1
00:00:00,000 --> 00:00:04,320
Stefan Seltz-Axmacher: this is
a, super key problem and is a

2
00:00:04,320 --> 00:00:08,640
big part in our belief of why
the robotics industry sucks so

3
00:00:08,640 --> 00:00:09,000
hard.

4
00:00:09,000 --> 00:00:12,530
Audrow Nash: We all know
robotics is hard. One reason

5
00:00:12,530 --> 00:00:16,727
it's hard is because it's hard
to find off the shelf components

6
00:00:16,727 --> 00:00:20,657
you can build a business on.
Because of this, most robotics

7
00:00:20,657 --> 00:00:24,720
companies end up reinventing the
wheel. For example, building

8
00:00:24,720 --> 00:00:28,783
their own mobility stack. And
this is a bummer because it's a

9
00:00:28,783 --> 00:00:32,647
huge barrier to entry for
robotics startups. Specifically,

10
00:00:32,647 --> 00:00:36,643
you may have to raise millions
of dollars and work for a few

11
00:00:36,643 --> 00:00:40,507
years to solve undifferentiating
problems. Like Autonomous

12
00:00:40,507 --> 00:00:44,503
Navigation. Before you can even
start on your startup's core

13
00:00:44,503 --> 00:00:48,167
problem. This is why what
Polymath Robotics is doing is

14
00:00:48,167 --> 00:00:51,830
exciting. They're creating a
platform that solves robot

15
00:00:51,830 --> 00:00:56,093
navigation for a large number of
industrial use cases so you can

16
00:00:56,093 --> 00:01:00,356
just build on top of it and not
reinvent the wheel. They're also

17
00:01:00,356 --> 00:01:04,020
doing a better job than you
probably would because it's

18
00:01:04,020 --> 00:01:07,683
their core problem and you might
cut corners because of

19
00:01:07,683 --> 00:01:11,680
deadlines. I really like where
this is going and expect more

20
00:01:11,680 --> 00:01:15,677
companies to follow Polymath
Robotics lead and make services

21
00:01:15,677 --> 00:01:19,873
for hard but undifferentiating
robotics problems so you can use

22
00:01:19,873 --> 00:01:23,670
them off the shelf. I think this
is a great thing for the

23
00:01:23,670 --> 00:01:27,733
robotics community as it makes
robotics startups a little bit

24
00:01:27,733 --> 00:01:31,597
or even a lot easier. And that
makes it so that we're more

25
00:01:31,597 --> 00:01:35,726
likely to see robots doing dull,
dirty, and dangerous tasks no

26
00:01:35,726 --> 00:01:40,056
one wants to do sooner. This was
a tremendously fun interview for

27
00:01:40,056 --> 00:01:44,119
me. I think you're going to like
it too, especially if you're

28
00:01:44,119 --> 00:01:47,916
interested in or curious about
robotics opportunities for

29
00:01:47,916 --> 00:01:51,846
mobile robots in industrial
spaces, if you're curious about

30
00:01:51,846 --> 00:01:55,643
new and potentially powerful
business models for robotics

31
00:01:55,643 --> 00:01:59,306
companies, or if you're a
hardcore nerd and love safety

32
00:01:59,306 --> 00:02:03,503
certifications and requirements
testing. I know some of you are

33
00:02:03,503 --> 00:02:07,100
out there. You're gonna love it.
Check out the episode

34
00:02:07,100 --> 00:02:11,163
description for timestamps if
you cannot wait. Lastly, if you

35
00:02:11,163 --> 00:02:15,426
enjoy this conversation, want to
hear more from Stefan and Ilia,

36
00:02:15,426 --> 00:02:19,289
they have a podcast called
Automate It, which is available

37
00:02:19,289 --> 00:02:23,552
on all the podcasting platforms
you'd expect. It's a lot of fun,

38
00:02:23,552 --> 00:02:27,083
and I recommend you check it
out. After you watch the

39
00:02:27,083 --> 00:02:31,079
interview, I'd love to hear what
you think. Do you think the

40
00:02:31,079 --> 00:02:35,276
approach Polymath is taking is
good for the robotics community?

41
00:02:35,276 --> 00:02:38,873
Do you imagine more robotics
companies doing something

42
00:02:38,873 --> 00:02:43,003
similar? I'd love to hear what
you think in the comments or on

43
00:02:43,003 --> 00:02:46,000
X. Without further ado, here's
our interview.

44
00:02:46,000 --> 00:02:50,274
Audrow Nash: Hi everyone.
Stefan, would you introduce

45
00:02:50,274 --> 00:02:51,000
yourself?

46
00:02:51,000 --> 00:02:53,631
Stefan Seltz-Axmacher: Yeah, of
course. My name is Stefan

47
00:02:53,631 --> 00:02:56,538
Seltz-Axmacher and I'm CEO and
Co-founder of Polymath Robotics.

48
00:02:56,538 --> 00:02:59,262
At Polymath, we're making it
really easy to add autonomy to

49
00:02:59,262 --> 00:03:01,662
any off highway vehicle. And
before this I worked on

50
00:03:01,662 --> 00:03:04,292
self-driving trucks at, at
Starsky where we put the first

51
00:03:04,292 --> 00:03:06,000
driver out truck on a public
highway.

52
00:03:06,000 --> 00:03:09,000
Audrow Nash: And Ilia, would you
introduce yourself?

53
00:03:09,000 --> 00:03:12,203
Ilia Baranov: Hi, I'm Ilia
Baranov, co-founder and CTO here

54
00:03:12,203 --> 00:03:15,406
at Polymath. before this grand
adventure with Stefan, I was

55
00:03:15,406 --> 00:03:18,880
working at Amazon on their home
robot astro and have lots of fun

56
00:03:18,880 --> 00:03:22,029
stories to talk about there. and
then prior to that was at

57
00:03:22,029 --> 00:03:25,394
Clearpath Robotics was one of
the early folks there developing

58
00:03:25,394 --> 00:03:28,000
a bunch of the research and
industrial autonomy.

59
00:03:28,000 --> 00:03:32,000
Audrow Nash: Now Stefan, would
you tell me about Polymath?

60
00:03:32,000 --> 00:03:34,777
Stefan Seltz-Axmacher: At
Polymath, we're building a

61
00:03:34,777 --> 00:03:38,088
modular stack where we're doing
everything from sensor drivers

62
00:03:38,088 --> 00:03:41,186
all the way through throttle and
steering commands for the

63
00:03:41,186 --> 00:03:44,444
software that runs onboard a
vehicle to make it move from one

64
00:03:44,444 --> 00:03:47,541
point in space to another. So
we're not building hardware,

65
00:03:47,541 --> 00:03:50,906
we're not building apps to tell
robots what to, what to do, but

66
00:03:50,906 --> 00:03:53,630
the really, difficult but
undifferentiated software

67
00:03:53,630 --> 00:03:57,048
packages that make a robot move
through space, we've productized

68
00:03:57,048 --> 00:04:00,359
that and we're, we, have it on
more robots than we have people

69
00:04:00,359 --> 00:04:01,000
on our team.

70
00:04:01,000 --> 00:04:04,421
Audrow Nash: That's so cool. How
many people on your team and how

71
00:04:04,421 --> 00:04:06,000
many robots do you have it on?

72
00:04:06,000 --> 00:04:10,178
Stefan Seltz-Axmacher: 10, and
we're, I think 11 robots going

73
00:04:10,178 --> 00:04:11,000
on 13 or 14.

74
00:04:11,000 --> 00:04:13,000
Audrow Nash: That's exciting.
Hell yeah.

75
00:04:13,000 --> 00:04:15,903
Stefan Seltz-Axmacher: Yeah. And
they're all different types of

76
00:04:15,903 --> 00:04:18,714
robots too. It's not the same
thing. 10 times we're on things

77
00:04:18,714 --> 00:04:21,571
as varied as like bulldozers and
tractors and articulated dump

78
00:04:21,571 --> 00:04:23,000
trucks and, all sorts of stuff.

79
00:04:23,000 --> 00:04:27,710
Audrow Nash: That's so cool. And
how are, so you mentioned you're

80
00:04:27,710 --> 00:04:32,348
not doing hardware. How are you
interfacing with these different

81
00:04:32,348 --> 00:04:33,000
vehicles?

82
00:04:33,000 --> 00:04:36,340
Ilia Baranov: Yeah. So one of
the things we had talked about

83
00:04:36,340 --> 00:04:39,735
with Stefan early, on is that we
had thought kind of Robotics

84
00:04:39,735 --> 00:04:43,186
keeps reinventing the wheel. And
ROS, even ROS's conception or

85
00:04:43,186 --> 00:04:46,692
the robot operating system, was
that we wanna reduce the amount

86
00:04:46,692 --> 00:04:49,475
of reinventing the wheel by
having a common set of

87
00:04:49,475 --> 00:04:52,703
communication packages and
examples and the community. and

88
00:04:52,703 --> 00:04:55,987
that helps a lot. But we had
seen that a lot of people will

89
00:04:55,987 --> 00:04:59,327
build this kind of vertical
stack for their particular robot

90
00:04:59,327 --> 00:05:02,777
or particular application. and
we just decided that we want to

91
00:05:02,777 --> 00:05:06,006
abstract that away. So we have
apps, we call it a hardware

92
00:05:06,006 --> 00:05:09,234
abstraction layer, pretty
standard term for input sensors.

93
00:05:09,234 --> 00:05:12,351
And then we have the same
hardware abstraction layer for

94
00:05:12,351 --> 00:05:15,412
the output to the vehicle. And
that outputs the vehicle

95
00:05:15,412 --> 00:05:18,918
hardware Abstraction layer takes
care of things like, what is a

96
00:05:18,918 --> 00:05:22,091
kinematic model? Turn radius
maximum speed. What does the

97
00:05:22,091 --> 00:05:25,430
footprint look like? and so the
core piece between those two

98
00:05:25,430 --> 00:05:28,603
hardware abstraction layers
doesn't really change because

99
00:05:28,603 --> 00:05:31,887
all the complexities that's
unique per vehicle are captured

100
00:05:31,887 --> 00:05:33,000
by those two layers.

101
00:05:33,000 --> 00:05:35,976
Audrow Nash: Very cool. So you
have a good way of talking to,

102
00:05:35,976 --> 00:05:37,000
oh, go ahead, Stefan.

103
00:05:37,000 --> 00:05:40,161
Stefan Seltz-Axmacher: yeah,
because the problem that we're

104
00:05:40,161 --> 00:05:43,589
solving is if you were to leave
your day job to go start, say an

105
00:05:43,589 --> 00:05:46,857
autonomous tractor, the vast
majority of the code you'd write

106
00:05:46,857 --> 00:05:49,696
in the first two to four years
would be fundamentally

107
00:05:49,696 --> 00:05:52,857
undifferentiated from any of the
other 150 or so autonomous

108
00:05:52,857 --> 00:05:56,179
tractor companies. You'd just be
like working really hard on a

109
00:05:56,179 --> 00:05:59,500
really hard set of problems that
no one cares about for you if

110
00:05:59,500 --> 00:06:01,000
you solve them successfully.

111
00:06:01,000 --> 00:06:02,000
Audrow Nash: Yeah,

112
00:06:02,000 --> 00:06:05,131
Stefan Seltz-Axmacher: What
we're doing is we're giving you

113
00:06:05,131 --> 00:06:08,316
that as a product so you can
focus on integrating it into an

114
00:06:08,316 --> 00:06:11,288
actual product that's some
farmer somewhere or some mine

115
00:06:11,288 --> 00:06:14,525
operator somewhere, or someone,
whoever caress about. And you

116
00:06:14,525 --> 00:06:17,285
can focus on actually just
making robots valuable as

117
00:06:17,285 --> 00:06:20,416
opposed to being the 1500th team
to reinvent point-to-point

118
00:06:20,416 --> 00:06:21,000
navigation.

119
00:06:21,000 --> 00:06:25,967
Audrow Nash: So are you thinking
of it like, you have this

120
00:06:25,967 --> 00:06:31,191
software that handles mobility
for robots and you are selling

121
00:06:31,191 --> 00:06:36,671
it to businesses that make like
tractors or big vehicles. And so

122
00:06:36,671 --> 00:06:42,066
then they outfit their vehicles
with your software, and they're

123
00:06:42,066 --> 00:06:47,033
selling that complete package.
So you are selling them the

124
00:06:47,033 --> 00:06:52,000
autonomy that they use as a
service on their hardware in a

125
00:06:52,000 --> 00:06:54,664
Stefan Seltz-Axmacher: So we
might sell it to an OEM like

126
00:06:54,664 --> 00:06:57,374
you're describing. We might sell
it to some big industrial

127
00:06:57,374 --> 00:06:59,617
company who's really
sophisticated and can write

128
00:06:59,617 --> 00:07:02,000
their own software to tell
robots what to do. Yeah.

129
00:07:02,000 --> 00:07:07,227
Audrow Nash: I love it. So how
does it, I, guess Ilia, how does

130
00:07:07,227 --> 00:07:12,205
it work in terms of are you
connecting over the can bus with

131
00:07:12,205 --> 00:07:17,017
the vehicles or how are you
actually interfacing with the,

132
00:07:17,017 --> 00:07:21,000
what do we call it, the vehicle,
whatever it is.

133
00:07:21,000 --> 00:07:23,927
Ilia Baranov: The, vehicle, the
robot. we have this discussion

134
00:07:23,927 --> 00:07:26,901
back and forth is it a vehicle,
is it a robot, is it machinery?

135
00:07:26,901 --> 00:07:29,734
Is it like industrial equipment?
that's a marketing problem.

136
00:07:29,734 --> 00:07:32,000
Luckily I don't have to deal
with that too much.

137
00:07:32,000 --> 00:07:33,000
Audrow Nash: Yeah.

138
00:07:33,000 --> 00:07:36,664
Ilia Baranov: but, yeah, so,
yeah, as part of the interface

139
00:07:36,664 --> 00:07:40,576
layer, one interface we have is
socket CAN, which can then talk

140
00:07:40,576 --> 00:07:44,364
to a whole bunch of CAN based
equipment. but the most extreme

141
00:07:44,364 --> 00:07:48,151
case we have is actually one of
our hardware partners, called

142
00:07:48,151 --> 00:07:51,877
Hardline in Canada. They're now
part of Hexagon. they have a

143
00:07:51,877 --> 00:07:54,858
1996 bulldozer, which is
entirely mechanical and

144
00:07:54,858 --> 00:07:58,397
hydraulic. And so the only
interface we get to it is open

145
00:07:58,397 --> 00:08:02,185
this valve or close this valve
that's the most control we get

146
00:08:02,185 --> 00:08:05,787
to it. and so in that case, our
hardware abstraction layer

147
00:08:05,787 --> 00:08:09,450
abstracting away all the valve
controls up to a level where

148
00:08:09,450 --> 00:08:13,486
it's okay, we do positional loop
closure using all the sensors we

149
00:08:13,486 --> 00:08:17,212
have, and then we control this
thing like a, like an unknown

150
00:08:17,212 --> 00:08:20,814
plant that we throw valve
controls at and get speed out of

151
00:08:20,814 --> 00:08:21,000
it.

152
00:08:21,000 --> 00:08:24,556
Audrow Nash: So you are doing, I
thought you were not doing any

153
00:08:24,556 --> 00:08:27,774
hardware, but you are doing
hardware for custom solutions

154
00:08:27,774 --> 00:08:28,000
like

155
00:08:28,000 --> 00:08:29,000
Stefan Seltz-Axmacher: So we
don't, so

156
00:08:29,000 --> 00:08:30,000
Ilia Baranov: even in

157
00:08:30,000 --> 00:08:31,000
Audrow Nash: Oh, they're doing
it.

158
00:08:31,000 --> 00:08:34,177
Stefan Seltz-Axmacher: Yeah.
Yeah. so like an interesting

159
00:08:34,177 --> 00:08:37,689
thing and I like, and, I think
everyone who's done robotics has

160
00:08:37,689 --> 00:08:40,866
seen when you try to build the
vehicle or when you try to

161
00:08:40,866 --> 00:08:44,489
retrofit the vehicle, there's an
incredible amount of really hard

162
00:08:44,489 --> 00:08:48,056
custom one-off work to retrofit
this 1997 bulldozer or this 2023

163
00:08:48,056 --> 00:08:51,400
tractor or this whatever. And
that work is really hard to do

164
00:08:51,400 --> 00:08:54,856
when you're a five person team.
And the same people doing that

165
00:08:54,856 --> 00:08:57,754
mechatronic design are also
writing the code or also

166
00:08:57,754 --> 00:09:01,098
building the ML algorithms or
also building the app are also

167
00:09:01,098 --> 00:09:04,387
figuring out how to operate
tractors as a service. So there

168
00:09:04,387 --> 00:09:07,508
are, however, a whole bunch of
teams all over the world,

169
00:09:07,508 --> 00:09:10,239
including hardline up in, in
Canada, who are just

170
00:09:10,239 --> 00:09:13,639
mechatronics consulting shops
where they can show up, look at

171
00:09:13,639 --> 00:09:17,262
a vehicle, they have a pre fitt,
retrofit kit. We'll customize it

172
00:09:17,262 --> 00:09:20,607
for that specific vehicle, and
then we can talk to their API

173
00:09:20,607 --> 00:09:22,000
and issue commands to it.

174
00:09:22,000 --> 00:09:23,000
Audrow Nash: that's super cool.

175
00:09:23,000 --> 00:09:25,818
Stefan Seltz-Axmacher: So we get
to just be a software company

176
00:09:25,818 --> 00:09:28,000
and not go be in cold places to
bring up robots.

177
00:09:28,000 --> 00:09:31,371
Ilia Baranov: so hopefully I'm,
not like, I'm not overstating

178
00:09:31,371 --> 00:09:34,521
our position here, but the,
equivalency I like to draw is

179
00:09:34,521 --> 00:09:37,616
that you can think of us as
Microsoft in some sense, and

180
00:09:37,616 --> 00:09:41,042
there's a whole bunch of laptop
manufacturers in the world and

181
00:09:41,042 --> 00:09:43,971
like an ASUS laptop and a,
whatever, a Dell laptop or

182
00:09:43,971 --> 00:09:47,011
fundamentally different beasts,
but they comply to some

183
00:09:47,011 --> 00:09:50,547
underlying hardware, abstraction
layer standard that then our OS

184
00:09:50,547 --> 00:09:53,587
can just sit on top of and
control. It's stretching the

185
00:09:53,587 --> 00:09:56,903
analogy a little bit too far,
but that's the direction we're

186
00:09:56,903 --> 00:09:59,887
going where there's these
retrofit and service and oEM

187
00:09:59,887 --> 00:10:02,982
companies all over the planet,
very specialized in their

188
00:10:02,982 --> 00:10:06,297
particular vehicles, and all of
them know how to treat their

189
00:10:06,297 --> 00:10:09,889
particular vehicle, and all they
have to do is give us some level

190
00:10:09,889 --> 00:10:10,000
of

191
00:10:10,000 --> 00:10:14,620
Audrow Nash: You just need the
interface. Yeah. And are you

192
00:10:14,620 --> 00:10:19,163
doing it all with onboard
sensors for these, I suppose you

193
00:10:19,163 --> 00:10:23,000
probably. So how would you, is
it all GPS driven?

194
00:10:23,000 --> 00:10:24,000
Ilia Baranov: can

195
00:10:24,000 --> 00:10:26,980
Audrow Nash: trying to imagine.
Yeah. So you are actually, your

196
00:10:26,980 --> 00:10:29,770
approach is quite modular,
depending on the needs of it, is

197
00:10:29,770 --> 00:10:31,000
what it sounds like. Okay.

198
00:10:31,000 --> 00:10:34,000
Stefan Seltz-Axmacher: because
sometimes GPS solves all your

199
00:10:34,000 --> 00:10:36,000
problems and sometimes it
doesn't exist.

200
00:10:36,000 --> 00:10:40,264
Audrow Nash: yeah. And then,
yeah, and sometimes it doesn't

201
00:10:40,264 --> 00:10:44,311
exist and sometimes you have
crazy reflections if you're

202
00:10:44,311 --> 00:10:48,430
using RTK and all sorts of
things. Okay. So you have this

203
00:10:48,430 --> 00:10:52,549
modular layer, you're the
software you can run on a whole

204
00:10:52,549 --> 00:10:56,523
bunch of different hardware. I
like that approach. what

205
00:10:56,523 --> 00:11:00,498
exactly, being that you're
modular, I suppose there's a

206
00:11:00,498 --> 00:11:04,762
bunch of things that you have
the ability to do, but, what,

207
00:11:04,762 --> 00:11:09,098
are the core competencies of
the, I dunno, your, what, do we

208
00:11:09,098 --> 00:11:13,000
call, your software? Like what,
how do we refer to it?

209
00:11:13,000 --> 00:11:15,441
Ilia Baranov: I, we've been, I
don't know, again, if this is a

210
00:11:15,441 --> 00:11:17,921
marketing problem, we just call
it the autonomy core, that kind

211
00:11:17,921 --> 00:11:18,000
of

212
00:11:18,000 --> 00:11:20,000
Audrow Nash: The autonomy core.
I'll just call it

213
00:11:20,000 --> 00:11:23,239
Ilia Baranov: the core. Then the
layers around it or the core.

214
00:11:23,239 --> 00:11:26,425
And Yeah. maybe that's a good
movie too. Anyway, there's, the

215
00:11:26,425 --> 00:11:29,455
hardware abstraction layers
around it, but the core is the

216
00:11:29,455 --> 00:11:32,537
piece that we really put a lot
of effort into and we try to

217
00:11:32,537 --> 00:11:34,000
change as little as possible

218
00:11:34,000 --> 00:11:35,000
Audrow Nash: So what

219
00:11:35,000 --> 00:11:37,000
Stefan Seltz-Axmacher: think of
that as everything.

220
00:11:37,000 --> 00:11:38,000
Audrow Nash: Oh, go ahead
Stefan.

221
00:11:38,000 --> 00:11:41,444
Stefan Seltz-Axmacher: yeah, so
think of, the functionality of

222
00:11:41,444 --> 00:11:44,556
that as everything from
literally running the drivers on

223
00:11:44,556 --> 00:11:47,833
whichever sensors we happen to
have to make sure that we're

224
00:11:47,833 --> 00:11:51,167
getting data from the LIDAR,
we're getting data from the RTK

225
00:11:51,167 --> 00:11:54,000
GPS, we're getting data from
the, radar or whatever

226
00:11:54,000 --> 00:11:57,556
combination of things we happen
to have. Turning those into some

227
00:11:57,556 --> 00:12:00,722
preset, things like point
clouds, turning that into, cost

228
00:12:00,722 --> 00:12:01,000
maps.

229
00:12:01,000 --> 00:12:02,000
Audrow Nash: on a lot of things.

230
00:12:02,000 --> 00:12:04,000
Stefan Seltz-Axmacher: And use
over and over

231
00:12:04,000 --> 00:12:05,000
Ilia Baranov: lot of the,

232
00:12:05,000 --> 00:12:06,000
Stefan Seltz-Axmacher: And then
do,

233
00:12:06,000 --> 00:12:08,634
Ilia Baranov: hardware
abstraction layer is basically

234
00:12:08,634 --> 00:12:11,615
turning into common data types
and having as few of those as

235
00:12:11,615 --> 00:12:14,000
possible that are as widely
defined as possible.

236
00:12:14,000 --> 00:12:16,639
Audrow Nash: Makes sense. That's
probably the best way to scale,

237
00:12:16,639 --> 00:12:18,000
is what I would imagine for
this.

238
00:12:18,000 --> 00:12:21,242
Stefan Seltz-Axmacher: basically
our secret evil plan. And

239
00:12:21,242 --> 00:12:24,764
honestly it's part of our sales
strategy as well, is we tend to

240
00:12:24,764 --> 00:12:28,286
just talk to our customers about
how we've built things. And it

241
00:12:28,286 --> 00:12:31,472
tends to be how they wish they
could have gotten to build

242
00:12:31,472 --> 00:12:34,938
things if they didn't have a six
month deadline when they were

243
00:12:34,938 --> 00:12:38,404
first building their whatever.
And 'cause the reality was when

244
00:12:38,404 --> 00:12:41,814
they, thought they were gonna
start whatever autonomy project

245
00:12:41,814 --> 00:12:45,056
they're currently on, they
thought, this time I'm gonna do

246
00:12:45,056 --> 00:12:48,019
it right. It's gonna be
reusable. and then, some jerk

247
00:12:48,019 --> 00:12:51,429
squad, CEO changes the deadline
or some customer says, I want

248
00:12:51,429 --> 00:12:54,559
some additional feature or
whatever. And the engineering

249
00:12:54,559 --> 00:12:57,689
team cranks out a bunch of
garbage spaghetti code in the

250
00:12:57,689 --> 00:13:01,043
last three weeks that then works
for the rest of the robot's

251
00:13:01,043 --> 00:13:04,509
life. everyone wants to build
robots this way. We've been free

252
00:13:04,509 --> 00:13:07,640
of the original sin of being
tied to a specific use case

253
00:13:07,640 --> 00:13:10,770
that's enabled us to build a
highly reusable, adaptable,

254
00:13:10,770 --> 00:13:12,000
modular autonomy code.

255
00:13:12,000 --> 00:13:15,969
Audrow Nash: Yeah, so that
sounds awesome to me. And you

256
00:13:15,969 --> 00:13:19,937
guys are using the robot
operating system pretty heavily

257
00:13:19,937 --> 00:13:24,189
with this, so I'm, the way that
I am thinking of this now is

258
00:13:24,189 --> 00:13:28,370
it's like it's an additional
abstraction on top of ROS that

259
00:13:28,370 --> 00:13:30,000
manages a lot of things

260
00:13:30,000 --> 00:13:31,000
Stefan Seltz-Axmacher: Yep.

261
00:13:31,000 --> 00:13:33,916
Audrow Nash: Okay. Is it, could
you almost say it's like a

262
00:13:33,916 --> 00:13:37,084
behavioral level where you say,
I wanted to do these things? Or

263
00:13:37,084 --> 00:13:40,000
how, or is it just unifying a
lot of things? Or maybe it's

264
00:13:40,000 --> 00:13:43,259
Ilia Baranov: behavior control
is a lot of what we do. So we do

265
00:13:43,259 --> 00:13:46,207
have a behavior tree and a
behavior engine, especially an

266
00:13:46,207 --> 00:13:49,517
API layer to command our vehicle
and how that all interacts. but

267
00:13:49,517 --> 00:13:52,362
yeah, I get that a lot of
roboticists, what people say.

268
00:13:52,362 --> 00:13:55,000
what you're describing is
basically ROS, which like

269
00:13:55,000 --> 00:13:59,937
Audrow Nash: well, higher level,
easier to input stuff and get it

270
00:13:59,937 --> 00:14:01,000
to do what you

271
00:14:01,000 --> 00:14:03,000
Ilia Baranov: and again. Yeah,
and I

272
00:14:03,000 --> 00:14:04,000
Audrow Nash: Probably does
resource

273
00:14:03,000 --> 00:14:04,000
Ilia Baranov: I give is

274
00:14:04,000 --> 00:14:06,000
Audrow Nash: all sorts of
things.

275
00:14:06,000 --> 00:14:09,200
Ilia Baranov: exactly, and, the
comparison I give is ROS 2 is a,

276
00:14:09,200 --> 00:14:12,200
is an excellent open source
toolbox and we're providing just

277
00:14:12,200 --> 00:14:15,250
the whole house. And yes, you
could use the toolbox and build

278
00:14:15,250 --> 00:14:18,300
the house, or you could just go
to us and we'll build you the

279
00:14:18,300 --> 00:14:21,000
house, or we have a house that
we can just copy paste.

280
00:14:21,000 --> 00:14:24,300
Audrow Nash: Do you guys know in
web development, there's like

281
00:14:24,300 --> 00:14:27,707
Ruby on Rails or something like
this, which streamlines a lot of

282
00:14:27,707 --> 00:14:30,954
things, would you guys, would
that be a fair comparison? Ruby

283
00:14:30,954 --> 00:14:34,148
on Rails, but for Robotics in
some sense? Or is it not quite

284
00:14:34,148 --> 00:14:35,000
broad enough and

285
00:14:35,000 --> 00:14:36,000
Ilia Baranov: I, would

286
00:14:36,000 --> 00:14:41,000
Audrow Nash: on Rails, but four
mobile robots I suppose,

287
00:14:41,000 --> 00:14:45,256
Ilia Baranov: for mobile robots,
I'd say it's, even one layer

288
00:14:45,256 --> 00:14:47,000
above that. I, would say,

289
00:14:47,000 --> 00:14:50,514
Stefan Seltz-Axmacher: So, our,
API, we have a relatively

290
00:14:50,514 --> 00:14:54,336
straightforward rest, API, where
like commands might be, go to

291
00:14:54,336 --> 00:14:57,911
this GPS coordinate with this
heading, and then we just do

292
00:14:57,911 --> 00:15:01,240
that. So like we have a
customer, for example, who's a

293
00:15:01,240 --> 00:15:05,000
big conglomerate. they have an
in-house built ERP, written by

294
00:15:05,000 --> 00:15:07,000
Audrow Nash: an ERP?

295
00:15:07,000 --> 00:15:11,047
Stefan Seltz-Axmacher: an
enterprise resource, program. So

296
00:15:11,047 --> 00:15:14,465
basically it's a software
program that runs their

297
00:15:14,465 --> 00:15:18,512
operation, their multi-billion
dollar operation. it's been

298
00:15:18,512 --> 00:15:22,419
written by this guy, like
literally a guy, over the last

299
00:15:22,419 --> 00:15:25,000
20 years as I think a monolith.
In C#

300
00:15:25,000 --> 00:15:27,000
Audrow Nash: That's horrible.

301
00:15:27,000 --> 00:15:30,626
Stefan Seltz-Axmacher: can issue
commands to our robots. doesn't

302
00:15:30,626 --> 00:15:34,140
know much about geospatial data,
he knows nothing about safety

303
00:15:34,140 --> 00:15:37,539
engineering. He knows nothing
about machine learning. He can

304
00:15:37,539 --> 00:15:41,109
issue commands to robots that we
drive in their fleet. so that,

305
00:15:41,109 --> 00:15:44,679
for example, the first time he,
we were doing a, an integration

306
00:15:44,679 --> 00:15:47,909
test, he was like, huh, I'm
sending messages and you guys

307
00:15:47,909 --> 00:15:51,195
aren't moving. You've been
making all these promises about

308
00:15:51,195 --> 00:15:54,652
your robots working and what's
going on. and we looked at the

309
00:15:54,652 --> 00:15:58,051
message and the GPS coordinate
he was sending us to was zero

310
00:15:58,051 --> 00:16:01,451
comma zero, which for those of
you not, familiar with GPS is

311
00:16:01,451 --> 00:16:04,341
the middle of the Atlantic
Ocean. we can enable our

312
00:16:04,341 --> 00:16:07,684
customers to just send simple
commands, and even if they're

313
00:16:07,684 --> 00:16:10,971
unsafe, we won't do unsafe
things. Like we won't leave the

314
00:16:10,971 --> 00:16:14,257
geofence in that use case. If
you try to order us to drive

315
00:16:14,257 --> 00:16:17,374
into an obstacle, we won't,
unless you for some reason,

316
00:16:17,374 --> 00:16:20,717
configured us to do that So now
a regular software engineer

317
00:16:20,717 --> 00:16:21,000
Yeah,

318
00:16:21,000 --> 00:16:22,000
Audrow Nash: saved something.
Yes.

319
00:16:22,000 --> 00:16:24,000
Stefan Seltz-Axmacher: a VC
during a demo.

320
00:16:24,000 --> 00:16:28,583
Audrow Nash: Ah, ha That's
hilarious. Okay. Yeah. Ilia,

321
00:16:28,583 --> 00:16:31,000
were you gonna say something?

322
00:16:31,000 --> 00:16:35,274
Ilia Baranov: Yeah. So I think I
was gonna expand a little bit on

323
00:16:35,274 --> 00:16:39,088
the API layer. our kind of
mission is to make the API just

324
00:16:39,088 --> 00:16:43,033
behave as any other web API
standard request and response so

325
00:16:43,033 --> 00:16:47,175
that anybody, any web developer
out there that you can hire for

326
00:16:47,175 --> 00:16:51,252
20 bucks an hour can suddenly
hook up whatever system you have

327
00:16:51,252 --> 00:16:55,000
in house to a robot and do so
without damaging equipment.

328
00:16:55,000 --> 00:16:58,673
Audrow Nash: I like that. How do
you, the biggest question that

329
00:16:58,673 --> 00:17:02,287
comes to my mind with that is
how do you manage state? Because

330
00:17:02,287 --> 00:17:05,726
it's complex. you tell the
robot, go here, you have to wait

331
00:17:05,726 --> 00:17:09,224
till it's there. There's some
sort of like underlying action

332
00:17:09,224 --> 00:17:12,839
like thing where, and then with
arrest API, it's very you send

333
00:17:12,839 --> 00:17:16,278
something, it sends something
back. how do you manage that?

334
00:17:16,278 --> 00:17:20,009
Does it open up a socket or is,
but then that gets more complex,

335
00:17:20,009 --> 00:17:21,000
or I guess how do

336
00:17:21,000 --> 00:17:24,875
Ilia Baranov: So we, actually,
yeah, so there's two layers to

337
00:17:24,875 --> 00:17:28,814
that. The first layer of your
question on, robot management is

338
00:17:28,814 --> 00:17:32,753
really using the behavior tree.
we fund, Steve Macenski's work

339
00:17:32,753 --> 00:17:36,438
in part in Nav2. And, fans of
it. We use big chunks of it.

340
00:17:36,438 --> 00:17:40,568
yeah, absolutely. It's fantastic
work. And his, he has a behavior

341
00:17:40,568 --> 00:17:44,571
tree internal to Nav2. We have
one, a kind of a larger one that

342
00:17:44,571 --> 00:17:48,510
controls other stuff outside of
Nav2, but very similar kind of

343
00:17:48,510 --> 00:17:52,576
concepts, similar, inspiration.
And that manages things like the

344
00:17:52,576 --> 00:17:55,943
client has requested five
different goals and halfway

345
00:17:55,943 --> 00:18:00,072
through it they interrupt us and
give us a new set of four goals.

346
00:18:00,072 --> 00:18:03,440
Should we then resume the
original goals or should we

347
00:18:03,440 --> 00:18:07,315
change, or should, what logic
state should we take? So that's

348
00:18:07,315 --> 00:18:11,000
all on robot and that's all
function of our behavior tree.

349
00:18:11,000 --> 00:18:14,239
Audrow Nash: if I send, two, two
positions or something, it

350
00:18:14,239 --> 00:18:17,643
creates a queue. How do you, but
so to me. That's probably the

351
00:18:17,643 --> 00:18:21,157
behavior expected, but it may be
like, oh no, I actually changed

352
00:18:21,157 --> 00:18:24,451
my mind. I want to go to this
one. I don't want to go to the

353
00:18:24,451 --> 00:18:25,000
first one.

354
00:18:25,000 --> 00:18:27,000
Ilia Baranov: so our rest, API,

355
00:18:27,000 --> 00:18:28,000
Audrow Nash: Okay. Yeah,

356
00:18:27,000 --> 00:18:30,393
Ilia Baranov: rest, API, when
you send commands, you have

357
00:18:30,393 --> 00:18:32,000
either a preempt or a queue

358
00:18:32,000 --> 00:18:33,000
Audrow Nash: Oh, you guys have
thought

359
00:18:33,000 --> 00:18:36,526
Ilia Baranov: we just provide
all the, we, yeah, we provide

360
00:18:36,526 --> 00:18:40,352
all of those knobs. And actually
a big challenge of building our

361
00:18:40,352 --> 00:18:44,117
API is we have hundreds of knobs
we can adjust under underneath

362
00:18:44,117 --> 00:18:47,166
our layer and being very
judicious on which ones we

363
00:18:47,166 --> 00:18:50,752
expose. So our API doesn't just
become this enormous surface

364
00:18:50,752 --> 00:18:54,099
area that we have to test
thousands of options, but just

365
00:18:54,099 --> 00:18:57,864
the minimum possible set that is
intelligent enough to, work in

366
00:18:57,864 --> 00:18:59,000
most circumstances.

367
00:18:59,000 --> 00:19:03,398
Audrow Nash: Yeah. Oh, I like
that a lot. Okay, so I really

368
00:19:03,398 --> 00:19:08,093
like the idea of these, Ilia, is
this your kind of, are you the

369
00:19:08,093 --> 00:19:11,000
master architect of all of this,
or how

370
00:19:11,000 --> 00:19:12,000
Ilia Baranov: I wish

371
00:19:12,000 --> 00:19:13,000
Audrow Nash: how is

372
00:19:13,000 --> 00:19:16,509
Ilia Baranov: gonna nod, but
it's actually, the engineering

373
00:19:16,509 --> 00:19:20,316
team. I'd say I, I'd say I, I do
the least amount of engineering

374
00:19:20,316 --> 00:19:23,766
outta the team, thankfully. I
think they're usually pretty

375
00:19:23,766 --> 00:19:27,335
happy when I'm not messing in
the code because my code tends

376
00:19:27,335 --> 00:19:29,000
to be experimental and tests

377
00:19:29,000 --> 00:19:31,000
Audrow Nash: You're prototyping,
you're trying to get it

378
00:19:31,000 --> 00:19:34,486
Ilia Baranov: I, I'm per, I'm,
very much that sort of CTO and

379
00:19:34,486 --> 00:19:38,029
engineer Exactly. And I can do
things quickly and kind of test

380
00:19:38,029 --> 00:19:41,457
things out. But really it's a
huge credit to the engineering

381
00:19:41,457 --> 00:19:44,657
team who really get it to a
quality level where it works

382
00:19:44,657 --> 00:19:47,000
every time and is safe and is
functional.

383
00:19:47,000 --> 00:19:48,000
Audrow Nash: I like

384
00:19:48,000 --> 00:19:51,491
Ilia Baranov: and as much as I'd
like to be that kind of person,

385
00:19:51,491 --> 00:19:54,000
I'm, not nearly enough of that
kind of person.

386
00:19:54,000 --> 00:19:58,237
Audrow Nash: the thing is, so
I've, seen companies where they

387
00:19:58,237 --> 00:20:02,475
have the master architect who
builds everything. and they, or

388
00:20:02,475 --> 00:20:06,712
it could be build or it could be
designs, everything. But the

389
00:20:06,712 --> 00:20:11,089
result that I've often seen is
that some of the deci, like many

390
00:20:11,089 --> 00:20:14,562
of the decisions might be
amazing, but some of the

391
00:20:14,562 --> 00:20:18,800
decisions were not, and they
create unbelievable chafing. And

392
00:20:18,800 --> 00:20:22,620
then it's also very hard to
onboard new people into the

393
00:20:22,620 --> 00:20:26,719
system. So if you're working
with a team, that's probably a

394
00:20:26,719 --> 00:20:30,956
very good way to distribute it
and get some sort of consensus

395
00:20:30,956 --> 00:20:35,194
around the decisions, especially
about the things that you're

396
00:20:35,194 --> 00:20:37,000
exposing, I would imagine.

397
00:20:37,000 --> 00:20:40,596
Ilia Baranov: Yeah. And I'd say
really my function is in a lot

398
00:20:40,596 --> 00:20:44,076
of ways to keep refocusing the
team on, let's make a general

399
00:20:44,076 --> 00:20:47,497
purpose of a piece of software
as we can. 'cause often when

400
00:20:47,497 --> 00:20:50,977
we're working with a client,
it's a very natural tendency to

401
00:20:50,977 --> 00:20:54,457
say, the client has requested X,
I'm gonna build a function,

402
00:20:54,457 --> 00:20:57,995
right? I'm gonna build X and
I'll often come in and re remind

403
00:20:57,995 --> 00:21:01,533
people like, X is A, subset of
this larger function that more

404
00:21:01,533 --> 00:21:05,070
people would want. Let's build
that larger function and teach

405
00:21:05,070 --> 00:21:08,608
the client to use that 'cause.
Then it immediately applies to

406
00:21:08,608 --> 00:21:10,000
our whole fleet because.

407
00:21:10,000 --> 00:21:13,495
Audrow Nash: and you guys had,
was it eight, eight people at

408
00:21:13,495 --> 00:21:16,000
the company? Or something around
there. 10.

409
00:21:16,000 --> 00:21:17,000
Stefan Seltz-Axmacher: Yep. Yep.

410
00:21:17,000 --> 00:21:20,717
Audrow Nash: And, so 10 people
are, is almost everyone an

411
00:21:20,717 --> 00:21:23,000
engineer for this kind of thing
or?

412
00:21:23,000 --> 00:21:24,000
Stefan Seltz-Axmacher: everyone.
Eight

413
00:21:24,000 --> 00:21:25,000
Audrow Nash: eight out of 10.
Yeah.

414
00:21:25,000 --> 00:21:27,000
Stefan Seltz-Axmacher: Yeah.

415
00:21:27,000 --> 00:21:30,746
Audrow Nash: Okay. And how long
have you guys been going Just

416
00:21:30,746 --> 00:21:34,000
to, like how long has this
company been around? Okay.

417
00:21:34,000 --> 00:21:36,000
Stefan Seltz-Axmacher: Two,
maybe. Two and a quarter.

418
00:21:36,000 --> 00:21:38,000
Ilia Baranov: Yeah, two and a
quarter about right.

419
00:21:38,000 --> 00:21:44,000
Audrow Nash: How, how are you
funding, the initiative so far?

420
00:21:44,000 --> 00:21:47,688
Stefan Seltz-Axmacher: So we've,
we've raised some money, a

421
00:21:47,688 --> 00:21:51,063
decent amount. we haven't
announced it publicly. more,

422
00:21:51,063 --> 00:21:53,000
than a little, less than a lot.

423
00:21:53,000 --> 00:21:54,000
Ilia Baranov: one.

424
00:21:54,000 --> 00:21:58,660
Stefan Seltz-Axmacher: and, we
also have this weird thing of

425
00:21:58,660 --> 00:22:02,000
customers paying us and SaaS
margins, Yeah.

426
00:22:02,000 --> 00:22:03,000
Audrow Nash: What a thing.

427
00:22:03,000 --> 00:22:06,779
Stefan Seltz-Axmacher: Yeah.
it's, crazy after trying to do a

428
00:22:06,779 --> 00:22:10,124
big monolithic, vertical
Robotics company before where

429
00:22:10,124 --> 00:22:13,717
the, lift to get to enough
robots to be profitable, was a,

430
00:22:13,717 --> 00:22:17,558
okay, we need $40 million. But
there's a very real world where

431
00:22:17,558 --> 00:22:21,212
we're just like a profitable
company this year, and I don't

432
00:22:21,212 --> 00:22:25,239
know how to grok that. My inner,
like my roboticist brain is just

433
00:22:25,239 --> 00:22:28,894
very upset with what do we do
when more money comes in than

434
00:22:28,894 --> 00:22:31,000
goes out? I think buy
Lamborghinis

435
00:22:31,000 --> 00:22:33,000
Audrow Nash: get into hardware
and then you can

436
00:22:33,000 --> 00:22:34,000
Stefan Seltz-Axmacher: Oh.

437
00:22:33,000 --> 00:22:36,000
Audrow Nash: You lambing you get
into,

438
00:22:36,000 --> 00:22:38,914
Ilia Baranov: that's the money
pit. That's the money pit we

439
00:22:38,914 --> 00:22:40,000
shovel the money into.

440
00:22:40,000 --> 00:22:42,000
Audrow Nash: yep, Okay. I really
like

441
00:22:42,000 --> 00:22:44,876
Ilia Baranov: it's, funny, on
that note, just a, small

442
00:22:44,876 --> 00:22:47,592
anecdote, Stefan and I keep
trading emails with our

443
00:22:47,592 --> 00:22:50,467
bookkeeper. 'cause we're always
like, am I miscounting

444
00:22:50,467 --> 00:22:51,000
something,

445
00:22:51,000 --> 00:22:52,000
Stefan Seltz-Axmacher: this
can't be right.

446
00:22:52,000 --> 00:22:56,604
Ilia Baranov: with our business
plan? it should be doing this

447
00:22:56,604 --> 00:23:00,000
and it's like doing this. I
don't understand.

448
00:23:00,000 --> 00:23:02,333
Audrow Nash: so exciting for
those listening pointing up

449
00:23:02,333 --> 00:23:03,000
instead of down.

450
00:23:03,000 --> 00:23:04,000
Stefan Seltz-Axmacher: Yeah.

451
00:23:04,000 --> 00:23:06,000
Ilia Baranov: Yes. yeah,

452
00:23:06,000 --> 00:23:08,260
Audrow Nash: Very exciting to
hear for robotics companies.

453
00:23:08,260 --> 00:23:09,000
Hell yeah. Love it.

454
00:23:09,000 --> 00:23:11,000
Ilia Baranov: yeah,

455
00:23:11,000 --> 00:23:14,833
Audrow Nash: what, I guess
another kind of fundamental

456
00:23:14,833 --> 00:23:18,951
question. I look on the website,
why are you guys not open

457
00:23:18,951 --> 00:23:23,281
source? is it because of the
kind of competitive advantage of

458
00:23:23,281 --> 00:23:27,185
this, but why not? why not? I
guess there were probably

459
00:23:27,185 --> 00:23:31,373
discussions around this 'cause
you guys like ROS and things

460
00:23:31,373 --> 00:23:34,000
like this, but what are you
thinking?

461
00:23:34,000 --> 00:23:39,549
Ilia Baranov: I, took this, from
a good friend of mentor of mine,

462
00:23:39,549 --> 00:23:41,000
Ryan Gariepy, at,

463
00:23:41,000 --> 00:23:42,000
Audrow Nash: He's wonderful.
Yeah.

464
00:23:42,000 --> 00:23:45,725
Ilia Baranov: he had told me
this. Yeah, he's great. and he

465
00:23:45,725 --> 00:23:49,702
had set this a long time ago at
Clearpath and I, quite like the

466
00:23:49,702 --> 00:23:53,616
logic behind it, which is, we'll
open source, anything that is

467
00:23:53,616 --> 00:23:57,467
useful to the community and not
directly harmful to us. So we

468
00:23:57,467 --> 00:24:01,444
actually open source as much as
we can. a lot of the Nav2 stuff

469
00:24:01,444 --> 00:24:05,231
that we're, trying to help fund
of course is open source. we

470
00:24:05,231 --> 00:24:08,893
have a big discussion going on
how to fuse. GPS into Locus

471
00:24:08,893 --> 00:24:12,554
Robotics is fuse factor graph
localization system. there's

472
00:24:12,554 --> 00:24:16,153
some efforts we're doing there.
We have an open container

473
00:24:16,153 --> 00:24:20,130
initiative that's fairly public
about a bunch of the containers

474
00:24:20,130 --> 00:24:23,728
we build use for build and
infrastructure. the specifics,

475
00:24:23,728 --> 00:24:27,832
for example, of how our behavior
tree dynamically generates nodes

476
00:24:27,832 --> 00:24:31,683
and makes sure that they're all
talking to the same system is

477
00:24:31,683 --> 00:24:35,533
useless for anybody else because
we already have built enough

478
00:24:35,533 --> 00:24:39,384
that it's weird and special and
so even if we did open source

479
00:24:39,384 --> 00:24:43,425
it, nobody would care. and also
would cause more problems for us

480
00:24:43,425 --> 00:24:46,897
than it solves. so when,
whenever it hits those two, we

481
00:24:46,897 --> 00:24:50,558
try to, we open source anything
that doesn't hit those two

482
00:24:50,558 --> 00:24:51,000
things.

483
00:24:51,000 --> 00:24:55,209
Audrow Nash: Makes sense. I
would wonder I don't know, there

484
00:24:55,209 --> 00:24:59,558
are, it seems like there's a few
companies that are in kind of

485
00:24:59,558 --> 00:25:03,907
similar space. I'd like a lot
how you guys have carved out the

486
00:25:03,907 --> 00:25:07,836
mobile robots space. I think
that was a very smart thing

487
00:25:07,836 --> 00:25:12,255
rather than we're gonna do all
Robotics. That, that to me seems

488
00:25:12,255 --> 00:25:16,464
like a good decision. I'd like
to talk about that later. but

489
00:25:16,464 --> 00:25:20,463
some of them are going open
source, hoping for additional

490
00:25:20,463 --> 00:25:24,742
adoption, I guess maybe Stefan,
how, do you think about this?

491
00:25:24,742 --> 00:25:28,881
would it accelerate adoption if
you made certain parts open

492
00:25:28,881 --> 00:25:33,230
source or even you could even
have a high level Python library

493
00:25:33,230 --> 00:25:37,439
that just calls your API or
something. why, are you thinking

494
00:25:37,439 --> 00:25:38,000
about Or

495
00:25:38,000 --> 00:25:40,000
Ilia Baranov: be clear, the API
is open source, by

496
00:25:40,000 --> 00:25:41,000
Audrow Nash: Oh, okay. I just
don't see

497
00:25:41,000 --> 00:25:44,268
Ilia Baranov: and at talking to
the Oh, yeah, No, it's there.

498
00:25:44,268 --> 00:25:47,536
we, we've gotta do a better job
of marketing it, but like any

499
00:25:47,536 --> 00:25:50,750
documentation and code on, and
I've even posted some YouTube

500
00:25:50,750 --> 00:25:54,125
videos on here's how you talk to
our robots, here's how you use

501
00:25:54,125 --> 00:25:56,964
them. All of that. That would
actually make sense for

502
00:25:56,964 --> 00:26:00,179
customers. All open source,
throw and pull requests. We love

503
00:26:00,179 --> 00:26:02,000
to review them. Yeah, no
question.

504
00:26:02,000 --> 00:26:03,000
Audrow Nash: Okay. That's

505
00:26:02,000 --> 00:26:05,023
Ilia Baranov: it's the
underlying ma, the underlying

506
00:26:05,023 --> 00:26:07,000
pieces that are weird and
special.

507
00:26:07,000 --> 00:26:10,745
Audrow Nash: as someone who's
very lazy, I would expect at the

508
00:26:10,745 --> 00:26:14,248
footer on your website that
there's a GitHub link and this

509
00:26:14,248 --> 00:26:16,000
kind of thing. That's exactly

510
00:26:16,000 --> 00:26:19,333
Stefan Seltz-Axmacher: good,
maybe by the time you publish

511
00:26:19,333 --> 00:26:21,000
this podcast that will exist.

512
00:26:21,000 --> 00:26:25,631
Audrow Nash: Okay. Because that
would be great. okay, so I

513
00:26:25,631 --> 00:26:30,262
misread the website in that
case. that seems awesome. And,

514
00:26:30,262 --> 00:26:35,133
okay, so I guess, Why are you
carving out the niche of mobile

515
00:26:35,133 --> 00:26:39,924
robots only? Why aren't you
doing arms? Why aren't you doing

516
00:26:39,924 --> 00:26:42,000
all sorts of other things?

517
00:26:42,000 --> 00:26:46,130
Stefan Seltz-Axmacher: Yeah. so
I think that has to do with a

518
00:26:46,130 --> 00:26:50,464
lot with what we've done before.
I'm not an armed guy. I'm not a

519
00:26:50,464 --> 00:26:54,662
quadcopter guy. There's also a
lot of people wanting to do arm

520
00:26:54,662 --> 00:26:58,521
stuff like this. There's a lot
of people doing quadcopter

521
00:26:58,521 --> 00:27:02,313
stuff. in both of our past
experiences, we rebuilt stuff

522
00:27:02,313 --> 00:27:06,511
that we had already built and
other people had already built a

523
00:27:06,511 --> 00:27:10,303
whole bunch of times and then
found it wasn't incredibly

524
00:27:10,303 --> 00:27:14,366
reusable. so previously, when I
had a, vertical self-driving

525
00:27:14,366 --> 00:27:18,632
truck startup, every time we'd
get some big piece of press, I'd

526
00:27:18,632 --> 00:27:22,423
get a, bunch of inbound from
random industrial companies

527
00:27:22,423 --> 00:27:26,621
saying, Hey, air quotes, you've
solved autonomy. If we pay you

528
00:27:26,621 --> 00:27:30,278
$8 million, can you automate
this super niche piece of

529
00:27:30,278 --> 00:27:34,273
equipment that only eight of in
the world exists and no one

530
00:27:34,273 --> 00:27:38,200
else, uses, but is really
valuable to us and we can't hire

531
00:27:38,200 --> 00:27:42,263
operators And because of how
we're architected, as like some

532
00:27:42,263 --> 00:27:46,190
big monolith vertical thing,
we'd basically always have to

533
00:27:46,190 --> 00:27:50,049
say no to that. It would be an
entire pivot. It would be,

534
00:27:50,049 --> 00:27:54,180
throwing out 75% of the code
base. And if you look at the big

535
00:27:54,180 --> 00:27:58,513
OEMs who built autonomy multiple
times, like some of 'em, one of

536
00:27:58,513 --> 00:28:02,779
them who likes to color green,
has, has built autonomy at least

537
00:28:02,779 --> 00:28:06,842
like 15 or so times. Never, none
of those use, none of those

538
00:28:06,842 --> 00:28:11,175
rebuilds have used more than 25%
of any existing code base. None

539
00:28:11,175 --> 00:28:15,441
of them have cost less than $10
million. None of them have been

540
00:28:15,441 --> 00:28:19,233
completed in less than three
years. This is a, super key

541
00:28:19,233 --> 00:28:23,092
problem and is a big part in my,
in our belief of why the

542
00:28:23,092 --> 00:28:27,426
Robotics industry sucks so hard.
Because like people want robots

543
00:28:27,426 --> 00:28:31,488
to grade construction sites so
they can start building or to

544
00:28:31,488 --> 00:28:35,145
till fields or to, spray,
fertilizer on crops or to do

545
00:28:35,145 --> 00:28:39,478
whatever. And to even look at a
use case like that pre Polymath,

546
00:28:39,478 --> 00:28:43,202
you might have to hire a
consulting firm five to 50 for

547
00:28:43,202 --> 00:28:47,197
five to $50 million to build
something that might only work

548
00:28:47,197 --> 00:28:50,854
for a demo day. And that does
not build a good, strong

549
00:28:50,854 --> 00:28:54,645
industry. So like the arm people
have done cool stuff in

550
00:28:54,645 --> 00:28:58,776
manufacturing. The drone people
have done some cool stuff in,

551
00:28:58,776 --> 00:29:02,974
in, in toys and inspection and
whatever. But for mobile robots

552
00:29:02,974 --> 00:29:06,969
that move through the earth,
like we need to make it not so

553
00:29:06,969 --> 00:29:09,000
hard to make, to create value.

554
00:29:09,000 --> 00:29:13,186
Audrow Nash: I think that's a
great idea. And it's so cool to

555
00:29:13,186 --> 00:29:16,000
me that, so going into the
founding story

556
00:29:16,000 --> 00:29:17,000
Stefan Seltz-Axmacher: Yeah.

557
00:29:17,000 --> 00:29:20,896
Audrow Nash: coming to Polymath,
Robotics, you guys, you, so

558
00:29:20,896 --> 00:29:25,052
Stefan you were doing a, highway
autonomous car up or autonomous

559
00:29:25,052 --> 00:29:27,000
truck, or what was it exactly?

560
00:29:27,000 --> 00:29:30,811
Stefan Seltz-Axmacher: Yeah. So
Starsky Robotics, we were doing

561
00:29:30,811 --> 00:29:34,743
self-driving trucks. We did, we,
we started with a similar thesis

562
00:29:34,743 --> 00:29:38,252
as Polymath that robots are
hard. the way we, that we were

563
00:29:38,252 --> 00:29:41,942
overcoming robots being hard is
we do autonomy on the highway

564
00:29:41,942 --> 00:29:45,209
where it's a relatively
constrained environment. We do

565
00:29:45,209 --> 00:29:48,838
tele op in the first and last
mile where it's kind of chaos.

566
00:29:48,838 --> 00:29:52,347
and then to make that problem
easier, we'd control our own

567
00:29:52,347 --> 00:29:56,279
routes by being our own trucking
company. and at the time when we

568
00:29:56,279 --> 00:29:59,667
were making a lot of hard
technical decisions, 2017 ish,

569
00:29:59,667 --> 00:30:03,296
basically no vendors, would give
you any part, unless you're

570
00:30:03,296 --> 00:30:07,047
willing to sign something that
said you never rely on it for a

571
00:30:07,047 --> 00:30:10,556
driver route test. So as a
result, what starsky became was

572
00:30:10,556 --> 00:30:14,367
this. Yeah. Yeah. So we talked
to LIDAR companies, we talked to

573
00:30:14,367 --> 00:30:16,000
the Velodyne, we talked to,

574
00:30:16,000 --> 00:30:19,448
Audrow Nash: stuff, but don't
use our stuff for real things.

575
00:30:19,448 --> 00:30:21,000
Is that what basically was?

576
00:30:21,000 --> 00:30:25,306
Stefan Seltz-Axmacher: Yeah,
yeah. legitimately that was what

577
00:30:25,306 --> 00:30:27,000
it was like. So we ended

578
00:30:27,000 --> 00:30:30,250
Audrow Nash: They don't wanna be
affiliated with it, I suppose if

579
00:30:30,250 --> 00:30:31,000
it goes poorly,

580
00:30:31,000 --> 00:30:35,106
Stefan Seltz-Axmacher: it wasn't
illegal, it was just fear. I

581
00:30:35,106 --> 00:30:39,413
think it's important to note in
on-road autonomy. I haven't done

582
00:30:39,413 --> 00:30:43,519
a, an actual count in a while. I
think there's five companies

583
00:30:43,519 --> 00:30:47,221
ever that have taken a person
out on public roads. like

584
00:30:47,221 --> 00:30:51,327
something like that. And there's
in the range of probably $25

585
00:30:51,327 --> 00:30:55,433
billion invested in the space.
It is mostly not real industry

586
00:30:55,433 --> 00:30:59,269
or I think actually the number
might be a hundred billion

587
00:30:59,269 --> 00:31:03,106
dollars invested. but so,
previously I'd done that and we

588
00:31:03,106 --> 00:31:06,606
had, because of robots being
hard, we grew into this

589
00:31:06,606 --> 00:31:10,308
frankensteinian monolith
vertical company. Where we had

590
00:31:10,308 --> 00:31:14,548
to do a, an incredibly not built
here syndrome, hardware stack,

591
00:31:14,548 --> 00:31:18,317
an incredibly not built here
syndrome, autonomy, tele op

592
00:31:18,317 --> 00:31:21,750
safety stack software for
running your own trucking

593
00:31:21,750 --> 00:31:25,721
company and then running a
trucking company. I'd have to go

594
00:31:25,721 --> 00:31:27,000
from meetings where

595
00:31:27,000 --> 00:31:28,000
Ilia Baranov: Yeah.

596
00:31:27,000 --> 00:31:31,000
Stefan Seltz-Axmacher: just
really, really hard. And I don't

597
00:31:31,000 --> 00:31:34,867
know if most people can be
successful in that, in, in that

598
00:31:34,867 --> 00:31:37,000
type of business if you reinvent

599
00:31:37,000 --> 00:31:39,706
Audrow Nash: have to be like
four CEOs at once to do that

600
00:31:39,706 --> 00:31:42,601
kind of thing. It seems like so
many different businesses and

601
00:31:42,601 --> 00:31:45,402
industries that you're, and I
don't know what your trucking

602
00:31:45,402 --> 00:31:48,155
experience is, but like you have
to be a trucking business

603
00:31:48,155 --> 00:31:50,908
manager as well for all these
things. So that sounds to me

604
00:31:50,908 --> 00:31:52,000
tremendously difficult.

605
00:31:52,000 --> 00:31:55,146
Stefan Seltz-Axmacher: I, think
the re one of the main reasons

606
00:31:55,146 --> 00:31:58,140
that company didn't raise a
Series B was because I was left

607
00:31:58,140 --> 00:32:01,337
less good at talking about the
financials of a trucking company

608
00:32:01,337 --> 00:32:04,179
than a Fortune 500 trucking
company CFO was, which was a

609
00:32:04,179 --> 00:32:07,325
hard thing to be good at talking
about right after explaining,

610
00:32:07,325 --> 00:32:09,000
safety engineering to someone
for

611
00:32:09,000 --> 00:32:12,754
Audrow Nash: yeah, totally.
Yeah. That's so interesting.

612
00:32:12,754 --> 00:32:16,777
Okay, and in that experience,
you saw a bunch of these, it'd

613
00:32:16,777 --> 00:32:20,665
be like, okay, you get press.
Then, it's oh, you've solved

614
00:32:20,665 --> 00:32:24,687
autonomy for these problems.
Some people see it and then go,

615
00:32:24,687 --> 00:32:28,575
Hey, we really need something
similar to that. And you saw

616
00:32:28,575 --> 00:32:32,397
these over and over again, and
this is how you've come to

617
00:32:32,397 --> 00:32:33,000
Polymath,

618
00:32:33,000 --> 00:32:35,982
Stefan Seltz-Axmacher:
Basically, big enterprise

619
00:32:35,982 --> 00:32:39,896
believes in autonomy. Like big
mining companies, big farms, big

620
00:32:39,896 --> 00:32:43,686
factories, they all have labor
shortages. They've seen enough

621
00:32:43,686 --> 00:32:47,476
headlines about how self-driving
is here, and they're looking

622
00:32:47,476 --> 00:32:51,142
around and thinking like, I
can't hire people to work in my

623
00:32:51,142 --> 00:32:55,180
lumber mill. I can't hire people
to work in my port. I can't hire

624
00:32:55,180 --> 00:32:58,908
people to work in my rail yard.
maybe some robots could move

625
00:32:58,908 --> 00:33:02,636
this stuff from A to B. And
before Polymath, maybe there was

626
00:33:02,636 --> 00:33:06,178
someone who happened to be
building the right solution on

627
00:33:06,178 --> 00:33:09,657
the right vehicle, like a,
perfect unicorn of a vertical

628
00:33:09,657 --> 00:33:13,571
lineup. Or maybe you had to hire
a consulting firm to build you

629
00:33:13,571 --> 00:33:15,000
something from scratch.

630
00:33:15,000 --> 00:33:19,420
Audrow Nash: So you are allowing
companies to build up around

631
00:33:19,420 --> 00:33:23,406
specific vehicles that big
companies are using, and you

632
00:33:23,406 --> 00:33:27,899
provide a lot of the research
and basically the infrastructure

633
00:33:27,899 --> 00:33:30,000
that they need to build to do

634
00:33:30,000 --> 00:33:32,967
Stefan Seltz-Axmacher: We give
them a code base that will just

635
00:33:32,967 --> 00:33:35,933
work. we can connect them to,
with people to do the retrofits.

636
00:33:35,933 --> 00:33:38,852
and we can take commands from
whatever their weird system for

637
00:33:38,852 --> 00:33:40,000
running their businesses

638
00:33:40,000 --> 00:33:44,552
Audrow Nash: Cool. Cool. That
seems so, I really, oh yeah, go

639
00:33:44,552 --> 00:33:45,000
ahead.

640
00:33:45,000 --> 00:33:48,569
Ilia Baranov: a I'll, give you a
fun, illustrative example, which

641
00:33:48,569 --> 00:33:51,809
is my favorite one. We haven't
done this yet. I keep hoping

642
00:33:51,809 --> 00:33:55,104
that eventually we'll get to
this mission, but there's this,

643
00:33:55,104 --> 00:33:58,673
problem in metal recycling where
if you recycle a bunch of metal,

644
00:33:58,673 --> 00:34:02,133
you end up with this big molten
pot with a bunch of slag on the

645
00:34:02,133 --> 00:34:04,000
top. Once you dump out the,
useful

646
00:34:04,000 --> 00:34:09,752
Stefan Seltz-Axmacher: and by
like big molten pot. imagine

647
00:34:09,752 --> 00:34:16,000
maybe, a cauldron with a, radius
of like 10 feet and like maybe

648
00:34:16,000 --> 00:34:17,000
Ilia Baranov: a backyard pool.

649
00:34:17,000 --> 00:34:19,647
Audrow Nash: like something a
villain would fall into at the

650
00:34:19,647 --> 00:34:20,000
end of a

651
00:34:20,000 --> 00:34:22,000
Ilia Baranov: yeah, exactly.
Exactly.

652
00:34:22,000 --> 00:34:24,000
Stefan Seltz-Axmacher: what
kills the bad guy in Terminator

653
00:34:24,000 --> 00:34:26,000
Audrow Nash: Yep. Yep.

654
00:34:26,000 --> 00:34:29,620
Ilia Baranov: Yeah. So it's this
enormous crucible. And what they

655
00:34:29,620 --> 00:34:32,961
have to do is they have to pick
it up, drive it outside to a

656
00:34:32,961 --> 00:34:36,581
particular dump site and dump it
over And there's a purpose built

657
00:34:36,581 --> 00:34:40,145
machine. There's a purpose-built
machine that drives this thing.

658
00:34:40,145 --> 00:34:43,487
And it's a heinously dangerous
job because if you accelerate

659
00:34:43,487 --> 00:34:46,717
too quickly or suddenly break or
have a bump, liquid metal

660
00:34:46,717 --> 00:34:49,000
splashes the vehicle and catches
on fire.

661
00:34:49,000 --> 00:34:50,000
Stefan Seltz-Axmacher: and you
die.

662
00:34:49,000 --> 00:34:53,231
Ilia Baranov: the vehicle's pur
and you die. Yeah. The vehicle's

663
00:34:53,231 --> 00:34:57,066
purpose built that, the like
liquid metal fire death is at

664
00:34:57,066 --> 00:35:01,165
the back and the propane tanks
running the vehicle are all the

665
00:35:01,165 --> 00:35:05,066
way at the front to make sure
that as far away as possible,

666
00:35:05,066 --> 00:35:08,901
there's probably like 200 of
these vehicles on the planet,

667
00:35:08,901 --> 00:35:12,537
but they're so ludicrously
dangerous that, like perfect

668
00:35:12,537 --> 00:35:15,843
case for autonomy, same spot
every day, same exact

669
00:35:15,843 --> 00:35:19,744
conditions, perfectly designed
space, but the guy's getting

670
00:35:19,744 --> 00:35:21,000
paid a ton of money

671
00:35:21,000 --> 00:35:22,000
Audrow Nash: I was gonna say
they

672
00:35:21,000 --> 00:35:23,000
Ilia Baranov: because it's
dangerous.

673
00:35:23,000 --> 00:35:24,000
Audrow Nash: year or something

674
00:35:24,000 --> 00:35:25,579
Stefan Seltz-Axmacher: they
also, a thing Ilia, I don't know

675
00:35:25,579 --> 00:35:26,000
if I've told you

676
00:35:26,000 --> 00:35:28,000
Audrow Nash: at a high rate.

677
00:35:28,000 --> 00:35:31,622
Stefan Seltz-Axmacher: at,
another thing I learned about

678
00:35:31,622 --> 00:35:35,697
this use case recently is it's
also incredibly carcinogenic. So

679
00:35:35,697 --> 00:35:39,383
even though the guys might be
making 300k a year after 10

680
00:35:39,383 --> 00:35:43,393
years, they have all sorts of
unpleasant cancer. so if you, if

681
00:35:43,393 --> 00:35:47,080
you think about like vertical
autonomy for that solution,

682
00:35:47,080 --> 00:35:51,154
there's 200 of those vehicles in
the in the world. Does it make

683
00:35:51,154 --> 00:35:55,164
sense to spend $50 million to
learn how to automate Just that,

684
00:35:55,164 --> 00:35:59,303
it literally might be cheaper to
just give people cancer than to

685
00:35:59,303 --> 00:36:03,184
build one off autonomy for that.
use case. Yeah. the mar the

686
00:36:03,184 --> 00:36:07,000
market has accepted cancer over
the status quo in Robotics.

687
00:36:07,000 --> 00:36:08,000
Audrow Nash: Oh my gosh.

688
00:36:08,000 --> 00:36:11,270
Ilia Baranov: But meanwhile,
it's the same steering wheel,

689
00:36:11,270 --> 00:36:14,653
the same gas and brake pedals,
like a human's driving it the

690
00:36:14,653 --> 00:36:18,205
same way that we drive our test
tractor. So there's zero reason

691
00:36:18,205 --> 00:36:21,475
we can't stick a few sensors and
actuators on there with a

692
00:36:21,475 --> 00:36:24,858
retrofit partner and then have
day one perfect autonomy just

693
00:36:24,858 --> 00:36:27,000
like we have on the rest of our
fleet.

694
00:36:27,000 --> 00:36:30,972
Audrow Nash: That's super cool.
I don't, so one thing that I

695
00:36:30,972 --> 00:36:34,877
don't fully understand is, the
sensors, so you said earlier

696
00:36:34,877 --> 00:36:38,253
that you use a lot of the
sensors that are on these

697
00:36:38,253 --> 00:36:41,894
vehicles. I would imagine you
want additional ones, for

698
00:36:41,894 --> 00:36:45,667
autonomy. So I would think
vision would be really useful.

699
00:36:45,667 --> 00:36:49,837
Maybe LIDAR would be useful. You
might want an IMU or whatever.

700
00:36:49,837 --> 00:36:53,610
so do you have a little box of
sensors you put onto these

701
00:36:53,610 --> 00:36:55,000
vehicles sometimes or

702
00:36:55,000 --> 00:36:58,708
Ilia Baranov: I wouldn't say a
box, I would say. Yeah, case by

703
00:36:58,708 --> 00:37:02,237
case, but generally speaking,
the main ingredients are some

704
00:37:02,237 --> 00:37:05,586
sort of feedback from our
vehicles. Just even a, live or

705
00:37:05,586 --> 00:37:09,175
not a live state is useful. So
on that 1997 bulldozer, we at

706
00:37:09,175 --> 00:37:12,943
least know the engines turning
over. So we have some concept of

707
00:37:12,943 --> 00:37:16,471
health. but even better than
that, if it has a speedometer,

708
00:37:16,471 --> 00:37:20,000
feeding in the speedometer is a
nice ground truthing for us

709
00:37:20,000 --> 00:37:21,000
Audrow Nash: Okay. And would, do
you do

710
00:37:21,000 --> 00:37:24,259
Ilia Baranov: If we're standing
still, we're standing still. So

711
00:37:24,259 --> 00:37:27,569
even that layer of data is, even
though it's coarse, it's useful

712
00:37:27,569 --> 00:37:30,000
for us as a, error checker for
everything else.

713
00:37:30,000 --> 00:37:33,721
Audrow Nash: Do you tap the
vehicle system in that case, or

714
00:37:33,721 --> 00:37:37,000
do you put a camera to just look
at it or, oh, okay.

715
00:37:37,000 --> 00:37:40,248
Ilia Baranov: we tend to just
use the CAN bus if it's there.

716
00:37:40,248 --> 00:37:43,659
if it's not, we'll put a little
encoder on the, again, it's not

717
00:37:43,659 --> 00:37:46,745
us, right? It's our partner
teams, our integrators, put a

718
00:37:46,745 --> 00:37:50,156
little encoder on, part of the
drive shaft or camera staring at

719
00:37:50,156 --> 00:37:53,350
the dashboard. Worst case. But
that one, usually the update

720
00:37:53,350 --> 00:37:54,000
rate is, eh,

721
00:37:54,000 --> 00:37:57,424
Stefan Seltz-Axmacher: and then
we'll have a bunch of different

722
00:37:57,424 --> 00:38:00,848
sensors added, depending on the
use case. So it could be LIDAR,

723
00:38:00,848 --> 00:38:03,891
it could be stereo cameras,
could be radar, it could be,

724
00:38:03,891 --> 00:38:07,098
whatever. an interesting thing
that we found is people have

725
00:38:07,098 --> 00:38:10,413
really strong opinions about
sensors. if we were just a LIDAR

726
00:38:10,413 --> 00:38:13,620
shop, there are, we would have
fewer customers than we have

727
00:38:13,620 --> 00:38:16,935
right now. if we were a never
LIDAR shop, we would have fewer

728
00:38:16,935 --> 00:38:19,000
customers than we are right now.
And I

729
00:38:19,000 --> 00:38:21,000
Audrow Nash: gotta support em
all. Yeah.

730
00:38:21,000 --> 00:38:24,927
Stefan Seltz-Axmacher: Yeah. And
like the thing is when everyone

731
00:38:24,927 --> 00:38:28,240
builds their robot as like a
monolith, what happens is

732
00:38:28,240 --> 00:38:31,922
there's so many crappy parts of
your code that at some point

733
00:38:31,922 --> 00:38:35,665
some demo was ruined because a
velodyne LIDAR wasn't behaving

734
00:38:35,665 --> 00:38:39,408
properly. Or at some other point
a demo was saved because, oh

735
00:38:39,408 --> 00:38:42,967
look, we replaced that LIDAR
with an ouster or replaced it

736
00:38:42,967 --> 00:38:46,894
with a Zed camera. so now every
robot forever needs a Zed camera

737
00:38:46,894 --> 00:38:50,391
on it. And we need to be able
to, work with both of those

738
00:38:50,391 --> 00:38:54,134
strong opinions. And my hunch is
every listener here has some

739
00:38:54,134 --> 00:38:58,000
sensor out there that they hate
and another one that they love.

740
00:38:58,000 --> 00:39:01,977
Audrow Nash: Yeah. That's so
funny. Yeah. And you guys, it

741
00:39:01,977 --> 00:39:06,434
makes sense. Catering to, or not
catering to, but allowing all of

742
00:39:06,434 --> 00:39:10,000
the different sensors that
people might want to use.

743
00:39:10,000 --> 00:39:12,000
Stefan Seltz-Axmacher: Yep.

744
00:39:12,000 --> 00:39:15,581
Audrow Nash: Okay. So that's
really cool. And I like, so a

745
00:39:15,581 --> 00:39:19,286
thing that is very interesting
to me about this is that what

746
00:39:19,286 --> 00:39:23,300
you guys are doing, I. It's like
okay, so you mentioned those 200

747
00:39:23,300 --> 00:39:26,881
vehicles that are doing the
super dangerous metal hauling,

748
00:39:26,881 --> 00:39:30,586
liquid metal terror. That just
sounds awful. But so you have

749
00:39:30,586 --> 00:39:34,229
200 vehicles that are doing
that. That problem does not it,

750
00:39:34,229 --> 00:39:38,057
like some companies want to pay
a lot of money to do that, but

751
00:39:38,057 --> 00:39:41,268
it doesn't make sense for a
venture capitalist to be

752
00:39:41,268 --> 00:39:45,034
investing in a company that does
that because you have 200 of

753
00:39:45,034 --> 00:39:48,739
these vehicles, it's not gonna
grow that much. Maybe you can

754
00:39:48,739 --> 00:39:52,382
apply it to other things. but
they're looking at the market

755
00:39:52,382 --> 00:39:56,334
that's gonna buy right now. And
so it's really interesting to me

756
00:39:56,334 --> 00:39:59,791
'cause I've seen, I feel like
with a lot of the Robotics

757
00:39:59,791 --> 00:40:03,373
companies that I see is they
pick one little vertical like

758
00:40:03,373 --> 00:40:07,016
that and then they go build
something in that. And it's not

759
00:40:07,016 --> 00:40:10,659
that attractive to venture
capitalists because they want to

760
00:40:10,659 --> 00:40:14,363
buy in at a fairly late stage
where they go, oh you are, you

761
00:40:14,363 --> 00:40:18,130
clearly have market traction,
whatever, and you need this big

762
00:40:18,130 --> 00:40:21,711
injection of cash for doing
something, but I want to see a

763
00:40:21,711 --> 00:40:25,416
10 or a hundred times return in
five years or something like

764
00:40:25,416 --> 00:40:29,059
this. You guys are doing it
clever where you're agnostic to

765
00:40:29,059 --> 00:40:32,517
all these problems and you let
other companies go and be

766
00:40:32,517 --> 00:40:36,345
profitable on those vehicles.
but there's a lot of growth that

767
00:40:36,345 --> 00:40:39,000
you guys can do, which is quite
exciting, I

768
00:40:39,000 --> 00:40:42,833
Stefan Seltz-Axmacher: yep. And
we can make a lot of betts in a

769
00:40:42,833 --> 00:40:46,241
lot of different industries,
where if our, customers are

770
00:40:46,241 --> 00:40:50,074
successful, we'll be successful.
but we can simultaneously make

771
00:40:50,074 --> 00:40:53,907
betts in metal recycling as well
as forestry, as well as ports,

772
00:40:53,907 --> 00:40:57,376
as well as this, that, and the
other. And as long as some

773
00:40:57,376 --> 00:41:01,270
people build robots that create
real value, we can make sure the

774
00:41:01,270 --> 00:41:04,921
robots work. But in a lot of
these use cases, the real value

775
00:41:04,921 --> 00:41:08,815
comes from what they're told to
do. So if you think about robots

776
00:41:08,815 --> 00:41:12,648
in a port, for example, driving
a yard truck with a trailer, is

777
00:41:12,648 --> 00:41:16,177
not special. There's something
like five to 15, yard truck

778
00:41:16,177 --> 00:41:19,585
autonomy companies that are
doing just that. There's, an

779
00:41:19,585 --> 00:41:22,992
incredible, diversity of yard
truck manufacturers. So if

780
00:41:22,992 --> 00:41:26,460
you're vertical, you have to
figure out which ones you're

781
00:41:26,460 --> 00:41:30,354
gonna integrate ones, which ones
you don't. The really important

782
00:41:30,354 --> 00:41:33,823
thing that yard trucks do is
they get commanded around by

783
00:41:33,823 --> 00:41:37,474
some system that manages a yard,
management system. And that

784
00:41:37,474 --> 00:41:41,124
system might be configured so
that, maybe there's five lanes

785
00:41:41,124 --> 00:41:44,897
going from north to south, and
on Wednesday morning a big ship

786
00:41:44,897 --> 00:41:48,608
is coming in. So four of those
lanes are coming away from the

787
00:41:48,608 --> 00:41:52,198
ship where one lane is going
towards the ship. And, an hour

788
00:41:52,198 --> 00:41:55,849
later you reverse the traffic
the other way around. and that

789
00:41:55,849 --> 00:41:59,561
way you can get the ship empty
two hours earlier, which means

790
00:41:59,561 --> 00:42:03,151
you can have five more ships
visit per year and blah, blah,

791
00:42:03,151 --> 00:42:06,619
blah. That's the stuff where
you're actually gonna make a

792
00:42:06,619 --> 00:42:10,392
billion dollars with robots. If
you're reinventing our part of

793
00:42:10,392 --> 00:42:11,000
the stack,

794
00:42:11,000 --> 00:42:12,000
Audrow Nash: Oh,

795
00:42:12,000 --> 00:42:13,000
Stefan Seltz-Axmacher: you're
missing

796
00:42:13,000 --> 00:42:14,000
Audrow Nash: you mean.

797
00:42:14,000 --> 00:42:17,665
Stefan Seltz-Axmacher: because a
lot of teams doing yard trucks

798
00:42:17,665 --> 00:42:21,156
are spending all their effort
building what we build and not

799
00:42:21,156 --> 00:42:24,531
like intelligent yard management
based on machine learning

800
00:42:24,531 --> 00:42:28,255
statistics and blah, blah, blah.
And that's really the cool part

801
00:42:28,255 --> 00:42:30,000
of being a yard truck company.

802
00:42:30,000 --> 00:42:35,057
Audrow Nash: Yeah. So yeah, I
agree. So if you do the super

803
00:42:35,057 --> 00:42:39,771
big tasks made by these super
big consumers, you can be

804
00:42:39,771 --> 00:42:45,000
rewarded handsomely and be a
billion dollar Robotics company,

805
00:42:45,000 --> 00:42:47,692
Stefan Seltz-Axmacher: Yep. And
if you're, and if you build and

806
00:42:47,692 --> 00:42:50,000
if your autonomy stack sucks,
you don't get any of it.

807
00:42:50,000 --> 00:42:53,605
Audrow Nash: Yeah. And you have
to build that. And that's a

808
00:42:53,605 --> 00:42:57,149
heavy investment initially. So
you guys are helping enable

809
00:42:57,149 --> 00:43:00,876
that, which is quite cool. Now,
I would love to hear what are

810
00:43:00,876 --> 00:43:04,664
your ambitions for this company
for, and I'd love to hear both

811
00:43:04,664 --> 00:43:08,513
of your perspective on this, to
like, where do you want this to

812
00:43:08,513 --> 00:43:12,057
go? Because it seems like a
slightly new business model to

813
00:43:12,057 --> 00:43:15,845
me, which is quite cool. maybe
we start with Ilia, like where,

814
00:43:15,845 --> 00:43:19,511
are we going in the long term?
Like how, do you see you guys

815
00:43:19,511 --> 00:43:20,000
growing?

816
00:43:20,000 --> 00:43:23,677
Ilia Baranov: Yeah. From, the
technical standpoint, I think

817
00:43:23,677 --> 00:43:27,415
the interesting part of our
business is, again, drawing that

818
00:43:27,415 --> 00:43:30,843
operating system kind of
comparison. you buy Windows to

819
00:43:30,843 --> 00:43:34,519
then run Excel or games or
whatever, right? I think because

820
00:43:34,519 --> 00:43:37,822
we're really focusing on A
software and B only motion

821
00:43:37,822 --> 00:43:41,499
software. and the edges of what
motion means will blur over

822
00:43:41,499 --> 00:43:45,113
time. concrete example really
clearly, really quickly that

823
00:43:45,113 --> 00:43:48,852
bulldozer will control, the
bulldozer will control the blade

824
00:43:48,852 --> 00:43:52,902
somewhat, but if a builder says,
here's a site plan, dig the site

825
00:43:52,902 --> 00:43:56,329
plan autonomously, we'll
probably partner with somebody

826
00:43:56,329 --> 00:44:00,131
who specializes in translating
site plans to blade control to

827
00:44:00,131 --> 00:44:04,119
whatever, and we'll drive it and
work with them. So I think the,

828
00:44:04,119 --> 00:44:07,858
technical interesting part is
we're building a foundation to

829
00:44:07,858 --> 00:44:11,472
then have. a web store. If, you
talk about what, universal

830
00:44:11,472 --> 00:44:15,148
robots did, I really like what
they did for arms, right? Is

831
00:44:15,148 --> 00:44:18,700
here's the arm, here's the
foundation, but like here's 50

832
00:44:18,700 --> 00:44:22,564
different companies that will
take our arm and then make Apple

833
00:44:22,564 --> 00:44:26,116
picking or make CNC machine
tending or whatever. It's the

834
00:44:26,116 --> 00:44:30,166
same thing is, take our autonomy
core and then build on top of it

835
00:44:30,166 --> 00:44:33,780
and here's a bunch of vendors
that will sell you pre-baked

836
00:44:33,780 --> 00:44:37,395
modules to do agriculture with
autonomy, to do mining with

837
00:44:37,395 --> 00:44:41,134
Polymath autonomy. We can do
those pieces, but I think we'll

838
00:44:41,134 --> 00:44:44,000
never outcompete thousands of
people using our

839
00:44:44,000 --> 00:44:45,818
Stefan Seltz-Axmacher: And I,
and we don't know how to run a

840
00:44:45,818 --> 00:44:47,667
mine. We don't know how to run a
forestry operation. We don't

841
00:44:47,667 --> 00:44:48,000
know how to

842
00:44:48,000 --> 00:44:50,000
Audrow Nash: saw this with your
trucking company. I

843
00:44:50,000 --> 00:44:53,231
Stefan Seltz-Axmacher: Yeah. I
had to learn how to run a

844
00:44:53,231 --> 00:44:56,000
trucking company, which turns
out is not simple.

845
00:44:56,000 --> 00:44:59,474
Audrow Nash: No, everything's
hard when you get into it, I

846
00:44:59,474 --> 00:45:02,947
think. Yeah. Okay. So that, I
really like that point. It's

847
00:45:02,947 --> 00:45:06,780
interesting to me how, it seems
like I. So many of the companies

848
00:45:06,780 --> 00:45:10,194
that I've talked to on this
podcast, it's like the vision

849
00:45:10,194 --> 00:45:13,727
eventually scales to being
something like an app store with

850
00:45:13,727 --> 00:45:17,141
this kind of thing where you
allow the functionality, you

851
00:45:17,141 --> 00:45:20,794
provide the core platform, and
then people can build off. And

852
00:45:20,794 --> 00:45:24,387
that sounds like where this may
go. and it's just, to me, it

853
00:45:24,387 --> 00:45:27,981
feels like an exciting time in
Robotics that maybe this will

854
00:45:27,981 --> 00:45:31,454
actually occur. 'cause it seems
like a lot more people are

855
00:45:31,454 --> 00:45:35,287
working on it recently, and have
good go-to market strategies to

856
00:45:35,287 --> 00:45:38,281
get there for this kind of
thing. At least from my

857
00:45:38,281 --> 00:45:39,000
perspective.

858
00:45:39,000 --> 00:45:41,000
Stefan Seltz-Axmacher: Yeah,
absolutely.

859
00:45:41,000 --> 00:45:44,229
Audrow Nash: and Stefan, what do
you hope, like where are you,

860
00:45:44,229 --> 00:45:46,000
what are your ambitions with
this?

861
00:45:46,000 --> 00:45:49,464
Stefan Seltz-Axmacher: yeah, so
my last company, start was my

862
00:45:49,464 --> 00:45:52,927
first ever startup, and it was
really hard. And at the time I

863
00:45:52,927 --> 00:45:56,278
didn't know how much, it was
hard because startups are hard

864
00:45:56,278 --> 00:45:59,401
and how much, it was hard
because I was building robots

865
00:45:59,401 --> 00:46:02,694
and running a trucking company
and writing regulations and

866
00:46:02,694 --> 00:46:04,000
blah, blah, blah, blah,

867
00:46:04,000 --> 00:46:07,000
Audrow Nash: You were five CEOs
at once? Yeah, for sure.

868
00:46:07,000 --> 00:46:11,009
Stefan Seltz-Axmacher: and then
after that I went and hung out

869
00:46:11,009 --> 00:46:14,954
with some SaaS companies and
holy crap. like, assuming Audrow

870
00:46:14,954 --> 00:46:18,511
that you've never worked at a,
at a, just a normal SaaS

871
00:46:18,511 --> 00:46:22,262
company, DocuSign 5.0 SaaS
Company. The way that you start

872
00:46:22,262 --> 00:46:26,400
a SaaS company like that is you
have a producty salesy, CEO. You

873
00:46:26,400 --> 00:46:30,022
have a generalist, 50th
percentile engineer who can talk

874
00:46:30,022 --> 00:46:33,772
to people but can mostly write
glue code. That product guy

875
00:46:33,772 --> 00:46:37,588
figures out like what you should
be building. The glue code

876
00:46:37,588 --> 00:46:40,821
engineer finds a bunch of
microservices, ties them

877
00:46:40,821 --> 00:46:44,637
together to build it. And like
theoretically in five years,

878
00:46:44,637 --> 00:46:48,581
they're each, worth a couple
hundred million dollars. Whereas

879
00:46:48,581 --> 00:46:52,461
in Robotics it looks almost
like, it looks like the world of

880
00:46:52,461 --> 00:46:56,536
PCs back in the seventies where
screw Steve Jobs. You need like

881
00:46:56,536 --> 00:47:00,480
Steve Wozniak who can like code
and binary and design his own

882
00:47:00,480 --> 00:47:04,490
circuit boards and build gooeys
and make them useful. And like

883
00:47:04,490 --> 00:47:08,046
without an army of Steve
Wozniaks, you have no product.

884
00:47:08,046 --> 00:47:12,056
Like the Robotics industry will
never get beyond that phase if

885
00:47:12,056 --> 00:47:16,065
everyone has to keep on building
point-to-point navigation. If

886
00:47:16,065 --> 00:47:19,881
Polymath can enable the market
to just focus on things that

887
00:47:19,881 --> 00:47:24,019
work and it's still not gonna be
as easy as DocuSign 5.0. But if

888
00:47:24,019 --> 00:47:27,641
Polymath makes it so that
building a robot is as easy as

889
00:47:27,641 --> 00:47:31,262
making a website in 1998, then
like we're gonna start to

890
00:47:31,262 --> 00:47:34,884
actually have robots. We're
gonna actually start to have

891
00:47:34,884 --> 00:47:38,634
cheaper, better food. We're
gonna have, cheaper materials.

892
00:47:38,634 --> 00:47:42,579
We're gonna have cheaper cost of
living. We're gonna enable a

893
00:47:42,579 --> 00:47:46,394
whole lot of the world to live
as nice as we do in, in the,

894
00:47:46,394 --> 00:47:49,951
west. And I think that like
along the way, if we enable

895
00:47:49,951 --> 00:47:54,025
that, we'll make a whole bunch
of money and build a cool thing.

896
00:47:54,025 --> 00:47:57,000
So that's what I think, that's
what I'm hoping

897
00:47:57,000 --> 00:48:01,823
Audrow Nash: Good for you. And,
exciting problems. 'cause to me

898
00:48:01,823 --> 00:48:06,263
what it seems like it's not,
it's a lot more exciting than

899
00:48:06,263 --> 00:48:11,239
this, but it sounds a little bit
like a lifestyle startup in some

900
00:48:11,239 --> 00:48:13,000
way. And what I mean by

901
00:48:13,000 --> 00:48:15,000
Stefan Seltz-Axmacher: What do
you mean?

902
00:48:15,000 --> 00:48:18,688
Audrow Nash: so what I mean is
that you get to leverage the

903
00:48:18,688 --> 00:48:22,563
power of the, SaaS model in some
sense where you don't have to

904
00:48:22,563 --> 00:48:26,000
be, you're, only one CEO at this
time. You're not five.

905
00:48:26,000 --> 00:48:27,000
Stefan Seltz-Axmacher: yeah,

906
00:48:26,000 --> 00:48:29,961
Audrow Nash: And so it, it seems
like it's a bit easier. You guys

907
00:48:29,961 --> 00:48:33,799
are profitable early, so you're
not like lighting the ground on

908
00:48:33,799 --> 00:48:37,577
fire behind you and running like
you would be if you're taking

909
00:48:37,577 --> 00:48:40,989
investment. maybe lifestyle
isn't the way to say it, but

910
00:48:40,989 --> 00:48:43,000
like it's more on your terms in
a

911
00:48:43,000 --> 00:48:45,000
Stefan Seltz-Axmacher: Yeah. we
get to be,

912
00:48:45,000 --> 00:48:48,233
Ilia Baranov: you're a, you,
accidentally slandered, Stefan

913
00:48:48,233 --> 00:48:49,000
there by using

914
00:48:49,000 --> 00:48:50,000
Stefan Seltz-Axmacher: I, I,

915
00:48:50,000 --> 00:48:53,562
Ilia Baranov: lifestyle. No
kidding, kidding, Purely

916
00:48:53,562 --> 00:48:57,740
kidding. VCs will use lifestyle
business as like something we

917
00:48:57,740 --> 00:49:00,000
shouldn't invest in, which is
not

918
00:49:00,000 --> 00:49:03,000
Stefan Seltz-Axmacher: like, why
would you ever want to just be

919
00:49:03,000 --> 00:49:05,000
rich and have a nice life that's
terrible.

920
00:49:05,000 --> 00:49:06,000
Audrow Nash: You,

921
00:49:06,000 --> 00:49:08,795
Ilia Baranov: No, you have to
make them rich first. That's

922
00:49:08,795 --> 00:49:10,000
your goal. Yeah. exactly.

923
00:49:10,000 --> 00:49:13,000
Audrow Nash: real good, you do
our 10, 10 CEOs. Yeah,

924
00:49:13,000 --> 00:49:16,296
Stefan Seltz-Axmacher: what's
funny is compared to a normal

925
00:49:16,296 --> 00:49:19,592
full stack vertical Robotics
company, this feels incredibly

926
00:49:19,592 --> 00:49:23,000
easy compared to my friends in
SaaS. This is still incredibly

927
00:49:23,000 --> 00:49:25,000
Audrow Nash: hard.

928
00:49:25,000 --> 00:49:29,296
Stefan Seltz-Axmacher: and I
think like even, in 10 years or

929
00:49:29,296 --> 00:49:33,234
so, the, some of the modules
that we sell as individual

930
00:49:33,234 --> 00:49:37,888
products that like, as a part of
our stack, there will be billion

931
00:49:37,888 --> 00:49:41,969
dollar companies that just do
things like SLAM or just do

932
00:49:41,969 --> 00:49:46,551
really clever path planning for
diff drive, vehicles. there will

933
00:49:46,551 --> 00:49:50,919
be industries like that, but
right now it's rather radical to

934
00:49:50,919 --> 00:49:55,000
say, oh, we'll just take the
autonomy core off your hands

935
00:49:55,000 --> 00:49:58,000
Audrow Nash: You mean make it so
they don't have to do that?

936
00:49:58,000 --> 00:49:59,000
Stefan Seltz-Axmacher: Yeah.

937
00:49:58,000 --> 00:49:59,000
Audrow Nash: you can

938
00:49:59,000 --> 00:50:02,150
Stefan Seltz-Axmacher: Right now
that, seems radically easy. In

939
00:50:02,150 --> 00:50:04,000
10 years it will seem radically
hard.

940
00:50:04,000 --> 00:50:07,932
Audrow Nash: Yeah. It is amazing
how that seems. Like building a

941
00:50:07,932 --> 00:50:11,495
website now is still difficult.
if you wanna build a fully

942
00:50:11,495 --> 00:50:15,181
featured SaaS stuff, like a lot
of my side projects are SaaS

943
00:50:15,181 --> 00:50:19,174
stuff. and like it's still a lot
of work. but it's so much easier

944
00:50:19,174 --> 00:50:22,000
than Robotics projects, which is
just bonkers.

945
00:50:22,000 --> 00:50:25,095
Stefan Seltz-Axmacher: let,
lemme tell you a version of

946
00:50:25,095 --> 00:50:28,641
that. I had a side project, an
e-commerce side project a couple

947
00:50:28,641 --> 00:50:32,242
months ago. I've never actually
publicly in any setting admitted

948
00:50:32,242 --> 00:50:35,000
was, me. So this is like an
Audrow Nash exclusive

949
00:50:35,000 --> 00:50:36,000
Audrow Nash: Cool. You heard it
here.

950
00:50:36,000 --> 00:50:40,606
Stefan Seltz-Axmacher: So yeah,
so early last year, I got wind

951
00:50:40,606 --> 00:50:44,914
on a, WhatsApp group for YC
founders that, something weird

952
00:50:44,914 --> 00:50:49,000
was going on at SVB. and like a
panic was happening and

953
00:50:49,000 --> 00:50:51,195
Audrow Nash: That was Silicon
Valley Bank, that big one that

954
00:50:51,195 --> 00:50:52,000
collapsed a while ago.

955
00:50:52,000 --> 00:50:55,849
Stefan Seltz-Axmacher: yep. And
I panicked early, like I

956
00:50:55,849 --> 00:50:59,699
panicked on that Wednesday
night, which meant we got our

957
00:50:59,699 --> 00:51:03,961
money out like relatively soon.
But while I panicked, I had a,

958
00:51:03,961 --> 00:51:07,948
like a dark joke that like the
Silicon Valley Bank run was

959
00:51:07,948 --> 00:51:12,209
kinda like a marathon run. so I
went on Fiverr and I basically

960
00:51:12,209 --> 00:51:16,334
paid a Fiverr, I think 20 bucks
to make a t-shirt design for

961
00:51:16,334 --> 00:51:20,733
what a bank run marathon T-shirt
would look like. so 25 bucks, I

962
00:51:20,733 --> 00:51:25,133
bought the domain name, I think
SVB run 23, which sounded like a

963
00:51:25,133 --> 00:51:29,188
good marathon run. so that was
10, $15. And then I set up a

964
00:51:29,188 --> 00:51:33,313
store on Teespring, and I don't
know if you're familiar with

965
00:51:33,313 --> 00:51:34,000
Teespring,

966
00:51:34,000 --> 00:51:37,000
Audrow Nash: Yeah. Open.
Robotics uses it

967
00:51:37,000 --> 00:51:40,662
Stefan Seltz-Axmacher: Oh, cool.
So for those of, for the

968
00:51:40,662 --> 00:51:44,260
listeners who don't know,
Teespring, Teespring will make

969
00:51:44,260 --> 00:51:48,115
it easy to do like a screen
printed shirt business where you

970
00:51:48,115 --> 00:51:52,034
can just upload a shirt design
and they will print the shirts

971
00:51:52,034 --> 00:51:55,824
and ship them and do all of
that. So I paid for the premium

972
00:51:55,824 --> 00:51:59,936
version of Teespring for I think
20 bucks. And for $75, I had an

973
00:51:59,936 --> 00:52:03,983
e-commerce site with a line of
nine different takes on SVB Bank

974
00:52:03,983 --> 00:52:07,709
run marathon shirts, which I
posted in a handful of links.

975
00:52:07,709 --> 00:52:11,885
And I suddenly had an e-commerce
site that got me, I was a top of

976
00:52:11,885 --> 00:52:16,061
1% Teespring creator that month.
and if that was a real business,

977
00:52:16,061 --> 00:52:19,852
that could be a thing that I
did. And, it has nothing to do

978
00:52:19,852 --> 00:52:23,000
with robotics in terms of how
hard it is. It is a

979
00:52:23,000 --> 00:52:25,891
Ilia Baranov: I add a little bit
of, can I add a little bit of

980
00:52:25,891 --> 00:52:28,922
late breaking news? This will be
a surprise to Stefan too, in the

981
00:52:28,922 --> 00:52:31,813
full circle of e-commerce. You
are also immediately ripped off

982
00:52:31,813 --> 00:52:32,000
by a

983
00:52:32,000 --> 00:52:34,000
Stefan Seltz-Axmacher: yeah.

984
00:52:34,000 --> 00:52:35,000
Audrow Nash: Oh,

985
00:52:34,000 --> 00:52:36,000
Ilia Baranov: This is your exact
design,

986
00:52:36,000 --> 00:52:37,000
Audrow Nash: no way.

987
00:52:37,000 --> 00:52:38,000
Ilia Baranov: seller

988
00:52:37,000 --> 00:52:40,000
Stefan Seltz-Axmacher: that's on
Etsy.

989
00:52:40,000 --> 00:52:41,000
Ilia Baranov: Etsy.

990
00:52:41,000 --> 00:52:44,600
Stefan Seltz-Axmacher: what a
schmuck. There was also, there's

991
00:52:44,600 --> 00:52:48,142
also another person ripped me
off on, Teespring and Teespring

992
00:52:48,142 --> 00:52:50,000
refused to do anything about it.

993
00:52:50,000 --> 00:52:52,000
Ilia Baranov: yeah.

994
00:52:52,000 --> 00:52:53,000
Stefan Seltz-Axmacher: But, I
think

995
00:52:52,000 --> 00:52:54,000
Ilia Baranov: of e-Commerce
Life.

996
00:52:54,000 --> 00:52:56,950
Stefan Seltz-Axmacher: but like
you can literally set up an

997
00:52:56,950 --> 00:52:59,900
e-commerce store without like
even opening up Photoshop, to

998
00:52:59,900 --> 00:53:02,900
make your own products. Like
Robotics is not gonna look like

999
00:53:02,900 --> 00:53:05,700
that for at least 20 to 30
years, but that's. If, robots

1000
00:53:05,700 --> 00:53:08,850
are gonna be a real part of our
life, if robots are gonna bring

1001
00:53:08,850 --> 00:53:11,700
us the promise that I think
everyone in this conversation

1002
00:53:11,700 --> 00:53:14,800
wants them to bring, they need
to get a lot easier. And that's

1003
00:53:14,800 --> 00:53:16,000
what we're hoping to do.

1004
00:53:16,000 --> 00:53:21,626
Audrow Nash: Hell yeah. Okay. I
really like that. what's the, so

1005
00:53:21,626 --> 00:53:24,000
you guys are 10 people now.

1006
00:53:24,000 --> 00:53:26,000
Stefan Seltz-Axmacher: Damn Etsy
store.

1007
00:53:26,000 --> 00:53:29,916
Audrow Nash: I know that's so
fricking funny. you made it that

1008
00:53:29,916 --> 00:53:32,000
you got ripped off. That's
great.

1009
00:53:32,000 --> 00:53:35,523
Ilia Baranov: I think we need to
open an Etsy store for Polymath

1010
00:53:35,523 --> 00:53:39,046
shirts. By the way, like the one
that Stefan's wearing right now

1011
00:53:39,046 --> 00:53:42,569
for the listeners, he's wearing
a shirt that just says Total ROS

1012
00:53:42,569 --> 00:53:45,817
bag on it. because we had an
internal joke at Polymath that

1013
00:53:45,817 --> 00:53:49,174
oh, anytime anybody would do
something bad, it'd be like, ah,

1014
00:53:49,174 --> 00:53:50,000
you ROS Bag. So

1015
00:53:50,000 --> 00:53:52,000
Stefan Seltz-Axmacher: It
doesn't work as well with mcap.

1016
00:53:52,000 --> 00:53:57,083
Ilia Baranov: you. Yeah. we also
have MCAP hats. We have MCAP

1017
00:53:57,083 --> 00:53:59,000
hats too. Yeah. Anyway,

1018
00:53:59,000 --> 00:54:03,479
Audrow Nash: Hell yeah. let's
see, I lost my train of thought

1019
00:54:03,479 --> 00:54:07,958
with this. That's very funny.
All the swag at everything. Oh,

1020
00:54:07,958 --> 00:54:12,584
so what's the, growth plan for
this? so you guys are, 10 people

1021
00:54:12,584 --> 00:54:16,916
now you're, that line is going
up, so it sounds like you're

1022
00:54:16,916 --> 00:54:20,000
profitable or becoming like
you're, you'll

1023
00:54:20,000 --> 00:54:21,419
Stefan Seltz-Axmacher: we're not
profitable yet, but it seems

1024
00:54:21,419 --> 00:54:22,000
like we're, it seems like

1025
00:54:22,000 --> 00:54:25,623
Audrow Nash: on the trajectory,
two years and starting to move

1026
00:54:25,623 --> 00:54:29,188
in the right direction. That's
wonderful. how, do you imagine

1027
00:54:29,188 --> 00:54:32,987
growing or do you need many more
people for the work you're doing

1028
00:54:32,987 --> 00:54:36,494
or will it be like, there'll be
like system integrators that

1029
00:54:36,494 --> 00:54:40,000
help push it out and maybe you
can focus on a few promising,

1030
00:54:40,000 --> 00:54:41,000
Stefan Seltz-Axmacher: Yeah.

1031
00:54:41,000 --> 00:54:44,282
Audrow Nash: areas or like what
are you thinking for how to grow

1032
00:54:44,282 --> 00:54:45,000
in the future?

1033
00:54:45,000 --> 00:54:47,439
Stefan Seltz-Axmacher: So it's
funny, like right now more of

1034
00:54:47,439 --> 00:54:50,000
our problems are like business
problems than they are, Robotics

1035
00:54:50,000 --> 00:54:51,000
Audrow Nash: Technical

1036
00:54:51,000 --> 00:54:54,714
Stefan Seltz-Axmacher: is also
weird. Like I didn't think that

1037
00:54:54,714 --> 00:54:58,429
could happen in Robotics. So we
need, like, we're, essentially

1038
00:54:58,429 --> 00:55:02,143
we might go raise again or we
might just grow to profitability

1039
00:55:02,143 --> 00:55:05,857
depending on how the VC market's
at and which could be a whole

1040
00:55:05,857 --> 00:55:09,332
other conversation. but if we
suddenly had 10 more people,

1041
00:55:09,332 --> 00:55:13,106
probably half of them would be
business people doing stuff like

1042
00:55:13,106 --> 00:55:16,581
account management and like
splitting functions into other

1043
00:55:16,581 --> 00:55:17,000
things.

1044
00:55:17,000 --> 00:55:20,274
Audrow Nash: It's interesting to
me 'cause it feels like with a

1045
00:55:20,274 --> 00:55:23,444
lot of hardware companies, say
like medical device companies,

1046
00:55:23,444 --> 00:55:26,250
as the company moves across
time, the staffing changes

1047
00:55:26,250 --> 00:55:29,317
pretty significantly. So you
have a lot of like researchers

1048
00:55:29,317 --> 00:55:32,071
and our engineers at the
beginning and then once they

1049
00:55:32,071 --> 00:55:35,137
make something it's okay, now
you need way less of them and

1050
00:55:35,137 --> 00:55:38,359
now you need like a big legal
team or certification or I don't

1051
00:55:38,359 --> 00:55:41,633
know, whatever. Then marketing
and who knows. but it seems like

1052
00:55:41,633 --> 00:55:44,803
you guys are going through that
as well where, okay, we had a

1053
00:55:44,803 --> 00:55:47,869
bunch of engineers, eight of the
10 were engineers. Now you

1054
00:55:47,869 --> 00:55:50,000
probably need another 10
business people.

1055
00:55:50,000 --> 00:55:52,000
Stefan Seltz-Axmacher: another
five.

1056
00:55:52,000 --> 00:55:54,000
Audrow Nash: Okay, something
like

1057
00:55:54,000 --> 00:55:57,403
Ilia Baranov: and I think the
key metric that I convinced

1058
00:55:57,403 --> 00:56:00,388
Stefan over time to start
tracking is our robot to

1059
00:56:00,388 --> 00:56:02,000
employee ratio. I really my

1060
00:56:02,000 --> 00:56:03,000
Audrow Nash: it right at

1061
00:56:02,000 --> 00:56:07,118
Ilia Baranov: is to get to 10 to
a hundred Yeah. Robots to

1062
00:56:07,118 --> 00:56:08,000
employees.

1063
00:56:08,000 --> 00:56:11,148
Stefan Seltz-Axmacher: and also
when IA says that I, IA actually

1064
00:56:11,148 --> 00:56:14,000
means different types of robots,
not even specific robots.

1065
00:56:14,000 --> 00:56:17,030
Ilia Baranov: not even copies.
Yeah. 'cause copies to me are

1066
00:56:17,030 --> 00:56:19,000
like, eh, whatever. It's the
exact same

1067
00:56:19,000 --> 00:56:21,000
Audrow Nash: they're free. You
already have it working,

1068
00:56:21,000 --> 00:56:22,000
Ilia Baranov: yeah, exactly.

1069
00:56:22,000 --> 00:56:26,255
Audrow Nash: but I guess those
1% cases will come up more, for

1070
00:56:26,255 --> 00:56:29,000
this kind of thing. So not
totally free.

1071
00:56:29,000 --> 00:56:32,000
Ilia Baranov: No, yeah. I'm,
big, a

1072
00:56:32,000 --> 00:56:35,000
Audrow Nash: Facetious. Yeah,
for sure.

1073
00:56:35,000 --> 00:56:38,745
Ilia Baranov: but yeah, no, I
think as we grow, I think we

1074
00:56:38,745 --> 00:56:42,361
will still need quite a
significant engineering presence

1075
00:56:42,361 --> 00:56:46,235
because one of the reasons we
sell ourselves as SaaS is that

1076
00:56:46,235 --> 00:56:49,722
we're continuously improving
every piece of our stack.

1077
00:56:49,722 --> 00:56:53,790
especially actually we can get
into this in technical detail in

1078
00:56:53,790 --> 00:56:57,987
a bit, but, the machine learning
part of what we do is actually a

1079
00:56:57,987 --> 00:57:01,603
very small, thin. Sugar
sprinkling on top of the cake in

1080
00:57:01,603 --> 00:57:05,735
a lot of ways where it's, a key
differentiator in some ways, but

1081
00:57:05,735 --> 00:57:09,868
in other ways actually, there's
so much fundamental to do at the

1082
00:57:09,868 --> 00:57:13,613
underlying layer that we gotta
get that large and expanded

1083
00:57:13,613 --> 00:57:14,000
first.

1084
00:57:14,000 --> 00:57:15,000
Audrow Nash: What's the, what is
the

1085
00:57:15,000 --> 00:57:16,000
Stefan Seltz-Axmacher: I think
to,

1086
00:57:16,000 --> 00:57:17,000
Audrow Nash: of your

1087
00:57:17,000 --> 00:57:18,000
Stefan Seltz-Axmacher: yeah.

1088
00:57:18,000 --> 00:57:20,000
Audrow Nash: about that at all.
I, don't know what it

1089
00:57:20,000 --> 00:57:23,216
Ilia Baranov: I'll give you a
con, I'll give you a concrete

1090
00:57:23,216 --> 00:57:26,105
example. So we have an
application in forestry and we

1091
00:57:26,105 --> 00:57:29,321
have a LIDAR system. And you can
imagine in a dense forest,

1092
00:57:29,321 --> 00:57:32,210
you're not getting very much
GPS. So there's a lot of

1093
00:57:32,210 --> 00:57:34,990
complexity on localization,
orientation and heading

1094
00:57:34,990 --> 00:57:38,478
management, obstacle avoidance,
those kind of things. but from a

1095
00:57:38,478 --> 00:57:41,748
LIDAR point cloud, you can't
really tell if the thing you're

1096
00:57:41,748 --> 00:57:44,964
seeing is a bush, which you're
supposed to just drive right

1097
00:57:44,964 --> 00:57:48,180
over 'cause you're a multi ton
vehicle. Or is it a boulder,

1098
00:57:48,180 --> 00:57:51,614
which if you try to drive over,
you'll damage your drive train,

1099
00:57:51,614 --> 00:57:54,885
right? And that's sort of stuff
that's where ML will come to

1100
00:57:54,885 --> 00:57:58,101
play a role with, image data or
LIDAR data, which classical

1101
00:57:58,101 --> 00:58:01,153
algorithms can't do quite as
well. You can think of some

1102
00:58:01,153 --> 00:58:04,642
heuristics maybe, but honestly,
like any heuristic you can think

1103
00:58:04,642 --> 00:58:07,912
of, a bush in a boulder will
look similar at some point from

1104
00:58:07,912 --> 00:58:11,237
some angle. whereas camera data
and humans can tell 'em apart

1105
00:58:11,237 --> 00:58:12,000
pretty easily.

1106
00:58:12,000 --> 00:58:13,000
Audrow Nash: So the.

1107
00:58:12,000 --> 00:58:15,202
Ilia Baranov: that's a case
where that extra guidance, even

1108
00:58:15,202 --> 00:58:18,512
though the, all the underlying
pieces to navigate and pathway

1109
00:58:18,512 --> 00:58:21,713
and obstacle void, the extra
guidance allows us to not make

1110
00:58:21,713 --> 00:58:25,078
suboptimal decisions of avoiding
a bush when we can just drive

1111
00:58:25,078 --> 00:58:26,000
right through it.

1112
00:58:26,000 --> 00:58:29,405
Audrow Nash: Gotcha. Yeah. So
these are in your modular

1113
00:58:29,405 --> 00:58:33,305
framework. These are components
you can add, that might improve

1114
00:58:33,305 --> 00:58:37,019
applications. Maybe you say,
okay, I have this path planning

1115
00:58:37,019 --> 00:58:40,114
and this path planning
effectively has a plugin or

1116
00:58:40,114 --> 00:58:44,076
something that says, is Bush or
not is bush or rock or something

1117
00:58:44,076 --> 00:58:47,790
like this. and you can have
machine learning determine which

1118
00:58:47,790 --> 00:58:51,690
one it is. So that's interesting
that you guys are getting into

1119
00:58:51,690 --> 00:58:52,000
that.

1120
00:58:52,000 --> 00:58:55,969
Stefan Seltz-Axmacher: but like
a relatively weird opinion.

1121
00:58:55,969 --> 00:58:59,870
Yeah. But like a relatively
weird opinion that we have. Is

1122
00:58:59,870 --> 00:59:03,637
in Robotics. there's a lot more
ML in MO in the world of

1123
00:59:03,637 --> 00:59:07,336
Robotics and the market of
Robotics and the products of

1124
00:59:07,336 --> 00:59:11,709
Robotics that you can buy. ML is
a lot more commodified than like

1125
00:59:11,709 --> 00:59:15,677
really good Robotics software
that is reliable and turns on

1126
00:59:15,677 --> 00:59:19,780
every time and spits out metrics
for why it doesn't work. And

1127
00:59:19,780 --> 00:59:23,547
deployment infrastructure so
that you don't have to send

1128
00:59:23,547 --> 00:59:27,112
people to west, west
Nowheresville, to go fix a robot

1129
00:59:27,112 --> 00:59:31,148
in the middle of the day night
with a, eight hour notice or,

1130
00:59:31,148 --> 00:59:35,251
controls engineering that can
smoothly drive vehicles of very

1131
00:59:35,251 --> 00:59:39,152
various different sizes. And I
think, I think there's this

1132
00:59:39,152 --> 00:59:43,256
weird market dynamic going on
where the VCs are really stoked

1133
00:59:43,256 --> 00:59:47,359
about like machine learning.
'cause they just got to see what

1134
00:59:47,359 --> 00:59:51,596
like LLMs can do and like a way
that they can directly interact

1135
00:59:51,596 --> 00:59:55,901
with and they're pretty amazed.
Whereas I don't know, I'm pretty

1136
00:59:55,901 --> 00:59:59,937
stoked by robots that turn on
every time. And I know how few

1137
00:59:59,937 --> 01:00:03,637
large Robotics organizations
that is true for, so we're

1138
01:00:03,637 --> 01:00:07,000
focusing a lot more on those
kind of fundamentals.

1139
01:00:07,000 --> 01:00:11,421
Audrow Nash: Yeah, no, I think
that's smart. a lot of the, I

1140
01:00:11,421 --> 01:00:15,326
don't know, it's very
interesting to me because there

1141
01:00:15,326 --> 01:00:19,674
are some things that they, like
machine learning seems well

1142
01:00:19,674 --> 01:00:24,463
suited for. but there's a lot of
things where we think it's good,

1143
01:00:24,463 --> 01:00:28,958
but then it isn't when you try
it. And, like the problem with

1144
01:00:28,958 --> 01:00:33,600
Robotics now really seems to be
a lot of exceptions pop up when

1145
01:00:33,600 --> 01:00:37,726
you bring robots into a less
structured environment. and

1146
01:00:37,726 --> 01:00:41,926
those, I just, it's very hard to
imagine a lot of machine

1147
01:00:41,926 --> 01:00:46,421
learning, doing very well with a
lot of those. I just I don't

1148
01:00:46,421 --> 01:00:50,695
know, I'm doing some thorny
coding, for work and the, it's

1149
01:00:50,695 --> 01:00:55,263
it's complex. There's all sorts
of hidden stuff that's hard to

1150
01:00:55,263 --> 01:00:59,979
understand and, I'm asking LLMs
advice on how to do this kind of

1151
01:00:59,979 --> 01:01:04,547
thing and. All their advice is
garbage for this kind of thing.

1152
01:01:04,547 --> 01:01:08,821
It's like I've actually gone to,
looking at these and just

1153
01:01:08,821 --> 01:01:13,021
Googling, because it is le it's
a, it's like it will look

1154
01:01:13,021 --> 01:01:17,368
convincing and it looks like
it's gonna work and then I try

1155
01:01:17,368 --> 01:01:21,863
it and it just doesn't, and I
can imagine that this pops up a

1156
01:01:21,863 --> 01:01:24,000
lot in Robotics applications.

1157
01:01:24,000 --> 01:01:27,369
Stefan Seltz-Axmacher: so the
exceptions thing is actually

1158
01:01:27,369 --> 01:01:31,029
part of why off highway or like
large vehicles are really neat.

1159
01:01:31,029 --> 01:01:34,340
So basically all of these
different industries, from like

1160
01:01:34,340 --> 01:01:38,000
ag to mining to whatever suffer
labor shortages they like, they

1161
01:01:38,000 --> 01:01:40,000
Audrow Nash: everything is
suffering labor

1162
01:01:40,000 --> 01:01:43,250
Stefan Seltz-Axmacher: spend
enough money to hire enough

1163
01:01:43,250 --> 01:01:46,674
people to do the work. So as a
result, they need, a smaller

1164
01:01:46,674 --> 01:01:50,388
number of people to do the same
amount of work, which means that

1165
01:01:50,388 --> 01:01:53,000
machines get bigger and bigger.
when machines

1166
01:01:53,000 --> 01:01:56,155
Audrow Nash: It's like shipping
containers that are just

1167
01:01:56,155 --> 01:01:57,000
absurdly large.

1168
01:01:57,000 --> 01:01:58,000
Stefan Seltz-Axmacher: yep.

1169
01:01:58,000 --> 01:02:02,000
Audrow Nash: Yeah. So they make
one person do what 20 used to do

1170
01:02:02,000 --> 01:02:03,000
Stefan Seltz-Axmacher: so
there's

1171
01:02:02,000 --> 01:02:05,677
Audrow Nash: 20 times larger
machine, this kind of thing.

1172
01:02:05,677 --> 01:02:06,000
Okay.

1173
01:02:06,000 --> 01:02:09,862
Stefan Seltz-Axmacher: And
what's cool about that? First of

1174
01:02:09,862 --> 01:02:13,527
all, the machine does more
valuable work. You, can spend

1175
01:02:13,527 --> 01:02:17,651
more money on sensors for it. So
that's useful for us. but more

1176
01:02:17,651 --> 01:02:21,513
than that, big machines are
really dangerous, so they drive

1177
01:02:21,513 --> 01:02:24,000
more cautiously. a 500 ton dump
truck.

1178
01:02:24,000 --> 01:02:25,000
Audrow Nash: interesting.

1179
01:02:25,000 --> 01:02:29,000
Stefan Seltz-Axmacher: yep.
Basically the most valuable

1180
01:02:29,000 --> 01:02:33,000
machines in the world to
automate are among the easiest

1181
01:02:33,000 --> 01:02:35,000
Audrow Nash: Aha. That's
wonderful. Yeah. 'cause Oh,

1182
01:02:35,000 --> 01:02:39,881
Ilia Baranov: giant mining truck
or the molten metal transporter

1183
01:02:39,881 --> 01:02:44,000
is way easier than astro in your
house, like orders of

1184
01:02:44,000 --> 01:02:47,986
Stefan Seltz-Axmacher: order.
Yeah. and like the 500 ton, dump

1185
01:02:47,986 --> 01:02:51,843
truck, I think Caterpillar gets
a million dollars a year for

1186
01:02:51,843 --> 01:02:53,000
each one of those.

1187
01:02:53,000 --> 01:02:54,000
Ilia Baranov: recurring,

1188
01:02:53,000 --> 01:02:55,000
Stefan Seltz-Axmacher: recurring
SaaS.

1189
01:02:55,000 --> 01:02:58,007
Audrow Nash: Wow. That's so
crazy. What a funny thing that

1190
01:02:58,007 --> 01:03:01,170
just, it makes complete sense,
but I hadn't thought about it.

1191
01:03:01,170 --> 01:03:04,489
Where the ones that are the most
valuable and then also like you

1192
01:03:04,489 --> 01:03:07,807
could strap expensive sensors on
them, and relative to the total

1193
01:03:07,807 --> 01:03:09,000
cost of the thing, it's

1194
01:03:09,000 --> 01:03:12,643
Ilia Baranov: And they're big,
so there's a lot of space for

1195
01:03:12,643 --> 01:03:16,164
them. There's a lot of power,
there's a lot of storage for

1196
01:03:16,164 --> 01:03:19,503
compute. Like all of your
problems get easier at larger

1197
01:03:19,503 --> 01:03:22,721
scales. There's also an
interesting, to harken back a

1198
01:03:22,721 --> 01:03:25,999
little bit to where we see
things going as our type of

1199
01:03:25,999 --> 01:03:29,702
approaches are hopefully us get
more and more successful. You

1200
01:03:29,702 --> 01:03:33,588
don't need these large vehicles
anymore. 'cause one of the labor

1201
01:03:33,588 --> 01:03:36,805
shortages You can go back to
small. and which kind of

1202
01:03:36,805 --> 01:03:40,266
dovetails nicely with sensors
getting smaller and compute

1203
01:03:40,266 --> 01:03:43,787
getting more efficient over
time. but there's a funny kind

1204
01:03:43,787 --> 01:03:47,490
of story I heard from the South
African mining industry where

1205
01:03:47,490 --> 01:03:50,890
they use, they're, very active
in mining. They have good

1206
01:03:50,890 --> 01:03:54,654
regulatory body for autonomy,
but they have a black market for

1207
01:03:54,654 --> 01:03:58,418
tires for these large, vehicles
where getting a new tire if it

1208
01:03:58,418 --> 01:04:02,000
blows a tire is like a four
month process. And in that time

1209
01:04:02,000 --> 01:04:03,000
Stefan Seltz-Axmacher: And

1210
01:04:03,000 --> 01:04:04,000
Ilia Baranov: of dollars.

1211
01:04:04,000 --> 01:04:07,107
Stefan Seltz-Axmacher: that
truck might be responsible for

1212
01:04:07,107 --> 01:04:10,589
moving $10 million in material a
day. So if it's outta commission

1213
01:04:10,589 --> 01:04:13,000
for 90 days, yeah. We've lost a
lot of money.

1214
01:04:13,000 --> 01:04:15,735
Ilia Baranov: Enormous, money.
And so it's, the incentives are

1215
01:04:15,735 --> 01:04:18,338
so high that they have people
running around stealing tires

1216
01:04:18,338 --> 01:04:19,000
from each other

1217
01:04:19,000 --> 01:04:20,000
Audrow Nash: Ah,

1218
01:04:20,000 --> 01:04:23,375
Ilia Baranov: like at that
point, you'll pay a group a

1219
01:04:23,375 --> 01:04:27,125
million dollars to get you a
tire, no problem. so if you're,

1220
01:04:27,125 --> 01:04:31,000
if, and these size vehicles are
so large that they get shipped

1221
01:04:31,000 --> 01:04:33,000
in pieces to the site, assembled

1222
01:04:33,000 --> 01:04:34,000
Audrow Nash: then assembled by

1223
01:04:34,000 --> 01:04:35,000
Ilia Baranov: their entire life
on site,

1224
01:04:35,000 --> 01:04:37,000
Audrow Nash: Yeah.

1225
01:04:37,000 --> 01:04:40,191
Ilia Baranov: and then left
there, and then left there. When

1226
01:04:40,191 --> 01:04:42,000
the site is done, the vehicles
are

1227
01:04:42,000 --> 01:04:43,000
Audrow Nash: They're not even
worth

1228
01:04:43,000 --> 01:04:44,000
Ilia Baranov: You don't have to

1229
01:04:44,000 --> 01:04:45,000
Stefan Seltz-Axmacher: Yeah.

1230
01:04:45,000 --> 01:04:47,971
Ilia Baranov: yet, you don't
have to do any of that. If you

1231
01:04:47,971 --> 01:04:51,194
have a smaller vehicle, you just
put 'em on the bat of a flatbed

1232
01:04:51,194 --> 01:04:54,014
truck and you just have a,
instead of one giant one, you

1233
01:04:54,014 --> 01:04:57,036
have 10 medium or, 30 small
size. If one of them breaks, you

1234
01:04:57,036 --> 01:04:59,000
only lose a 30th of your
effectiveness.

1235
01:04:59,000 --> 01:05:01,829
Stefan Seltz-Axmacher: and when
they get smaller, they get

1236
01:05:01,829 --> 01:05:03,000
easier to electrify too.

1237
01:05:03,000 --> 01:05:07,048
Audrow Nash: Oh, that's a cool
point. Okay. Which that's

1238
01:05:07,048 --> 01:05:11,096
desirable because it's just
clean, easy to maintain. You

1239
01:05:11,096 --> 01:05:15,000
don't need big, like diesel,
high pressure, hydraulic,

1240
01:05:15,000 --> 01:05:19,858
Stefan Seltz-Axmacher: I'm, if
you like the idea of sometimes

1241
01:05:19,858 --> 01:05:24,000
there being snow, more electric
vehicles seems nice.

1242
01:05:24,000 --> 01:05:26,000
Audrow Nash: is that? That
doesn't make.

1243
01:05:26,000 --> 01:05:28,000
Stefan Seltz-Axmacher: Sorry. if
you like the idea of not

1244
01:05:28,000 --> 01:05:29,000
Ilia Baranov: change

1245
01:05:29,000 --> 01:05:31,000
Audrow Nash: facetious. Ah,

1246
01:05:31,000 --> 01:05:32,000
Stefan Seltz-Axmacher: Sorry.

1247
01:05:32,000 --> 01:05:34,704
Audrow Nash: it went right over
my head. Yeah. Yeah. Still don't

1248
01:05:34,704 --> 01:05:35,000
get it.

1249
01:05:35,000 --> 01:05:39,000
Stefan Seltz-Axmacher: wasn't my
best climate change joke?

1250
01:05:39,000 --> 01:05:42,601
Audrow Nash: One thing that's
interesting to me, so Electric

1251
01:05:42,601 --> 01:05:46,442
Sheep, the first interview that
I had on this podcast, they also

1252
01:05:46,442 --> 01:05:49,623
had a similar thing where they
were. Instead like the

1253
01:05:49,623 --> 01:05:52,864
constraint, was exactly as
you've said where they were

1254
01:05:52,864 --> 01:05:56,466
trying to do lawn mowing. And
companies have fewer people so

1255
01:05:56,466 --> 01:05:59,407
they get bigger mowers and
bigger mowers are more

1256
01:05:59,407 --> 01:06:02,828
dangerous, but they probably
also go fast for that 'cause

1257
01:06:02,828 --> 01:06:06,549
they gotta mow quickly. and so
what they said is we can remove

1258
01:06:06,549 --> 01:06:10,330
that assumption and have a whole
bunch of small mowers. so that

1259
01:06:10,330 --> 01:06:13,511
seems like where it goes
eventually, but you guys are

1260
01:06:13,511 --> 01:06:16,872
preparing for the near term
where you can automate these

1261
01:06:16,872 --> 01:06:20,653
single large vehicles and get a
lot of bang for the buck there.

1262
01:06:20,653 --> 01:06:24,195
And then as you automate, I
suppose then we can remove that

1263
01:06:24,195 --> 01:06:27,556
restriction as we build more
robots and things like this

1264
01:06:27,556 --> 01:06:31,337
Yeah. So you build, basically we
need more deployed and the new

1265
01:06:31,337 --> 01:06:34,758
generation can start to be
smaller ones that can then use

1266
01:06:34,758 --> 01:06:38,539
the thing, but we get to start
on the big ones and it's easier,

1267
01:06:38,539 --> 01:06:41,000
which is just glorious. Like
what a great

1268
01:06:41,000 --> 01:06:44,680
Stefan Seltz-Axmacher: you need
to, and you need to deploy on

1269
01:06:44,680 --> 01:06:48,299
the equipment people have today.
Like I've heard of autonomy

1270
01:06:48,299 --> 01:06:51,979
programs that. Like for some use
case, have gone to customers

1271
01:06:51,979 --> 01:06:55,598
we've talked to and said, all
right, cool for us to save you

1272
01:06:55,598 --> 01:06:59,277
$500,000 of labor cost per year
or $750,000 of labor cost per

1273
01:06:59,277 --> 01:07:02,414
year. Step one is you need to
buy $15 million of new

1274
01:07:02,414 --> 01:07:06,154
equipment. 'cause we only can
work with this one specific type

1275
01:07:06,154 --> 01:07:09,291
of vehicle. And yeah, weirdly
enough, those autonomy

1276
01:07:09,291 --> 01:07:13,031
deployments never happened. So
like we have to be able to meet

1277
01:07:13,031 --> 01:07:16,650
the market where it is today on
the vehicles that it's using

1278
01:07:16,650 --> 01:07:20,390
today, and then as, the needs
around having physically present

1279
01:07:20,390 --> 01:07:23,768
labor go away, the market can
move towards smaller, more

1280
01:07:23,768 --> 01:07:26,000
electric, more nimble vehicles
goals.

1281
01:07:26,000 --> 01:07:30,207
Audrow Nash: I like that a lot.
Yeah. And then you don't have, I

1282
01:07:30,207 --> 01:07:34,283
guess then you can remove the
black market for these tires and

1283
01:07:34,283 --> 01:07:38,228
things like this. 'cause smaller
ones are probably easier to

1284
01:07:38,228 --> 01:07:42,106
procure. That's great. Okay.
What a cool thing. What, so we

1285
01:07:42,106 --> 01:07:45,788
touched a little on AI and
machine learning. What do you

1286
01:07:45,788 --> 01:07:49,601
guys generally like, just
because it's so topical, what do

1287
01:07:49,601 --> 01:07:53,545
you guys generally think about
all the generative artificial

1288
01:07:53,545 --> 01:07:57,555
intelligence stuff? And also,
there's a lot of other exciting

1289
01:07:57,555 --> 01:08:01,368
work where it's like learning
from demonstration. There've

1290
01:08:01,368 --> 01:08:05,509
been some cool demos lately. and
just, and even like end-to-end

1291
01:08:05,509 --> 01:08:09,520
Robotics learning stuff where
they do a lot in simulation and

1292
01:08:09,520 --> 01:08:13,661
then learn the final things in
the real world. what do you guys

1293
01:08:13,661 --> 01:08:17,737
think there? Maybe start with
Ilia or Ilia. Do you wanna think

1294
01:08:17,737 --> 01:08:18,000
for,

1295
01:08:18,000 --> 01:08:21,000
Ilia Baranov: I think,

1296
01:08:21,000 --> 01:08:24,000
Audrow Nash: thoughtful for.

1297
01:08:24,000 --> 01:08:28,120
Ilia Baranov: I think there's a
lot of usefulness there as if I

1298
01:08:28,120 --> 01:08:31,456
put it on purpose, eh? No, I
think there's a lot of

1299
01:08:31,456 --> 01:08:35,249
usefulness there for, high
variability tasks. Like working

1300
01:08:35,249 --> 01:08:39,108
in the home again, or, high
degrees of freedom manipulators

1301
01:08:39,108 --> 01:08:43,097
where you can't buy inspection
or, algorithmically test every

1302
01:08:43,097 --> 01:08:46,694
possible outcome. So image
generation, for example, why

1303
01:08:46,694 --> 01:08:50,553
it's so interesting in, machine
learning, or gen generative

1304
01:08:50,553 --> 01:08:54,542
networks for images in general,
is that the image space is so

1305
01:08:54,542 --> 01:08:58,597
enormous that there's no kind of
algorithmic approach that can

1306
01:08:58,597 --> 01:09:02,586
generate enough data. So you
have to have these probabilistic

1307
01:09:02,586 --> 01:09:05,791
machine learning based
approaches to actually get

1308
01:09:05,791 --> 01:09:09,715
anything reasonable. there's no
good way to do it otherwise.

1309
01:09:09,715 --> 01:09:13,574
'cause the, the data content of
an image that you're seeing

1310
01:09:13,574 --> 01:09:17,498
right now is incredible, right?
and so for those kind of use

1311
01:09:17,498 --> 01:09:21,291
cases, especially in Robotics, I
think it's very promising

1312
01:09:21,291 --> 01:09:25,215
because there, we don't have
enough compute or algorithms to

1313
01:09:25,215 --> 01:09:29,139
solve it any other reasonable
way today. However, again, in,

1314
01:09:29,139 --> 01:09:32,802
in our space where it's very
much the opposite of a very

1315
01:09:32,802 --> 01:09:36,660
well-defined kinematic models in
con in somewhat controlled

1316
01:09:36,660 --> 01:09:40,715
spaces, I find the ML approaches
are a little bit overkill and

1317
01:09:40,715 --> 01:09:44,705
they tend to over fit to their
space. kinda the answer I give

1318
01:09:44,705 --> 01:09:48,040
people is, they say, you should
use, you should use

1319
01:09:48,040 --> 01:09:51,768
photorealistic simulation and
test all your stuff in sim,

1320
01:09:51,768 --> 01:09:55,888
which we do use simulation, but
we put almost no effort to make

1321
01:09:55,888 --> 01:09:57,000
it photorealistic

1322
01:09:57,000 --> 01:09:58,000
Audrow Nash: You don't need it.

1323
01:09:58,000 --> 01:10:01,663
Ilia Baranov: The amount of
money, the, yeah. A, we don't

1324
01:10:01,663 --> 01:10:05,582
need it, but B, the gap between
us and somebody like Tesla is

1325
01:10:05,582 --> 01:10:08,988
several tens of billions of
dollars and they have the

1326
01:10:08,988 --> 01:10:12,265
resources and are at that
bleeding edge where every

1327
01:10:12,265 --> 01:10:16,056
incremental dollar makes a
difference. Whereas us, we don't

1328
01:10:16,056 --> 01:10:19,976
need a billion dollars to make a
90% difference in what we're

1329
01:10:19,976 --> 01:10:24,024
doing. We only need a fraction
of that. And so it just wouldn't

1330
01:10:24,024 --> 01:10:27,880
be a wise investment for us at
this time in our use cases to

1331
01:10:27,880 --> 01:10:30,000
put the bulk of our effort at
ML.

1332
01:10:30,000 --> 01:10:34,127
Stefan Seltz-Axmacher: I also
think with the types of vehicles

1333
01:10:34,127 --> 01:10:37,854
that we're driving, they're
supposed to do very discreet

1334
01:10:37,854 --> 01:10:41,981
things. They're not supposed to
do fuzzy things. They might be

1335
01:10:41,981 --> 01:10:45,908
like told to do what they're
supposed to do in a fuzzy way,

1336
01:10:45,908 --> 01:10:49,835
right? Like it might be that
someone says, Hey, yard truck,

1337
01:10:49,835 --> 01:10:54,162
can you bring this over to crane
72 right now? Or they might say,

1338
01:10:54,162 --> 01:10:58,222
yard truck 14, bring this to
crane 72. Like both of those two

1339
01:10:58,222 --> 01:11:02,149
things should lead to the same
discrete action, but the way

1340
01:11:02,149 --> 01:11:06,076
that we drive that yard truck
through that port, the speeds

1341
01:11:06,076 --> 01:11:10,203
that we drive at the way that we
interact at a, a. And if it's

1342
01:11:10,203 --> 01:11:12,000
not fixed, it's terrifying.

1343
01:11:12,000 --> 01:11:13,000
Audrow Nash: Oh yeah. Totally.

1344
01:11:13,000 --> 01:11:16,991
Stefan Seltz-Axmacher: You don't
wanna wonder like, why is that

1345
01:11:16,991 --> 01:11:21,045
bulldozer doing that thing it's
doing towards me? Like, it needs

1346
01:11:21,045 --> 01:11:24,529
to be like, oh, it goes exactly
there and if it's going

1347
01:11:24,529 --> 01:11:27,000
somewhere else, something's
really bad.

1348
01:11:27,000 --> 01:11:28,000
Audrow Nash: Yeah.

1349
01:11:28,000 --> 01:11:32,050
Stefan Seltz-Axmacher: So I
think for our application, where

1350
01:11:32,050 --> 01:11:36,099
I'm really excited about, about
the broader, what people are

1351
01:11:36,099 --> 01:11:40,284
currently calling AI, what I'm
really excited about it is for,

1352
01:11:40,284 --> 01:11:43,659
the UI, is for the interaction
between our highly,

1353
01:11:43,659 --> 01:11:47,911
deterministic systems and like
fairly non-deterministic people.

1354
01:11:47,911 --> 01:11:52,298
So whether that's voice commands
for robots to do things, whether

1355
01:11:52,298 --> 01:11:56,686
that is using an LLM to scan, an
organization's, manual to figure

1356
01:11:56,686 --> 01:12:00,870
out how their facility operates
and to try to turn that into a

1357
01:12:00,870 --> 01:12:04,920
list of Boolean that we could
have a person read through and

1358
01:12:04,920 --> 01:12:06,000
see if they make

1359
01:12:06,000 --> 01:12:08,000
Audrow Nash: starting point.
Yeah.

1360
01:12:08,000 --> 01:12:09,000
Ilia Baranov: Yeah.

1361
01:12:08,000 --> 01:12:12,040
Stefan Seltz-Axmacher: that's
that probably like that second

1362
01:12:12,040 --> 01:12:16,416
thing I just said would probably
save 20 hours of stupid meetings

1363
01:12:16,416 --> 01:12:20,658
that no one wants to be in for
implementing a hundred robots in

1364
01:12:20,658 --> 01:12:24,900
a multi-billion dollar facility.
But I don't think, us figuring

1365
01:12:24,900 --> 01:12:28,940
out a novel path, based on the
vibes of the, the transformer

1366
01:12:28,940 --> 01:12:33,115
model, across, that facility,
that's probably not gonna be the

1367
01:12:33,115 --> 01:12:35,000
right strategy anytime soon.

1368
01:12:35,000 --> 01:12:39,080
Audrow Nash: what a thing. Yeah,
it's the thing that's really

1369
01:12:39,080 --> 01:12:42,893
interesting to me with it is a
lot of times, so these big

1370
01:12:42,893 --> 01:12:47,040
companies where they are, they
have this huge handbook and it,

1371
01:12:47,040 --> 01:12:50,719
they tell you how to do
something and they want it done

1372
01:12:50,719 --> 01:12:54,666
in that way. and if you have an
LLM, go and make an initial

1373
01:12:54,666 --> 01:12:58,746
thing, you still have to pay the
expert to go and confirm its

1374
01:12:58,746 --> 01:13:02,358
results for these kinds of
things. It's been, thinking

1375
01:13:02,358 --> 01:13:06,639
about this more broadly and it's
I feel like programmers, I feel

1376
01:13:06,639 --> 01:13:10,385
like copywriters, all these
different things that we all

1377
01:13:10,385 --> 01:13:14,666
thought were gonna be automated
by ChatGPT. I think they're safe

1378
01:13:14,666 --> 01:13:18,746
for a while because you want
that person to validate it. like

1379
01:13:18,746 --> 01:13:22,826
you, you need that check on it
to make sure it's reliable and

1380
01:13:22,826 --> 01:13:26,706
does what you want. And then
with, programming especially,

1381
01:13:26,706 --> 01:13:30,987
it's and if you wanna build off
of what you've done now you have

1382
01:13:30,987 --> 01:13:35,000
to give it back to the thing.
And if it says it can't do it,

1383
01:13:35,000 --> 01:13:37,000
Stefan Seltz-Axmacher: Do it

1384
01:13:37,000 --> 01:13:39,653
Audrow Nash: and, but it's still
doing the wrong thing. It's just

1385
01:13:39,653 --> 01:13:42,143
interesting 'cause a programmer
could probably figure it out.

1386
01:13:42,143 --> 01:13:44,592
But an LLM may just get stuck
like it is on the problem that

1387
01:13:44,592 --> 01:13:45,000
I'm trying

1388
01:13:45,000 --> 01:13:48,649
Stefan Seltz-Axmacher: though.
what's useful for us is, like

1389
01:13:48,649 --> 01:13:52,359
the guidebooks for a lot of our
machines are written, they're

1390
01:13:52,359 --> 01:13:55,704
written like booleans for
idiots. 'cause oftentimes the

1391
01:13:55,704 --> 01:13:59,293
managers of these facilities
don't have the world's highest

1392
01:13:59,293 --> 01:14:02,942
level of respect for the vehicle
operators. like I've read a

1393
01:14:02,942 --> 01:14:05,922
manual for a, particular
facility. Where, there's

1394
01:14:05,922 --> 01:14:09,389
vehicles that lift up 30,000
pounds and then drive around

1395
01:14:09,389 --> 01:14:13,281
with 30,000 pounds and then move
it. and there's a list of maybe

1396
01:14:13,281 --> 01:14:16,687
like 50 rules of operation for
when you're driving these

1397
01:14:16,687 --> 01:14:20,397
vehicles. And one of them was
like, if you have lifted a load

1398
01:14:20,397 --> 01:14:23,986
and you're in reverse and you
see someone standing within a

1399
01:14:23,986 --> 01:14:27,513
hundred meters behind you, turn
off the engine, get on the

1400
01:14:27,513 --> 01:14:31,223
radio, ask for help. If you're
driving from here to there and

1401
01:14:31,223 --> 01:14:34,934
this gate is open, turn off the
engine, get on the radio call

1402
01:14:34,934 --> 01:14:38,522
for help. and there are rules
written like that like an LLM

1403
01:14:38,522 --> 01:14:42,232
could make quick work of. and
that we could deploy into a use

1404
01:14:42,232 --> 01:14:46,064
case really easily. But there's
also the, rules that you'd only

1405
01:14:46,064 --> 01:14:49,774
learn by sitting down with all
of the management staff of the

1406
01:14:49,774 --> 01:14:53,605
site for three weeks in, in 14
different meetings that are each

1407
01:14:53,605 --> 01:14:57,133
scheduled for an hour and a
half, but none of them go less

1408
01:14:57,133 --> 01:15:00,782
than two and a half hours. Where
in the 15th meeting someone

1409
01:15:00,782 --> 01:15:04,188
happens to say, oh, yeah, you
can't drive in that corner

1410
01:15:04,188 --> 01:15:07,959
'cause that's the corner where
this thing drains and all these

1411
01:15:07,959 --> 01:15:11,000
people died once. So you gotta
go around that. And

1412
01:15:11,000 --> 01:15:12,000
Audrow Nash: these

1413
01:15:12,000 --> 01:15:14,205
Stefan Seltz-Axmacher: that's
written down somewhere, the LLM

1414
01:15:14,205 --> 01:15:15,000
won, I won't learn it.

1415
01:15:15,000 --> 01:15:18,906
Audrow Nash: Yeah, that context.
Yeah, it's interesting 'cause

1416
01:15:18,906 --> 01:15:22,811
there's a lot that's implicit I
think in how people do things.

1417
01:15:22,811 --> 01:15:26,717
And it may not be in a manual,
just has to be learned for this

1418
01:15:26,717 --> 01:15:30,811
kind of thing. Interesting. Any,
so Has it changed? Oh, go ahead,

1419
01:15:30,811 --> 01:15:31,000
ia.

1420
01:15:31,000 --> 01:15:33,000
Stefan Seltz-Axmacher: go ahead.

1421
01:15:33,000 --> 01:15:36,148
Ilia Baranov: yeah, there's, a
bigger meta point here on

1422
01:15:36,148 --> 01:15:39,746
machine learning in general and
Robotics and our particular view

1423
01:15:39,746 --> 01:15:43,175
on it. I was, careful earlier to
say that machine learning is

1424
01:15:43,175 --> 01:15:46,772
this kind of sugar res sprinkle
on top to, to really solve these

1425
01:15:46,772 --> 01:15:50,257
corner use cases because in a
lot of our vehicles, we actually

1426
01:15:50,257 --> 01:15:52,000
want to safety rate our system.

1427
01:15:52,000 --> 01:15:53,000
Audrow Nash: Oh,

1428
01:15:52,000 --> 01:15:55,636
Ilia Baranov: And again, safety
rating is something where if you

1429
01:15:55,636 --> 01:15:58,534
want to do an actual formal
certification, it's not

1430
01:15:58,534 --> 01:16:01,716
impossible with machine
learning, but the bar is so much

1431
01:16:01,716 --> 01:16:04,841
higher right than, very
deterministic systems, which we

1432
01:16:04,841 --> 01:16:08,477
can prove out a hundred percent
exactly what it will do in every

1433
01:16:08,477 --> 01:16:11,943
scenario. and so the lower down
the stack you get towards the

1434
01:16:11,943 --> 01:16:15,239
underlying safety layer, the
more and more rigid we become

1435
01:16:15,239 --> 01:16:18,648
and the more and more kind of
rules driven we become so that

1436
01:16:18,648 --> 01:16:22,114
machine learning side at the top
can like maybe differentiate

1437
01:16:22,114 --> 01:16:25,295
shape between a tree and a
boulder, but it's never gonna

1438
01:16:25,295 --> 01:16:27,000
command the vehicles directly.

1439
01:16:27,000 --> 01:16:29,000
Audrow Nash: Yeah. it has to

1440
01:16:29,000 --> 01:16:30,818
Ilia Baranov: because we need to
deploy these things in real

1441
01:16:30,818 --> 01:16:31,000
world.

1442
01:16:31,000 --> 01:16:35,084
Audrow Nash: Wow. I love that.
I, can't believe I didn't think

1443
01:16:35,084 --> 01:16:39,235
about talking about all of the,
safety certification stuff, but

1444
01:16:39,235 --> 01:16:43,188
that makes a lot of sense, which
is you guys can, because so

1445
01:16:43,188 --> 01:16:47,272
these companies, they come and
they try to build their own and

1446
01:16:47,272 --> 01:16:51,225
they spend, you said like $50
million or some bonkers amount

1447
01:16:51,225 --> 01:16:55,375
of money trying to build their
autonomy stack, whereas you guys

1448
01:16:55,375 --> 01:16:59,591
are making a business out of it.
And then what you can do is you

1449
01:16:59,591 --> 01:17:03,742
can invest as you need or a you
can invest so that you actually

1450
01:17:03,742 --> 01:17:07,365
safety certify things so they
can just drop in a safety

1451
01:17:07,365 --> 01:17:10,000
certified system. That's really
amazing.

1452
01:17:10,000 --> 01:17:11,000
Stefan Seltz-Axmacher: Yep.

1453
01:17:11,000 --> 01:17:14,025
Audrow Nash: that's probably a
huge competitive advantage and

1454
01:17:14,025 --> 01:17:17,000
probably a very good moat, I
would imagine for you guys as a

1455
01:17:17,000 --> 01:17:18,000
Stefan Seltz-Axmacher: Yep.

1456
01:17:18,000 --> 01:17:19,000
Ilia Baranov: Yep.

1457
01:17:18,000 --> 01:17:21,355
Stefan Seltz-Axmacher: Safety's
really hard and like safety's

1458
01:17:21,355 --> 01:17:24,820
really hard. It's also hard to
build a culture where people are

1459
01:17:24,820 --> 01:17:28,285
smart about safety, where people
are, thinking about safety the

1460
01:17:28,285 --> 01:17:31,420
right way. And essentially
because we use the same safety

1461
01:17:31,420 --> 01:17:34,940
core on all of our machines, we
are basically constantly getting

1462
01:17:34,940 --> 01:17:37,965
learnings. We're constantly
getting hours of testing to

1463
01:17:37,965 --> 01:17:40,000
prove that the system actually
works.

1464
01:17:40,000 --> 01:17:44,167
Audrow Nash: Yeah. So part of it
is putting the case together and

1465
01:17:44,167 --> 01:17:47,885
because you have deployed on
systems and your systems have

1466
01:17:47,885 --> 01:17:51,731
monitoring, and then you can
log, so that you can prove that

1467
01:17:51,731 --> 01:17:55,449
this has happened for some
amount of time. So you can make

1468
01:17:55,449 --> 01:17:59,359
these claims. That's nice. And
so you can solve, Val, you can

1469
01:17:59,359 --> 01:18:03,141
solve cases where they don't
require the certification, but

1470
01:18:03,141 --> 01:18:06,923
you can use those to bootstrap
justification for additional

1471
01:18:06,923 --> 01:18:10,000
safety certifications. That's
clever. Hell yeah.

1472
01:18:10,000 --> 01:18:12,000
Stefan Seltz-Axmacher: yeah.

1473
01:18:12,000 --> 01:18:14,511
Ilia Baranov: Yeah. And that's
from a business side, that's

1474
01:18:14,511 --> 01:18:16,000
also a big competitive moat for
us,

1475
01:18:16,000 --> 01:18:17,000
Audrow Nash: that's what I was

1476
01:18:17,000 --> 01:18:20,549
Ilia Baranov: over time, because
we really, yeah, we really try

1477
01:18:20,549 --> 01:18:23,930
to, again, keep that autonomy
core the same and the same for

1478
01:18:23,930 --> 01:18:27,479
the safety subsystem. So we'll
build up hours of operation that

1479
01:18:27,479 --> 01:18:30,915
are directly applicable across
units faster because we deploy

1480
01:18:30,915 --> 01:18:34,014
it across a much more varied
fleet. And of course, as a

1481
01:18:34,014 --> 01:18:37,563
roboticist, you can imagine the
more varied use cases you have,

1482
01:18:37,563 --> 01:18:41,056
The much better. quality actual
proof you have of safety. It's

1483
01:18:41,056 --> 01:18:44,662
not just the same test a hundred
times, it's a hundred different

1484
01:18:44,662 --> 01:18:45,000
tests.

1485
01:18:45,000 --> 01:18:49,615
Audrow Nash: That's super cool.
Okay. What a thing. I really

1486
01:18:49,615 --> 01:18:54,154
like that. Do you have, safety
certifications yet? Or which

1487
01:18:54,154 --> 01:18:58,000
ones do you have and what are
you working towards?

1488
01:18:58,000 --> 01:19:00,641
Ilia Baranov: Yeah, we're
working. So there's two

1489
01:19:00,641 --> 01:19:03,928
different threads we're working
towards. One is internally we

1490
01:19:03,928 --> 01:19:07,000
wanna certify to UL 4600, which
is a fairly new standard.

1491
01:19:07,000 --> 01:19:08,000
Audrow Nash: Yeah, I'm not
familiar with

1492
01:19:08,000 --> 01:19:09,903
Ilia Baranov: in particular
we're targeting that one. Yeah.

1493
01:19:09,903 --> 01:19:10,000
So,

1494
01:19:10,000 --> 01:19:11,000
Stefan Seltz-Axmacher: so

1495
01:19:11,000 --> 01:19:15,000
Ilia Baranov: past the, the, ISO
26262

1496
01:19:15,000 --> 01:19:16,000
Audrow Nash: Oh, okay. I know
that one. Okay.

1497
01:19:16,000 --> 01:19:17,000
Stefan Seltz-Axmacher: yep.

1498
01:19:17,000 --> 01:19:21,061
Ilia Baranov: So what's nice
about it is ISO 26262, if you go

1499
01:19:21,061 --> 01:19:24,789
through the certification at
some point it assumes human

1500
01:19:24,789 --> 01:19:28,849
takeover, whereas UL 4600 does
not. And that's that's one but

1501
01:19:28,849 --> 01:19:32,511
kind of key difference. And UL
4600 is more a, umbrella

1502
01:19:32,511 --> 01:19:36,771
standard of how to design a safe
autonomous system without being

1503
01:19:36,771 --> 01:19:39,834
prescriptive on, like
specifically follow this

1504
01:19:39,834 --> 01:19:43,894
pattern. But it itself suggests
things like IEC 61508 for sub

1505
01:19:43,894 --> 01:19:47,822
electrical systems and those
kind of things, which build up

1506
01:19:47,822 --> 01:19:51,284
proof to get to a UL 4600
certification standard. So

1507
01:19:51,284 --> 01:19:55,344
again, it allows us to be more
flexible and, if we want to go

1508
01:19:55,344 --> 01:19:59,538
that direction, I can talk about
how I'm really excited to turn

1509
01:19:59,538 --> 01:20:03,000
safety certification into
another CICD code problem.

1510
01:20:03,000 --> 01:20:05,625
Audrow Nash: I am, I would love
to talk about that. No, I would

1511
01:20:05,625 --> 01:20:07,000
love that. That sounds
wonderful.

1512
01:20:07,000 --> 01:20:08,000
Ilia Baranov: Okay.

1513
01:20:08,000 --> 01:20:11,580
Audrow Nash: I love safety
certification stuff. I, host

1514
01:20:11,580 --> 01:20:15,355
spaces on X. I actually have one
tonight. and you guys are

1515
01:20:15,355 --> 01:20:19,260
welcome to come anytime, but
they're like an hour and a half

1516
01:20:19,260 --> 01:20:22,710
every week or something like
this. We had one, before

1517
01:20:22,710 --> 01:20:26,615
Christmas sometime, and it was
an hour and a half of talking

1518
01:20:26,615 --> 01:20:30,391
about safety certifications and
after it's like I found my

1519
01:20:30,391 --> 01:20:34,556
people, this is just wonderful,
Okay. Tell me it's, I would love

1520
01:20:34,556 --> 01:20:37,941
to hear how you're thinking of
it. Like a continuous

1521
01:20:37,941 --> 01:20:41,000
integration system, testing this
kind of thing.

1522
01:20:41,000 --> 01:20:44,596
Ilia Baranov: yeah. Yeah. So,
having gone through this kind of

1523
01:20:44,596 --> 01:20:48,308
process before. A lot of it is a
very manual, like engineer sits

1524
01:20:48,308 --> 01:20:51,381
down, writes a whole bunch of
failure modes, get that

1525
01:20:51,381 --> 01:20:54,919
reviewed, then they get testing,
they get, that reviewed. And

1526
01:20:54,919 --> 01:20:58,515
then maybe a small piece of it
is, there's some requirement on

1527
01:20:58,515 --> 01:21:01,995
code test coverage and maybe
that piece is part of the CICD,

1528
01:21:01,995 --> 01:21:05,707
but I want to actually wrap the
entire everything I've described

1529
01:21:05,707 --> 01:21:09,012
as part of a CICD so that every
build we, we run for our,

1530
01:21:09,012 --> 01:21:12,724
specifically our safety system.
So we try to keep the scope very

1531
01:21:12,724 --> 01:21:14,000
small and very simple.

1532
01:21:14,000 --> 01:21:15,000
Audrow Nash: Of course.

1533
01:21:15,000 --> 01:21:18,730
Ilia Baranov: Not only does unit
testing, but also auto generates

1534
01:21:18,730 --> 01:21:22,000
a lot of the okay to be in
compliance with this part of U

1535
01:21:22,000 --> 01:21:23,000
Audrow Nash: Oh, I love it.

1536
01:21:22,000 --> 01:21:24,920
Ilia Baranov: have this
certification that refers to

1537
01:21:24,920 --> 01:21:28,009
this thing, here's in this
particular vehicle, which is

1538
01:21:28,009 --> 01:21:31,547
configured this way. So when we
do a build for this vehicle, it

1539
01:21:31,547 --> 01:21:34,861
has these characteristics which
we can pull in as their own

1540
01:21:34,861 --> 01:21:38,342
proof for their use cases. So
this vehicle goes at this speed,

1541
01:21:38,342 --> 01:21:41,993
hence we need this speed of look
ahead, here's our proof of that,

1542
01:21:41,993 --> 01:21:45,587
look ahead, here's our proof of
the data and to auto generate as

1543
01:21:45,587 --> 01:21:49,125
much of that as we possibly can
so that the end requirement, of

1544
01:21:49,125 --> 01:21:52,382
course you still need a human
review, But a lot of that is

1545
01:21:52,382 --> 01:21:53,000
auto built.

1546
01:21:53,000 --> 01:21:55,000
Audrow Nash: Yeah. That's

1547
01:21:55,000 --> 01:21:57,000
Ilia Baranov: rather than hand
calculating all of it.

1548
01:21:57,000 --> 01:22:00,623
Audrow Nash: is laborious for
sure. That's awesome. I, so it

1549
01:22:00,623 --> 01:22:04,428
would be like, effectively, it'd
be like you have requirements.

1550
01:22:04,428 --> 01:22:08,111
So you, could take a standard,
you could parse it and turn it

1551
01:22:08,111 --> 01:22:11,553
into a bunch of requirements.
The requirements might have

1552
01:22:11,553 --> 01:22:15,116
parameters, and then you could
use these to define specific

1553
01:22:15,116 --> 01:22:18,498
tests that then you run that are
part of your continuous

1554
01:22:18,498 --> 01:22:22,000
integration and testing. is
that, oh, that sounds so cool.

1555
01:22:22,000 --> 01:22:25,189
Ilia Baranov: UL 4600 and yeah.
and UL 4600 is particular, is

1556
01:22:25,189 --> 01:22:28,011
written as this like giant
spanning tree of a bunch of

1557
01:22:28,011 --> 01:22:31,095
things. to be compliant with
this, you have to be compliant

1558
01:22:31,095 --> 01:22:34,075
with this, and then you go to
those. to be compliant with

1559
01:22:34,075 --> 01:22:37,264
that, you have to be compliant
with this, So it's basically a

1560
01:22:37,264 --> 01:22:40,452
big tree traversal problem of,
to hit this, here's my, all my

1561
01:22:40,452 --> 01:22:43,693
branches and here's proofs for
everyone that then build on top

1562
01:22:43,693 --> 01:22:45,000
of each other to get that

1563
01:22:45,000 --> 01:22:46,000
Audrow Nash: Oh, that's so

1564
01:22:46,000 --> 01:22:47,000
Ilia Baranov: That's What I

1565
01:22:47,000 --> 01:22:48,000
Audrow Nash: a moat that will

1566
01:22:48,000 --> 01:22:51,000
Ilia Baranov: that's what we're
building.

1567
01:22:51,000 --> 01:22:52,000
Audrow Nash: Ah-Huh?

1568
01:22:51,000 --> 01:22:54,000
Ilia Baranov: actually I do want
to open source as much

1569
01:22:54,000 --> 01:22:58,000
Audrow Nash: I want that to be
open sourced. Yeah.

1570
01:22:58,000 --> 01:23:00,565
Ilia Baranov: At least, the
learnings from it and the

1571
01:23:00,565 --> 01:23:03,516
pitfalls we find and those kind
of things. I do want to build

1572
01:23:03,516 --> 01:23:06,565
that out because I think for too
long safety has been this like

1573
01:23:06,565 --> 01:23:09,516
pay us X millions of dollars and
we're a safety certification

1574
01:23:09,516 --> 01:23:12,419
organization. We'll give you the
check mark, like here's the

1575
01:23:12,419 --> 01:23:14,597
golden check mark. I wanna
actually build the

1576
01:23:14,597 --> 01:23:17,403
infrastructure to let other
groups do that themselves. And

1577
01:23:17,403 --> 01:23:20,403
of course provide that service,
we will as well. But I just as

1578
01:23:20,403 --> 01:23:23,403
basic autonomy, you should be
able to build this yourself. But

1579
01:23:23,403 --> 01:23:26,452
if you want a shortcut, you work
with us because we know how to

1580
01:23:26,452 --> 01:23:28,000
do it. We have the proof points.

1581
01:23:28,000 --> 01:23:32,520
Audrow Nash: Yeah, actually that
could be a, that could even be

1582
01:23:32,520 --> 01:23:37,040
like a side business or a side
thing within your business where

1583
01:23:37,040 --> 01:23:40,987
it's like you basically have all
this infrastructure to

1584
01:23:40,987 --> 01:23:45,004
facilitate. UL 4600, 4,600, you,
you could help Robotics

1585
01:23:45,004 --> 01:23:49,022
companies certify their,
sophomore easily. 'cause it's a

1586
01:23:49,022 --> 01:23:53,112
pain that people have to go and
learn all of these safety

1587
01:23:53,112 --> 01:23:57,058
certifications and all the
details of it. And if it was

1588
01:23:57,058 --> 01:24:01,363
like, here, hook up to your
robot to our, like some, sort of

1589
01:24:01,363 --> 01:24:05,883
interface. and we're gonna see
if it passes in our environment,

1590
01:24:05,883 --> 01:24:10,260
this kind of thing. That would
be just amazing. That would be

1591
01:24:10,260 --> 01:24:14,709
so cool. That would, accelerate
all of Robotics I would think.

1592
01:24:14,709 --> 01:24:16,000
Like all different

1593
01:24:16,000 --> 01:24:19,665
Ilia Baranov: yeah, and I think
an additional detail is that

1594
01:24:19,665 --> 01:24:22,842
while that's the umbrella
standard we've picked, the

1595
01:24:22,842 --> 01:24:26,202
reason we picked it is it
because a high, it has a high

1596
01:24:26,202 --> 01:24:29,562
enough barrier of proof and
demonstration, that kind of

1597
01:24:29,562 --> 01:24:33,166
stuff that I, I believe at
least, and we're working through

1598
01:24:33,166 --> 01:24:36,831
the proof exactly. We, the same
techniques we've used on U 4

1599
01:24:36,831 --> 01:24:40,252
8600, we can then deploy to
whatever people need because

1600
01:24:40,252 --> 01:24:43,429
mining standards in South
Africa. Are different than

1601
01:24:43,429 --> 01:24:46,850
Australia, are different than
the US. Instead of killing

1602
01:24:46,850 --> 01:24:50,576
ourselves, trying to comply with
everyone, we'll generate the

1603
01:24:50,576 --> 01:24:53,997
products and then go to
specialists on site and say, for

1604
01:24:53,997 --> 01:24:57,541
your mine, you know how to
certify. Here's all the data we

1605
01:24:57,541 --> 01:25:01,328
have. What are we missing? What
do you need That isn't part of

1606
01:25:01,328 --> 01:25:02,000
this stack.

1607
01:25:02,000 --> 01:25:05,899
Audrow Nash: will be really
cool. Yeah, I really like that a

1608
01:25:05,899 --> 01:25:09,733
lot. I think, yeah, making
certifications. it's interesting

1609
01:25:09,733 --> 01:25:13,503
because if we made them very
algorithmic in a way, and you

1610
01:25:13,503 --> 01:25:17,207
can parse them so they're in
PDFs or whatever and you can

1611
01:25:17,207 --> 01:25:21,366
parse them to make them, like a,
tree of things that you have to

1612
01:25:21,366 --> 01:25:25,200
go check and then you have some
document that shows all the

1613
01:25:25,200 --> 01:25:29,294
requirements that would just be
so valuable. I had to, so I was

1614
01:25:29,294 --> 01:25:33,324
the ROS boss for Humble for ROS
2 and built a tool that helped

1615
01:25:33,324 --> 01:25:37,353
us with requirements. and so I
did a pretty good investigation

1616
01:25:37,353 --> 01:25:41,382
on the like requirements, tools,
space to try to see okay, you

1617
01:25:41,382 --> 01:25:45,151
have these requirements, they
lead to these test cases and

1618
01:25:45,151 --> 01:25:49,180
whatever, and how do you do it?
And they were all Horrible. So

1619
01:25:49,180 --> 01:25:51,000
we ended up building our own

1620
01:25:51,000 --> 01:25:55,240
Ilia Baranov: all awful And
super expensive. and like

1621
01:25:55,240 --> 01:25:57,000
ludicrously expensive.

1622
01:25:57,000 --> 01:26:00,529
Audrow Nash: many of them had
like free trials or something,

1623
01:26:00,529 --> 01:26:03,941
so I would try them and they're
just terrible. Like awful,

1624
01:26:03,941 --> 01:26:07,588
Everything. Awful and expensive.
but I think if you could make

1625
01:26:07,588 --> 01:26:10,765
that good, that would be
fantastic. I. that would be a

1626
01:26:10,765 --> 01:26:14,294
big value to the whole Robotics
community. I would expect. I

1627
01:26:14,294 --> 01:26:17,765
would love, and if you open
source that please let me know.

1628
01:26:17,765 --> 01:26:20,000
I would love to see it too. Hell
yeah.

1629
01:26:20,000 --> 01:26:22,595
Ilia Baranov: Yeah. Yeah, no,
for sure. you know what? After

1630
01:26:22,595 --> 01:26:25,276
this send me that requirements.
If you have any notes on that,

1631
01:26:25,276 --> 01:26:27,784
I'd love to see if we can
integrate that. That'd be really

1632
01:26:27,784 --> 01:26:28,000
cool.

1633
01:26:28,000 --> 01:26:29,000
Audrow Nash: Oh hell yeah. Okay.

1634
01:26:29,000 --> 01:26:32,333
Ilia Baranov: If there's
anything public about it, throw

1635
01:26:32,333 --> 01:26:34,000
it my way. I'll take a look.

1636
01:26:34,000 --> 01:26:38,105
Audrow Nash: the tool I ended up
making was a bit simpler and it

1637
01:26:38,105 --> 01:26:42,017
was basically for, it's all open
source and it has aged a bit

1638
01:26:42,017 --> 01:26:45,994
'cause I haven't had maintenance
time for it, but we are still

1639
01:26:45,994 --> 01:26:49,907
using it for every ROS 2 and
Gazebo release. but yeah, I, can

1640
01:26:49,907 --> 01:26:53,819
send it to you. It's called yet
another test case manager was

1641
01:26:53,819 --> 01:26:56,000
the dumb name that I gave it.
But,

1642
01:26:56,000 --> 01:26:58,000
Ilia Baranov: I love it.

1643
01:26:58,000 --> 01:27:02,258
Audrow Nash: it was fun to
build. And then, so we're coming

1644
01:27:02,258 --> 01:27:06,804
to the end, but I would love to,
we didn't talk about it at all

1645
01:27:06,804 --> 01:27:11,351
about your podcast, so I'd love
to talk a bit about that. yeah.

1646
01:27:11,351 --> 01:27:12,000
go ahead.

1647
01:27:12,000 --> 01:27:16,275
Ilia Baranov: before we jump
into that, can I do, one quick

1648
01:27:16,275 --> 01:27:17,000
diversion?

1649
01:27:17,000 --> 01:27:19,172
Stefan Seltz-Axmacher: Is there,
a chat feature in this? Can we

1650
01:27:19,172 --> 01:27:20,000
send, you Audrow a link?

1651
01:27:20,000 --> 01:27:23,000
Audrow Nash: Yeah. it's in the
bottom right

1652
01:27:23,000 --> 01:27:24,000
Stefan Seltz-Axmacher: Cool.

1653
01:27:23,000 --> 01:27:27,658
Ilia Baranov: All right. So for
our listeners, you'll have less

1654
01:27:27,658 --> 01:27:32,169
fun than, we do currently. but
I, just to describe it, I sent

1655
01:27:32,169 --> 01:27:36,754
Audrow a link to our live test
tractor in Modesto, California,

1656
01:27:36,754 --> 01:27:41,338
and I'm gonna share my screen so
you can see a video feed from

1657
01:27:41,338 --> 01:27:44,000
it. now what Audrow's gonna do
see a

1658
01:27:44,000 --> 01:27:47,537
Stefan Seltz-Axmacher: let,
lemme let me, lemme, let me walk

1659
01:27:47,537 --> 01:27:50,661
you through a little bit of
what's going on. So, live

1660
01:27:50,661 --> 01:27:54,198
streaming here on the screen is
a first person view from our

1661
01:27:54,198 --> 01:27:57,381
test tractor in Modesto,
California. There's no one in

1662
01:27:57,381 --> 01:28:00,918
this tractor, no one on site.
The closest member of our team

1663
01:28:00,918 --> 01:28:04,632
is about a hundred miles away.
The last time anyone in our team

1664
01:28:04,632 --> 01:28:08,404
was physically near this tractor
was probably Ilia and I back in

1665
01:28:08,404 --> 01:28:11,882
September. Audrow, did that,
link work for you? Do you have

1666
01:28:11,882 --> 01:28:12,000
it

1667
01:28:12,000 --> 01:28:15,589
Audrow Nash: I have a, it's like
a Google Maps or something, and

1668
01:28:15,589 --> 01:28:18,000
I click the location, so it
looks like it's

1669
01:28:18,000 --> 01:28:19,000
Stefan Seltz-Axmacher: Oh,

1670
01:28:18,000 --> 01:28:21,875
Ilia Baranov: you, you, clicked
a location. Yep. There you go.

1671
01:28:21,875 --> 01:28:22,000
So

1672
01:28:22,000 --> 01:28:23,000
Stefan Seltz-Axmacher: you just,

1673
01:28:23,000 --> 01:28:25,000
Ilia Baranov: our robot.

1674
01:28:25,000 --> 01:28:28,000
Stefan Seltz-Axmacher: so the
app that, re

1675
01:28:28,000 --> 01:28:29,000
Audrow Nash: Oh, so

1676
01:28:29,000 --> 01:28:33,202
Stefan Seltz-Axmacher: was. Was
written by a buddy of mine who

1677
01:28:33,202 --> 01:28:37,064
is a self-described 50th
percentile engineer. he's just a

1678
01:28:37,064 --> 01:28:40,927
front end engineer at a big
boring SaaS company. He knows

1679
01:28:40,927 --> 01:28:44,519
nothing about controls, he knows
nothing about safety

1680
01:28:44,519 --> 01:28:48,111
engineering. He knows nothing
about machine learning,

1681
01:28:48,111 --> 01:28:52,109
mechatronics any of that. He
basically, in 20 hours he took

1682
01:28:52,109 --> 01:28:55,972
map box's, API to pull in
satellite data of the field and

1683
01:28:55,972 --> 01:28:59,699
took in our API to be able to
see location of where the

1684
01:28:59,699 --> 01:29:03,833
tractor is and issue, commands
to the tractor. And from that,

1685
01:29:03,833 --> 01:29:08,102
Audrow seems to have now sent us
on yet another, point that the

1686
01:29:08,102 --> 01:29:10,000
tractor's driving itself to.

1687
01:29:10,000 --> 01:29:13,000
Audrow Nash: That's super cool.
I love that.

1688
01:29:13,000 --> 01:29:16,648
Ilia Baranov: it's doing a three
point turn and then it's 'cause

1689
01:29:16,648 --> 01:29:20,012
it will, so for listeners, Aras
clicking in a little yellow

1690
01:29:20,012 --> 01:29:23,318
geofence and we're seeing a live
feed of the vehicle drive

1691
01:29:23,318 --> 01:29:26,568
around. What the vehicle will
always try to do is, end up

1692
01:29:26,568 --> 01:29:29,817
oriented from where it started
to where it is. So usually

1693
01:29:29,817 --> 01:29:33,238
that'll involve a three point
turn. So right now it's facing

1694
01:29:33,238 --> 01:29:36,202
north. If Audrow clicks
somewhere south, we're gonna

1695
01:29:36,202 --> 01:29:39,508
turn around and then end up
facing south. and this kind of

1696
01:29:39,508 --> 01:29:43,214
easy to use interface. There you
go. and this kind of easy to use

1697
01:29:43,214 --> 01:29:46,634
interface where you just click
it and tell it where to go is

1698
01:29:46,634 --> 01:29:50,055
just one example app. But the
underlying rest API is what we

1699
01:29:50,055 --> 01:29:53,475
actually use to control our
vehicles. And this tractor could

1700
01:29:53,475 --> 01:29:56,952
easily be a mining vehicle or a
forestry vehicle and it would

1701
01:29:56,952 --> 01:30:00,430
behave completely differently.
But Audrow's clicking would be

1702
01:30:00,430 --> 01:30:01,000
identical.

1703
01:30:01,000 --> 01:30:05,093
Stefan Seltz-Axmacher: Just the
other week we were co-hosting a

1704
01:30:05,093 --> 01:30:08,992
happy hour at CES and we were
using the same app, at 9:00 PM

1705
01:30:08,992 --> 01:30:12,955
Pacific time to issue commands
to, an AGV that our autonomy's

1706
01:30:12,955 --> 01:30:16,659
on in Australia, made by a
partner of ours, BIA5. So like

1707
01:30:16,659 --> 01:30:20,752
the reason for this and like the
whole thought process here, if

1708
01:30:20,752 --> 01:30:24,196
you wanna automate a tractor,
the special thing about

1709
01:30:24,196 --> 01:30:28,029
automating a tractor is not the
hardware that goes onto the

1710
01:30:28,029 --> 01:30:31,798
tractor. It's not like the
point-to-point navigation. It's

1711
01:30:31,798 --> 01:30:35,696
like figuring out like the five
screens and four buttons per

1712
01:30:35,696 --> 01:30:39,790
screen that Farmer John needs to
see and, interact with to make

1713
01:30:39,790 --> 01:30:43,623
something valuable for Farmer
John happen. And I don't know

1714
01:30:43,623 --> 01:30:47,782
what Farmer John needs, 'cause I
don't know if Farmer John is in

1715
01:30:47,782 --> 01:30:51,680
Ohio and grows wheat, or if it's
really Father Juan in Spain

1716
01:30:51,680 --> 01:30:55,514
growing whatever they grow in
Spain. either farmer probably

1717
01:30:55,514 --> 01:30:58,762
needs a different ui. They
probably need different

1718
01:30:58,762 --> 01:31:02,726
behaviors to happen. We let our
customers build that. We just

1719
01:31:02,726 --> 01:31:05,000
solved the point to point nav
part.

1720
01:31:05,000 --> 01:31:09,109
Audrow Nash: Huh. That's super
cool. Did not know we were gonna

1721
01:31:09,109 --> 01:31:11,000
get a live demo. I love that.

1722
01:31:11,000 --> 01:31:13,653
Stefan Seltz-Axmacher: Yeah, we,
did that demo I think a thousand

1723
01:31:13,653 --> 01:31:15,000
times last year is a real
number.

1724
01:31:15,000 --> 01:31:18,000
Ilia Baranov: Literally
something. Yeah.

1725
01:31:18,000 --> 01:31:21,239
Stefan Seltz-Axmacher: like if
you listener are on a call with

1726
01:31:21,239 --> 01:31:24,269
me and it's during business
hours, You'll almost certainly

1727
01:31:24,269 --> 01:31:25,000
end up getting

1728
01:31:25,000 --> 01:31:26,000
Audrow Nash: ha.

1729
01:31:26,000 --> 01:31:28,752
Stefan Seltz-Axmacher: and
that's part of our secret evil

1730
01:31:28,752 --> 01:31:31,455
plan to prove that our robots
just work is by like doing

1731
01:31:31,455 --> 01:31:34,303
constant demos. of this is not a
big deal. This's not a big

1732
01:31:34,303 --> 01:31:37,007
production. We're not even
treating you all that special

1733
01:31:37,007 --> 01:31:39,759
right now, Audrow, we're just
doing the thing we do every

1734
01:31:39,759 --> 01:31:40,000
time.

1735
01:31:40,000 --> 01:31:44,200
Audrow Nash: That's great. It's
great to not be treated special

1736
01:31:44,200 --> 01:31:46,000
with the get the good demo.

1737
01:31:46,000 --> 01:31:49,493
Stefan Seltz-Axmacher: I know
how you make, I know how to make

1738
01:31:49,493 --> 01:31:50,000
you feel.

1739
01:31:50,000 --> 01:31:54,345
Audrow Nash: Yeah. Okay. Do you
think So it's very interesting

1740
01:31:54,345 --> 01:31:58,340
to me. The web connects all of
us. it's just like all our

1741
01:31:58,340 --> 01:32:02,616
devices are running web stuff,
everything is making requests.

1742
01:32:02,616 --> 01:32:06,751
I, feel like in the robotics
community we've looked on rest

1743
01:32:06,751 --> 01:32:11,096
APIs, which I don't know, they
have their git and put and post

1744
01:32:11,096 --> 01:32:15,512
and whatever their requests. and
we've looked at them and gone,

1745
01:32:15,512 --> 01:32:19,086
that will never work for
robotics. And you guys are

1746
01:32:19,086 --> 01:32:23,151
making these nice demos and you
have these simple rest API

1747
01:32:23,151 --> 01:32:27,287
endpoints for querying. Tell me
a bit about that. is it the

1748
01:32:27,287 --> 01:32:31,212
future of how we're going to
interact with robots as you

1749
01:32:31,212 --> 01:32:35,136
think like a high level rest,
API will manage things and

1750
01:32:35,136 --> 01:32:39,552
you'll handle all the complexity
on the robot side, but keep it

1751
01:32:39,552 --> 01:32:43,757
simple for the user. Is that
kind of how it will be? Or tell

1752
01:32:43,757 --> 01:32:46,000
me about rest APIs and why them?

1753
01:32:46,000 --> 01:32:49,712
Ilia Baranov: I think, yeah, I
think we, when we were just

1754
01:32:49,712 --> 01:32:52,976
getting started, we wanted
something that every web

1755
01:32:52,976 --> 01:32:57,072
developer and every language out
there has a easy to use getting

1756
01:32:57,072 --> 01:33:00,656
started guide on. So I think
like Python or C or Rust or

1757
01:33:00,656 --> 01:33:02,000
HTTP, Ruby, whatever.

1758
01:33:02,000 --> 01:33:03,000
Audrow Nash: Whatever it is.
Yeah.

1759
01:33:03,000 --> 01:33:06,826
Ilia Baranov: And REST was this
very simple to understand that

1760
01:33:06,826 --> 01:33:10,404
you could even run from a
command line if you wanted to or

1761
01:33:10,404 --> 01:33:14,353
run from a browser. Exactly. You
can curl it. and it is a polled

1762
01:33:14,353 --> 01:33:18,055
interface in a lot of ways. So
if, for example, we wanted to

1763
01:33:18,055 --> 01:33:21,943
webstream, you notice the video
I showed was using forint as an

1764
01:33:21,943 --> 01:33:25,768
actual web backend. We're big
fans of formint we use them for,

1765
01:33:25,768 --> 01:33:29,594
this sort of stuff. REST is not
the path. Yeah. So REST is not

1766
01:33:29,594 --> 01:33:32,000
the path for streaming large
amounts of

1767
01:33:32,000 --> 01:33:33,000
Audrow Nash: No,

1768
01:33:33,000 --> 01:33:36,096
Ilia Baranov: it at all. but
Right, but the right tool for

1769
01:33:36,096 --> 01:33:39,511
the right process. If, you have
a large fleet of vehicles, let's

1770
01:33:39,511 --> 01:33:42,980
say a hundred, and you just need
to be able to tell them, vehicle

1771
01:33:42,980 --> 01:33:46,129
A, go to this position, vehicle
B go to this position, it's

1772
01:33:46,129 --> 01:33:49,545
actually scales relatively well
for that purpose. without having

1773
01:33:49,545 --> 01:33:52,694
to set up unique connections for
each one without having to

1774
01:33:52,694 --> 01:33:55,949
manage web RTC stuff. It's just
spiel out a bunch of commands

1775
01:33:55,949 --> 01:33:59,045
from anywhere you are in the
world and they'll get to your

1776
01:33:59,045 --> 01:34:02,354
robots. because it's built on
standard web technology, you get

1777
01:34:02,354 --> 01:34:05,663
a lot of that security, you get
a lot of that multi-threading,

1778
01:34:05,663 --> 01:34:08,812
multi-user agent stuff is all
well figured out. So we don't

1779
01:34:08,812 --> 01:34:11,000
have to reinvent any of that
stuff either

1780
01:34:11,000 --> 01:34:14,138
Stefan Seltz-Axmacher: And, it
can work both for this fun demo,

1781
01:34:14,138 --> 01:34:17,375
which is a neat thing to do from
a distance, but it can also work

1782
01:34:17,375 --> 01:34:20,264
if everything is on the same
site and there is no external

1783
01:34:20,264 --> 01:34:23,303
internet connection. The same
API can work for the vehicle in

1784
01:34:23,303 --> 01:34:24,000
that use case.

1785
01:34:24,000 --> 01:34:26,000
Audrow Nash: Oh, that's a big

1786
01:34:26,000 --> 01:34:29,099
Ilia Baranov: the API's hosted
locally, so the API's on the

1787
01:34:29,099 --> 01:34:32,040
vehicle. And so the API could
even be self triggered. So

1788
01:34:32,040 --> 01:34:34,929
something we don't have ready
yet, but I mentioned that

1789
01:34:34,929 --> 01:34:38,081
behavior tree. the long-term
goal here is to be able to give

1790
01:34:38,081 --> 01:34:41,180
our customers the ability to
generate little SubT trees and

1791
01:34:41,180 --> 01:34:44,541
inject them back onto the robot
for itself. concrete example, if

1792
01:34:44,541 --> 01:34:47,535
they're building their apple
picking robot, they focus on

1793
01:34:47,535 --> 01:34:50,897
their arm to pick the apple and
then they onboard trigger an API

1794
01:34:50,897 --> 01:34:52,000
function, move to the

1795
01:34:52,000 --> 01:34:53,000
Stefan Seltz-Axmacher: forward a
foot.

1796
01:34:53,000 --> 01:34:56,526
Ilia Baranov: which they've, or,
move to the next tree, which

1797
01:34:56,526 --> 01:34:59,936
they've defined as move until
you see a barrier under these

1798
01:34:59,936 --> 01:35:03,000
conditions, whatever. Like a
little subtree of logic.

1799
01:35:03,000 --> 01:35:07,626
Audrow Nash: I like that. Yeah,
so it can self trigger. So you,

1800
01:35:07,626 --> 01:35:12,326
can basically have a system make
its own calls from another call

1801
01:35:12,326 --> 01:35:16,952
so it can cascade in this kind
of thing. That seems powerful to

1802
01:35:16,952 --> 01:35:21,505
me. very interesting. Seems a
lot more complex, but seems like

1803
01:35:21,505 --> 01:35:25,911
you could do some cool things or
maybe you can get, behavior

1804
01:35:25,911 --> 01:35:30,391
that's more responsive with that
kind of thing. 'cause it can

1805
01:35:30,391 --> 01:35:34,724
trigger something right then I
bet. Is that the big benefit

1806
01:35:34,724 --> 01:35:37,000
that you get with it? Ah, cool.

1807
01:35:37,000 --> 01:35:40,109
Ilia Baranov: it's, local
interactivity is definitely one

1808
01:35:40,109 --> 01:35:43,273
of them. So, again, and this
isn't fully ready yet, so our

1809
01:35:43,273 --> 01:35:46,655
current API, we can publish the
API specs. They're online, but

1810
01:35:46,655 --> 01:35:50,036
they don't, they won't mention
this subtree behavior yet. But,

1811
01:35:50,036 --> 01:35:53,309
one of our early clients, they
just wanted a button that you

1812
01:35:53,309 --> 01:35:56,473
could press. To press a button
and the robot does x And we

1813
01:35:56,473 --> 01:35:59,745
thought, oh, we could write a
custom ROS interface. We could

1814
01:35:59,745 --> 01:36:02,800
do all these things to, to get
it to behave. But then we

1815
01:36:02,800 --> 01:36:06,182
thought we're already running an
API, let's just loop back the

1816
01:36:06,182 --> 01:36:09,236
API call to itself. And that
they just have their little

1817
01:36:09,236 --> 01:36:12,509
button do a little programmatic,
REST API call. Which sounds

1818
01:36:12,509 --> 01:36:15,509
funny, but it actually ended up
being the most flexible

1819
01:36:15,509 --> 01:36:16,000
approach.

1820
01:36:16,000 --> 01:36:20,020
Audrow Nash: Yeah, I would
imagine it would be the most

1821
01:36:20,020 --> 01:36:24,260
flexible approach. That seems
great. and you, do you think

1822
01:36:24,260 --> 01:36:29,011
that, do you think this is gonna
be a larger trend in Robotics? I

1823
01:36:29,011 --> 01:36:33,104
know like a lot of companies,
they end up wanting to put

1824
01:36:33,104 --> 01:36:37,198
things into some sort of
interface you can view from the

1825
01:36:37,198 --> 01:36:41,510
web. and I guess this often
means putting, making endpoints

1826
01:36:41,510 --> 01:36:45,603
for something. and sometimes
they'll use web sockets and

1827
01:36:45,603 --> 01:36:49,916
things for streaming video or
whatever it might be. but do,

1828
01:36:49,916 --> 01:36:54,375
you think that we're gonna see a
larger trend of robots doing

1829
01:36:54,375 --> 01:36:58,760
more, being, interacted with
through a REST API or something

1830
01:36:58,760 --> 01:37:03,000
like this? is that something
that you think will catch on?

1831
01:37:03,000 --> 01:37:06,691
Ilia Baranov: I, think, if I
could jump in, sorry, Stefan,

1832
01:37:06,691 --> 01:37:10,000
but, not necessarily REST in
particular, but I think

1833
01:37:10,000 --> 01:37:11,000
Audrow Nash: Just web standards.

1834
01:37:11,000 --> 01:37:13,798
Ilia Baranov: in general has.
like web, different web

1835
01:37:13,798 --> 01:37:16,913
websites, whatever it is. If
there's a new thing next, next

1836
01:37:16,913 --> 01:37:20,345
month, ROS special communication
protocol, like happy to write an

1837
01:37:20,345 --> 01:37:23,618
interface layer to it. 'cause we
don't care ultimately, right?

1838
01:37:23,618 --> 01:37:26,786
Like none of our actual safety
or behavior happens up there.

1839
01:37:26,786 --> 01:37:29,742
It's just a translator. and I
think we are moving in the

1840
01:37:29,742 --> 01:37:32,857
direction where there's more
connectivity for lower cost on

1841
01:37:32,857 --> 01:37:36,183
more devices. there's a lot of
this edge computing that's taken

1842
01:37:36,183 --> 01:37:39,562
hold, a lot of cloud computing.
so I think covering our eyes and

1843
01:37:39,562 --> 01:37:42,835
pretending that doesn't happen
in Robotics isn't the right way

1844
01:37:42,835 --> 01:37:45,000
forwards. I think we have to
accept that.

1845
01:37:45,000 --> 01:37:47,667
Audrow Nash: Definitely. And we
might as well leverage

1846
01:37:47,667 --> 01:37:49,000
everything that exists too.

1847
01:37:49,000 --> 01:37:50,000
Stefan Seltz-Axmacher: Yeah.

1848
01:37:50,000 --> 01:37:53,742
Audrow Nash: there's just so
much there where huge sums of

1849
01:37:53,742 --> 01:37:57,677
money have been invested into
improving these systems. and we

1850
01:37:57,677 --> 01:38:01,871
might as well not reinvent that.
So the web is great for a lot of

1851
01:38:01,871 --> 01:38:05,806
things like that, is my opinion
on it. Hell yeah. What a cool

1852
01:38:05,806 --> 01:38:08,000
thing. Thank you for sharing
that.

1853
01:38:08,000 --> 01:38:09,000
Stefan Seltz-Axmacher:
Absolutely.

1854
01:38:09,000 --> 01:38:13,742
Audrow Nash: Yeah. I heard the
podcast episode that I listened

1855
01:38:13,742 --> 01:38:18,484
to, that Nic shared, is, or you
guys talked about the demoing,

1856
01:38:18,484 --> 01:38:23,150
like instead of the kind of your
culture where it's just demo

1857
01:38:23,150 --> 01:38:27,969
constantly. Therefore it's not
like you do a big demo once in a

1858
01:38:27,969 --> 01:38:32,405
while and it breaks and this
kind of thing. And it, I, did

1859
01:38:32,405 --> 01:38:36,000
not know that I would be, seeing
a demo now. So

1860
01:38:36,000 --> 01:38:37,000
Stefan Seltz-Axmacher: I didn't
know that was

1861
01:38:36,000 --> 01:38:37,000
Audrow Nash: see. You're still

1862
01:38:37,000 --> 01:38:39,000
Stefan Seltz-Axmacher: you
listened to, so that worked out.

1863
01:38:39,000 --> 01:38:41,000
Audrow Nash: Hell yeah.

1864
01:38:41,000 --> 01:38:44,000
Ilia Baranov: CICD stands for
continuous demo by

1865
01:38:44,000 --> 01:38:45,000
Audrow Nash: Ah ha.

1866
01:38:44,000 --> 01:38:45,000
Ilia Baranov: just to be

1867
01:38:45,000 --> 01:38:49,140
Audrow Nash: Oh yeah. Yeah.
Instead of deployment. I love

1868
01:38:49,140 --> 01:38:53,570
it. Continuous demo. For you
guys it might be demos. Yeah. So

1869
01:38:53,570 --> 01:38:57,782
what do you, tell me about the
podcast. I, like at a, high

1870
01:38:57,782 --> 01:39:02,140
level, what is the Automate It
podcast and how, what are you

1871
01:39:02,140 --> 01:39:06,642
guys trying to achieve with it?
Or, I don't know, what do you,

1872
01:39:06,642 --> 01:39:11,000
what kind of questions do you
guys investigate? Maybe Stefan

1873
01:39:11,000 --> 01:39:15,710
Stefan Seltz-Axmacher: co of
course. Yeah. So every one to 12

1874
01:39:15,710 --> 01:39:20,498
weeks. we release a podcast on
our podcast called Automate It.

1875
01:39:20,498 --> 01:39:25,131
and, basically automate it is
the podcast version of the fun

1876
01:39:25,131 --> 01:39:29,996
Robotics conversations you have
with your friends at the end of

1877
01:39:29,996 --> 01:39:31,000
the long day.

1878
01:39:31,000 --> 01:39:32,000
Audrow Nash: does seem

1879
01:39:31,000 --> 01:39:34,716
Stefan Seltz-Axmacher: two
sections of it, where the first

1880
01:39:34,716 --> 01:39:38,495
is we each draw a random card to
represent a technology and

1881
01:39:38,495 --> 01:39:42,531
represent a business case, and
then we have to spitball into an

1882
01:39:42,531 --> 01:39:46,439
existence, a robot that uses
that technology in that business

1883
01:39:46,439 --> 01:39:50,603
case. And we try to make them be
stupid, sometimes ends up making

1884
01:39:50,603 --> 01:39:54,511
them be good. Often make them
stupid, but representative of a

1885
01:39:54,511 --> 01:39:58,611
real company, which is troubling
every time. and then the second

1886
01:39:58,611 --> 01:40:02,775
half, we'll talk about different
challenges in Robotics, like why

1887
01:40:02,775 --> 01:40:06,683
indoor localization is hard, or
why demoing often is good, or

1888
01:40:06,683 --> 01:40:10,398
any number of these kind of
normal things that happen when

1889
01:40:10,398 --> 01:40:12,000
you're working on robots.

1890
01:40:12,000 --> 01:40:18,000
Audrow Nash: Let's see. Ilia,
anything to add?

1891
01:40:18,000 --> 01:40:21,983
Ilia Baranov: yeah, I would just
add that generally we do this

1892
01:40:21,983 --> 01:40:26,031
talk after a beer or two, so it
tends to be a little bit it off

1893
01:40:26,031 --> 01:40:29,950
the rails at the start. I think
my favorite was we invented a

1894
01:40:29,950 --> 01:40:33,547
bartending robot that was a
bunch of arms on electrified

1895
01:40:33,547 --> 01:40:37,659
rails that would fly around the
bar to pour drinks or bring them

1896
01:40:37,659 --> 01:40:41,000
to people. And, yeah, very
economically viable. It's

1897
01:40:41,000 --> 01:40:44,496
Stefan Seltz-Axmacher: Yeah,
it'd be, a great business. but

1898
01:40:44,496 --> 01:40:48,230
in terms of why we started it,
so weird thing about our company

1899
01:40:48,230 --> 01:40:51,785
is there's, I'm sure there's
more personalities in this, but

1900
01:40:51,785 --> 01:40:55,222
maybe there's three archetypes
of types of people starting

1901
01:40:55,222 --> 01:40:59,015
Robotics projects. There's like
the optimistic producty business

1902
01:40:59,015 --> 01:41:02,807
guy who thinks robots just work
and let's figure out this, let's

1903
01:41:02,807 --> 01:41:05,000
build a robot around this idea
that I

1904
01:41:05,000 --> 01:41:07,000
Audrow Nash: it's an opportunity
basically.

1905
01:41:07,000 --> 01:41:11,009
Stefan Seltz-Axmacher: Yeah,
this is a great opportunity.

1906
01:41:11,009 --> 01:41:15,580
there's the like young, dumb and
incredibly confident, roboticist

1907
01:41:15,580 --> 01:41:19,870
who maybe they just got out of
CMU, maybe they just left some

1908
01:41:19,870 --> 01:41:23,598
big Robotics lab and like that
Robotics lab, they did

1909
01:41:23,598 --> 01:41:27,888
everything slightly wrong. And
as an individual roboticist, I

1910
01:41:27,888 --> 01:41:32,178
one time made a project that
looked neat for a minute. So I'm

1911
01:41:32,178 --> 01:41:36,257
gonna rebuild everything in
robotics myself to solve every

1912
01:41:36,257 --> 01:41:39,000
single problem and it's gonna go
great.

1913
01:41:39,000 --> 01:41:40,000
Audrow Nash: I've been

1914
01:41:39,000 --> 01:41:42,970
Stefan Seltz-Axmacher: kind of
third archetype who either of

1915
01:41:42,970 --> 01:41:46,808
those first two become after
robots are jerks for a while.

1916
01:41:46,808 --> 01:41:51,043
like the type of roboticist you
are when you once had dreams of,

1917
01:41:51,043 --> 01:41:54,947
solving general problems. And
now you're just hoping God, I

1918
01:41:54,947 --> 01:41:58,586
hope it turns on tomorrow and
can do something and that

1919
01:41:58,586 --> 01:42:02,622
people, let me keep on working
on this for Christ's sake. Why

1920
01:42:02,622 --> 01:42:06,658
won't it turn on? and that third
group is they tend to be our

1921
01:42:06,658 --> 01:42:10,562
best customer base. because the
people who think that robot

1922
01:42:10,562 --> 01:42:14,533
Robotics is solved, the people
who think that they alone can

1923
01:42:14,533 --> 01:42:18,503
solve it, they're not gonna let
us solve auto point to point

1924
01:42:18,503 --> 01:42:22,473
autonomy for them. 'cause that
will take them like a weekend

1925
01:42:22,473 --> 01:42:26,377
probably. it's really the people
who have been beaten up by

1926
01:42:26,377 --> 01:42:30,545
robots, chewed up, spat out, who
really wanna make a robot work

1927
01:42:30,545 --> 01:42:34,317
this time and, create real
value. That's, who we sell the

1928
01:42:34,317 --> 01:42:38,552
best to. So as a result, we made
a podcast to help people get to

1929
01:42:38,552 --> 01:42:41,000
that stage of, their robotics
journey

1930
01:42:41,000 --> 01:42:45,320
Audrow Nash: Oh, that's really
cool. Yeah. So basically you're

1931
01:42:45,320 --> 01:42:49,570
discussing all the things and
you're letting people in on the

1932
01:42:49,570 --> 01:42:53,680
process, and the idea is it
helps people become educated in

1933
01:42:53,680 --> 01:42:57,861
Robotics sooner about how to
make things and that they can't

1934
01:42:57,861 --> 01:42:58,000
do

1935
01:42:58,000 --> 01:43:01,528
Ilia Baranov: see the back to
see the backstage pain and

1936
01:43:01,528 --> 01:43:05,244
suffering. we were talking at
CSS just recently and we were

1937
01:43:05,244 --> 01:43:09,213
joking that we should have named
it actually Robotics Anonymous

1938
01:43:09,213 --> 01:43:12,866
as as a, as like a just a
support group for here's all the

1939
01:43:12,866 --> 01:43:14,000
pain we've, faced.

1940
01:43:14,000 --> 01:43:18,083
Audrow Nash: Yeah. Yeah. In the
demo one you guys talked about

1941
01:43:18,083 --> 01:43:22,034
some of your past experiences
and they seemed super painful,

1942
01:43:22,034 --> 01:43:25,920
with spinning parts that were
hot and it was super cold out

1943
01:43:25,920 --> 01:43:29,476
and it was just crazy. yeah, I
feel the, thing that is

1944
01:43:29,476 --> 01:43:33,361
interesting to me is I was
definitely stage two or whatever

1945
01:43:33,361 --> 01:43:37,312
of what you were doing while I
was in grad school. like I, I

1946
01:43:37,312 --> 01:43:41,000
tried to build ROS at some
point, and then by the time I

1947
01:43:41,000 --> 01:43:43,000
Stefan Seltz-Axmacher: you are
gonna do it the right way.

1948
01:43:43,000 --> 01:43:47,106
Audrow Nash: Yeah, I know, but,
and then I got to ar I needed

1949
01:43:47,106 --> 01:43:50,674
something like Rviz and
something like ROS2bag and it

1950
01:43:50,674 --> 01:43:54,848
was like, I guess I'll just use
it and this kind of thing. But

1951
01:43:54,848 --> 01:43:58,685
now, it's been interesting
'cause I've been mostly on the

1952
01:43:58,685 --> 01:44:02,523
software side. and so I'm trying
to do some Robotics side

1953
01:44:02,523 --> 01:44:06,831
projects occasionally and buy a
bunch of stuff from Sparkfun and

1954
01:44:06,831 --> 01:44:10,803
hook it up to ROS 2 and things
like this. And it's a lot of

1955
01:44:10,803 --> 01:44:14,707
work. and the thing that is
really becoming clear to me is

1956
01:44:14,707 --> 01:44:18,881
to do a good job at robotics and
to like actually build value.

1957
01:44:18,881 --> 01:44:23,190
You need a lot of people. Which
is interesting 'cause there's so

1958
01:44:23,190 --> 01:44:26,825
many things to do, where you
need the CEO who goes and

1959
01:44:26,825 --> 01:44:30,528
figures out the market and
knows, and you need a domain

1960
01:44:30,528 --> 01:44:34,163
expert and you need the
technical guy and you need all

1961
01:44:34,163 --> 01:44:38,000
the manufacturing knowledge and
there's a million things.

1962
01:44:38,000 --> 01:44:41,750
Stefan Seltz-Axmacher: Yeah. And
like I think I would argue that

1963
01:44:41,750 --> 01:44:45,266
like for robotics to be the
space, we all want it to be you.

1964
01:44:45,266 --> 01:44:48,898
You need to build a product, you
need to build a \company, you

1965
01:44:48,898 --> 01:44:52,648
build an organization where you
only need to be good at one part

1966
01:44:52,648 --> 01:44:53,000
of it.

1967
01:44:53,000 --> 01:44:54,000
Audrow Nash: Definitely.

1968
01:44:54,000 --> 01:44:57,608
Stefan Seltz-Axmacher: maybe
you're doing. the hardware,

1969
01:44:57,608 --> 01:45:01,537
maybe you're doing the retrofit.
Maybe you are getting really

1970
01:45:01,537 --> 01:45:05,274
smart about what farmers in Ohio
need as a ui, ux, and, as

1971
01:45:05,274 --> 01:45:09,332
Robotics behaviors, and you're
using somebody like us to do the

1972
01:45:09,332 --> 01:45:13,004
driving and someone else to
retrofit the vehicles and If,

1973
01:45:13,004 --> 01:45:17,063
you're willing to do just that,
it's still gonna turn out to be

1974
01:45:17,063 --> 01:45:21,186
incredibly hard. 'cause it turns
out like as much as like we all

1975
01:45:21,186 --> 01:45:25,245
might make fun of DocuSign 5.0,
it's hard to build DocuSign 5.0

1976
01:45:25,245 --> 01:45:28,852
or how to not for cats like
those are hard businesses to

1977
01:45:28,852 --> 01:45:32,589
build in their own right. And
like by focusing just on the

1978
01:45:32,589 --> 01:45:36,583
specific part of the value that
you have unique insight to, we

1979
01:45:36,583 --> 01:45:38,000
can all go make robots

1980
01:45:38,000 --> 01:45:43,000
Audrow Nash: I like that a lot.
Ilia, anything to add?

1981
01:45:43,000 --> 01:45:46,958
Ilia Baranov: I, I would say,
from the content creator

1982
01:45:46,958 --> 01:45:51,723
perspective, so Nicole, our kind
of marketing guru suggested this

1983
01:45:51,723 --> 01:45:56,194
to Stefan and I, and to start a
podcast. And at first kind of

1984
01:45:56,194 --> 01:45:57,000
both Stefan

1985
01:45:57,000 --> 01:45:58,000
Stefan Seltz-Axmacher: Hard

1986
01:45:58,000 --> 01:46:02,000
Ilia Baranov: eh, a little
leery, eh, I got enough.

1987
01:46:02,000 --> 01:46:04,000
Audrow Nash: You are doing a
lot.

1988
01:46:04,000 --> 01:46:05,000
Stefan Seltz-Axmacher: Who talks
on

1989
01:46:04,000 --> 01:46:08,337
Ilia Baranov: at them this week
at some point. Yeah. but but

1990
01:46:08,337 --> 01:46:12,964
realistically, it ended up being
A super fun for us. And I think

1991
01:46:12,964 --> 01:46:17,301
that kind of unstructured part
helps get us in the mood and,

1992
01:46:17,301 --> 01:46:21,277
talk about random, crazy
Robotics anonymous topics. and

1993
01:46:21,277 --> 01:46:25,759
secondly, I think it ended up
being hilariously effective as a

1994
01:46:25,759 --> 01:46:28,000
marketing and re outreach tool.

1995
01:46:28,000 --> 01:46:29,000
Audrow Nash: Oh

1996
01:46:29,000 --> 01:46:31,614
Ilia Baranov: it's, not a, we
don't have a thousand, we don't

1997
01:46:31,614 --> 01:46:34,186
even have a thousand listeners.
we don't have a lot, it's, a

1998
01:46:34,186 --> 01:46:35,000
very small listener

1999
01:46:35,000 --> 01:46:36,000
Stefan Seltz-Axmacher: It's a
niche podcast.

2000
01:46:36,000 --> 01:46:39,389
Ilia Baranov: is, yeah. it's
very niche, right? But the funny

2001
01:46:39,389 --> 01:46:42,889
thing is, any conference we'll
go to, invariably, at least one,

2002
01:46:42,889 --> 01:46:46,333
if not multiple people will come
up and be like, oh yeah, you,

2003
01:46:46,333 --> 01:46:48,000
I've listened to your podcast.

2004
01:46:48,000 --> 01:46:49,000
Stefan Seltz-Axmacher: And
they'll have,

2005
01:46:48,000 --> 01:46:52,000
Ilia Baranov: the people we
wanna work with. yep,

2006
01:46:52,000 --> 01:46:54,828
Stefan Seltz-Axmacher: yep. And
they'll be specifically yeah,

2007
01:46:54,828 --> 01:46:57,517
maybe you can help us. What
we're doing is really hard. It

2008
01:46:57,517 --> 01:47:00,252
seems like it's been hard for
you before too, and maybe you

2009
01:47:00,252 --> 01:47:03,172
can do some things better than
we can. Yeah, we'd love to, we'd

2010
01:47:03,172 --> 01:47:06,000
love to, commiserate with you on
robots and help automate it.

2011
01:47:06,000 --> 01:47:11,575
Audrow Nash: Yes. Oh, that's so
funny. Yeah, I I have that too,

2012
01:47:11,575 --> 01:47:16,000
where it's a small fan base, but
they are awesome.

2013
01:47:16,000 --> 01:47:17,000
Stefan Seltz-Axmacher: Yep.

2014
01:47:17,000 --> 01:47:20,657
Audrow Nash: I get recognized
just a handful of times at every

2015
01:47:20,657 --> 01:47:24,314
conference. They're like, oh, I
love the pod, whatever. Or you

2016
01:47:24,314 --> 01:47:28,088
should talk to this person. It's
just, I don't know. It's such a

2017
01:47:28,088 --> 01:47:31,745
cool, wonderful thing. I feel
like talking to a small audience

2018
01:47:31,745 --> 01:47:34,871
that really likes the content
that you make is really

2019
01:47:34,871 --> 01:47:38,410
wonderful. there's something
just glorious in it, like very,

2020
01:47:38,410 --> 01:47:39,000
enjoyable.

2021
01:47:39,000 --> 01:47:41,776
Stefan Seltz-Axmacher: I think
it goes back to a basic like

2022
01:47:41,776 --> 01:47:44,600
Paul Graham of make a hundred
people love, or make 10 people

2023
01:47:44,600 --> 01:47:47,000
love you, as opposed to a
thousand people like you.

2024
01:47:47,000 --> 01:47:51,609
Audrow Nash: ah, yeah. It's a
good way to look at it, Paul.

2025
01:47:51,609 --> 01:47:52,000
Yeah.

2026
01:47:52,000 --> 01:47:53,000
Stefan Seltz-Axmacher: Yeah.

2027
01:47:53,000 --> 01:47:57,323
Audrow Nash: Hell yeah. that's
super cool. I love the reason

2028
01:47:57,323 --> 01:48:01,573
for doing the podcast. and it's
funny that it can bring you

2029
01:48:01,573 --> 01:48:05,752
people to commiserate with and
help educate to move people

2030
01:48:05,752 --> 01:48:10,075
towards this. robots can be
really cool and they can solve a

2031
01:48:10,075 --> 01:48:14,182
lot of things, but they're
really hard and I dunno. Okay.

2032
01:48:14,182 --> 01:48:18,000
So you gotta approach it that
way, like they're super

2033
01:48:18,000 --> 01:48:21,000
Stefan Seltz-Axmacher: Yep.

2034
01:48:21,000 --> 01:48:25,414
Audrow Nash: Let's see. where do
people find your podcast? Is it

2035
01:48:25,414 --> 01:48:27,000
on all of the different

2036
01:48:27,000 --> 01:48:30,641
Stefan Seltz-Axmacher: so we're
on all the major, outlets,

2037
01:48:30,641 --> 01:48:34,157
Spotify, Apple Podcasts,
whatever, just call to automate

2038
01:48:34,157 --> 01:48:37,547
it. And there's a nice
cartoonish image, of Ilia and I

2039
01:48:37,547 --> 01:48:41,000
both donning facial hair and
looking incredibly, dapper

2040
01:48:41,000 --> 01:48:44,229
Audrow Nash: Nice. Yeah, I think
what, it's funny 'cause I've

2041
01:48:44,229 --> 01:48:47,088
listened to a couple of your
episodes and they're, the

2042
01:48:47,088 --> 01:48:50,000
beginning is very much
complimenting each other's hair,

2043
01:48:50,000 --> 01:48:51,000
Stefan Seltz-Axmacher: hair.

2044
01:48:50,000 --> 01:48:53,000
Audrow Nash: this kind of thing.
And it's oh, your

2045
01:48:53,000 --> 01:48:54,000
Stefan Seltz-Axmacher: for
anyone

2046
01:48:54,000 --> 01:48:56,000
Audrow Nash: today. A little
goofy.

2047
01:48:56,000 --> 01:48:59,137
Stefan Seltz-Axmacher: for
anyone, anyone viewing this,

2048
01:48:59,137 --> 01:49:02,673
that just makes obvious sense.
And for the people who are just

2049
01:49:02,673 --> 01:49:05,981
listening, imagine the most
incredibly good looking people

2050
01:49:05,981 --> 01:49:09,460
ever. And that's exactly what we
look like, especially in the

2051
01:49:09,460 --> 01:49:11,000
hair department. Just yeah,

2052
01:49:11,000 --> 01:49:14,000
Audrow Nash: And then do not go
onto the YouTube. And

2053
01:49:14,000 --> 01:49:19,164
Ilia Baranov: Stefan starts the
episodes critiquing. My, Stefan

2054
01:49:19,164 --> 01:49:24,000
usually starts by critiquing
whatever I'm wearing that day,

2055
01:49:24,000 --> 01:49:25,000
Audrow Nash: really,

2056
01:49:25,000 --> 01:49:28,843
Ilia Baranov: always a good
start to our podcast episode.

2057
01:49:28,843 --> 01:49:31,000
It's ah, same shirt. I see. Huh.

2058
01:49:31,000 --> 01:49:34,000
Audrow Nash: I wear a black
shirt in every episode, but

2059
01:49:34,000 --> 01:49:37,526
Stefan Seltz-Axmacher: This
sounds a lot like my wife who

2060
01:49:37,526 --> 01:49:40,000
tells me I look the best with a
mask on.

2061
01:49:40,000 --> 01:49:45,217
Audrow Nash: how funny. Not
gonna touch that one. Let's see,

2062
01:49:45,217 --> 01:49:50,522
we're, coming to the end of the
time. I can talk a little bit

2063
01:49:50,522 --> 01:49:55,913
more. I don't know if you guys
have a hard cutoff or anything.

2064
01:49:55,913 --> 01:50:01,304
but we're at the end of, okay.
Okay. I, wanted to see what you

2065
01:50:01,304 --> 01:50:06,696
guys have for advice for people,
wanting to get into Robotics.

2066
01:50:06,696 --> 01:50:11,304
say, it's someone who's in
their, the middle of their

2067
01:50:11,304 --> 01:50:16,522
university or something, what
should they know or what would

2068
01:50:16,522 --> 01:50:20,000
save them time? maybe start with
Stefan.

2069
01:50:20,000 --> 01:50:22,333
Stefan Seltz-Axmacher: yeah, so
I'll, maybe I'll answer this

2070
01:50:22,333 --> 01:50:24,511
more. For people who are
interested in Robotics products

2071
01:50:24,511 --> 01:50:27,000
and Ilia can talk about it more
from an engineering perspective.

2072
01:50:27,000 --> 01:50:28,000
Audrow Nash: it.

2073
01:50:28,000 --> 01:50:32,355
Stefan Seltz-Axmacher: I think
the, when you're, when you have

2074
01:50:32,355 --> 01:50:36,780
not built a robot yourself and
you're looking in a very common,

2075
01:50:36,780 --> 01:50:40,994
intellectual fallacy, or flaw
that you fall into is that you

2076
01:50:40,994 --> 01:50:45,419
see things that have been built
before and you assume that's no

2077
01:50:45,419 --> 01:50:49,493
big deal. And I, I'll just, I'll
grab an arm from a Toyota

2078
01:50:49,493 --> 01:50:53,707
factory, I'll grab a Waymo car
and then I'll have a car that

2079
01:50:53,707 --> 01:50:57,922
can drive around and put mail in
mailboxes and it's gonna be

2080
01:50:57,922 --> 01:51:01,644
great. the, weirdest and,
shittiest and hardest thing

2081
01:51:01,644 --> 01:51:05,789
about Robotics from a, product
and a business and investing

2082
01:51:05,789 --> 01:51:09,933
perspective is that unlike the
world of SaaS apps where any

2083
01:51:09,933 --> 01:51:13,655
feature is purchasable as a
microservice in Robotics,

2084
01:51:13,655 --> 01:51:18,080
basically every feature that you
see, every supporting piece of

2085
01:51:18,080 --> 01:51:22,084
tooling, every supporting piece
of infrastructure, almost

2086
01:51:22,084 --> 01:51:26,509
everything that you see has to
be reinvented from scratch. with

2087
01:51:26,509 --> 01:51:30,794
the exception of some of the
stuff in ROS that is in general,

2088
01:51:30,794 --> 01:51:35,008
mostly a lot harder to set up
than any piece of tooling that

2089
01:51:35,008 --> 01:51:39,293
you would get for starting an
e-commerce site. So like a good

2090
01:51:39,293 --> 01:51:43,296
Robotics business, a good
Robotics product is a matter of

2091
01:51:43,296 --> 01:51:47,651
what is a thing that is super
expensive for humans to do. Like

2092
01:51:47,651 --> 01:51:52,076
just there's an incredible labor
shortage around it and is drop

2093
01:51:52,076 --> 01:51:56,361
dead simple for a robot to do.
'cause it needs the same thing

2094
01:51:56,361 --> 01:52:00,715
to be done over and over again.
And start with that. And don't

2095
01:52:00,715 --> 01:52:05,000
just start by combining robots
you've seen on, in TechCrunch.

2096
01:52:05,000 --> 01:52:12,000
Audrow Nash: love it. yeah.
Ilia, what do you think?

2097
01:52:12,000 --> 01:52:15,145
Ilia Baranov: So on, on my end,
my thoughts are summarized. So

2098
01:52:15,145 --> 01:52:18,138
I, really love XKCD I'm sure a
lot of people listening have

2099
01:52:18,138 --> 01:52:19,000
heard of that web

2100
01:52:19,000 --> 01:52:21,000
Audrow Nash: Yeah, they're
great. I.

2101
01:52:21,000 --> 01:52:24,112
Ilia Baranov: there's a
particular one that's There's

2102
01:52:24,112 --> 01:52:27,694
one that's oh, we built this
robot that like shoots lightning

2103
01:52:27,694 --> 01:52:31,335
and is also an air balloon and
whatever. And somebody ask what

2104
01:52:31,335 --> 01:52:35,094
is that for? And it's search and
rescue. that's like the default

2105
01:52:35,094 --> 01:52:38,558
answer for robot projects. It's
oh, search and rescue. so I

2106
01:52:38,558 --> 01:52:41,906
think the trap that engineers
fall into is solutionism or

2107
01:52:41,906 --> 01:52:45,605
robots for robot's sake. And I
love to do that for my own life.

2108
01:52:45,605 --> 01:52:48,541
Like I build lots of fun
trinkets just for fun for

2109
01:52:48,541 --> 01:52:51,947
myself, but that is not a
business. And so as an engineer,

2110
01:52:51,947 --> 01:52:55,647
really try to step back and to
give it the question of, is this

2111
01:52:55,647 --> 01:52:59,053
the most effective way I can
solve this problem? Is robots

2112
01:52:59,053 --> 01:53:02,517
really the answer here? If not,
you're gonna bang your head

2113
01:53:02,517 --> 01:53:06,334
against that problem, regardless
of how good your tech is. You're

2114
01:53:06,334 --> 01:53:09,740
just not gonna solve it. I'll
give a really quick concrete

2115
01:53:09,740 --> 01:53:12,559
example. there's a mining
project, which they're

2116
01:53:12,559 --> 01:53:16,317
interested in building autonomy
and stuff like that, but, in the

2117
01:53:16,317 --> 01:53:19,606
short term, they're saying
their, immediate term goal is

2118
01:53:19,606 --> 01:53:23,012
just to reduce the amount of
idling that their fleet does,

2119
01:53:23,012 --> 01:53:26,535
because across their fleet, when
their drivers just idle the

2120
01:53:26,535 --> 01:53:30,000
engine, they waste like multiple
millions of dollars a year

2121
01:53:30,000 --> 01:53:33,563
Stefan Seltz-Axmacher: $8
million a year of diesel wasted

2122
01:53:33,563 --> 01:53:37,313
because the operators won't,
turn the engine off. because if

2123
01:53:37,313 --> 01:53:41,000
the engine's turned off, it
looks like they're not working.

2124
01:53:41,000 --> 01:53:42,000
Audrow Nash: huh?

2125
01:53:42,000 --> 01:53:45,433
Ilia Baranov: Or, 'cause they
want to keep the AC on, right?

2126
01:53:45,433 --> 01:53:48,638
And, as an engineer, my, my
thing is oh, we can automate

2127
01:53:48,638 --> 01:53:52,185
their vehicles and we can do all
this thing, we can put in all

2128
01:53:52,185 --> 01:53:55,676
these monitoring systems. And
then the answer is or you could

2129
01:53:55,676 --> 01:53:59,166
just have a 10 cent circuit that
if you're idling from longer

2130
01:53:59,166 --> 01:54:02,599
than X automatically turns off
your engine. And like that 50

2131
01:54:02,599 --> 01:54:05,460
cents or that $2 solution is
worth $8 million. Not

2132
01:54:05,460 --> 01:54:09,065
particularly interesting. It's a
it's a 555 timer and a switch,

2133
01:54:09,065 --> 01:54:12,270
but worth $8 million off the
bat. so really when you're,

2134
01:54:12,270 --> 01:54:15,646
being an engineer, really try to
think do you actually need

2135
01:54:15,646 --> 01:54:18,850
robots here? And if the answer's
no and you want to be a

2136
01:54:18,850 --> 01:54:22,283
roboticist look for a different
idea, you're, not gonna tech

2137
01:54:22,283 --> 01:54:24,000
your way through this problem,

2138
01:54:24,000 --> 01:54:26,000
Stefan Seltz-Axmacher: Yeah,

2139
01:54:26,000 --> 01:54:27,000
Audrow Nash: oh, go ahead,
Stefan.

2140
01:54:27,000 --> 01:54:30,962
Stefan Seltz-Axmacher: I was
gonna say, so an interesting

2141
01:54:30,962 --> 01:54:35,272
thing, when I was working at my
self-driving trucking company,

2142
01:54:35,272 --> 01:54:39,442
the point where being better at
writing software to manage a

2143
01:54:39,442 --> 01:54:43,682
trucking fleet would make us
more profit than solving general

2144
01:54:43,682 --> 01:54:47,436
autonomy. like we were
autonomous, like 90, 95% of the

2145
01:54:47,436 --> 01:54:51,398
time when we're on highways
teleoperated for the last 5%.

2146
01:54:51,398 --> 01:54:55,777
And like being more clever about
how we route trucks being more

2147
01:54:55,777 --> 01:55:00,226
clever about like automatically
telling more brokers that trucks

2148
01:55:00,226 --> 01:55:04,327
are available for a load. That
type of stuff would add more

2149
01:55:04,327 --> 01:55:08,637
cents per mile of profit than
$10 billion AGI solution type of

2150
01:55:08,637 --> 01:55:13,155
thing. and that realization made
my board member think, or, maybe

2151
01:55:13,155 --> 01:55:16,700
I should scratch that, made
senior financial people

2152
01:55:16,700 --> 01:55:20,731
connected to us. Think, why do
you guys just let go of the

2153
01:55:20,731 --> 01:55:24,485
Robotics team? Just focus on
being a software, focused

2154
01:55:24,485 --> 01:55:28,934
trucking company. And like when
robot trucks get solved, you buy

2155
01:55:28,934 --> 01:55:33,174
whoever makes the best ones.
Which I think is gonna happen to

2156
01:55:33,174 --> 01:55:37,136
a lot of the vertical folks
where it's gonna become like,

2157
01:55:37,136 --> 01:55:40,542
oh. running a, day-to-day
business in this, heavy

2158
01:55:40,542 --> 01:55:44,782
industry. These companies suck a
lot Just by bringing in like

2159
01:55:44,782 --> 01:55:48,258
standard Silicon Valley
practices of how to manage

2160
01:55:48,258 --> 01:55:52,150
resources would make you more
efficient than most of the

2161
01:55:52,150 --> 01:55:55,000
world's biggest industrial
players. Yeah.

2162
01:55:55,000 --> 01:55:59,242
Audrow Nash: That's bonkers.
What a thing. That was probably

2163
01:55:59,242 --> 01:56:02,000
a hard time, I would imagine,
with that

2164
01:56:02,000 --> 01:56:05,081
Stefan Seltz-Axmacher: Yeah.
Yeah. Yeah. That was a like,

2165
01:56:05,081 --> 01:56:08,162
huh, I can't tell a lot of
people on the team about this,

2166
01:56:08,162 --> 01:56:10,000
or morale would drop pretty
sadly.

2167
01:56:10,000 --> 01:56:13,841
Audrow Nash: Yeah. What a thing.
How long, how much longer what

2168
01:56:13,841 --> 01:56:17,622
actually, I know we're late in
the thing, but what happened to

2169
01:56:17,622 --> 01:56:20,000
that company and where are they
at now?

2170
01:56:20,000 --> 01:56:23,827
Stefan Seltz-Axmacher: Starsky
ran outta money. We, we failed

2171
01:56:23,827 --> 01:56:27,653
pretty, in like late 2019, right
before 2020. Around that. By

2172
01:56:27,653 --> 01:56:31,292
that time, we, to my knowledge,
were the most commercially

2173
01:56:31,292 --> 01:56:34,930
successful L four trucking
company. We were doing about $7

2174
01:56:34,930 --> 01:56:38,694
million in annualized revenue
run rate in freight that we're

2175
01:56:38,694 --> 01:56:42,018
hauling mostly with regular
trucks. But our customers

2176
01:56:42,018 --> 01:56:45,970
thought we were hauling freight
with like unmanned self-driving

2177
01:56:45,970 --> 01:56:49,923
trucks. We did the world's first
ever driver out autonomous run

2178
01:56:49,923 --> 01:56:53,875
on a live highway. So we drove a
truck for seven miles, like in

2179
01:56:53,875 --> 01:56:57,701
regular traffic with no person
in it. and we, went to VCs and

2180
01:56:57,701 --> 01:57:01,465
were like, Hey, look the tech
works like we need. Here's our

2181
01:57:01,465 --> 01:57:05,229
safety testing plan that will
allow us In about 12 months to

2182
01:57:05,229 --> 01:57:09,118
be having five trucks driving
about full-time and no person in

2183
01:57:09,118 --> 01:57:12,945
it. The business more or less
works like, here are the things

2184
01:57:12,945 --> 01:57:16,646
that we need to do to make it be
outright profitable and an

2185
01:57:16,646 --> 01:57:20,347
interesting and like very cookie
cutter, running a trucking

2186
01:57:20,347 --> 01:57:24,236
company, types of things that we
need to do. We've made, we've

2187
01:57:24,236 --> 01:57:28,000
proven all the stuff, let's go,
give us $50 million dollars.

2188
01:57:28,000 --> 01:57:29,000
Audrow Nash: Yeah, we did the
hard

2189
01:57:28,000 --> 01:57:31,202
Stefan Seltz-Axmacher: Super
profitable. And VCs looked at

2190
01:57:31,202 --> 01:57:34,681
us, they're like, wow. So you're
working really hard. running a

2191
01:57:34,681 --> 01:57:37,000
trucking company is hard. Robots
are hard.

2192
01:57:37,000 --> 01:57:38,000
Ilia Baranov: hard.

2193
01:57:38,000 --> 01:57:41,650
Stefan Seltz-Axmacher: and
you're saying that with another

2194
01:57:41,650 --> 01:57:45,238
$50 million, maybe you can be a
trucking company with 50%

2195
01:57:45,238 --> 01:57:47,000
margins. That kind of sucks.

2196
01:57:47,000 --> 01:57:48,000
Audrow Nash: Oh

2197
01:57:47,000 --> 01:57:51,460
Stefan Seltz-Axmacher: I would
way rather give $50 million to

2198
01:57:51,460 --> 01:57:55,407
DocuSign 5.0 where for $50
million, they'll create $40

2199
01:57:55,407 --> 01:58:00,086
million of revenue, but they'll
have 90% margins on it. And, and

2200
01:58:00,086 --> 01:58:04,473
anyone in San Francisco can run
that company. Like you don't

2201
01:58:04,473 --> 01:58:09,225
need some special weird unicorn.
Like anyone in the world can run

2202
01:58:09,225 --> 01:58:13,392
a unicorn SaaS company. this,
this autonomy thing isn't a

2203
01:58:13,392 --> 01:58:15,000
great business, is it?

2204
01:58:15,000 --> 01:58:21,100
Audrow Nash: Oh, what a thing.
Okay. And then, so after that,

2205
01:58:21,100 --> 01:58:25,000
then you effectively picked
yourself up

2206
01:58:25,000 --> 01:58:26,000
Stefan Seltz-Axmacher: Yep.

2207
01:58:26,000 --> 01:58:27,000
Ilia Baranov: Yep,

2208
01:58:26,000 --> 01:58:31,061
Audrow Nash: started, Polymath
with this and used a lot of the

2209
01:58:31,061 --> 01:58:34,000
lessons and like Interesting. I
see.

2210
01:58:34,000 --> 01:58:37,000
Stefan Seltz-Axmacher: So like
I, COVID happened. I went and

2211
01:58:37,000 --> 01:58:40,050
hung out with a bunch of SaaS
companies. I had an interesting

2212
01:58:40,050 --> 01:58:43,150
experience where, while I was
hanging out with SaaS companies,

2213
01:58:43,150 --> 01:58:46,150
I took a deep look at like the
NetSuite ecosystem, if you're

2214
01:58:46,150 --> 01:58:47,000
familiar with it.

2215
01:58:47,000 --> 01:58:51,000
Audrow Nash: Is that like
Microsoft what? What's NetSuite?

2216
01:58:51,000 --> 01:58:54,147
Stefan Seltz-Axmacher: so
netSuite is this big, crappy

2217
01:58:54,147 --> 01:58:57,818
piece of enterprise software to
do accounting set up in a super

2218
01:58:57,818 --> 01:59:01,605
special way. So if you've worked
at a company with more, than 500

2219
01:59:01,605 --> 01:59:04,868
people, they probably use
NetSuite or something like it,

2220
01:59:04,868 --> 01:59:08,423
think of it like QuickBooks, but
a couple million dollars and

2221
01:59:08,423 --> 01:59:12,211
then a couple million dollars to
implement it a super special way

2222
01:59:12,211 --> 01:59:15,648
because Coke likes to do, its
accounting slightly different

2223
01:59:15,648 --> 01:59:19,261
than Pepsi does, who likes to do
it slightly different than, I

2224
01:59:19,261 --> 01:59:22,641
don't know, Snapple or whatever.
and when I looked at that

2225
01:59:22,641 --> 01:59:26,195
ecosystem, I saw software that
didn't just work, right? There

2226
01:59:26,195 --> 01:59:29,342
was software that like, it
wasn't like Gmail where you

2227
01:59:29,342 --> 01:59:33,013
swipe a credit card and you have
an inter and you have an email

2228
01:59:33,013 --> 01:59:36,568
account. It wasn't like Uber
where you download an app, hit a

2229
01:59:36,568 --> 01:59:40,064
button and a car shows up. it
was software where you buy it,

2230
01:59:40,064 --> 01:59:43,269
you spend a bunch of time
integrating it, not even just

2231
01:59:43,269 --> 01:59:46,241
business to business, but you
spend a bunch of time

2232
01:59:46,241 --> 01:59:49,853
configuring it to work just for
you. And then it is one of the

2233
01:59:49,853 --> 01:59:53,000
most valuable pieces of software
in your entire stack.

2234
01:59:53,000 --> 01:59:54,000
Audrow Nash: whoa.

2235
01:59:54,000 --> 01:59:57,232
Stefan Seltz-Axmacher: And I
don't think Robotics is ready to

2236
01:59:57,232 --> 02:00:00,411
be Gmail yet, and I don't think
Robotics is ready to be Uber

2237
02:00:00,411 --> 02:00:03,536
yet. But I think Robotics is
ready to be something that you

2238
02:00:03,536 --> 02:00:06,662
buy, configure, and then is a
key part of how your business

2239
02:00:06,662 --> 02:00:10,000
operates. And that's why we're
building Polymath the way we are

2240
02:00:10,000 --> 02:00:13,333
Audrow Nash: I love it. Almost
like a, I don't know, Zapier or

2241
02:00:13,333 --> 02:00:15,000
something, like some automation

2242
02:00:15,000 --> 02:00:18,966
Stefan Seltz-Axmacher: like,
like Zapier without a nice UI

2243
02:00:18,966 --> 02:00:23,000
with Zapier, but you have to
hire consultants to set it up.

2244
02:00:23,000 --> 02:00:25,000
Audrow Nash: Yeah. Cool.

2245
02:00:25,000 --> 02:00:28,841
Stefan Seltz-Axmacher: Zapier's
pitch was probably like we make

2246
02:00:28,841 --> 02:00:32,439
it so you can do Oracle type
integrations without a team of

2247
02:00:32,439 --> 02:00:35,000
consultants was probably their
seed pitch.

2248
02:00:35,000 --> 02:00:39,000
Audrow Nash: huh. Yeah. What a
thing. Okay.

2249
02:00:39,000 --> 02:00:41,000
Stefan Seltz-Axmacher: Okay.

2250
02:00:41,000 --> 02:00:46,239
Audrow Nash: let's see. So last,
there's one more thing I wanted

2251
02:00:46,239 --> 02:00:51,314
to ask. So we had, advice, you
guys both gave your advice. Oh,

2252
02:00:51,314 --> 02:00:56,389
Ilia, I wanted to ask, 'cause
you mentioned side projects, you

2253
02:00:56,389 --> 02:01:01,219
build your own trinkets for
these kinds of things. I wanted

2254
02:01:01,219 --> 02:01:06,294
to see, what you think of side
projects. I think they're super

2255
02:01:06,294 --> 02:01:10,715
important, as a way of growing
your skills and staying

2256
02:01:10,715 --> 02:01:15,790
relevant, but I wanna see what
your perspective is on them and

2257
02:01:15,790 --> 02:01:18,000
from a CTO perspective too.

2258
02:01:18,000 --> 02:01:22,428
Ilia Baranov: Yeah. yeah, no,
for sure. I, two things. One is

2259
02:01:22,428 --> 02:01:26,856
my side projects keep me sane,
because, especially as you get

2260
02:01:26,856 --> 02:01:31,574
into the, more CTO roles, like I
would love for most of my day to

2261
02:01:31,574 --> 02:01:36,147
be coding, but like maybe 10% of
it is realistically and 90% is

2262
02:01:36,147 --> 02:01:40,430
everything else. so doing the
side projects keeps me happy.

2263
02:01:40,430 --> 02:01:44,568
but the other funny thing is
that my side projects end up

2264
02:01:44,568 --> 02:01:48,560
being, aligned with Polymath
some of the time, which is

2265
02:01:48,560 --> 02:01:52,988
interesting. so early, on we
didn't really have a good way to

2266
02:01:52,988 --> 02:01:56,980
emergency stop our vehicles
remotely. So as the demo we

2267
02:01:56,980 --> 02:02:01,408
gave, we have this completely
remote vehicle, so like none of

2268
02:02:01,408 --> 02:02:05,909
the line of sight e stops will
work for us, right? And fort's,

2269
02:02:05,909 --> 02:02:10,264
wireless over network stuff
wasn't ready and nobody else was

2270
02:02:10,264 --> 02:02:14,547
providing anything of real
value. so because I had built up

2271
02:02:14,547 --> 02:02:18,685
some stuff in like Arduino land
basically, I was able to,

2272
02:02:18,685 --> 02:02:22,605
quickly throw together a
prototype emergency stop that

2273
02:02:22,605 --> 02:02:25,000
ended up being relatively
robust.

2274
02:02:25,000 --> 02:02:27,000
Audrow Nash: Nice

2275
02:02:27,000 --> 02:02:30,621
Ilia Baranov: and those sort of
things, I, think. TLDR is that

2276
02:02:30,621 --> 02:02:33,775
side projects are critical,
especially I'd say there's

2277
02:02:33,775 --> 02:02:37,279
almost like a bathtub curve when
you're junior. They let you

2278
02:02:37,279 --> 02:02:40,958
iterate quickly, and when you're
very senior, they let you keep

2279
02:02:40,958 --> 02:02:44,054
your sanity because seniority
tends to correlate with

2280
02:02:44,054 --> 02:02:47,733
meetings. It, not always, but in
a lot of ways. so they let you

2281
02:02:47,733 --> 02:02:51,237
like just quickly prototype
something and test it out. and I

2282
02:02:51,237 --> 02:02:54,391
find the risk for, I would say
not as much in robotics

2283
02:02:54,391 --> 02:02:57,662
companies, but for software
companies in general is over

2284
02:02:57,662 --> 02:03:00,874
time as they become larger,
there's this growing divide

2285
02:03:00,874 --> 02:03:04,553
between people making decisions
and people writing code. And if

2286
02:03:04,553 --> 02:03:07,882
the people making decisions
don't make a habit of writing

2287
02:03:07,882 --> 02:03:11,561
code and being connected to the
system and like basically being

2288
02:03:11,561 --> 02:03:15,066
a chaos monkey is another thing
that I do quite often that I

2289
02:03:15,066 --> 02:03:16,000
highly recommend

2290
02:03:16,000 --> 02:03:17,000
Audrow Nash: I dunno what it

2291
02:03:17,000 --> 02:03:20,273
Ilia Baranov: put out an API and
I'll just write like the worst

2292
02:03:20,273 --> 02:03:21,000
possible code.

2293
02:03:21,000 --> 02:03:22,000
Audrow Nash: Yeah.

2294
02:03:22,000 --> 02:03:25,410
Ilia Baranov: yeah, it's just
I'll write the worst possible

2295
02:03:25,410 --> 02:03:28,819
code to talk to our API and
break a whole bunch of stuff by

2296
02:03:28,819 --> 02:03:32,286
accident, but by doing that,
we'll figure out like, oh, like

2297
02:03:32,286 --> 02:03:36,043
we didn't handle this exception.
Or a, dumb user like me would've

2298
02:03:36,043 --> 02:03:39,626
thought it works this way, but
it works that way. And like all

2299
02:03:39,626 --> 02:03:43,035
of this stuff. and, all of that
all flows into keeping your

2300
02:03:43,035 --> 02:03:45,000
sanity and having a good
business.

2301
02:03:45,000 --> 02:03:49,441
Audrow Nash: Hell yeah. Yeah. I,
think it's an interesting thing.

2302
02:03:49,441 --> 02:03:53,471
I feel like, I've been doing
side projects really regularly

2303
02:03:53,471 --> 02:03:57,638
for a few years now. Like every
evening I'm working on a side

2304
02:03:57,638 --> 02:04:01,327
project, which I absolutely
love. It's like some of my

2305
02:04:01,327 --> 02:04:05,563
favorite time each day. but or,
so one of the things with that

2306
02:04:05,563 --> 02:04:09,252
is I feel like, you learn
through your own pain how to

2307
02:04:09,252 --> 02:04:12,668
make good architectural
decisions. And I feel like

2308
02:04:12,668 --> 02:04:16,630
that's a very hard thing unless
you're going end to end in

2309
02:04:16,630 --> 02:04:20,934
something you're building. And
it's, it like if I just focus on

2310
02:04:20,934 --> 02:04:25,033
the one specific narrow thing
that I'm working on in work, I

2311
02:04:25,033 --> 02:04:29,405
don't understand the end-to-end
process, but when I'm working on

2312
02:04:29,405 --> 02:04:33,572
that one specific thing, it's
easier to make better decisions

2313
02:04:33,572 --> 02:04:37,535
if I understand the larger
context through doing the whole

2314
02:04:37,535 --> 02:04:41,702
process through side projects.
That's been one of my thoughts

2315
02:04:41,702 --> 02:04:43,000
on it recently, but

2316
02:04:43,000 --> 02:04:44,000
Ilia Baranov: totally,

2317
02:04:43,000 --> 02:04:44,000
Stefan Seltz-Axmacher: of sense.

2318
02:04:44,000 --> 02:04:48,683
Audrow Nash: hell yeah. alright.
So questions. what are you guys

2319
02:04:48,683 --> 02:04:53,146
excited about in Robotics? what
are the things you're looking

2320
02:04:53,146 --> 02:04:56,000
forward to? Ilia, let's start
with you.

2321
02:04:56,000 --> 02:04:59,503
Ilia Baranov: I'll come back to
my experience at least with

2322
02:04:59,503 --> 02:05:02,828
Amazon, where, the Astro
program, I think, and for, your

2323
02:05:02,828 --> 02:05:06,332
listeners, if they haven't seen,
so Amazon has a home robot

2324
02:05:06,332 --> 02:05:09,835
called Astro. You can find it
online, theoretically you can

2325
02:05:09,835 --> 02:05:13,397
buy it. I don't know, I don't
make their business decisions,

2326
02:05:13,397 --> 02:05:17,079
but you can actually play with
this thing. It works reasonably

2327
02:05:17,079 --> 02:05:20,523
well. We have one at the office
just for fun. but them and

2328
02:05:20,523 --> 02:05:24,263
Samsung and Sony and basically
everybody and their dog has made

2329
02:05:24,263 --> 02:05:27,885
these kind of human interactive
robots. But the, little piece

2330
02:05:27,885 --> 02:05:30,973
that they're missing is the
question, the very small

2331
02:05:30,973 --> 02:05:34,714
question of what the hell does
it do, which nobody has answered

2332
02:05:34,714 --> 02:05:38,276
yet. And the killer app for
robots in the home seems to have

2333
02:05:38,276 --> 02:05:41,780
been vacuuming from like the
1990s till today. And nobody's

2334
02:05:41,780 --> 02:05:45,402
come up with a better idea. And
I'm still waiting for what is

2335
02:05:45,402 --> 02:05:49,261
the next killer app for Robotics
in the home. putting away dishes

2336
02:05:49,261 --> 02:05:52,824
or laundry, like everybody,
that's the first thing everybody

2337
02:05:52,824 --> 02:05:55,971
jumps towards. But manipulation
is very expensive and

2338
02:05:55,971 --> 02:05:59,355
complicated. And even if ML
Magic fixes it, like just the

2339
02:05:59,355 --> 02:06:03,036
actuators are very expensive.
Nobody's gonna spend even $5,000

2340
02:06:03,036 --> 02:06:06,777
on this, right? And $5,000 for
manipulators ridiculously cheap,

2341
02:06:06,777 --> 02:06:10,518
right? For a good one that moves
around and stuff. anyway, so I

2342
02:06:10,518 --> 02:06:13,843
think we're still missing the
killer app in the home for

2343
02:06:13,843 --> 02:06:17,465
robots, and I'm excited to see
what that ends up being. Maybe

2344
02:06:17,465 --> 02:06:21,087
it's an embodiment of these.
Assistance machine learned based

2345
02:06:21,087 --> 02:06:24,709
approaches of just shout in your
room. Hey, find me a path to

2346
02:06:24,709 --> 02:06:28,272
this theater, and this screen
magically comes over and talks

2347
02:06:28,272 --> 02:06:31,894
to you and tells you what to do.
I don't know if I knew maybe

2348
02:06:31,894 --> 02:06:35,397
we'd do that company, but for
now we're in industrial space

2349
02:06:35,397 --> 02:06:37,000
and, a different, business.

2350
02:06:37,000 --> 02:06:40,926
Stefan Seltz-Axmacher: This
whole thing sounded like what we

2351
02:06:40,926 --> 02:06:44,721
did at the beginning of our
podcast where we'll start like

2352
02:06:44,721 --> 02:06:48,517
coming up with these things and
be yeah, but wait, I could

2353
02:06:48,517 --> 02:06:52,050
already ask that question
anywhere in my house. And it

2354
02:06:52,050 --> 02:06:55,976
gets answered for me, and the
screen that pops up is my cell

2355
02:06:55,976 --> 02:06:59,510
phone. And it just happens
already. So why do I need a

2356
02:06:59,510 --> 02:07:03,698
screen on an arm in each room of
my house? I'd say what I'm most

2357
02:07:03,698 --> 02:07:07,558
excited about, so for better or
worse, the self-driving car

2358
02:07:07,558 --> 02:07:11,419
industry threw a lot of money
and interest into robotics. I

2359
02:07:11,419 --> 02:07:15,280
think like more than other
spaces, I, know there's a lot of

2360
02:07:15,280 --> 02:07:19,402
quad copter people, but from my
own perspective, it seemed like

2361
02:07:19,402 --> 02:07:23,328
way more people got way more
excited about self-driving cars

2362
02:07:23,328 --> 02:07:27,123
than quad copers. And now
there's a lot of people who have

2363
02:07:27,123 --> 02:07:30,984
worked in and around Robotics
and like some of those people

2364
02:07:30,984 --> 02:07:35,041
have been siphoned off to LLM
rappers for XYZ and like some of

2365
02:07:35,041 --> 02:07:38,509
those will become real
businesses or not. And there's

2366
02:07:38,509 --> 02:07:42,697
like a lot of, there's a similar
flavor of self-driving car hype

2367
02:07:42,697 --> 02:07:46,689
that's happening in humanoid
Robotics right now. And my hunch

2368
02:07:46,689 --> 02:07:50,811
is it will go roughly the same
as it went with Robotaxis. which

2369
02:07:50,811 --> 02:07:54,606
you could decide how much
sarcasm you wanna read into. but

2370
02:07:54,606 --> 02:07:58,532
I think we now have a critical
mass of people who care about

2371
02:07:58,532 --> 02:08:02,655
robots who, are hooked on these
sorts of problems in such a way

2372
02:08:02,655 --> 02:08:06,712
where Dropbox 5.0 is no longer
an appealing job. And that now,

2373
02:08:06,712 --> 02:08:10,834
like they need to build robots
for something. And many of those

2374
02:08:10,834 --> 02:08:14,564
people will work on nonsense,
like the robots for robot's

2375
02:08:14,564 --> 02:08:18,032
sake. But I think we're gonna
have more real Robotics

2376
02:08:18,032 --> 02:08:21,893
companies doing valuable and
interesting things in the next

2377
02:08:21,893 --> 02:08:26,015
founded, in the next five years
than we had the last 100 years.

2378
02:08:26,015 --> 02:08:29,680
and I think like right now.
Probably the most successful

2379
02:08:29,680 --> 02:08:33,475
Robotics company of all time
outside of defense and mining

2380
02:08:33,475 --> 02:08:37,270
is, is probably iRobot. And I
think you'll have five to 10

2381
02:08:37,270 --> 02:08:41,000
companies as successful as
iRobot in the next five years.

2382
02:08:41,000 --> 02:08:42,000
Audrow Nash: Cool.

2383
02:08:42,000 --> 02:08:47,531
Stefan Seltz-Axmacher: even 50.
And that's gonna be fucking

2384
02:08:47,531 --> 02:08:48,000
cool.

2385
02:08:48,000 --> 02:08:52,354
Audrow Nash: Hell yeah. That's
awesome. I think you're probably

2386
02:08:52,354 --> 02:08:56,500
right with that, and I think
that is such an exciting thing,

2387
02:08:56,500 --> 02:09:00,646
and that matches my observation
too, that there's a bunch of

2388
02:09:00,646 --> 02:09:05,000
people that are hooked on this
and they want to be using it for

2389
02:09:05,000 --> 02:09:08,261
Stefan Seltz-Axmacher: man,
boring SaaS just is not

2390
02:09:08,261 --> 02:09:12,417
something like I desperately did
not wanna start another Robotics

2391
02:09:12,417 --> 02:09:16,445
company. Like I was so wanting
to do something simple that like

2392
02:09:16,445 --> 02:09:20,601
I was even working on like a, an
e-commerce for planning funerals

2393
02:09:20,601 --> 02:09:24,182
idea for six months. That seems
more cheerful than go be

2394
02:09:24,182 --> 02:09:27,890
miserable in the mud with
robots. And the thing is, I just

2395
02:09:27,890 --> 02:09:32,046
can't work on anything else now.
And I think like a lot of people

2396
02:09:32,046 --> 02:09:35,627
feel the same as me, so maybe
right now some of them are

2397
02:09:35,627 --> 02:09:39,463
doing, open AI for asking your
plumber questions and there's

2398
02:09:39,463 --> 02:09:43,363
money in that right now. But
there's a lot of people at those

2399
02:09:43,363 --> 02:09:47,263
companies thinking, how do I get
back to Robotics? Where do I

2400
02:09:47,263 --> 02:09:51,228
find a Robotics application that
is fundable and buildable and

2401
02:09:51,228 --> 02:09:55,000
valuable? And I can't wait for
that to start bearing fruit.

2402
02:09:55,000 --> 02:09:58,442
Audrow Nash: too. Hell yeah. I
love that answer. Okay. Hell

2403
02:09:58,442 --> 02:10:02,000
yeah. thank you both. It's been
so much fun having you on the

2404
02:10:02,000 --> 02:10:05,000
Stefan Seltz-Axmacher: Thanks so
much, Audrow.

2405
02:10:05,000 --> 02:10:07,000
Audrow Nash: Hell yeah. All
right. Bye everyone.

2406
02:10:07,000 --> 02:10:09,224
Stefan Seltz-Axmacher: It was a
pleasure hanging out. Thanks so

2407
02:10:09,224 --> 02:10:10,000
much for joining us. ​

2408
02:10:10,000 --> 02:10:14,299
Audrow Nash: You made it! I
should be quiet just in case

2409
02:10:14,299 --> 02:10:19,135
you're sleeping peacefully. In
any case, I hope you enjoyed the

2410
02:10:19,135 --> 02:10:23,434
interview. What did you think?
Is Polymath Robotic on to

2411
02:10:23,434 --> 02:10:28,193
something? Would you use their
service? Did you like the demo?

2412
02:10:28,193 --> 02:10:33,106
I'd love to hear what you think
in the comments or on X. Also, I

2413
02:10:33,106 --> 02:10:37,328
host a weekly space on X on
Thursday evenings, US time,

2414
02:10:37,328 --> 02:10:41,550
where we talk about all things
robotics, including this

2415
02:10:41,550 --> 02:10:46,309
interview. Feel free to come and
chat. Okay, that's all. Happy

2416
02:10:46,309 --> 02:10:47,000
building.
