53,342
Audrow Nash: Was that the goal
from the beginning? To get the

540
00:30:53,342 --> 00:30:56,521
large world model? Or was it
like, as you're in it, you're

541
00:30:56,521 --> 00:30:59,644
like, we could do this. And then
it's like, maybe this is

542
00:30:59,644 --> 00:31:02,000
feasible. or how, how did that
come around?

543
00:31:02,000 --> 00:31:05,672
Nag Murthy: I think the goal was
to always, we knew the value of

544
00:31:05,672 --> 00:31:09,000
data, right? Like that it was
true, it was clear that data

545
00:31:09,000 --> 00:31:12,270
would sort of lead to an AI
revolution. The question was,

546
00:31:12,270 --> 00:31:15,828
how do you time it, right? And
what business model do you need

547
00:31:15,828 --> 00:31:16,000
to?

548
00:31:16,000 --> 00:31:18,241
Audrow Nash: Oh, definitely.
Yeah. The timing is the really

549
00:31:18,241 --> 00:31:19,000
tricky thing. Right?

550
00:31:19,000 --> 00:31:21,926
Nag Murthy: Exactly. But then
you can, you can sort of

551
00:31:21,926 --> 00:31:25,177
abstract away the timing if you,
and this is what we started

552
00:31:25,177 --> 00:31:28,266
with. We said, look, the
technology is going to catch up.

553
00:31:28,266 --> 00:31:31,517
You know, all of this is going
to catch up, but if you solve

554
00:31:31,517 --> 00:31:34,823
for profit and revenue, and if
you can solve for a profitable

555
00:31:34,823 --> 00:31:38,074
way to get the future on your
side, exactly, then, you know,

556
00:31:38,074 --> 00:31:41,325
whatever is the latest advance
in technology. Can sort of be

557
00:31:41,325 --> 00:31:44,685
folded into what you're doing.
Right. So you don't take on too

558
00:31:44,685 --> 00:31:47,990
much technique, technical debt
at the beginning by committing

559
00:31:47,990 --> 00:31:51,241
to like, you know, this one
approach, right? There's profits

560
00:31:51,241 --> 00:31:54,547
are always good. Like revenue is
always good. Right. Deployed

561
00:31:54,547 --> 00:31:57,852
robots are always good. So solve
for that. And then all these

562
00:31:57,852 --> 00:32:01,103
other things kind of like sort
for themselves. that's always

563
00:32:01,103 --> 00:32:03,000
been our philosophy. Since day
one.

564
00:32:03,000 --> 00:32:06,995
Audrow Nash: Yeah, I love that.
It is aggressively pragmatic, as

565
00:32:06,995 --> 00:32:10,802
you were saying. I think that's
wonderful. Cause yeah, if you

566
00:32:10,802 --> 00:32:14,672
just have the business alive,
like if you, if you just keep it

567
00:32:14,672 --> 00:32:18,293
going for a bunch of years and
it's doing well and you are

568
00:32:18,293 --> 00:32:21,913
making profit, like. You can
always invest in whatever you

569
00:32:21,913 --> 00:32:25,159
think the best thing is the
whole time. and that's a

570
00:32:25,159 --> 00:32:28,654
wonderful thing. versus going
and heavily betting on one

571
00:32:28,654 --> 00:32:32,462
technology and not optimizing
for profit. Like, this is going

572
00:32:32,462 --> 00:32:36,082
to be big, but one day it's
going to pay off. And, then, I

573
00:32:36,082 --> 00:32:39,889
don't know, funding dries up
like it is now. Or, any of these

574
00:32:39,889 --> 00:32:43,135
things, you don't find good
customer fit right away,

575
00:32:43,135 --> 00:32:46,818
whatever, and then the company
dies. So just optimizing for

576
00:32:46,818 --> 00:32:50,688
profit seems really, really good
to me. and Mike, when did you

577
00:32:50,688 --> 00:32:51,000
join?

578
00:32:51,000 --> 00:32:55,686
Michael Laskey: So I joined
about, a year and maybe like

579
00:32:55,686 --> 00:33:01,126
three months ago. Wow. Yeah. and
I remember in the interview, Nog

580
00:33:01,126 --> 00:33:06,565
was like, I want you to make the
mind of a sheep. And I was like,

581
00:33:06,565 --> 00:33:11,000
seems like a really interesting,
like, place to join.

582
00:33:11,000 --> 00:33:14,736
Audrow Nash: Okay, hell yeah.
Just, okay, so that's the

583
00:33:14,736 --> 00:33:18,743
context. So you've been working
on this. Now for a year and

584
00:33:18,743 --> 00:33:22,751
three months, tell me a bit
about, so we were talking about

585
00:33:22,751 --> 00:33:27,166
the sim to real gap for this. So
we, we generated some data, from

586
00:33:27,166 --> 00:33:31,106
the first two years of trying
things. And then, so you add

587
00:33:31,106 --> 00:33:35,113
that to seed things with, we
have examples of mowing and it

588
00:33:35,113 --> 00:33:39,189
working. How, how do we get from
there to more sophisticated

589
00:33:39,189 --> 00:33:43,468
behavior? How do we fold that in
so that you have these learned

590
00:33:43,468 --> 00:33:47,000
controllers effectively? Like,
how does it all work?

591
00:33:47,000 --> 00:33:50,773
Michael Laskey: Yeah, so I think
the data that we started with

592
00:33:50,773 --> 00:33:54,424
was really, like, it kind of,
you know, you start realizing,

593
00:33:54,424 --> 00:33:57,527
like, the breadth of the
problem, the scope, right,

594
00:33:57,527 --> 00:34:00,874
because you've seen the
diversity of mowing. It's a lot

595
00:34:00,874 --> 00:34:04,768
more complex than what you might
think at first, especially when

596
00:34:04,768 --> 00:34:08,115
you're trying to solve it
generically with no teaching.

597
00:34:08,115 --> 00:34:11,522
But just given my background, I
was always drawn towards

598
00:34:11,522 --> 00:34:14,930
synthetic data to start, you
know, building models from.

599
00:34:14,930 --> 00:34:18,763
There's a few reasons why first
pre training synthetic data can

600
00:34:18,763 --> 00:34:22,414
be a nice advantage to a model,
but then actually having the

601
00:34:22,414 --> 00:34:25,822
model go out and then learn from
the real data, from the

602
00:34:25,822 --> 00:34:29,230
interactive data. So that's
generally the paradigm we've

603
00:34:29,230 --> 00:34:32,941
been taking now, is using like
the landscaping companies that

604
00:34:32,941 --> 00:34:36,471
we acquired as like an RL
sandbox to like refine the model

605
00:34:36,471 --> 00:34:40,000
and teach it about the long
tail. But where we start is...

606
00:34:40,000 --> 00:34:42,784
Audrow Nash: What do You mean
the exceptions? You mean the

607
00:34:42,784 --> 00:34:45,520
things that don't happen
frequently? You're... What's the

608
00:34:45,520 --> 00:34:46,000
long tail?

609
00:34:46,000 --> 00:34:49,791
Michael Laskey: Yeah, the long
tail of the real world. So, what

610
00:34:49,791 --> 00:34:53,462
happens is... Like a kid's toy
being left in the lawn or this

611
00:34:53,462 --> 00:34:57,313
kind of thing? Kid's toys, like
interesting properties where you

612
00:34:57,313 --> 00:35:00,623
have like very unclear
boundaries of like grass. One of

613
00:35:00,623 --> 00:35:03,993
the properties that we're
actually looking at, there was

614
00:35:03,993 --> 00:35:07,543
like grass that led right up to
the pool side. So the robot

615
00:35:07,543 --> 00:35:11,154
would actually have to know, oh,
I can't put my cask outside

616
00:35:11,154 --> 00:35:14,644
because I would fall down into
the water, right? And that,

617
00:35:14,644 --> 00:35:17,894
yeah, exactly. Like the
interactions with the physical

618
00:35:17,894 --> 00:35:21,384
world and that world model. We
didn't necessarily cover in

619
00:35:21,384 --> 00:35:25,115
simulation, right? Thinking like
we would be mowing like right

620
00:35:25,115 --> 00:35:28,605
next to a pool and it was a
complete like drop off. so you

621
00:35:28,605 --> 00:35:32,276
can never really, I think, you
know, like McCarthy has like a

622
00:35:32,276 --> 00:35:35,887
really good quote in his Tesla's
talk of like, you can never

623
00:35:35,887 --> 00:35:39,136
really simulate the world
because you don't understand

624
00:35:39,136 --> 00:35:42,927
like the true long tail. So you
need this like interactive data

625
00:35:42,927 --> 00:35:46,658
to cover like the last. Because
honestly, when you think about

626
00:35:46,658 --> 00:35:50,329
robotics, 90 percent and even 99
percent is garbage. What you

627
00:35:50,329 --> 00:35:54,000
really need is like 6 9's. When
you talk about these systems,

628
00:35:54,000 --> 00:35:57,605
Audrow Nash: oh, I get what you
mean. I wasn't quite sure what

629
00:35:57,605 --> 00:36:00,860
you were saying. So you're
saying that if you solve only

630
00:36:00,860 --> 00:36:04,233
95%, it doesn't really, you, you
have failure all the time

631
00:36:04,233 --> 00:36:07,663
because of that 5%. If you have
99%, you still, if you have

632
00:36:07,663 --> 00:36:11,384
thousands of robots running for
thousands of hours, you're gonna

633
00:36:11,384 --> 00:36:14,698
come up into these really.
Uncommon events that are going

634
00:36:14,698 --> 00:36:18,360
to be showstoppers all the time.
and that's going to be a giant

635
00:36:18,360 --> 00:36:19,000
pain. Okay.

636
00:36:19,000 --> 00:36:22,314
Michael Laskey: Yeah. And then
that's why there is this long

637
00:36:22,314 --> 00:36:25,573
tail problem where in like a
more like SaaS like framework,

638
00:36:25,573 --> 00:36:28,555
you can ship an 80 percent
model. Like if you actually

639
00:36:28,555 --> 00:36:31,759
upheld ChatGPT to like a safety
critical system of like...

640
00:36:31,759 --> 00:36:34,907
That'd be terrifying. Right,
yeah. Absolutely terrifying.

641
00:36:34,907 --> 00:36:38,000
Like don't mess up with the
blade. You know, yeah, like.

642
00:36:38,000 --> 00:36:41,797
Nag Murthy: This is a bit of a,
you know, like a abstract sort

643
00:36:41,797 --> 00:36:45,472
of question maybe for both of
you guys. Like, do you see the

644
00:36:45,472 --> 00:36:49,209
analogy between sort of ChatGPT?
you know, ChatGPT is sort of

645
00:36:49,209 --> 00:36:53,129
this... It's great at Q&amp; A,
right? And there's a bit of like

646
00:36:53,129 --> 00:36:56,742
Cleverhand syndrome going on
over there, where the human in

647
00:36:56,742 --> 00:37:00,478
the loop is sort of saying, yep,
that's great. And that's not

648
00:37:00,478 --> 00:37:03,541
great. And so you attribute a
lot of smartness and

649
00:37:03,541 --> 00:37:07,155
intelligence and agency to the
model at the end of the day,

650
00:37:07,155 --> 00:37:11,013
right? Like, so you think it's
amazing, but it's really, you're

651
00:37:11,013 --> 00:37:14,872
sort of discarding all the bad
stuff it gave you, but like only

652
00:37:14,872 --> 00:37:18,792
latching on to the good stuff it
gave you, right? Robots, on the

653
00:37:18,792 --> 00:37:22,712
other hand, okay, they're not as
clever sounding as ChatGPT. But

654
00:37:22,712 --> 00:37:26,571
they have to execute a series of
actions and tasks that have to

655
00:37:26,571 --> 00:37:30,368
converge to like high levels of
reliability, right? And you're

656
00:37:30,368 --> 00:37:34,104
seeing the same challenges that,
that are happening out there

657
00:37:34,104 --> 00:37:37,779
when it comes to things like
auto GPT and like some of these

658
00:37:37,779 --> 00:37:41,270
other sort of agent driven
things that people are trying,

659
00:37:41,270 --> 00:37:45,007
right? They're running into the
same issues, the long tail of

660
00:37:45,007 --> 00:37:48,804
issues that robotics runs into,
but they're running into it in

661
00:37:48,804 --> 00:37:52,418
the software realm. And if you
ask a roboticist, they'll be

662
00:37:52,418 --> 00:37:56,215
like, yeah, it's dumb to like,
ask an ML model to like, act as

663
00:37:56,215 --> 00:37:59,890
an agent. You know, it's just
you need to like, I guess, put

664
00:37:59,890 --> 00:38:03,688
another way. You can't just RL
or like expect like good output

665
00:38:03,688 --> 00:38:07,485
or you can't like train it to
just give you a good answer to a

666
00:38:07,485 --> 00:38:11,283
question. You have to train it
to converge to perform a series

667
00:38:11,283 --> 00:38:14,590
of tasks, plan and sort of
reason across the series of

668
00:38:14,590 --> 00:38:18,388
tasks. And that's what we mean
by the long tail. Over here, in

669
00:38:18,388 --> 00:38:19,000
many ways.

670
00:38:19,000 --> 00:38:22,780
Audrow Nash: Interesting. Yeah,
I like that. That it's, it's you

671
00:38:22,780 --> 00:38:26,442
doing actions over time, and it
has to constrain to finish the

672
00:38:26,442 --> 00:38:30,104
actions before it can proceed.
whereas ChatGPT can just have a

673
00:38:30,104 --> 00:38:33,944
witty one time response that had
no, like the response before was

674
00:38:33,944 --> 00:38:37,488
garbage. And you just discard
it, and it doesn't matter. But

675
00:38:37,488 --> 00:38:41,150
this one, it's very sequential
in nature when you do robotics.

676
00:38:41,150 --> 00:38:44,576
And I totally have seen what
you're saying with like agent

677
00:38:44,576 --> 00:38:48,060
GPT and things like this where
they try to do a sequence of

678
00:38:48,060 --> 00:38:51,663
things and like in the middle
they write a Python script that

679
00:38:51,663 --> 00:38:55,089
has an error in it and then it
fails and then it's like, I

680
00:38:55,089 --> 00:38:58,397
don't know what to do now,
failing, like trying 10 other

681
00:38:58,397 --> 00:39:02,000
things, all equally failing, and
then it doesn't do anything.

682
00:39:02,000 --> 00:39:05,301
Nag Murthy: Yeah, it's a loose
analogy, but that's kind of

683
00:39:05,301 --> 00:39:08,374
like, and I think it's, it's
similar to, you know, the

684
00:39:08,374 --> 00:39:09,000
challenges.

685
00:39:09,000 --> 00:39:11,987
Audrow Nash: I agree. It's, it's
interacting with complex

686
00:39:11,987 --> 00:39:15,341
unstructured systems. Like if it
goes and tries to pull the web,

687
00:39:15,341 --> 00:39:18,590
the web has a bunch of formats
and a bunch of exceptions. Yup.

688
00:39:18,590 --> 00:39:21,000
Robotics, the world has a bunch
of exceptions.

689
00:39:21,000 --> 00:39:24,806
Nag Murthy: Exactly. Exactly.
that's, that may be the stronger

690
00:39:24,806 --> 00:39:28,734
analogy than like, I don't know,
like it, it just stands, right?

691
00:39:28,734 --> 00:39:32,049
Like, one of the really
fascinating examples we saw in

692
00:39:32,049 --> 00:39:35,916
the Sim2Real sort of translation
was the, the notion, like, you

693
00:39:35,916 --> 00:39:39,476
know, how do you model the sun?
And this was all like mind

694
00:39:39,476 --> 00:39:43,404
blowing to me when like Mike was
talking about it, right? How do

695
00:39:43,404 --> 00:39:46,780
you model for sun glare? And
until we actually deployed

696
00:39:46,780 --> 00:39:50,463
across like so many different
geographies at different times

697
00:39:50,463 --> 00:39:54,207
of the day, like we could not
get a true sense for how do you

698
00:39:54,207 --> 00:39:58,136
sort of, what's the range of. So
there's a lot of values you use

699
00:39:58,136 --> 00:40:01,941
to sweep for sort of modeling
the sun and sin that is relevant

700
00:40:01,941 --> 00:40:05,870
to like what we do as a company,
which I thought was like really

701
00:40:05,870 --> 00:40:09,000
interesting, you know, that
example sticks with me.

702
00:40:09,000 --> 00:40:14,091
Audrow Nash: Oh yeah. Mike, any
thoughts with, NAGs, uh,

703
00:40:14,091 --> 00:40:19,727
connection to GPT or auto GPT
and all those things, or I don't

704
00:40:19,727 --> 00:40:22,000
know, any thoughts on it?

705
00:40:22,000 --> 00:40:25,909
Michael Laskey: He's definitely
right though. Like when you see

706
00:40:25,909 --> 00:40:29,818
the consequence of an action in
an agent, then you really start

707
00:40:29,818 --> 00:40:32,797
evaluating its true
intelligence. When it's this

708
00:40:32,797 --> 00:40:36,209
kind of these zero shot text
prompts, you fill in a lot

709
00:40:36,209 --> 00:40:39,560
because you come in with your
sense of the world, your

710
00:40:39,560 --> 00:40:42,973
understanding, and it's easy to
ascribe intelligence to

711
00:40:42,973 --> 00:40:46,447
something, but with like a
robot, it's, psychologically,

712
00:40:46,447 --> 00:40:50,046
it's, it's kind of like, you
know, when the robot fails to

713
00:40:50,046 --> 00:40:53,583
mow, as humans, we're like, why
can't you, yeah, it's the

714
00:40:53,583 --> 00:40:57,058
classic like, more of that,
like, why can't you do these

715
00:40:57,058 --> 00:41:01,091
simple things, like, why is this
so hard? yeah, the failure modes

716
00:41:01,091 --> 00:41:05,000
of robots just, they hurt extra
more, I feel like, in that way.

717
00:41:05,000 --> 00:41:08,442
Audrow Nash: Yeah, and there's
risks, and it could be

718
00:41:08,442 --> 00:41:12,662
expensive, and like, I mean, you
do a terrible job mowing, or you

719
00:41:12,662 --> 00:41:16,818
like, run over a sprinkler, or,
like, all sorts of things, like,

720
00:41:16,818 --> 00:41:20,714
it can be expensive, it can ruin
the client's day, like, all

721
00:41:20,714 --> 00:41:24,545
these things, there's real
consequences. For these kinds of

722
00:41:24,545 --> 00:41:25,000
things.

723
00:41:25,000 --> 00:41:27,831
Nag Murthy: not so, yeah, sorry,
not as bad as like the

724
00:41:27,831 --> 00:41:31,125
consequences for an autonomous,
like self driving car, like, you

725
00:41:31,125 --> 00:41:32,000
know, yeah, yeah.

726
00:41:32,000 --> 00:41:36,000
Michael Laskey: A lot of it's
more like just sad moments.

727
00:41:36,000 --> 00:41:40,924
Audrow Nash: You're like, damn
it. Oh, what a pain. Yeah, for

728
00:41:40,924 --> 00:41:45,767
sure. Or like the robot breaks
itself. Like if it, if it did

729
00:41:45,767 --> 00:41:50,691
just go right into the pool,
it's like, Oh, got to get it out

730
00:41:50,691 --> 00:41:55,453
of the pool. All the hardware's
broken, whatever, for sure.

731
00:41:55,453 --> 00:42:00,135
Yeah, that's funny. At this
scale, it's just sad. It's not

732
00:42:00,135 --> 00:42:04,897
catastrophic or anything. so
how, how do you build in these

733
00:42:04,897 --> 00:42:09,740
safety systems? To what you're
doing. And I don't know who's

734
00:42:09,740 --> 00:42:12,000
the best one to answer this.

735
00:42:12,000 --> 00:42:15,253
Nag Murthy: Yeah, I can talk
about it because like my

736
00:42:15,253 --> 00:42:18,568
background is in building
medical devices before this,

737
00:42:18,568 --> 00:42:22,251
right? Like that's kind of, I
guess I, yeah, maybe I'll give

738
00:42:22,251 --> 00:42:25,811
you a bit of a segue, right?
Like I built my first company

739
00:42:25,811 --> 00:42:29,494
out of Stanford grad school. And
that was a low cost medical

740
00:42:29,494 --> 00:42:33,361
devices company that, you know,
it's like being like, it's been

741
00:42:33,361 --> 00:42:36,982
in the market for a long time
now. It's saved like 300, 000

742
00:42:36,982 --> 00:42:40,419
lives across 22 countries. And
it's been a great sort of

743
00:42:40,419 --> 00:42:43,980
learning opportunity to really
understand how do you build

744
00:42:43,980 --> 00:42:47,908
safety critical systems? Because
you know, the devices that like

745
00:42:47,908 --> 00:42:51,345
I designed before, you put
premature babies in them. And

746
00:42:51,345 --> 00:42:55,028
like, that's the most fragile,
like, life form you can think

747
00:42:55,028 --> 00:42:58,895
of, right? Yeah, really. You put
in, like, a device. And, like,

748
00:42:58,895 --> 00:43:02,517
so this approach of, like,
safety first, trying to be very,

749
00:43:02,517 --> 00:43:06,077
sort of, risk driven, risk
minimization driven, always has

750
00:43:06,077 --> 00:43:10,066
been the, at the core of what we
do as a company. and so, when we

751
00:43:10,066 --> 00:43:13,872
first started, right, we started
by automating a lot of these,

752
00:43:13,872 --> 00:43:17,616
Automating these large 60 inch,
72 inch, zero ton mowers, and

753
00:43:17,616 --> 00:43:21,238
which is what a lot of our
competitors are also still doing

754
00:43:21,238 --> 00:43:24,982
today. Yeah, like Greensy is
doing that kind of thing. Right,

755
00:43:24,982 --> 00:43:28,726
yeah. And like, great companies,
right? Like, really respect,

756
00:43:28,726 --> 00:43:32,225
like, these guys. The core
concern that we've had, right,

757
00:43:32,225 --> 00:43:35,847
with these systems, and we
built, like, we probably had the

758
00:43:35,847 --> 00:43:39,714
largest deployed fleet of, like,
these large machines two years

759
00:43:39,714 --> 00:43:43,703
ago. Right. And we were, we were
about to like scale them like in

760
00:43:43,703 --> 00:43:47,509
an order of magnitude more, but
then we were working with like

761
00:43:47,509 --> 00:43:50,578
a, you know, like a IEC, you
know, standard safety

762
00:43:50,578 --> 00:43:54,199
consulting, 6150 at safety
consulting, you know, consultant

763
00:43:54,199 --> 00:43:57,944
to sort of figure out how does
this apply, you know, when you

764
00:43:57,944 --> 00:44:01,565
scale, right. When you don't
have sitters on the ground who

765
00:44:01,565 --> 00:44:04,941
are watching these, you know,
and safety sitters on the

766
00:44:04,941 --> 00:44:08,194
ground, how does it really
scale? The short answer is

767
00:44:08,194 --> 00:44:11,632
nobody really has an answer
today, right? Because you're

768
00:44:11,632 --> 00:44:12,000
doing,

769
00:44:12,000 --> 00:44:13,000
Audrow Nash: you're going zero
to one in a sense. So

770
00:44:13,000 --> 00:44:16,120
Nag Murthy: there's right,
exactly. There is no model. 80

771
00:44:16,120 --> 00:44:19,568
companies have solved for this
by like, you know, self insuring

772
00:44:19,568 --> 00:44:22,852
or whatever. Right. Like, or, I
mean, you know, daddy Google

773
00:44:22,852 --> 00:44:26,026
like puts in a lot of money,
like, you know, there's like,

774
00:44:26,026 --> 00:44:29,419
there's like all that money to
blow. Right. So they, they self

775
00:44:29,419 --> 00:44:32,648
insure and they get away with
it. Tesla has taken the other

776
00:44:32,648 --> 00:44:35,987
approach where, you know, they
put a guy behind, like there's

777
00:44:35,987 --> 00:44:38,997
implicitly a guy behind full
self driving all the time,

778
00:44:38,997 --> 00:44:42,226
right? and he's paying for the
privilege of like doing self

779
00:44:42,226 --> 00:44:45,237
driving testing, which is
incredible if you think about

780
00:44:45,237 --> 00:44:48,247
it, right? And so they've
solved, they've addressed the

781
00:44:48,247 --> 00:44:51,531
safety problem that way. They
always have a sitter, but then

782
00:44:51,531 --> 00:44:54,924
like companies in the robotics
space, they are, you know, they

783
00:44:54,924 --> 00:44:58,263
only have like. This notion of,
okay, let's put a SICK LiDAR,

784
00:44:58,263 --> 00:45:01,601
like, and SICK has a monopoly in
this market, right? Or maybe

785
00:45:01,601 --> 00:45:04,502
Hokuyo is like some other
company that does this, but

786
00:45:04,502 --> 00:45:07,786
let's put a SICK LiDAR and let's
cover our machine with SICK

787
00:45:07,786 --> 00:45:11,179
LiDARs, right? And like that way
you can ensure safety. At the

788
00:45:11,179 --> 00:45:12,000
end of the day.

789
00:45:12,000 --> 00:45:15,060
Audrow Nash: And so, yeah, they
make sure it doesn't bump

790
00:45:15,060 --> 00:45:18,067
anything and it's, but then it's
a huge data problem and

791
00:45:18,067 --> 00:45:21,396
everything. Like I think of like
cruise or Waymo or something,

792
00:45:21,396 --> 00:45:24,779
and they just have like sensor,
sensor, sensor, sensor, sensor,

793
00:45:24,779 --> 00:45:28,000
big rack on top with a gigantic
LiDAR on top and everything.

794
00:45:28,000 --> 00:45:31,364
Nag Murthy: Exactly. So I think
the, then we came to the

795
00:45:31,364 --> 00:45:35,029
realization that look, maybe
there's a different way to think

796
00:45:35,029 --> 00:45:38,814
about safety, which is you still
want to build in redundancies.

797
00:45:38,814 --> 00:45:42,418
You still want to go through,
you know, like a risk analysis

798
00:45:42,418 --> 00:45:45,842
and all that. And that's a
given. But then if you look at

799
00:45:45,842 --> 00:45:48,966
like, how do you impact like
the. You can reduce the

800
00:45:48,966 --> 00:45:52,691
probability of occurrence and
you can also reduce the severity

801
00:45:52,691 --> 00:45:55,995
of occurrence. right. You can
reduce the probability of

802
00:45:55,995 --> 00:45:59,780
occurrence by building in like
some redundancy. So you can have

803
00:45:59,780 --> 00:46:03,324
a mechanical bumper in addition
to like your stereo camera,

804
00:46:03,324 --> 00:46:06,809
right. To sort of not hit
objects. But then you can reduce

805
00:46:06,809 --> 00:46:10,353
the severity of it by making
these machines smaller. Right?

806
00:46:10,353 --> 00:46:14,258
And by making them lower powered
than some of these large beastly

807
00:46:14,258 --> 00:46:16,000
sort of 60, 72 inch machines.

808
00:46:16,000 --> 00:46:18,868
Audrow Nash: The ones that
people ride on and this kind of

809
00:46:18,868 --> 00:46:21,736
thing. And like ride on for
industrial things where it's a

810
00:46:21,736 --> 00:46:24,703
huge fan beneath them and this
kind of thing cutting all the

811
00:46:24,703 --> 00:46:25,000
grass.

812
00:46:25,000 --> 00:46:29,101
Nag Murthy: Exactly. So then now
the question is, do you need

813
00:46:29,101 --> 00:46:33,067
that paradigm? Because that
paradigm, right, has evolved to

814
00:46:33,067 --> 00:46:36,832
meet the needs of efficiency
when it's being driven by a

815
00:46:36,832 --> 00:46:41,000
human, right? Bigger, better,
faster, right? Because there's a

816
00:46:41,000 --> 00:46:45,370
dude driving it. . Right. Robots
don't need that paradigm. Right.

817
00:46:45,370 --> 00:46:49,538
Because you can swarm a robot
and you can have a lot of little

818
00:46:49,538 --> 00:46:53,706
like lighter electric basically.
Yeah. Electric sheep. So it's

819
00:46:53,706 --> 00:46:57,739
safe. Exactly. Safety is driven
by like a change in thinking

820
00:46:57,739 --> 00:47:01,303
about how you think about form.
That's the key thing.

821
00:47:01,303 --> 00:47:05,471
Interesting. I like that, you
know, and then you sort of layer

822
00:47:05,471 --> 00:47:09,437
in like redundancies that you
always have to, you know, but

823
00:47:09,437 --> 00:47:13,672
that's kind of it really at the
end of the day, and, and that's

824
00:47:13,672 --> 00:47:17,840
also explains like why we took
this sort of turn as a company,

825
00:47:17,840 --> 00:47:21,739
because we're like, we can scale
a hundred of like the big

826
00:47:21,739 --> 00:47:25,975
machines, but really, is there a
future, right? For like these.

827
00:47:25,975 --> 00:47:30,076
Yeah, and there's, it's clearly
like the writings on the wall

828
00:47:30,076 --> 00:47:33,975
that machine learning, cheap
cameras, small robots, that's

829
00:47:33,975 --> 00:47:37,000
going to dominate. So let's just
go that way.

830
00:47:37,000 --> 00:47:40,922
Audrow Nash: Interesting. Yeah.
I like how you've looked at a

831
00:47:40,922 --> 00:47:44,973
lot of forces that are kind of
shaping like where technology is

832
00:47:44,973 --> 00:47:48,895
going and what's working and
picked what you think to bet on,

833
00:47:48,895 --> 00:47:52,367
like, stereo cameras. But, so,
okay, safety first. You

834
00:47:52,367 --> 00:47:55,968
basically, you've made it so
that the form is more, much

835
00:47:55,968 --> 00:47:59,890
safer. And so your robots, just
looking at the videos online,

836
00:47:59,890 --> 00:48:03,555
they're basically like push
lawnmowers that someone would

837
00:48:03,555 --> 00:48:07,477
have. and it's like, it's like
an electric, robot push mower,

838
00:48:07,477 --> 00:48:11,464
basically. They even have the
handle so you can flip it up and

839
00:48:11,464 --> 00:48:15,000
someone can use it like a
regular lawnmower, I suppose.

840
00:48:15,000 --> 00:48:19,133
Nag Murthy: it's more to the
Fandle flipping over is really

841
00:48:19,133 --> 00:48:23,546
just to, it's to help the robot
be driven from one site back to

842
00:48:23,546 --> 00:48:27,539
the van. you know, it's for UX,
pull it back on a cart or

843
00:48:27,539 --> 00:48:31,602
something. Yeah. we don't like
encourage, or we don't even

844
00:48:31,602 --> 00:48:35,806
design it to be push, like dual
use. Because, one assumption

845
00:48:35,806 --> 00:48:40,219
about it. No, no, no, no. That's
not the idea at all. you know,

846
00:48:40,219 --> 00:48:43,932
because like these robots are
already fairly good and

847
00:48:43,932 --> 00:48:48,205
autonomous. Like they 80, 90
percent of a given yard. so, you

848
00:48:48,205 --> 00:48:52,618
know, we don't really need like,
you know, this thing to double

849
00:48:52,618 --> 00:48:55,000
up as like a secondary mode.
Yeah.

850
00:48:55,000 --> 00:48:59,214
Audrow Nash: Okay. So, that's
awesome. I love, I really, I

851
00:48:59,214 --> 00:49:03,356
feel like I need to think more
about the safety driven by

852
00:49:03,356 --> 00:49:08,006
change in form idea. I feel like
that's a very interesting, it's

853
00:49:08,006 --> 00:49:12,002
like you just change what you're
doing and that changes

854
00:49:12,002 --> 00:49:16,071
everything about the risk of
this kind of thing and that

855
00:49:16,071 --> 00:49:20,286
changes the safety. I think
that's a really cool idea. I'm

856
00:49:20,286 --> 00:49:24,645
going to have to think more
about that. Let's see. Mike. So,

857
00:49:24,645 --> 00:49:29,078
tell me about safety from the,
like. I don't know the AI side

858
00:49:29,078 --> 00:49:33,510
of things or how you put things
in place in addition to maybe

859
00:49:33,510 --> 00:49:38,087
the AI controller. I don't know
if you have like, here's the AI

860
00:49:38,087 --> 00:49:42,738
part of everything, um, and then
on top of that you have a thing

861
00:49:42,738 --> 00:49:47,097
that says, whoa, whoa, whoa,
terrible idea, or something, or

862
00:49:47,097 --> 00:49:51,239
how do you... From a software
architecture perspective, I

863
00:49:51,239 --> 00:49:54,000
guess, how are you thinking of
safety?

864
00:49:54,000 --> 00:49:57,354
Michael Laskey: Yeah, that's a
really good question. So,

865
00:49:57,354 --> 00:50:01,126
there's definitely... Like Nox
said, we reduced the probability

866
00:50:01,126 --> 00:50:04,719
of a severe life accident down
significantly. and then, from

867
00:50:04,719 --> 00:50:08,013
that, you know, we then said,
okay, well let's make the

868
00:50:08,013 --> 00:50:11,726
camera... The primary source of
safety, and this is where it's

869
00:50:11,726 --> 00:50:14,600
going to get a little
complicated of my thoughts

870
00:50:14,600 --> 00:50:18,433
about, different types of vision
problems and the robustness you

871
00:50:18,433 --> 00:50:21,966
obtain. So, when you look at the
safety system of a, a push

872
00:50:21,966 --> 00:50:25,619
firing robot, right, there's,
what's your, what's your number

873
00:50:25,619 --> 00:50:29,033
one concern? It's going to be
running into people, right,

874
00:50:29,033 --> 00:50:32,746
running into a kid or some other
human. now that is just pure,

875
00:50:32,746 --> 00:50:36,518
like, collision avoidance, and
that's what we would call, like,

876
00:50:36,518 --> 00:50:40,351
more, like, low level vision, in
a way. where, yeah, like, depth

877
00:50:40,351 --> 00:50:44,064
sensors, like stereo, are trying
to just solve those problems.

878
00:50:44,064 --> 00:50:47,477
And it's actually, it's not,
it's not a high level vision

879
00:50:47,477 --> 00:50:51,011
problem. You don't have to
consider semantics of the world.

880
00:50:51,011 --> 00:50:54,364
You just have to think about,
like, when you look at the

881
00:50:54,364 --> 00:50:57,897
fundamentals of stereo, you're
just looking at, like, where

882
00:50:57,897 --> 00:51:01,311
does this pixel match this
pixel? So you're actually just

883
00:51:01,311 --> 00:51:05,084
doing, like, a... A mapping of
two images, so you don't have to

884
00:51:05,084 --> 00:51:07,000
think about like true semantics.

885
00:51:07,000 --> 00:51:09,440
Audrow Nash: Yeah, so you're
inferring depth and you're using

886
00:51:09,440 --> 00:51:11,000
that depth to not hit stuff, I
suppose.

887
00:51:11,000 --> 00:51:14,830
Michael Laskey: Exactly. And
this is actually something that

888
00:51:14,830 --> 00:51:18,596
in the literature, it's been
shown time and time again that

889
00:51:18,596 --> 00:51:22,298
if you pre train on synthetic
data, the raw point cloud is

890
00:51:22,298 --> 00:51:26,000
actually extremely robust. so
you can get away with highly

891
00:51:26,000 --> 00:51:29,830
like You're not going to get
like LiDAR, you know, precision

892
00:51:29,830 --> 00:51:33,596
and accuracy, but your ability
to estimate like core shape,

893
00:51:33,596 --> 00:51:37,553
especially at the distance of
like two meters in front of you,

894
00:51:37,553 --> 00:51:41,128
is very reliable. We haven't
really seen this ever fail.

895
00:51:41,128 --> 00:51:44,000
Testing now for over a year on
our platforms.

896
00:51:44,000 --> 00:51:48,099
Audrow Nash: Do you ever like
someone has like a metal fence

897
00:51:48,099 --> 00:51:51,994
or a glass? Wall or something
like this or something like

898
00:51:51,994 --> 00:51:55,000
that. Are those pretty easy to
detect still?

899
00:51:55,000 --> 00:51:58,822
Michael Laskey: Yeah, and
actually with like Learn Stereo,

900
00:51:58,822 --> 00:52:02,709
especially when you train on
like heavily domain randomized

901
00:52:02,709 --> 00:52:06,860
images, I've had papers before
that show you can like precisely

902
00:52:06,860 --> 00:52:10,880
manipulate glass objects with
robotic grippers. That was some

903
00:52:10,880 --> 00:52:15,097
work coming out of like TRI, but
today we run it next to fences.

904
00:52:15,097 --> 00:52:19,050
There's not a lot of glass in
the outdoor world, sadly. But,

905
00:52:19,050 --> 00:52:23,333
you know, cars and things, which
are kind of optically tricky, it

906
00:52:23,333 --> 00:52:27,353
works very well on. we've never
had an incident where it like

907
00:52:27,353 --> 00:52:29,000
ran into an object. Cool.

908
00:52:29,000 --> 00:52:31,000
Audrow Nash: Nag, you look like
you have something to add.

909
00:52:31,000 --> 00:52:34,733
Nag Murthy: No, it's just, it's
just interesting, right? Like,

910
00:52:34,733 --> 00:52:38,104
when you design for safety,
like, there's, the standards

911
00:52:38,104 --> 00:52:41,536
really call for, like, harm.
Like, you do no harm when it

912
00:52:41,536 --> 00:52:44,907
comes to living. Like, it's
specifically humans, really,

913
00:52:44,907 --> 00:52:48,339
right? They don't even mention
dogs and cats and whatnot,

914
00:52:48,339 --> 00:52:51,951
right? So then, this is really a
question of... Sort of what

915
00:52:51,951 --> 00:52:55,443
you're designing towards for us,
sort of the do no harm is

916
00:52:55,443 --> 00:52:59,115
basically do no harm to living
objects, right? The non living

917
00:52:59,115 --> 00:53:02,968
objects, you can easily sort of
RL your way through, right? Like

918
00:53:02,968 --> 00:53:06,641
with more and more data points
as you, as you. So it's really

919
00:53:06,641 --> 00:53:10,313
from a product perspective, it's
really not as relevant to be

920
00:53:10,313 --> 00:53:14,166
solving for like, you know, uh,
even though like it works pretty

921
00:53:14,166 --> 00:53:18,019
well. Yeah, we have to solve for
the living things and then, you

922
00:53:18,019 --> 00:53:21,391
know. It does like do pretty
well on like the non living

923
00:53:21,391 --> 00:53:25,304
stuff and where it doesn't like,
you can just always like improve

924
00:53:25,304 --> 00:53:28,856
it, but that's the minimum bar
you want to meet when you're

925
00:53:28,856 --> 00:53:30,000
considering safety.

926
00:53:30,000 --> 00:53:34,573
Michael Laskey: I will say it's
more like the animal world, not

927
00:53:34,573 --> 00:53:39,000
necessarily plants. we've, I
mean, things that we definitely.

928
00:53:39,000 --> 00:53:41,000
Nag Murthy: Yeah. Okay. Plants,
were

929
00:53:41,000 --> 00:53:44,295
Michael Laskey: also living
Yeah, yeah. Like plants aren't

930
00:53:44,295 --> 00:53:46,000
gonna count, but like, yeah. .

931
00:53:46,000 --> 00:53:47,000
Audrow Nash: Yeah. I guess

932
00:53:46,000 --> 00:53:49,861
Nag Murthy: technically living.
Okay. I mean, we, we, we are

933
00:53:49,861 --> 00:53:53,980
cutting a plant at every single
second of operation, like grass.

934
00:53:53,980 --> 00:53:57,584
So, like , I mean, like, sure.
Okay. Gotta draw the line

935
00:53:57,584 --> 00:53:59,000
somewhere. Yeah. Yeah.

936
00:53:59,000 --> 00:54:01,432
Audrow Nash: Don't want to kill
all the roses in the yard or

937
00:54:01,432 --> 00:54:02,000
something too.

938
00:54:02,000 --> 00:54:04,000
Nag Murthy: There we go.

939
00:54:04,000 --> 00:54:07,968
Audrow Nash: Okay. So do you
have, in terms of architecture,

940
00:54:07,968 --> 00:54:11,672
is it like, here is like the
front level analysis of the

941
00:54:11,672 --> 00:54:15,376
sensor data. Like the first
thing it gets to is like, is

942
00:54:15,376 --> 00:54:19,608
this alive? And to do that, you
go, has it moved? Or is it warm?

943
00:54:19,608 --> 00:54:23,511
Like if you had, I guess you
probably just have your stereo

944
00:54:23,511 --> 00:54:27,413
one. I'm imagining if you had
like an infrared one, Oh, I'm

945
00:54:27,413 --> 00:54:29,000
getting the Apple thing.

946
00:54:29,000 --> 00:54:32,000
Nag Murthy: That's funny. How do
you do that?

947
00:54:32,000 --> 00:54:36,904
Audrow Nash: There's it's hand
gestures that make it do things.

948
00:54:36,904 --> 00:54:41,731
Let's see now that it's doing
it. Look at this. I am unable to

949
00:54:41,731 --> 00:54:45,000
turn this off. So now it just is
here, but

950
00:54:45,000 --> 00:54:48,358
Nag Murthy: Where you have AI
agents everywhere and they're

951
00:54:48,358 --> 00:54:51,715
like, you know, oh, can I get
you this like did you ask for

952
00:54:51,715 --> 00:54:52,000
that?

953
00:54:52,000 --> 00:54:56,028
Audrow Nash: I know, I know.
It's so silly. Also, like, if

954
00:54:56,028 --> 00:55:00,333
something is bad, let's see if
it'll do it. Oh, it's not doing

955
00:55:00,333 --> 00:55:04,778
it. Oh, there it is. You can do
all sorts of things. It's like a

956
00:55:04,778 --> 00:55:08,944
device level, too. It works in
all of the video things. It's

957
00:55:08,944 --> 00:55:13,319
not, but anyways. Yeah, crazy.
So, I know, I can't believe, the

958
00:55:13,319 --> 00:55:17,556
update, I didn't even know. It
just, like, I was talking to a

959
00:55:17,556 --> 00:55:22,000
friend, and I, like, did, like,
a thumbs up or something, and it

960
00:55:22,000 --> 00:55:26,167
showed some emoji with that.
It's like, ugh. I guess this is

961
00:55:26,167 --> 00:55:30,125
what it is now, but so you're
doing like a detecting life

962
00:55:30,125 --> 00:55:34,361
thing. Are you doing it right at
the beginning in a sense? So

963
00:55:34,361 --> 00:55:38,667
like the first thing is don't
hit living things and maybe that

964
00:55:38,667 --> 00:55:42,694
has some semantics built into
it. Like here's a human pose

965
00:55:42,694 --> 00:55:46,653
detector or just this thing
moved or how do you deal with

966
00:55:46,653 --> 00:55:47,000
this?

967
00:55:47,000 --> 00:55:50,796
Michael Laskey: Yeah, so you, we
don't explicitly like have a

968
00:55:50,796 --> 00:55:54,591
life detector. that'd be kind
of, that'd be interesting to go

969
00:55:54,591 --> 00:55:58,636
down. But, you know, so you can,
you can start off by just, like,

970
00:55:58,636 --> 00:56:02,493
looking at the geometry of the
world, right? So, like, how big

971
00:56:02,493 --> 00:56:06,227
are humans normally? What size
of objects would you have to,

972
00:56:06,227 --> 00:56:09,898
like, know you can detect as,
like, a geometric problem? so

973
00:56:09,898 --> 00:56:13,444
can you just say, like, there's
an object in front of the

974
00:56:13,444 --> 00:56:17,302
camera? So that'd be step one.
Because you know, like, in this

975
00:56:17,302 --> 00:56:21,160
space, the robot should never
mow into something that's, like,

976
00:56:21,160 --> 00:56:25,018
camera height, for example. So,
like, that's gonna pretty much

977
00:56:25,018 --> 00:56:28,689
eliminate anything that can
stand taller than, like, half a

978
00:56:28,689 --> 00:56:29,000
foot.

979
00:56:29,000 --> 00:56:31,980
Audrow Nash: Okay. Yeah, so you
don't need to do too much

980
00:56:31,980 --> 00:56:35,275
special for this. You just say,
don't run into things that are,

981
00:56:35,275 --> 00:56:37,000
like, camera height and
whatever.

982
00:56:37,000 --> 00:56:40,687
Michael Laskey: Yeah, so that's
gonna, I mean, that's what's

983
00:56:40,687 --> 00:56:44,374
nice about a lot of these tasks,
is, like, you can do pretty

984
00:56:44,374 --> 00:56:48,000
simple collision avoidance. But
then, it... Go ahead. I was

985
00:56:48,000 --> 00:56:51,000
Audrow Nash: going to see how
that... You go.

986
00:56:51,000 --> 00:56:54,548
Michael Laskey: Well, okay, so
that's like your, your

987
00:56:54,548 --> 00:56:58,699
foundational level. But then it
does get more challenging when

988
00:56:58,699 --> 00:57:02,849
you actually think about like
the semantics of mowing. Because

989
00:57:02,849 --> 00:57:07,000
like what stops our robot from
just running to the street? and

990
00:57:07,000 --> 00:57:10,950
that could also cause like a
huge thing. Here the robot now

991
00:57:10,950 --> 00:57:14,766
needs to know, Oh, I can mow
over this and I need to turn

992
00:57:14,766 --> 00:57:18,782
around this property. and that's
where we're at then, that's

993
00:57:18,782 --> 00:57:22,531
where the more advanced ML is
coming in, like high level

994
00:57:22,531 --> 00:57:23,000
vision.

995
00:57:23,000 --> 00:57:27,529
Audrow Nash: Gotcha. I think my
wife and dog just got home. He's

996
00:57:27,529 --> 00:57:31,562
at the door. let's see. Well
that sounds cool. So, I feel

997
00:57:31,562 --> 00:57:35,878
like there's a lot to discuss
with this. I really want to get

998
00:57:35,878 --> 00:57:40,124
into more of the, like, the
business model that you guys are

999
00:57:40,124 --> 00:57:44,440
going to. but is there anything
else? I mean, it seems like a

1000
00:57:44,440 --> 00:57:48,686
wonderful approach that you guys
are using. Your ES1 system,

1001
00:57:48,686 --> 00:57:52,790
you're learning a lot. You have
simple heuristics that are

1002
00:57:52,790 --> 00:57:57,177
robust for, don't hit stuff and
things like this, and then you

1003
00:57:57,177 --> 00:58:01,564
have a lot of. physical world
understanding that you're being,

1004
00:58:01,564 --> 00:58:05,527
that you're developing.
Actually, one thing before we go

1005
00:58:05,527 --> 00:58:10,126
on to the business side, tell me
more about how to generalize for

1006
00:58:10,126 --> 00:58:14,442
other tasks. Like how do you
imagine this mowing data teaches

1007
00:58:14,442 --> 00:58:18,334
us some sort of physics or
navigation or whatever it is

1008
00:58:18,334 --> 00:58:22,721
that we might have, and how do
we extend that to snow plowing?

1009
00:58:22,721 --> 00:58:26,684
And how do we extend that to
picking apples, eventually.

1010
00:58:26,684 --> 00:58:30,788
Like, how does that go? And
whoever, whoever wants to jump

1011
00:58:30,788 --> 00:58:31,000
in.

1012
00:58:31,000 --> 00:58:33,000
Nag Murthy: It's Mike.

1013
00:58:33,000 --> 00:58:36,560
Michael Laskey: Okay. Yeah. So,
okay, let's think about like

1014
00:58:36,560 --> 00:58:40,119
snow plowing, for example. It's
very similar to push mowing,

1015
00:58:40,119 --> 00:58:43,145
right? You have the
understanding of like, what can

1016
00:58:43,145 --> 00:58:47,001
you can, what can you move your,
you know, plow is basically like

1017
00:58:47,001 --> 00:58:50,679
a blade, but it's now kind of
vertical. what can you move that

1018
00:58:50,679 --> 00:58:54,298
over versus not? So there, the
idea is like deep awareness of

1019
00:58:54,298 --> 00:58:57,442
collision avoidance. So like
understanding like small

1020
00:58:57,442 --> 00:59:01,002
obstacles, manhole covers. All
of that comes into play. then

1021
00:59:01,002 --> 00:59:04,443
there's localization. You
obviously are trying to lay down

1022
00:59:04,443 --> 00:59:08,180
stripes efficiently. All of that
also comes into play. and then

1023
00:59:08,180 --> 00:59:11,740
there's the semantics of like,
where is snow versus not. and

1024
00:59:11,740 --> 00:59:15,181
that is very similar to like,
where is there like, mowable

1025
00:59:15,181 --> 00:59:18,859
grass versus not. Snowpine, I
think there's some nuances there

1026
00:59:18,859 --> 00:59:22,300
that's gonna refine the model
because you have significant

1027
00:59:22,300 --> 00:59:25,622
occlusions, right? You're gonna
have to have this, like,

1028
00:59:25,622 --> 00:59:28,945
somewhat of a prior in some
spaces or be able to detect,

1029
00:59:28,945 --> 00:59:32,504
like, the edge of boundaries.
Very hard. Yeah. Yeah, I think

1030
00:59:32,504 --> 00:59:36,242
that's gonna be one of the key
challenges, but when you look at

1031
00:59:36,242 --> 00:59:39,683
how humans do it, it's very
possible that with the central

1032
00:59:39,683 --> 00:59:42,946
stack we have, these could be
done. So we feel like the

1033
00:59:42,946 --> 00:59:46,802
intelligence is capable of doing
this. We just have to like test,

1034
00:59:46,802 --> 00:59:50,480
refine, and build. and then if
you want to get into, you know,

1035
00:59:50,480 --> 00:59:54,218
like more dexterous things, like
let's say apple picking. A lot

1036
00:59:54,218 --> 00:59:57,421
of it is just like geometric
motion planning, semantic

1037
00:59:57,421 --> 01:00:01,040
awareness, understanding, like,
how do you move your arm in a

1038
01:00:01,040 --> 01:00:04,659
free space? But then there's
also the contact physics, right?

1039
01:00:04,659 --> 01:00:08,219
Like, knowing, like, how to
grasp something and pull it. I'm

1040
01:00:08,219 --> 01:00:11,482
pretty hopeful because, you
know, like, coming from, UC

1041
01:00:11,482 --> 01:00:15,101
Berkeley, one of the papers I
worked on was, like, the DexNet

1042
01:00:15,101 --> 01:00:18,363
project, where we used
simulation to learn grasping. so

1043
01:00:18,363 --> 01:00:21,626
you definitely see there's
already like huge amounts of

1044
01:00:21,626 --> 01:00:24,652
research pointing that
simulation can learn contact

1045
01:00:24,652 --> 01:00:28,212
physics to a pretty reliable
degree. so I don't think that's

1046
01:00:28,212 --> 01:00:31,000
impossible for like our world
model to produce.

1047
01:00:31,000 --> 01:00:35,417
Nag Murthy: Yeah, I think we...
No, just one more thing, right?

1048
01:00:35,417 --> 01:00:39,625
Like a lot of this needs to be
evaluated also from a product

1049
01:00:39,625 --> 01:00:44,182
lens. I think this is where, you
know, the question is like, what

1050
01:00:44,182 --> 01:00:48,670
is, what is the true, like, you
care about generalization. Sure.

1051
01:00:48,670 --> 01:00:53,017
But like across what dimension,
right? wait, did we lose Mike?

1052
01:00:53,017 --> 01:00:57,435
So, you know, the, the one thing
I was sort of, you know, going

1053
01:00:57,435 --> 01:01:01,361
back to was this notion of.
Like, how do you, like this,

1054
01:01:01,361 --> 01:01:05,428
like, as we build this kernel,
right? Like you, it's, it's

1055
01:01:05,428 --> 01:01:09,846
already pretty generalizable to
a lot of surface coverage, even

1056
01:01:09,846 --> 01:01:13,352
without going to be, without
going to something as

1057
01:01:13,352 --> 01:01:17,559
complicated as snow. So you take
things like weed treatment,

1058
01:01:17,559 --> 01:01:21,275
right? Or like fertilizer
application on a lawn. Like

1059
01:01:21,275 --> 01:01:25,202
that's like exactly the same
motion on like turf. That's

1060
01:01:25,202 --> 01:01:29,619
something we're able to do right
away today. Right, so that's a

1061
01:01:29,619 --> 01:01:34,037
task that you get paid for, it's
a very high margin task. A lot

1062
01:01:34,037 --> 01:01:36,000
of interesting angles there.

1063
01:01:36,000 --> 01:01:39,196
Audrow Nash: So you're saying
you'll be pulled by different

1064
01:01:39,196 --> 01:01:42,554
markets as they are, like you'll
basically evaluate which ones

1065
01:01:42,554 --> 01:01:46,021
are the good things to go into,
in addition to which ones you're

1066
01:01:46,021 --> 01:01:49,000
well suited for. So it's a
pragmatic choice, basically.

1067
01:01:49,000 --> 01:01:52,965
Nag Murthy: Yeah, it's always a
pragmatic choice. And you know,

1068
01:01:52,965 --> 01:01:57,056
it's, it's, these are the things
that we do today, right, as part

1069
01:01:57,056 --> 01:02:00,895
of our existing business. We're
already collecting data ahead

1070
01:02:00,895 --> 01:02:04,671
for some of these other tasks.
That we need to do right with

1071
01:02:04,671 --> 01:02:08,322
the view. And what's also
interesting is like the, this 3d

1072
01:02:08,322 --> 01:02:11,343
point cloud that Mike talked
about, it's already

1073
01:02:11,343 --> 01:02:15,182
generalizable to things like,
you know, trimming and blowing.

1074
01:02:15,182 --> 01:02:18,895
which are, which are sort of
very different tasks than pure

1075
01:02:18,895 --> 01:02:22,483
surface coverage. But this point
cloud is already able to

1076
01:02:22,483 --> 01:02:26,259
predict, you know, sort of the
attributes. You can still use

1077
01:02:26,259 --> 01:02:29,594
it. Correct. So, and all
together, these sort of more

1078
01:02:29,594 --> 01:02:33,371
than add up to like 30 to 40
cents on the dollar of revenue.

1079
01:02:33,371 --> 01:02:36,329
So if you're making like a
dollar of revenue in

1080
01:02:36,329 --> 01:02:39,413
landscaping, all these
activities, right, mowing,

1081
01:02:39,413 --> 01:02:43,315
blowing, trimming, and then weed
application or weed treatment

1082
01:02:43,315 --> 01:02:47,217
application, they will total up
to more than. I'd say at least

1083
01:02:47,217 --> 01:02:50,867
like 40 cents of like labor cost
when it comes to your own

1084
01:02:50,867 --> 01:02:54,266
business. So now you're
effectively talking about sort

1085
01:02:54,266 --> 01:02:57,853
of 40 cents of automation impact
with just You know, with

1086
01:02:57,853 --> 01:03:01,503
basically the model that we
already have today. Do we want

1087
01:03:01,503 --> 01:03:05,091
to go into snow? Yes. You know,
should we, like, we could

1088
01:03:05,091 --> 01:03:08,741
probably go IPO and not even
touch snow. Right. Like that,

1089
01:03:08,741 --> 01:03:12,643
that's the way to think about
it. And again, like it goes back

1090
01:03:12,643 --> 01:03:16,671
to the philosophy of like, what
should drive a robotics company,

1091
01:03:16,671 --> 01:03:20,510
right? It has to be like, you
know, can you like get a margin

1092
01:03:20,510 --> 01:03:24,350
out of this? And if you can't
like, you know, that's research

1093
01:03:24,350 --> 01:03:28,000
and that's not engineering.
Really don't want to go there.

1094
01:03:28,000 --> 01:03:31,900
Audrow Nash: That's a great way
to put it. Let's see, uh, so

1095
01:03:31,900 --> 01:03:35,864
Mike, going back a little bit,
so you're mentioning all these

1096
01:03:35,864 --> 01:03:39,699
tasks and you kind of broke down
mowing into four different

1097
01:03:39,699 --> 01:03:43,663
components earlier. So is it
like each of these components is

1098
01:03:43,663 --> 01:03:47,498
a different model that you're
kind of putting together in a

1099
01:03:47,498 --> 01:03:51,397
smart way? Or is it, are they
all connected in some way? Are

1100
01:03:51,397 --> 01:03:55,297
they all implicitly handled as
like constraints and whatever

1101
01:03:55,297 --> 01:03:59,261
you're doing to solve? For your
current policy or how, how do

1102
01:03:59,261 --> 01:04:02,706
they fit when you have these
four modes and you add a

1103
01:04:02,706 --> 01:04:06,606
different additional modes when
you go to wheat or you go to

1104
01:04:06,606 --> 01:04:10,505
snow or whatever it is, how do
these all sit together, these

1105
01:04:10,505 --> 01:04:12,000
different capabilities?

1106
01:04:12,000 --> 01:04:16,041
Michael Laskey: Yeah, so on the
robot, there's one single neural

1107
01:04:16,041 --> 01:04:19,829
network that's producing all
these things. so it's just like

1108
01:04:19,829 --> 01:04:23,554
a single, that's why it's kind
of like a, like a, you know,

1109
01:04:23,554 --> 01:04:27,657
like a transformer. But, I'm not
going to read too much about the

1110
01:04:27,657 --> 01:04:31,382
architecture. It is heavily
optimized for embedded devices.

1111
01:04:31,382 --> 01:04:35,360
So it's not just like, it's not
like a completely like fat GPT.

1112
01:04:35,360 --> 01:04:39,463
Obviously, there's no 70 billion
weights on a Jetson. but it is a

1113
01:04:39,463 --> 01:04:43,378
single model that can produce
all these joy presentations. and

1114
01:04:43,378 --> 01:04:47,292
I think as we scale, so right
now we are very inclined to say,

1115
01:04:47,292 --> 01:04:51,206
like, let's produce, like, human
interpretable representations

1116
01:04:51,206 --> 01:04:54,742
that a robotics engineer could
use and maybe, like, code

1117
01:04:54,742 --> 01:04:58,530
heuristics on top of. I want to
get away from that, but in a

1118
01:04:58,530 --> 01:05:02,065
pragmatic way, to the point
where you do have much more,

1119
01:05:02,065 --> 01:05:05,853
like, learned representations
that just feed directly into a

1120
01:05:05,853 --> 01:05:08,000
learned pattern and execute
tasks.

1121
01:05:08,000 --> 01:05:12,840
Audrow Nash: Why do you want to
get away from that? Cause to me,

1122
01:05:12,840 --> 01:05:17,303
it's like, making the black box
bigger, which. Do you think

1123
01:05:17,303 --> 01:05:21,992
it'll be significant performance
gains or you just want things

1124
01:05:21,992 --> 01:05:26,000
opaque so I can't understand
anything or what's that?

1125
01:05:26,000 --> 01:05:29,266
Michael Laskey: I think the
more, the more you apply

1126
01:05:29,266 --> 01:05:33,159
learning, the more interesting,
like, feedback. Well, okay, so

1127
01:05:33,159 --> 01:05:36,488
if you're an end to end system,
one thing that's very

1128
01:05:36,488 --> 01:05:39,879
interesting is like your
feedback becomes really easy,

1129
01:05:39,879 --> 01:05:43,459
right? Because if you make a
mistake and you, like, crash

1130
01:05:43,459 --> 01:05:47,478
your robot, Then all you have to
do is say, hey, on that action,

1131
01:05:47,478 --> 01:05:50,995
you shouldn't have done that.
There's no human labeling.

1132
01:05:50,995 --> 01:05:54,575
There's no sense of like fine
tuning on some sort of like

1133
01:05:54,575 --> 01:05:57,903
weird point cloud. You can
literally just scale these

1134
01:05:57,903 --> 01:06:01,860
models and then like they will
Observe data. So the scalability

1135
01:06:01,860 --> 01:06:05,000
becomes way higher, in a way, if
that makes sense.

1136
01:06:05,000 --> 01:06:08,656
Audrow Nash: Ah, yeah, because
you are not labeling things. And

1137
01:06:08,656 --> 01:06:12,022
because you're not labeling
things, you're not limited by,

1138
01:06:12,022 --> 01:06:15,272
like, people labeling things
with efficient software for

1139
01:06:15,272 --> 01:06:18,000
labeling things. Like all the,
that difficulty.

1140
01:06:18,000 --> 01:06:21,608
Michael Laskey: So we don't
have, because we find, SIEM, our

1141
01:06:21,608 --> 01:06:25,396
labeling pipelines are just in
house, very small. We don't have

1142
01:06:25,396 --> 01:06:28,883
more than, like, you know, a
very small R&amp; D labeling.

1143
01:06:28,883 --> 01:06:32,611
But the more you can unlock true
unsupervised learning or like

1144
01:06:32,611 --> 01:06:36,098
reinforcement learning, the more
interesting it becomes at

1145
01:06:36,098 --> 01:06:37,000
digesting data.

1146
01:06:37,000 --> 01:06:42,247
Audrow Nash: Yep. Yeah. And so
you're using simple heuristics

1147
01:06:42,247 --> 01:06:45,000
like you bump it, it's not good.

1148
01:06:45,000 --> 01:06:48,661
Michael Laskey: Exactly. You
bump it, your point cloud is

1149
01:06:48,661 --> 01:06:52,385
bad. what we're definitely
starting to add to our model is

1150
01:06:52,385 --> 01:06:56,110
like action heads that would
synthesize things and then be

1151
01:06:56,110 --> 01:06:59,514
able to like say, so basically
through, you can learn

1152
01:06:59,514 --> 01:07:03,367
representations through data,
but you don't necessarily have

1153
01:07:03,367 --> 01:07:07,284
those override the like point
cloud. But through fine tuning,

1154
01:07:07,284 --> 01:07:11,202
they could adjust things. the
bumping into things is actually

1155
01:07:11,202 --> 01:07:15,183
not a problem. So I don't want
to give that impression. a wide

1156
01:07:15,183 --> 01:07:19,101
problem. I would say the main
problem right now is like, it's

1157
01:07:19,101 --> 01:07:23,083
actually, let's see, what's the
biggest problem? It's probably

1158
01:07:23,083 --> 01:07:26,936
just laying down these, like,
perfect lines for landscapers.

1159
01:07:26,936 --> 01:07:31,046
Like, no, like, perfect non zero
overlaps, no matter how a human

1160
01:07:31,046 --> 01:07:35,028
does. I would say that's one of
the more, like, core technical

1161
01:07:35,028 --> 01:07:39,009
challenges in mowing. to do it
with, like, almost zero overlap

1162
01:07:39,009 --> 01:07:41,000
is actually really interesting.

1163
01:07:41,000 --> 01:07:46,545
Audrow Nash: So, to train your
system to do that, do you look

1164
01:07:46,545 --> 01:07:51,545
at the result and kind of
evaluate it? Or is there some

1165
01:07:51,545 --> 01:07:56,909
way for you to avoid manual
labeling? so that you can train

1166
01:07:56,909 --> 01:08:00,000
this, or how do you approach
that?

1167
01:08:00,000 --> 01:08:03,444
Michael Laskey: Yeah, I mean,
that's what we're, it's actually

1168
01:08:03,444 --> 01:08:06,833
pretty tricky, because This is
an open problem at the moment.

1169
01:08:06,833 --> 01:08:10,333
Yeah, so like, we train heavily
in SIEM on like, where you have

1170
01:08:10,333 --> 01:08:13,778
ground truth, right? These kind
of performances. and then it's

1171
01:08:13,778 --> 01:08:17,222
really like, what's, cause yeah,
you could have people kind of

1172
01:08:17,222 --> 01:08:20,667
like, dictate and say like, this
line was bad. how do we like,

1173
01:08:20,667 --> 01:08:24,167
get that in a more unsupervised
way? I think it's kind of open.

1174
01:08:24,167 --> 01:08:27,389
the ROAS today lays like pretty
clean, straight lines, but

1175
01:08:27,389 --> 01:08:30,611
periodically it could mess up,
and that's I think the last

1176
01:08:30,611 --> 01:08:32,000
challenge of pushmilling.

1177
01:08:32,000 --> 01:08:35,405
Nag Murthy: This is actually a
really interesting question

1178
01:08:35,405 --> 01:08:38,751
again, right? Which sort of,
which goes back to the whole

1179
01:08:38,751 --> 01:08:41,863
product technology debate,
right? when you sell it to

1180
01:08:41,863 --> 01:08:45,502
landscapers, right? They will
like go on and on about sort of,

1181
01:08:45,502 --> 01:08:48,849
you know, perfect, precise
lines. And, you know, and then

1182
01:08:48,849 --> 01:08:52,489
you talk to the property owners
or the property managers, they

1183
01:08:52,489 --> 01:08:55,894
just don't care. Right. So now
the question is like, whose

1184
01:08:55,894 --> 01:08:59,357
problem should you try and solve
for at the end of the day?

1185
01:08:59,357 --> 01:09:02,586
Right. And there's a level of
engineering, right? Yeah,

1186
01:09:02,586 --> 01:09:06,167
exactly. But there's a level of
engineering pride to get like

1187
01:09:06,167 --> 01:09:09,631
these really straight lines.
Great and the research team or

1188
01:09:09,631 --> 01:09:13,329
in research time should pursue
all that. They can high five and

1189
01:09:13,329 --> 01:09:17,028
everything. Yeah, right But it's
not needed and the other thing

1190
01:09:17,028 --> 01:09:20,726
is, you know, okay So what like
if you miss a couple of stripes

1191
01:09:20,726 --> 01:09:24,366
You can always swarm the damn
thing right like oh like build a

1192
01:09:24,366 --> 01:09:27,889
bit bunch of redundancy into
like how you cover You know the

1193
01:09:27,889 --> 01:09:31,000
the ground and be done with it
At the end of the day.

1194
01:09:31,000 --> 01:09:34,348
Audrow Nash: Yeah, it is it is
so funny because we can fixate

1195
01:09:34,348 --> 01:09:37,530
on those Things that are now
performance measures like did

1196
01:09:37,530 --> 01:09:40,878
you get perfect straight lines?
There's some mowing guy who's

1197
01:09:40,878 --> 01:09:43,841
like the best mower get
impressed by how it does it or

1198
01:09:43,841 --> 01:09:47,079
is the customer like yeah,
here's money Like I'm happy with

1199
01:09:47,079 --> 01:09:49,000
what you did this kind of thing
But

1200
01:09:49,000 --> 01:10:00,000
Michael Laskey: here's the thing
right I feel like

1201
01:10:00,000 --> 01:10:03,423
Nag Murthy: That's what makes
for like robotics. It's such

1202
01:10:03,423 --> 01:10:06,786
fascinating field, right?
Because you have this interplay

1203
01:10:06,786 --> 01:10:10,327
of operational pride with like
business or product pride and

1204
01:10:10,327 --> 01:10:13,455
engineering pride. You don't
want like all of them to

1205
01:10:13,455 --> 01:10:17,054
converge, right? Like, I mean,
you want, you want to converge

1206
01:10:17,054 --> 01:10:20,890
them, but you don't want them to
lose their independence and lose

1207
01:10:20,890 --> 01:10:24,195
their voice. Because then you
have like, you know, three

1208
01:10:24,195 --> 01:10:27,853
shitty sort of like, you know,
endeavors all converging versus

1209
01:10:27,853 --> 01:10:31,217
three perfectionists who will
sort of vote for their, you

1210
01:10:31,217 --> 01:10:34,699
know, thing to happen. But
there's always be consensus that

1211
01:10:34,699 --> 01:10:37,000
forms, you know, at the end of
the day.

1212
01:10:37,000 --> 01:10:40,706
Audrow Nash: And if you put the
legs super close together, they

1213
01:10:40,706 --> 01:10:41,000
fall.

1214
01:10:41,000 --> 01:10:44,225
Nag Murthy: Exactly. That's a
great analogy. Yeah. That's what

1215
01:10:44,225 --> 01:10:46,983
I'm imagining as you're
speaking. That's such a great

1216
01:10:46,983 --> 01:10:50,000
analogy right there. No, it's a
really nice way to put it.

1217
01:10:50,000 --> 01:10:54,219
Audrow Nash: Yeah. And it's, I
really like, I feel like each of

1218
01:10:54,219 --> 01:10:58,170
you, so Mike, Nag, and then
you're, you're, like operations

1219
01:10:58,170 --> 01:11:02,054
founder. I feel like you guys
are probably the great three

1220
01:11:02,054 --> 01:11:05,000
legs voting for your different
perspectives.

1221
01:11:05,000 --> 01:11:08,000
Nag Murthy: That's pretty much
how it works. Yep, exactly.

1222
01:11:08,000 --> 01:11:13,095
Audrow Nash: Oh yeah. Let's see.
So that sounds really cool. the

1223
01:11:13,095 --> 01:11:17,473
simulation side is, I mean,
we're just, the approach is

1224
01:11:17,473 --> 01:11:22,488
very, very cool. one thing that
I wanted to make sure we got to

1225
01:11:22,488 --> 01:11:27,582
talk about, because Nag, when we
talked earlier, there were many

1226
01:11:27,582 --> 01:11:32,438
interesting things brought up.
And so, one thing that I would

1227
01:11:32,438 --> 01:11:37,453
love to hear your perspective on
again is... robot as a service

1228
01:11:37,453 --> 01:11:40,000
model and why you don't like it.

1229
01:11:40,000 --> 01:11:44,041
Nag Murthy: yeah, I think like,
it's not that we don't like it.

1230
01:11:44,041 --> 01:11:48,017
I think it's just lazy pattern
matching by a bunch of software

1231
01:11:48,017 --> 01:11:51,801
VCs, you know, who act to like
justify opportunity costs to

1232
01:11:51,801 --> 01:11:55,842
their LPs who ended up saying,
you know, I'll change one letter

1233
01:11:55,842 --> 01:11:59,497
and we'll call it RAS. As
opposed to SAS. Instead of SAS.

1234
01:11:59,497 --> 01:12:03,025
Yeah. Yeah. And they all high
fived each other. that's,

1235
01:12:03,025 --> 01:12:06,873
that's, that's basically what
happened. it is, it is a model

1236
01:12:06,873 --> 01:12:10,593
that you do want to get to in
the limit, right? Like when,

1237
01:12:10,593 --> 01:12:14,505
when you're, when you've built
like this sort of common sense

1238
01:12:14,505 --> 01:12:18,097
robot that can like navigate the
world, right? It really

1239
01:12:18,097 --> 01:12:22,073
understands what the world is.
It's kind of like OpenAI didn't

1240
01:12:22,073 --> 01:12:25,729
like start by like licensing
chat GPT, like from day one,

1241
01:12:25,729 --> 01:12:29,898
right? They took a bunch of time
to develop it to the point where

1242
01:12:29,898 --> 01:12:33,874
it was like, you could put an
API call to it and you know, you

1243
01:12:33,874 --> 01:12:37,787
could make some use out of it. I
think robotics, if you start

1244
01:12:37,787 --> 01:12:41,442
enforcing like a RAS like
approach. What that leads to is

1245
01:12:41,442 --> 01:12:45,419
you're forced to make a lot of
classical choices when it comes

1246
01:12:45,419 --> 01:12:48,754
to your stack. And it's clear
that the future is not

1247
01:12:48,754 --> 01:12:52,795
classical. Like, we should take
it as an article of faith that,

1248
01:12:52,795 --> 01:12:56,450
you know, I'm not saying
everybody should, we've taken it

1249
01:12:56,450 --> 01:13:00,427
as an article of faith. That the
future is in classical. So if

1250
01:13:00,427 --> 01:13:03,890
you buy that piece, right,
exactly. So if you buy that

1251
01:13:03,890 --> 01:13:07,931
argument, then you have to sort
of say, okay, fine. What, like,

1252
01:13:07,931 --> 01:13:11,907
why would you want to like go
with RAS knowing that, you know,

1253
01:13:11,907 --> 01:13:15,691
a few years from now, like a
fully ML based system is going

1254
01:13:15,691 --> 01:13:18,000
to blow everything out of the
water.

1255
01:13:18,000 --> 01:13:21,906
Audrow Nash: Yeah. The thing
that struck me when we talked

1256
01:13:21,906 --> 01:13:25,946
earlier about this perspective
was the alternative was. Just

1257
01:13:25,946 --> 01:13:30,121
have the robot do the job and
charge for the job like it's any

1258
01:13:30,121 --> 01:13:34,498
other job. So like, if I want my
lawn cut, I don't care that it's

1259
01:13:34,498 --> 01:13:38,673
robots or people cutting it, I
just want my damn lawn cut. And

1260
01:13:38,673 --> 01:13:42,916
if that's the case, what you can
do is you can charge the exact

1261
01:13:42,916 --> 01:13:47,226
way that the customer was taking
it before. It's on a job by job

1262
01:13:47,226 --> 01:13:51,333
basis, not a subscription. And
like a big benefit of that was

1263
01:13:51,333 --> 01:13:55,508
it's easier for the customer.
And they can try you out without

1264
01:13:55,508 --> 01:13:59,549
big commitments where it's a
robot as a service It's often a

1265
01:13:59,549 --> 01:14:03,724
big upfront payment or probably
is sometimes and then it's the

1266
01:14:03,724 --> 01:14:08,034
subscription cost where you say
I'm gonna do six months at least

1267
01:14:08,034 --> 01:14:12,141
And pay you several thousand
dollars for supporting the robot

1268
01:14:12,141 --> 01:14:16,451
and this kind of thing, which is
a lot more friction to adoption

1269
01:14:16,451 --> 01:14:18,000
for this kind of thing.

1270
01:14:18,000 --> 01:14:21,911
Nag Murthy: Yeah, I think we're
going like even more extreme

1271
01:14:21,911 --> 01:14:25,495
than that, right? We're
basically saying, you know, buy

1272
01:14:25,495 --> 01:14:29,080
a company that provides the
service. And, you know, all

1273
01:14:29,080 --> 01:14:33,056
like, we're not even saying by
job, we're saying by business,

1274
01:14:33,056 --> 01:14:36,836
right? Like, it's just, it's
like, take the extreme, like,

1275
01:14:36,836 --> 01:14:40,486
like argument, it's sort of,
ridiculous extremes, right?

1276
01:14:40,486 --> 01:14:44,332
Like, okay, classical, take it
to the ridiculous extreme of

1277
01:14:44,332 --> 01:14:48,438
like all ML, and then take the
business model and say, take RAS

1278
01:14:48,438 --> 01:14:52,284
and take it to the extreme of
saying, just get paid for the

1279
01:14:52,284 --> 01:14:56,194
service. right. and it's, yeah,
I think both of them sort of

1280
01:14:56,194 --> 01:14:59,844
work well together. And the
reason why we're doing this,

1281
01:14:59,844 --> 01:15:03,755
again, I don't think, I want to
say that, you know, it's not

1282
01:15:03,755 --> 01:15:07,535
true that we'll never sell a
robot to a landscaper, right?

1283
01:15:07,535 --> 01:15:11,576
Like, maybe that'll happen, you
know, when we're like a couple

1284
01:15:11,576 --> 01:15:15,617
billion in revenue, right? 10
years from now, we have like the

1285
01:15:15,617 --> 01:15:19,202
best model out there with the
deepest moats, then we'll

1286
01:15:19,202 --> 01:15:22,917
basically happily sort of
destroy the competition, right?

1287
01:15:22,917 --> 01:15:26,567
Like that, that's. Not a
problem, right? But step one is

1288
01:15:26,567 --> 01:15:30,413
to build that mode and like
build it as deep and as wide as

1289
01:15:30,413 --> 01:15:34,389
you can, right? with the data
because Machine learning again,

1290
01:15:34,389 --> 01:15:38,234
like it's like to the to the
person with the model goes the

1291
01:15:38,234 --> 01:15:42,080
spoils Like every everyone else
is going to build a wrapper

1292
01:15:42,080 --> 01:15:45,990
around chat GPT, right? But
OpenAI is gonna, OpenAI is gonna

1293
01:15:45,990 --> 01:15:49,771
eat everybody's lunch Like it's,
you know, like they, they

1294
01:15:49,771 --> 01:15:53,681
started to eat Midjourney's
lunch with like, you know, Dolly

1295
01:15:53,681 --> 01:15:57,657
3 and like, this is going to
like take, and like, you look at

1296
01:15:57,657 --> 01:16:01,568
what happened to Jasper and like
some of these other sort of

1297
01:16:01,568 --> 01:16:05,088
wrappers that were built on
OpenAI. They don't stand a

1298
01:16:05,088 --> 01:16:06,000
chance, right?

1299
01:16:06,000 --> 01:16:09,933
Audrow Nash: Yeah, because
they're not, they're not doing

1300
01:16:09,933 --> 01:16:14,003
the thing. They're just using
the thing. And so they can, I

1301
01:16:14,003 --> 01:16:18,143
feel like Apple does this too.
Where it's like, oh, they let

1302
01:16:18,143 --> 01:16:22,628
everyone make a notes taping app
and then they make the best, not

1303
01:16:22,628 --> 01:16:26,767
that their notes app is that
great, it's okay. But they keep

1304
01:16:26,767 --> 01:16:30,769
folding things in that other
people were doing, except for

1305
01:16:30,769 --> 01:16:35,116
that this thing. Let's see. I'm
on There it is everything that,

1306
01:16:35,116 --> 01:16:39,048
that one I could have lived
without, but maybe not, never

1307
01:16:39,048 --> 01:16:43,119
know But, so yeah, so you're
doing the thing versus, you're

1308
01:16:43,119 --> 01:16:46,983
creating rather than just
wrapping something else you're

1309
01:16:46,983 --> 01:16:51,260
going heavily in the Direction
of, and you're, you're building

1310
01:16:51,260 --> 01:16:55,193
that moat by getting out there,
by doing the model that's

1311
01:16:55,193 --> 01:16:59,401
easiest for customers for this
kind of thing. And you're also

1312
01:16:59,401 --> 01:17:03,748
just by buying businesses. Like
that is a very cool step. Cause

1313
01:17:03,748 --> 01:17:07,405
I've heard of like boring
businesses and this kind of

1314
01:17:07,405 --> 01:17:11,820
thing where you can buy, I don't
know, a landscaping company and

1315
01:17:11,820 --> 01:17:15,960
then you add some technology.
And what you guys are doing is

1316
01:17:15,960 --> 01:17:19,824
you're literally doing that,
where you buy a landscaping

1317
01:17:19,824 --> 01:17:24,101
company, and then you augment
them. With your robot platforms,

1318
01:17:24,101 --> 01:17:28,172
making them even more efficient
and even better, able to do

1319
01:17:28,172 --> 01:17:29,000
their tasks.

1320
01:17:29,000 --> 01:17:32,766
Nag Murthy: Exactly. That's the
goal. again, though the, you

1321
01:17:32,766 --> 01:17:36,720
know, the value of the company.
So if you think about like when

1322
01:17:36,720 --> 01:17:40,736
you buy a landscaping business,
right. You're sort of valuing it

1323
01:17:40,736 --> 01:17:44,377
on like an EBITDA multiple,
right? So you're buying it for

1324
01:17:44,377 --> 01:17:48,142
sort of X times its earnings.
but then like to us, the value

1325
01:17:48,142 --> 01:17:52,033
of the company is not just like
what we buy it for in terms of

1326
01:17:52,033 --> 01:17:55,862
its earnings. It's the data and
you're buying essentially the

1327
01:17:55,862 --> 01:17:59,565
services of all the operators
who can perform reinforcement

1328
01:17:59,565 --> 01:18:03,519
learning for you on the ground.
What a clever model. That's the

1329
01:18:03,519 --> 01:18:07,096
crux of the whole thing. And
again, this is all driven by

1330
01:18:07,096 --> 01:18:10,611
like this sort of, it's an
article of faith that machine

1331
01:18:10,611 --> 01:18:14,000
learning will eat everything,
you know, going forward.

1332
01:18:14,000 --> 01:18:18,197
Audrow Nash: I wonder, I mean,
it seems like it probably will

1333
01:18:18,197 --> 01:18:22,326
to me that just like all the
advances we've been seeing. And

1334
01:18:22,326 --> 01:18:26,523
I mean, like even, so you look
at like perception and now the

1335
01:18:26,523 --> 01:18:30,170
way everyone does it, it was
classical when I started

1336
01:18:30,170 --> 01:18:34,229
robotics and now it's all.
Learned like it's all perception

1337
01:18:34,229 --> 01:18:38,702
that's running through AI, which
is totally nuts And then you get

1338
01:18:38,702 --> 01:18:42,693
a pose estimation or whatever it
is out of it, and they're

1339
01:18:42,693 --> 01:18:44,000
they're all learned

1340
01:18:44,000 --> 01:18:47,726
Nag Murthy: You know, what's
also interesting is like this is

1341
01:18:47,726 --> 01:18:51,391
it's not it's not like it's not
like like we discovered this

1342
01:18:51,391 --> 01:18:55,178
religion You know in isolation,
right? It's because we tried a

1343
01:18:55,178 --> 01:18:59,026
bunch of other approaches and we
tried to scale hard with those

1344
01:18:59,026 --> 01:19:02,752
approaches, right? Like I built
like the first Like GPS based

1345
01:19:02,752 --> 01:19:06,723
more in my garage, right? Like I
built it on like a, like a drone

1346
01:19:06,723 --> 01:19:10,449
audio pilot stack. I learned
like, like literally, right. And

1347
01:19:10,449 --> 01:19:14,052
we got our first, like, you
know. A couple hundred thousand

1348
01:19:14,052 --> 01:19:17,962
dollar contract on the basis of
like this rickety thing that I'd

1349
01:19:17,962 --> 01:19:21,443
built in my garage. It was
crazy, right? And we're deeply

1350
01:19:21,443 --> 01:19:25,047
sort of aware of the pain,
painful sort of aspects of these

1351
01:19:25,047 --> 01:19:28,529
classical systems, whether
you're talking about GPS. Then

1352
01:19:28,529 --> 01:19:32,255
we found lidars and we're like,
okay, this is amazing, right?

1353
01:19:32,255 --> 01:19:36,103
You can map the whole world, you
can teach a path, but then you

1354
01:19:36,103 --> 01:19:39,401
start deploying those and you
realize how brittle that

1355
01:19:39,401 --> 01:19:42,761
approach is because the
environment is changing all the

1356
01:19:42,761 --> 01:19:46,304
time. Right? Unless you have
semantic awareness, you don't

1357
01:19:46,304 --> 01:19:49,663
have any hope of solving. Like a
robotic problem in the

1358
01:19:49,663 --> 01:19:53,206
unstructured outdoors to the
level of reliability where it

1359
01:19:53,206 --> 01:19:57,054
can actually, you know, generate
like significant value unlock.

1360
01:19:57,054 --> 01:20:00,597
and then what's interesting
about machine learning is that

1361
01:20:00,597 --> 01:20:04,323
you can get to an 80 percent
good solution pretty quickly. So

1362
01:20:04,323 --> 01:20:08,110
it's, it's not like a moonshot
bet where you, you have to wait

1363
01:20:08,110 --> 01:20:11,531
for it to become a hundred
percent. So these approaches,

1364
01:20:11,531 --> 01:20:15,134
our mowers today are generating
value, right? They're fully

1365
01:20:15,134 --> 01:20:18,677
like, they're all ML, but
they're generating value. And so

1366
01:20:18,677 --> 01:20:22,342
that allows us to sort of plow
back that value back into the

1367
01:20:22,342 --> 01:20:26,068
business, right? So you're not
trying to like take a deep hit

1368
01:20:26,068 --> 01:20:29,000
on the cost. In any case, when
it comes to this.

1369
01:20:29,000 --> 01:20:33,370
Audrow Nash: Yeah. Cause you're
staying profitable. Like you're

1370
01:20:33,370 --> 01:20:37,809
just, you're buying a business,
you are scaling and getting more

1371
01:20:37,809 --> 01:20:41,346
data and maybe getting
investment, but it's on your

1372
01:20:41,346 --> 01:20:45,438
terms cause you're profitable.
and it just, it seems like a

1373
01:20:45,438 --> 01:20:49,461
really good way to proceed to
me. Like the whole thing of,

1374
01:20:49,461 --> 01:20:53,484
cause one thing I really have,
I've seen so many companies

1375
01:20:53,484 --> 01:20:57,438
like, so I've been podcasting
for, I think it's almost 10

1376
01:20:57,438 --> 01:21:01,044
years now, which is nuts. So
I've talked to a ton of

1377
01:21:01,044 --> 01:21:04,929
companies for this. So Sense
Think Act, but then RoboHub

1378
01:21:04,929 --> 01:21:09,090
before that. and so I've talked
to, and the whole time I was

1379
01:21:09,090 --> 01:21:13,529
interested in startups and, I've
talked to so many companies and

1380
01:21:13,529 --> 01:21:17,344
a lot of them are out of
business, which is nuts. and a

1381
01:21:17,344 --> 01:21:21,714
big reason that that. Seems to
occur is they take on a bunch of

1382
01:21:21,714 --> 01:21:25,529
investment and then they go and
try to find a market or

1383
01:21:25,529 --> 01:21:29,690
something like this. Maybe they
found one, but then they are

1384
01:21:29,690 --> 01:21:34,060
forced to become profitable or
they're forced to try to cash in

1385
01:21:34,060 --> 01:21:38,360
so that the venture capitalists
can cash out on some timeline.

1386
01:21:38,360 --> 01:21:42,591
And a big way to fight against
that in a sense, like you want

1387
01:21:42,591 --> 01:21:46,684
investment to scale and you want
it for the connections and

1388
01:21:46,684 --> 01:21:50,845
things like this. So it's not
all bad. but by having revenue

1389
01:21:50,845 --> 01:21:55,145
by having kind of like a very
good long plan, you put yourself

1390
01:21:55,145 --> 01:21:59,376
in a much better position. To
accept investment on your terms

1391
01:21:59,376 --> 01:22:02,983
and by buying profitable
businesses that are already

1392
01:22:02,983 --> 01:22:07,145
profitable And then just making
them a little more efficient

1393
01:22:07,145 --> 01:22:11,306
like that that to me seems like
everything is going to be on

1394
01:22:11,306 --> 01:22:12,000
your terms

1395
01:22:12,000 --> 01:22:15,167
Nag Murthy: Exactly on your
timeline your terms and you

1396
01:22:15,167 --> 01:22:18,276
know, not everyone can raise
like A billion dollars in

1397
01:22:18,276 --> 01:22:21,788
charity investment like OpenAI
did, right? Like, we've got to

1398
01:22:21,788 --> 01:22:25,242
figure out, like, They're all
for profit, though, I think. I

1399
01:22:25,242 --> 01:22:28,582
don't know exactly about their
situation. Well, yeah, they

1400
01:22:28,582 --> 01:22:32,267
pulled a for benefit thing. But
what's interesting also is that,

1401
01:22:32,267 --> 01:22:35,721
you know, OpenAI could, like,
scrape the web and get all the

1402
01:22:35,721 --> 01:22:39,233
language data that they could.
Now they can watch YouTube and

1403
01:22:39,233 --> 01:22:42,803
they can build multimodal LLMs,
right? But where are you going

1404
01:22:42,803 --> 01:22:46,258
to get the data to build robots?
Yeah, it doesn't exist. And

1405
01:22:46,258 --> 01:22:49,655
sure, Google and DeepMind and
all these guys can, you know,

1406
01:22:49,655 --> 01:22:53,052
like rally together and like
collaborate to like, you know,

1407
01:22:53,052 --> 01:22:56,391
build this massive dataset. But
it's not enough to build a

1408
01:22:56,391 --> 01:22:59,903
dataset. You need to have like
boots on the ground to do that

1409
01:22:59,903 --> 01:23:03,358
sort of reinforcement learning
and provide the feedback. So,

1410
01:23:03,358 --> 01:23:06,697
without a boots on the ground
approach, there's no hope in

1411
01:23:06,697 --> 01:23:09,000
hell to build outdoor robotics
at scale.

1412
01:23:09,000 --> 01:23:12,251
Audrow Nash: And especially if
you're not in Google or

1413
01:23:12,251 --> 01:23:16,104
whatever, some huge company that
can fund all this. Like, if you

1414
01:23:16,104 --> 01:23:19,716
want to do it as an individual,
you need to probably do this

1415
01:23:19,716 --> 01:23:23,147
approach. I'm sure you've
thought about this quite a lot,

1416
01:23:23,147 --> 01:23:27,000
um, and this seems like the way
to scale. You have to bootstrap.

1417
01:23:27,000 --> 01:23:31,013
Nag Murthy: Yeah, it's the way
to fund yourself and, you know,

1418
01:23:31,013 --> 01:23:34,961
and honestly, there's another
angle to this, which is you end

1419
01:23:34,961 --> 01:23:38,197
up building something, like
robotics is a workflow

1420
01:23:38,197 --> 01:23:42,340
optimization problem and it's a
people plus machine optimization

1421
01:23:42,340 --> 01:23:46,029
problem and the only way you can
solve this is by sort of

1422
01:23:46,029 --> 01:23:50,042
injecting robots and trying to
inject robots in many different

1423
01:23:50,042 --> 01:23:54,249
ways into an existing operation,
right? Because it's hard to sort

1424
01:23:54,249 --> 01:23:58,133
of just say, here's the perfect
machine, I'll drop it in and

1425
01:23:58,133 --> 01:24:01,757
like, boom, like, you know,
everybody's like high fiving

1426
01:24:01,757 --> 01:24:05,900
each other. That doesn't happen.
Yeah. You know, it's very hard.

1427
01:24:05,900 --> 01:24:09,718
and so it's almost like you're,
you've sort of evolved this

1428
01:24:09,718 --> 01:24:13,537
robot, like fit like a glove.
into your existing operation.

1429
01:24:13,537 --> 01:24:17,550
And it is, the data engine is
not just for ML. The data engine

1430
01:24:17,550 --> 01:24:21,369
is also to influence all other
aspects of your product. you

1431
01:24:21,369 --> 01:24:25,188
know, and so it, it's, it's
just, it's just like a rational

1432
01:24:25,188 --> 01:24:27,000
way to do it. Yeah, exactly.

1433
01:24:27,000 --> 01:24:31,565
Audrow Nash: For sure. Yeah.
It's a very cool idea. So Mike,

1434
01:24:31,565 --> 01:24:35,826
how, so going, building this
data model for all sorts of

1435
01:24:35,826 --> 01:24:40,011
things, like how, how are we
doing that? Like how we're

1436
01:24:40,011 --> 01:24:44,348
collecting lots of data. How do
you make it so things can

1437
01:24:44,348 --> 01:24:49,065
generalize? From the mowers to
the snowblowers to, or the snow

1438
01:24:49,065 --> 01:24:53,859
pushers, whatever it is, all the
things. How do you, how do you

1439
01:24:53,859 --> 01:24:58,576
organize, how does all this data
work together, I suppose, and

1440
01:24:58,576 --> 01:25:02,000
I'm sure that's a very
interesting challenge.

1441
01:25:02,000 --> 01:25:05,552
Michael Laskey: Yeah, I mean,
so, I think we have to, the way

1442
01:25:05,552 --> 01:25:08,871
we're kind of structuring is
like, we're trying to build,

1443
01:25:08,871 --> 01:25:12,365
like, an intuitive physical
model of the world for the robot

1444
01:25:12,365 --> 01:25:15,568
that can have, like, key
concepts, like, I know where I

1445
01:25:15,568 --> 01:25:18,829
am, I know where I've been, I
understand the semantic 3D

1446
01:25:18,829 --> 01:25:22,148
structure of the world, I
understand what I can move over

1447
01:25:22,148 --> 01:25:25,641
and what I cannot move over.
Those core concepts, they apply

1448
01:25:25,641 --> 01:25:28,844
to all the tasks you just
mentioned. so, as we, what we

1449
01:25:28,844 --> 01:25:32,396
try to do these days is, like,
you know, this year we shipped

1450
01:25:32,396 --> 01:25:36,065
the mowing product, Next year,
we're going to want to see like,

1451
01:25:36,065 --> 01:25:39,558
okay, let's test our stack and
do something like trimming or

1452
01:25:39,558 --> 01:25:43,052
edging or blowing. We'll take
the same model and then put it

1453
01:25:43,052 --> 01:25:46,837
on a different platform and have
the same, basic deployment cycle

1454
01:25:46,837 --> 01:25:50,098
with that. But we keep it to
being trained with the same

1455
01:25:50,098 --> 01:25:53,534
model and the same model should
run on every robot. I think

1456
01:25:53,534 --> 01:25:57,202
that's very feasible, especially
because, as Nox said, a lot of

1457
01:25:57,202 --> 01:26:00,754
these tasks are so similar that
you would have to change like

1458
01:26:00,754 --> 01:26:04,365
almost nothing to do different
tasks right now. so cool. Yeah,

1459
01:26:04,365 --> 01:26:07,684
and what's gonna get really
interesting over the years is

1460
01:26:07,684 --> 01:26:11,352
when we get into like dexterous
manipulation, which is when you

1461
01:26:11,352 --> 01:26:14,904
look at landscaping, it's almost
like this goldmine of mobile

1462
01:26:14,904 --> 01:26:17,932
manipulation outdoor
unstructured environments. Like

1463
01:26:17,932 --> 01:26:21,601
you actually probably couldn't
think of a better benchmark than

1464
01:26:21,601 --> 01:26:25,153
landscaping because every day
you have a crew like, you know,

1465
01:26:25,153 --> 01:26:28,414
right now we have a hundred
people go out and do complex

1466
01:26:28,414 --> 01:26:31,674
manipulation in random parks,
random cities all over the

1467
01:26:31,674 --> 01:26:35,401
country, right? So, like, that's
not, you're not gonna get that,

1468
01:26:35,401 --> 01:26:38,779
like, Google, they have, like,
maybe 10 people demo to the

1469
01:26:38,779 --> 01:26:42,272
robot in a conference room. Same
conference room, same robot

1470
01:26:42,272 --> 01:26:45,766
every day. The type of data that
we're getting is very much,

1471
01:26:45,766 --> 01:26:49,318
like, real, unstructured mobile
manipulation in the wild. and

1472
01:26:49,318 --> 01:26:52,870
doing complex tasks, like, if
you ever see someone tree trim,

1473
01:26:52,870 --> 01:26:56,015
that's insane. Like, if a
humanoid was able to climb a

1474
01:26:56,015 --> 01:26:59,508
tree and cut a branch. That
would be like a huge AGI moment.

1475
01:26:59,508 --> 01:27:03,177
Oh my god, yeah. Yeah, so these
are some of those complex tasks

1476
01:27:03,177 --> 01:27:06,729
that you see people do daily.
And we can collect and train on

1477
01:27:06,729 --> 01:27:09,000
that data, which is really
interesting.

1478
01:27:09,000 --> 01:27:12,770
Audrow Nash: Huh, so for people
to do these manipulation tasks

1479
01:27:12,770 --> 01:27:16,419
and everything, are you just
having your crews wear cameras?

1480
01:27:16,419 --> 01:27:20,128
Or like, what's the, or is it
just your robots are going out,

1481
01:27:20,128 --> 01:27:23,473
is that what you mean? But
eventually, maybe you'll add

1482
01:27:23,473 --> 01:27:27,000
more sensors around the crew, so
you can get data on them?

1483
01:27:27,000 --> 01:27:30,421
Michael Laskey: The number one
data right now is like they

1484
01:27:30,421 --> 01:27:34,079
bring the robots with them and
deploy them every day. So we're

1485
01:27:34,079 --> 01:27:37,736
getting that like interactive
point of view perspective of the

1486
01:27:37,736 --> 01:27:41,275
robot. as we start scaling to
like dexterous, you would have

1487
01:27:41,275 --> 01:27:44,933
to ask the question like is the
robot platform we're trying to

1488
01:27:44,933 --> 01:27:48,413
do good enough that you can just
do like RL type data, like

1489
01:27:48,413 --> 01:27:51,952
interactive feedback, or would
you want to have passive data

1490
01:27:51,952 --> 01:27:55,433
collection? That's gonna be the
challenge. But then you can

1491
01:27:55,433 --> 01:27:59,031
also, like, you can use these
people as basically ops, right?

1492
01:27:59,031 --> 01:28:02,511
Like, every Roblox company has
an ops team. Our ops team is

1493
01:28:02,511 --> 01:28:05,933
likely the biggest ops team.
Because we have, like, entire

1494
01:28:05,933 --> 01:28:09,000
crews of people doing this work
for us, right? Yeah.

1495
01:28:09,000 --> 01:28:12,877
Audrow Nash: It seems a lot like
what Tesla did with their self

1496
01:28:12,877 --> 01:28:16,815
driving for everything, which is
they just kind of outsourced it

1497
01:28:16,815 --> 01:28:20,508
and then they get the benefit
from millions or billions or I

1498
01:28:20,508 --> 01:28:24,015
don't even know how much,
driving data. Billions of miles

1499
01:28:24,015 --> 01:28:25,000
of driving data.

1500
01:28:25,000 --> 01:28:28,729
Nag Murthy: It's, it's what
every company has done though,

1501
01:28:28,729 --> 01:28:32,393
right? Like if you take a look
at like, like Facebook and

1502
01:28:32,393 --> 01:28:36,379
Google or whoever else who sits
on all these like massive data

1503
01:28:36,379 --> 01:28:40,364
sort of, you know, data, what a
units, Which, you know, people

1504
01:28:40,364 --> 01:28:44,414
buy to train their LLMs that was
generated by users who sort of

1505
01:28:44,414 --> 01:28:48,593
paying in kind that data for the
privilege of using a Facebook or

1506
01:28:48,593 --> 01:28:52,321
a Google or a Reddit or a
Twitter, right? So it's the same

1507
01:28:52,321 --> 01:28:56,307
thing. Tesla's basically doing
the same thing for the physical

1508
01:28:56,307 --> 01:29:00,293
world, and we are doing the same
thing for landscaping. That's

1509
01:29:00,293 --> 01:29:01,000
super cool.

1510
01:29:01,000 --> 01:29:05,984
Audrow Nash: Yeah. What a cool
connection. so where, where do

1511
01:29:05,984 --> 01:29:11,132
you guys see yourself? Where do
you see yourselves going in the

1512
01:29:11,132 --> 01:29:15,872
next like. Like, what's the
future, you think, for like, I

1513
01:29:15,872 --> 01:29:20,693
don't know, five years, ten
years? What do you think? maybe

1514
01:29:20,693 --> 01:29:22,000
start with Mike?

1515
01:29:22,000 --> 01:29:25,959
Michael Laskey: Yeah, I mean, so
I'll speak more technology wise.

1516
01:29:25,959 --> 01:29:29,248
I think what we're wanting to
get to is like, into the

1517
01:29:29,248 --> 01:29:32,964
dexterous tasks, right? Like
we've definitely, where we're at

1518
01:29:32,964 --> 01:29:36,557
today, I feel like surface,
mechanized surface coverage, is

1519
01:29:36,557 --> 01:29:40,151
obtainable within the next
couple years. And then we really

1520
01:29:40,151 --> 01:29:43,745
want to have hardware platforms
where we could do dexterous

1521
01:29:43,745 --> 01:29:46,912
tasks. and I think you're
seeing, with hardware it's

1522
01:29:46,912 --> 01:29:50,384
actually really interesting
because you're just seeing so

1523
01:29:50,384 --> 01:29:53,977
many prototypes come out of like
hardware platforms. so you

1524
01:29:53,977 --> 01:29:57,632
could, you might even have like
some sort of commodity thing

1525
01:29:57,632 --> 01:30:01,165
that can sort of move in a
dexterous way and then apply AI

1526
01:30:01,165 --> 01:30:05,002
to it or we build our own. which
we're very capable of. I think

1527
01:30:05,002 --> 01:30:08,839
we have a pretty decent in house
like hardware team these days.

1528
01:30:08,839 --> 01:30:12,676
and then, we want to get to the
scale though, where it is like,

1529
01:30:12,676 --> 01:30:16,575
you know, every state we're in,
every like, county we're in, and

1530
01:30:16,575 --> 01:30:20,229
our robots are learning from
every data possible. and if you

1531
01:30:20,229 --> 01:30:24,066
look at, like, this space, it's
very possible. You have so much

1532
01:30:24,066 --> 01:30:27,782
fragmentation in the industry,
it's just waiting to be rolled

1533
01:30:27,782 --> 01:30:29,000
up in a massive way.

1534
01:30:29,000 --> 01:30:32,000
Audrow Nash: Oh yeah, what do
you think?

1535
01:30:32,000 --> 01:30:35,787
Nag Murthy: I think it's like in
five years, it's, you know,

1536
01:30:35,787 --> 01:30:39,259
given like our sort of
acquisition pipeline and sort of

1537
01:30:39,259 --> 01:30:42,983
the state of the industry, it's
very easy to see getting to

1538
01:30:42,983 --> 01:30:46,896
something in like the billion
dollars of revenue range when it

1539
01:30:46,896 --> 01:30:50,494
comes to rolling up this
industry. And then I think we'll

1540
01:30:50,494 --> 01:30:54,407
have this really mature model,
which we can then start to, and

1541
01:30:54,407 --> 01:30:58,068
like these robots that are fully
baked, right, for our own

1542
01:30:58,068 --> 01:31:01,918
operations, we can start opening
it up to other and like, you

1543
01:31:01,918 --> 01:31:05,516
know, other people in the
industry. But really the, would

1544
01:31:05,516 --> 01:31:09,366
be awesome to build a monopoly.
And just like, you know, just

1545
01:31:09,366 --> 01:31:12,838
like, just take the whole
industry. Cause I mean, yeah,

1546
01:31:12,838 --> 01:31:16,499
like, yeah, deliver the service,
build your own robots. It

1547
01:31:16,499 --> 01:31:20,223
doesn't get like more sort of,
you know, a bigger moat than

1548
01:31:20,223 --> 01:31:23,000
that. Sort of take the industry.
Yeah. Yeah.

1549
01:31:23,000 --> 01:31:28,496
Audrow Nash: Cause I mean, like,
related to that, I would love to

1550
01:31:28,496 --> 01:31:33,401
hear what advice you have. For
people other than join your

1551
01:31:33,401 --> 01:31:38,559
company like what what kinds of
what so with this kind of new

1552
01:31:38,559 --> 01:31:43,463
world of machine learning and
robotics and Data being very

1553
01:31:43,463 --> 01:31:48,537
valuable. how, how do you think
someone can skill up or best

1554
01:31:48,537 --> 01:31:53,779
position themselves or start a
company like what? I know it's,

1555
01:31:53,779 --> 01:31:59,022
it's a super broad question, but
what would you think would be

1556
01:31:59,022 --> 01:32:03,673
some of the most important
things to doing well in this

1557
01:32:03,673 --> 01:32:08,577
kind of new world? We're in,
maybe you start with Nag this

1558
01:32:08,577 --> 01:32:09,000
time.

1559
01:32:09,000 --> 01:32:12,771
Nag Murthy: Yeah, I'd say it
really comes down to like people

1560
01:32:12,771 --> 01:32:16,110
from different disciplines
really need to talk to each

1561
01:32:16,110 --> 01:32:19,325
other. So if you're a
roboticist, like don't talk to

1562
01:32:19,325 --> 01:32:23,096
other roboticists, go talk to
like the business folks. Right.

1563
01:32:23,096 --> 01:32:26,867
and if you're, yeah. And vice
versa, right? If you're someone

1564
01:32:26,867 --> 01:32:30,391
in the PE space or someone in
like the VCs, I don't know,

1565
01:32:30,391 --> 01:32:34,286
they've got their own sort of,
you know, their own sort of swim

1566
01:32:34,286 --> 01:32:37,810
lanes that they swim in. But I
think robotics is going to

1567
01:32:37,810 --> 01:32:41,519
benefit a lot from like the PE
folks talking to the, through

1568
01:32:41,519 --> 01:32:45,043
the roboticists. Again, there
are reasons why it'll never

1569
01:32:45,043 --> 01:32:48,753
happen, but I wish that like
more of it would happen. How do

1570
01:32:48,753 --> 01:32:49,000
you,

1571
01:32:49,000 --> 01:32:51,862
Audrow Nash: what are some of
the reasons and how can you kind

1572
01:32:51,862 --> 01:32:54,677
of best them for this? Or like
what, cause individuals can do

1573
01:32:54,677 --> 01:32:55,000
things.

1574
01:32:55,000 --> 01:32:58,674
Nag Murthy: Individuals should
and will talk to each other,

1575
01:32:58,674 --> 01:33:02,161
right? Like, it's already
happening, right? So you have,

1576
01:33:02,161 --> 01:33:06,021
well, it's already happening for
non robotics things is what I

1577
01:33:06,021 --> 01:33:10,068
mean, right? Like where you have
PE companies, you know, that are

1578
01:33:10,068 --> 01:33:13,555
looking to inject like LLMs, LLM
agents, you know, their

1579
01:33:13,555 --> 01:33:17,540
operations. We use LLMs to sort
of, you know, do bidding on like

1580
01:33:17,540 --> 01:33:21,214
new job sites and things like
that. So there's, yeah, it's,

1581
01:33:21,214 --> 01:33:25,075
it's fun how all that is shaping
out. but what's, what happens

1582
01:33:25,075 --> 01:33:28,998
with PE is typically they have,
they don't know how to evaluate

1583
01:33:28,998 --> 01:33:32,920
technical risk, private equity,
like, right. It's like, I think

1584
01:33:32,920 --> 01:33:36,532
the world today is sort of
siloed into people who know how

1585
01:33:36,532 --> 01:33:40,392
to evaluate technical risk and
people who know how to evaluate

1586
01:33:40,392 --> 01:33:44,191
financial risk, right? Private
equity is like financial risk.

1587
01:33:44,191 --> 01:33:47,802
So they know how to take a
mature business and then inject

1588
01:33:47,802 --> 01:33:51,600
a bunch of debt into it and then
blow it up to be bigger. The

1589
01:33:51,600 --> 01:33:55,274
VC's come in and they know how
to evaluate. Technical risk,

1590
01:33:55,274 --> 01:33:58,885
right? But it's for like, for
anyone who's trying to build

1591
01:33:58,885 --> 01:34:02,808
like something similar or wants
to like, you know, explore more

1592
01:34:02,808 --> 01:34:06,856
along these lines, it just feels
like there's got to be sort of a

1593
01:34:06,856 --> 01:34:10,592
way. Where these two models can
merge for robotics to really

1594
01:34:10,592 --> 01:34:14,078
take off because otherwise
you're going to have the same

1595
01:34:14,078 --> 01:34:17,690
sort of repeated cycle of
Companies that go public through

1596
01:34:17,690 --> 01:34:21,239
SPACs and like, you know, I
don't want to name names here

1597
01:34:21,239 --> 01:34:25,037
But they're trading on like, you
know, the stock exchange and

1598
01:34:25,037 --> 01:34:29,085
they're worth less than the cash
they have on their hand Which is

1599
01:34:29,085 --> 01:34:32,883
really weird and this is because
they're massively cash, cash

1600
01:34:32,883 --> 01:34:36,930
burning entities who are selling
robotics and AV and whatnot, but

1601
01:34:36,930 --> 01:34:40,417
they really don't have like a
way to generate cash flow.

1602
01:34:40,417 --> 01:34:44,091
because these things are so,
yeah, exactly. And, and that's

1603
01:34:44,091 --> 01:34:47,702
why Tesla is what it is. And
like, you know. Any other RAS

1604
01:34:47,702 --> 01:34:51,501
company is what it is, at its
core. And so really there needs

1605
01:34:51,501 --> 01:34:55,423
to be like a rethinking of how
you think about, like, what kind

1606
01:34:55,423 --> 01:34:59,222
of business model robotics will
need. And for people to start

1607
01:34:59,222 --> 01:35:02,895
talking, you know, with each
other to sort of. evolve that.

1608
01:35:02,895 --> 01:35:06,694
Because left to itself, like the
pattern matching was sort of

1609
01:35:06,694 --> 01:35:10,056
insured. Yeah, like, and it's
just going to be pattern

1610
01:35:10,056 --> 01:35:13,979
matching all the way down. And
it's really not, I mean, really,

1611
01:35:13,979 --> 01:35:17,902
yeah, like, it's not gonna lead
to anything exciting. You know,

1612
01:35:17,902 --> 01:35:21,638
you're not gonna change the
world with like, you know, these

1613
01:35:21,638 --> 01:35:25,000
same repeated ways of thinking.
It's not gonna happen.

1614
01:35:25,000 --> 01:35:29,575
Audrow Nash: Yeah, for sure. How
do you, so do you have any,

1615
01:35:29,575 --> 01:35:34,150
like, great Investment thinkers
or anyone's to recommend for

1616
01:35:34,150 --> 01:35:38,497
this kind of thing? Like, I
mean, I'm in technology and I

1617
01:35:38,497 --> 01:35:42,996
know very little about private
equity. I talked to a lot of

1618
01:35:42,996 --> 01:35:47,876
people through podcasting, but,
like who, who to learn more, who

1619
01:35:47,876 --> 01:35:52,527
to become, or even like, I mean,
you guys are, in my opinion,

1620
01:35:52,527 --> 01:35:57,179
from this interview, a great
example of kind of merging these

1621
01:35:57,179 --> 01:36:00,000
two, but. Who else can we learn
from?

1622
01:36:00,000 --> 01:36:02,767
Nag Murthy: It's, yeah. So I
think we should make a

1623
01:36:02,767 --> 01:36:05,967
distinction to say, you know,
private equity. I don't think

1624
01:36:05,967 --> 01:36:09,113
like you can have a private
equity company that merges the

1625
01:36:09,113 --> 01:36:12,368
technology company tomorrow.
Right? Like there are different

1626
01:36:12,368 --> 01:36:15,840
fund structures and whatnot, so
there's logistical challenges to

1627
01:36:15,840 --> 01:36:19,203
do that. But, but I think there
are examples out there like of

1628
01:36:19,203 --> 01:36:22,186
companies that are building
these vertically integrated

1629
01:36:22,186 --> 01:36:25,224
solutions like an Amazon or a
Tesla. And the way they've

1630
01:36:25,224 --> 01:36:28,370
approached building their
business is what I think like we

1631
01:36:28,370 --> 01:36:31,462
can learn to emulate. By
focusing on cash flow, right? By

1632
01:36:31,462 --> 01:36:34,771
focusing on sort of, like, by
asking yourself, like, what are

1633
01:36:34,771 --> 01:36:38,026
you getting paid for? And like,
where is the margin accruing

1634
01:36:38,026 --> 01:36:41,335
from? And then what are the jobs
to be done? And what can you

1635
01:36:41,335 --> 01:36:44,427
truly, do you truly need to
automate? And what can, what,

1636
01:36:44,427 --> 01:36:46,000
what can you automate, right?

1637
01:36:46,000 --> 01:36:49,758
Audrow Nash: Yeah. Okay. I
really like that. I mean, it's

1638
01:36:49,758 --> 01:36:52,000
just simple pragmatism in a
sense.

1639
01:36:52,000 --> 01:36:55,233
Nag Murthy: Cashflow. Exactly. I
mean, follow the cashflow,

1640
01:36:55,233 --> 01:36:58,630
right? Like follow the cashflow,
but also like you have to aim

1641
01:36:58,630 --> 01:37:02,137
for the moon here because again,
classical methods are not going

1642
01:37:02,137 --> 01:37:05,534
to solve it for you, but find a
way to like link the two, in a

1643
01:37:05,534 --> 01:37:08,438
way that works for your
industry. It may not work for

1644
01:37:08,438 --> 01:37:11,726
all industries, right? I think
it works for landscaping. You

1645
01:37:11,726 --> 01:37:15,178
know, we're growing, we'll find
out, right, as we grow further.

1646
01:37:15,178 --> 01:37:18,247
we can certainly tell you how
not to do things, and then

1647
01:37:18,247 --> 01:37:21,644
hopefully, like, in another year
or so, we can tell you, like,

1648
01:37:21,644 --> 01:37:24,000
you know, how to do things
really well, so.

1649
01:37:24,000 --> 01:37:28,624
Audrow Nash: Yeah, for sure.
It's iteration. And, Mike, how

1650
01:37:28,624 --> 01:37:33,641
do we keep up with all this LLM
stuff, all of the world language

1651
01:37:33,641 --> 01:37:38,657
model, all the AI, how, how, how
does one, I don't know, operate

1652
01:37:38,657 --> 01:37:43,046
in this or skill up or? Yeah.
Yeah. Just do well in this

1653
01:37:43,046 --> 01:37:47,278
coming world with, if machine
learning is going to eat

1654
01:37:47,278 --> 01:37:52,373
everything or everybody's lunch.
How do you, how do you be a part

1655
01:37:52,373 --> 01:37:53,000
of that?

1656
01:37:53,000 --> 01:37:56,573
Michael Laskey: Yeah, I mean
that's, keeping up with it is

1657
01:37:56,573 --> 01:38:00,270
definitely a chore that you have
to like, you know, by then,

1658
01:38:00,270 --> 01:38:03,844
like, your week is reading
papers. it's a very fast moving

1659
01:38:03,844 --> 01:38:07,848
field. I, you know, and the pure
software side, I, I'm definitely

1660
01:38:07,848 --> 01:38:11,422
not an expert in that. When I
think about robotics though,

1661
01:38:11,422 --> 01:38:14,995
like, how do you, how do you do
ML in robotics? How do you

1662
01:38:14,995 --> 01:38:18,630
actually, like, Build a career
and like actually ship these

1663
01:38:18,630 --> 01:38:22,265
products. The one lesson I've
learned, and this is actually

1664
01:38:22,265 --> 01:38:25,777
why Electrosheep is really
appealing, is like it's really

1665
01:38:25,777 --> 01:38:28,550
easy to be overestimate in
timelines, or like

1666
01:38:28,550 --> 01:38:32,308
overconfident, and saying like,
yeah, tomorrow the robot will

1667
01:38:32,308 --> 01:38:36,005
work, or like the next day the
robot's gonna work. These are

1668
01:38:36,005 --> 01:38:39,455
very subtle problems that
require a lot of iteration and

1669
01:38:39,455 --> 01:38:43,398
feedback. So you really do want
to think about like, what is the

1670
01:38:43,398 --> 01:38:46,972
cash flow of your business?
What's the runway? And what is

1671
01:38:46,972 --> 01:38:50,915
the realistic timeline for these
ML products to actually deliver

1672
01:38:50,915 --> 01:38:54,550
value? And then plan businesses
in a way that gives you the

1673
01:38:54,550 --> 01:38:58,000
R&amp; D runway to actually like
scale and ship product.

1674
01:38:58,000 --> 01:39:02,664
Audrow Nash: Yeah, I really like
that. So keeping the R&amp; D

1675
01:39:02,664 --> 01:39:07,178
runway, because that's how you
are differentiating yourself.

1676
01:39:07,178 --> 01:39:11,917
But you need to keep cash flow
positive so you don't have to do

1677
01:39:11,917 --> 01:39:16,657
things on bad terms or just get
bought for way less than you're

1678
01:39:16,657 --> 01:39:21,171
worth by some company because
you're in a bad moment and you

1679
01:39:21,171 --> 01:39:25,835
ran out of runway or all these
things. What, Mike, do you have

1680
01:39:25,835 --> 01:39:30,499
any other, like, I don't know,
like, advice for learning about

1681
01:39:30,499 --> 01:39:35,389
this or applying, learning about
AI and ML or, and applying it or

1682
01:39:35,389 --> 01:39:39,000
how do you get really good
technically, I guess?

1683
01:39:39,000 --> 01:39:42,570
Michael Laskey: I think for
robotics, the number one thing

1684
01:39:42,570 --> 01:39:46,264
that's really guided me is,
like, put it on a physical robot

1685
01:39:46,264 --> 01:39:49,834
and actually put it out there
and see the performance. The

1686
01:39:49,834 --> 01:39:53,590
benchmarks, they matter, but
what really matters is how's the

1687
01:39:53,590 --> 01:39:56,975
robot's doing. because that's
going to be your ultimate

1688
01:39:56,975 --> 01:40:00,854
benchmark, like does it really
work on a robot? Can you see the

1689
01:40:00,854 --> 01:40:04,486
physical embodiment of the AI?
that is your ultimate, like,

1690
01:40:04,486 --> 01:40:08,056
North Star when it comes to
shipping robotic products that

1691
01:40:08,056 --> 01:40:11,627
actually work. really easy to
get lost in like the Twitter

1692
01:40:11,627 --> 01:40:15,259
esque sphere of like what's hot,
but when you really see it

1693
01:40:15,259 --> 01:40:19,014
fundamentally on a robot and
delivering value, I think that's

1694
01:40:19,014 --> 01:40:22,769
when you know you've actually
like done ML research in a cool

1695
01:40:22,769 --> 01:40:24,000
way that's exciting.

1696
01:40:24,000 --> 01:40:29,527
Audrow Nash: Mm hmm. Yeah, when
you've actually provided value

1697
01:40:29,527 --> 01:40:34,519
using these methods, it's
awesome. Okay, well let's see,

1698
01:40:34,519 --> 01:40:40,136
we've gone very long. It's been
a blast talking to you guys. Do

1699
01:40:40,136 --> 01:40:45,484
you have any, links or contact
info you'd like me to include

1700
01:40:45,484 --> 01:40:47,000
with the episode?

1701
01:40:47,000 --> 01:40:50,577
Nag Murthy: it's just, we have a
new website, right? So, shiny

1702
01:40:50,577 --> 01:40:53,000
new website, like you should
include that.

1703
01:40:53,000 --> 01:40:56,371
Audrow Nash: Yes, for sure. Of
course, the shiny new website

1704
01:40:56,371 --> 01:40:58,000
will be included. It's great.

1705
01:40:58,000 --> 01:41:01,930
Michael Laskey: Yeah. Much more
information. We're doing a lot

1706
01:41:01,930 --> 01:41:05,732
of PR soon, so there'll be more
stuff, but for now, just the

1707
01:41:05,732 --> 01:41:07,000
shiny website. Yeah.

1708
01:41:07,000 --> 01:41:10,996
Nag Murthy: And then, my email
is nag at electricsheep. company

1709
01:41:10,996 --> 01:41:15,055
and Mike's is michael. lasky at
electricsheep. company, so. Both

1710
01:41:15,055 --> 01:41:18,670
of those, you know, we can
provide you and, yeah, so that

1711
01:41:18,670 --> 01:41:22,349
people can sort of get us, get
in touch with us there. and

1712
01:41:22,349 --> 01:41:26,091
then, yeah, we're based in San
Jose, you know, and, we keep

1713
01:41:26,091 --> 01:41:29,896
like doing demos all over the
Bay Area. So, anyone who wants

1714
01:41:29,896 --> 01:41:33,829
to come and hang with us. We'd
love to sort of, or when you're

1715
01:41:33,829 --> 01:41:37,000
in the Bay Area, you should come
and hang with us.

1716
01:41:37,000 --> 01:41:42,676
Michael Laskey: Yeah. You can
literally just pick a park and

1717
01:41:42,676 --> 01:41:44,000
we'll show up.

1718
01:41:44,000 --> 01:41:47,030
Audrow Nash: Yeah. I'll reach
out, once a quarter I'm there.

1719
01:41:47,030 --> 01:41:49,000
So I will reach out and we'll
hang out.

1720
01:41:49,000 --> 01:41:51,000
Nag Murthy: Yeah, absolutely.

1721
01:41:51,000 --> 01:41:53,813
Michael Laskey: Where are you
at? Where are you living right

1722
01:41:53,813 --> 01:41:54,000
now?

1723
01:41:54,000 --> 01:41:58,779
Audrow Nash: San Antonio, Texas.
I was in San Francisco. We moved

1724
01:41:58,779 --> 01:42:03,338
like a year and last, last June,
we moved here. So I like it a

1725
01:42:03,338 --> 01:42:07,603
lot. We don't have family or
anything. We thought it was a

1726
01:42:07,603 --> 01:42:09,000
nice place to live.

1727
01:42:09,000 --> 01:42:12,467
Michael Laskey: Oh, cool. Do you
have a massive like house? Like,

1728
01:42:12,467 --> 01:42:13,000
is it like

1729
01:42:13,000 --> 01:42:17,929
Audrow Nash: relatively, we were
in a, I was in a 400 square foot

1730
01:42:17,929 --> 01:42:22,631
apartment in San Francisco. And,
my mortgage is the same for a

1731
01:42:22,631 --> 01:42:27,030
four bedroom house that has a
yard and stuff. And we're in

1732
01:42:27,030 --> 01:42:31,580
like the best part of town or
one of the best parts of town.

1733
01:42:31,580 --> 01:42:35,903
It's bonkers. so I don't know. I
like it here. And I love

1734
01:42:35,903 --> 01:42:40,302
barbecue. I, my wife and I, we
went and looked at Austin a

1735
01:42:40,302 --> 01:42:44,549
while ago and, we had the
barbecue. And we were thinking

1736
01:42:44,549 --> 01:42:49,326
about where to live and, at some
point in the meal, I was like,

1737
01:42:49,326 --> 01:42:54,028
I'm ready to beg. I will. Like,
I want to live here. I want to

1738
01:42:54,028 --> 01:42:56,000
be close to this barbecue.

1739
01:42:56,000 --> 01:42:59,424
Nag Murthy: You know what's
interesting? Like, Roon put out,

1740
01:42:59,424 --> 01:43:02,963
like, the guy on Twitter, like,
who, you know, who's known for

1741
01:43:02,963 --> 01:43:06,502
sort of these takes on AI. He
put out a tweet like a couple of

1742
01:43:06,502 --> 01:43:09,869
days ago where he said, you
know, the future is going to be

1743
01:43:09,869 --> 01:43:13,065
sort of these immaculate sort of
living spaces. I'm just

1744
01:43:13,065 --> 01:43:16,604
paraphrasing. It's not the exact
tweet, right? Where it's like

1745
01:43:16,604 --> 01:43:20,143
all these built like gorgeous
outdoor and indoor living spaces

1746
01:43:20,143 --> 01:43:22,996
for us, right? There are
fractions of the cost and

1747
01:43:22,996 --> 01:43:26,421
anyone, right? Whether or not
they work in tech like is able

1748
01:43:26,421 --> 01:43:29,788
to afford it. I was like, yeah,
that is such a cool vision,

1749
01:43:29,788 --> 01:43:33,384
right? Because Like that's the
promise of robotics and that's a

1750
01:43:33,384 --> 01:43:37,037
promise of like these AI agents
that we're building, right? Just

1751
01:43:37,037 --> 01:43:40,233
massive improvements in
standards of living for everyone

1752
01:43:40,233 --> 01:43:43,600
at the end of the day Which
should be insanely cool to make

1753
01:43:43,600 --> 01:43:44,000
happen.

1754
01:43:44,000 --> 01:43:47,916
Audrow Nash: I mean even also
just like easing labor shortages

1755
01:43:47,916 --> 01:43:50,000
and stuff in the short term
Yeah,

1756
01:43:50,000 --> 01:43:54,013
Michael Laskey: which also means
like things become cheaper and

1757
01:43:54,013 --> 01:43:58,025
the abundance of labor. Yeah, it
just it's a whole new equation

1758
01:43:58,025 --> 01:44:00,000
When you have infinite labor or

1759
01:44:00,000 --> 01:44:04,876
Audrow Nash: I was listening to
this thing on chat GPT or it was

1760
01:44:04,876 --> 01:44:09,219
the Sam Altman interview with
Joe Rogan and he he was Sam

1761
01:44:09,219 --> 01:44:13,105
Altman was saying how if Like
their goal is to make

1762
01:44:13,105 --> 01:44:16,000
intelligence free Which is quite
cool.

1763
01:44:16,000 --> 01:44:18,980
Nag Murthy: So then it's like
all the knowledge work in a

1764
01:44:18,980 --> 01:44:22,065
sense, becomes free, which is
really interesting and robots

1765
01:44:22,065 --> 01:44:24,000
would do that for the physical
world,

1766
01:44:24,000 --> 01:44:28,143
Audrow Nash: which would be
quite cool. Yeah. Well, we got

1767
01:44:28,143 --> 01:44:30,000
to get towards our utopia.

1768
01:44:30,000 --> 01:44:33,093
Nag Murthy: Yeah, we got to get
you like permanent barbecue.

1769
01:44:33,093 --> 01:44:35,000
Like, you know, like that's the
goal.

1770
01:44:35,000 --> 01:44:37,850
Audrow Nash: And if you guys
come visit San Antonio, it's

1771
01:44:37,850 --> 01:44:40,850
lovely here. The barbecue is
wonderful and there's plenty of

1772
01:44:40,850 --> 01:44:42,000
grass to mow. Do. Sure.

1773
01:44:42,000 --> 01:44:45,258
Michael Laskey: Yeah. I think
it's, it's in the options of

1774
01:44:45,258 --> 01:44:47,000
places we'll be going to, yeah.

1775
01:44:47,000 --> 01:44:50,667
Nag Murthy: Yeah. Hell yeah. We
might be there sooner than you

1776
01:44:50,667 --> 01:44:54,392
think, like, we're looking at a
few, yeah, acquisitions in that

1777
01:44:54,392 --> 01:44:58,000
area, so like, we'll be there
sooner than you think. Awesome.

1778
01:44:58,000 --> 01:45:01,293
Audrow Nash: Yeah. Keep me
posted. Okay. Hell yeah. So

1779
01:45:01,293 --> 01:45:04,524
we'll wrap this up. great
talking to you guys and I'm

1780
01:45:04,524 --> 01:45:08,000
looking forward to following and
seeing what you do more.

1781
01:45:08,000 --> 01:45:11,441
Nag Murthy: Likewise. Thank you
so much for having us and it was

1782
01:45:11,441 --> 01:45:13,000
a pleasure chatting with you.

1783
01:45:13,000 --> 01:45:16,988
Audrow Nash: Hell yeah. All
right. Bye everyone. That's it!

1784
01:45:16,988 --> 01:45:21,044
If you made it this far, you're
either asleep or enjoyed the

1785
01:45:21,044 --> 01:45:25,033
interview. If you enjoyed it,
consider subscribing. I'm not

1786
01:45:25,033 --> 01:45:29,359
planning on keeping any kind of
regular schedule, so subscribing

1787
01:45:29,359 --> 01:45:33,550
or following me on X or LinkedIn
is the best way to hear about

1788
01:45:33,550 --> 01:45:37,741
new episodes. If you're asleep,
I hope it was restful and that

1789
01:45:37,741 --> 01:45:42,000
you wake up and build something.
Until next time, bye everyone!
