1
00:00:00,000 --> 00:00:02,630
Audrow Nash: There's a humanoid
kicking in a door with a holding

2
00:00:02,630 --> 00:00:03,000
a pistol.

3
00:00:03,000 --> 00:00:08,000
Amarjot Singh: Rather than you
going inside first, the humanoid

4
00:00:08,000 --> 00:00:13,000
can go in. It can go from riot
gear to defusing bombs to firing

5
00:00:13,000 --> 00:00:18,079
weapons to everything people do.
then you replace the human with

6
00:00:18,079 --> 00:00:23,000
the humanoid. Can you tell it's
an actual human or a humanoid?

7
00:00:23,000 --> 00:00:28,065
Audrow Nash: What a world we
live in. It's so freaking cool.

8
00:00:28,065 --> 00:00:33,552
is kind of the goal you think to
make like a super soldier or Hi!

9
00:00:33,552 --> 00:00:36,000
Would you introduce yourself?

10
00:00:36,000 --> 00:00:40,375
Amarjot Singh: Hi, Audrow. My
name is Amarjot Singh, and I'm

11
00:00:40,375 --> 00:00:43,000
the founder and CEO of Skylark
Labs.

12
00:00:43,000 --> 00:00:46,000
Audrow Nash: Hell yeah. Tell me
about Skylark Labs.

13
00:00:46,000 --> 00:00:49,679
Amarjot Singh: Yeah, so Skylark
is primarily a physical AI

14
00:00:49,679 --> 00:00:53,548
company, which is focused on
building the brain for different

15
00:00:53,548 --> 00:00:57,354
kind of machines going all the
way from drones to humanoids.

16
00:00:57,354 --> 00:01:00,462
And we build it for
mission-critical applications

17
00:01:00,462 --> 00:01:04,140
all the way from mines to
defense. And the key thing which

18
00:01:04,140 --> 00:01:08,136
we are focused on is building a
brain which has the capacity to

19
00:01:08,136 --> 00:01:12,069
learn from limited data. Data is
a big bottleneck in all these

20
00:01:12,069 --> 00:01:15,938
applications. And then once it
is deployed, it's always going

21
00:01:15,938 --> 00:01:19,617
to see new variations of the
data which you learn from, or

22
00:01:19,617 --> 00:01:23,486
totally novel things, and how
does the brain identify, label,

23
00:01:23,486 --> 00:01:27,418
and learn these new things on
the fly. So that's what we focus

24
00:01:27,418 --> 00:01:31,097
on, building this self-learning
AI brain for machines, for

25
00:01:31,097 --> 00:01:33,000
mission-critical applications.

26
00:01:33,000 --> 00:01:38,738
Audrow Nash: Shaw my background
mm-hmm awesome. hell yeah Before

27
00:01:38,738 --> 00:01:43,848
we dive into how you guys
actually do that would you tell

28
00:01:43,848 --> 00:01:49,317
me some of the details about
Skylark, where you're based, how

29
00:01:49,317 --> 00:01:54,159
many people you are, your
general round of funding and

30
00:01:54,159 --> 00:01:59,000
valuation things like this? Just
like general details.

31
00:01:59,000 --> 00:02:04,031
Amarjot Singh: Yeah, so Skylark
is headquartered in Menlo Park

32
00:02:04,031 --> 00:02:08,332
in California. We have roughly
70-plus people. So our

33
00:02:08,332 --> 00:02:13,282
headquarters is Menlo Park. Then
we have an office in Indiana

34
00:02:13,282 --> 00:02:17,907
where we do a lot of our
real-world testing. Then we have

35
00:02:17,907 --> 00:02:22,613
an office in Del Rio in Texas
where we deploy a lot of our

36
00:02:22,613 --> 00:02:27,401
solutions with DHS. And that
will be 23 people total in the

37
00:02:27,401 --> 00:02:32,675
U.S. And then we have another 50
people in Pune in India where we

38
00:02:32,675 --> 00:02:37,706
do all sorts of real-world data
gathering from mines to oceans

39
00:02:37,706 --> 00:02:42,656
to everywhere. And then we also
have a lot of our engineering

40
00:02:42,656 --> 00:02:47,606
and tech stack over there. And
we are in our pre-series A. We

41
00:02:47,606 --> 00:02:52,800
just finished our pre-series A.
We raised a total of 15 million.

42
00:02:52,800 --> 00:02:57,993
And valuation is 250 million. We
are going for our series A now.

43
00:02:57,993 --> 00:03:02,780
We are in the process. We are
raising it in two phases. The

44
00:03:02,780 --> 00:03:07,000
first one is 60 million and then
another 90 million.

45
00:03:07,000 --> 00:03:11,297
Audrow Nash: Hell yeah. And can
you speak on the valuation for

46
00:03:11,297 --> 00:03:14,000
your Series A that you're
expecting or?

47
00:03:14,000 --> 00:03:16,000
Amarjot Singh: It's going to be
close to a billion.

48
00:03:16,000 --> 00:03:21,631
Audrow Nash: That's so cool,
what a trip. Okay, so with that,

49
00:03:21,631 --> 00:03:26,615
mining, oceans, security, and
borders and things, some

50
00:03:26,615 --> 00:03:31,877
military applications. The
common part of all this is the

51
00:03:31,877 --> 00:03:37,785
learning on device and correct.
Can you tell me about that? Like

52
00:03:37,785 --> 00:03:42,954
what is kind of the core
approach that you all are using

53
00:03:42,954 --> 00:03:46,000
on all these different
platforms?

54
00:03:46,000 --> 00:03:52,917
Amarjot Singh: Yes. So I think
the core idea is that for a lot

55
00:03:52,917 --> 00:03:59,835
of these applications you want
to do certain mission sets. And

56
00:03:59,835 --> 00:04:06,864
those mission sets in humanoid
can be all the way from scanning

57
00:04:06,864 --> 00:04:13,781
for bombs. And I can also show
you a video of that. So you can

58
00:04:13,781 --> 00:04:20,364
see right up top you can see
this humanoid scanning for the

59
00:04:20,364 --> 00:04:26,946
bombs. It would pick up the
metal detector. It would go and

60
00:04:26,946 --> 00:04:33,083
scan. Or we can have share
another application. Make it

61
00:04:33,083 --> 00:04:39,219
easier. So you can also see
our... A little bit jumping

62
00:04:39,219 --> 00:04:40,000
around.

63
00:04:40,000 --> 00:04:42,000
Audrow Nash: No worries.

64
00:04:42,000 --> 00:04:48,617
Amarjot Singh: So you can see
our humanoids learning how to

65
00:04:48,617 --> 00:04:54,000
engage weapons. They can also...
Yeah, you could

66
00:04:54,000 --> 00:04:55,743
Audrow Nash: Humanoid with a
paintball gun is awesome. That's

67
00:04:55,743 --> 00:04:56,000
so funny.

68
00:04:56,000 --> 00:05:00,897
Amarjot Singh: You can go all
the way from dealing with the

69
00:05:00,897 --> 00:05:05,877
shields, which is mainly riot
gear, to scanning for bombs to

70
00:05:05,877 --> 00:05:10,856
logistics. And then the real
world mission sets which we are

71
00:05:10,856 --> 00:05:15,504
looking at, this would be one of
them. Where you want to

72
00:05:15,504 --> 00:05:19,820
basically take human out of
danger. So if you had to

73
00:05:19,820 --> 00:05:24,717
actually get a person out of
this kind of application where

74
00:05:24,717 --> 00:05:29,862
you have to engage the weapon. A
humanoid can get that mission

75
00:05:29,862 --> 00:05:35,008
set. Or this one where rather
than you going inside first, the

76
00:05:35,008 --> 00:05:37,000
humanoid can go in. Sir.

77
00:05:37,000 --> 00:05:39,710
Audrow Nash: Just for people
that are listening, there's a

78
00:05:39,710 --> 00:05:42,421
humanoid kicking in a door with
a holding a pistol. It's a

79
00:05:42,421 --> 00:05:45,178
little humanoid. Like it looks
like it's like two thirds as

80
00:05:45,178 --> 00:05:47,000
tall as people. Which humanoid
is this?

81
00:05:47,000 --> 00:05:50,000
Amarjot Singh: He is four and a
half feet. Just a little guy.

82
00:05:50,000 --> 00:05:52,192
Audrow Nash: Yeah, yeah, but
it's just, I mean, it's like

83
00:05:52,192 --> 00:05:54,692
pistol-butts the door to get the
door open in the video, which is

84
00:05:54,692 --> 00:05:55,000
awesome.

85
00:05:55,000 --> 00:05:58,000
Amarjot Singh: Yeah. But the
longer version of this is

86
00:05:58,000 --> 00:06:02,508
Audrow Nash: What is the
humanoid you're using in these

87
00:06:02,508 --> 00:06:03,000
shots?

88
00:06:03,000 --> 00:06:05,814
Amarjot Singh: Quite interesting
because in this mission set or

89
00:06:05,814 --> 00:06:08,583
let's say Where you are holding
the shield and you are hitting

90
00:06:08,583 --> 00:06:11,263
or when you are grabbing the
metal detector or even the box.

91
00:06:11,263 --> 00:06:14,077
When you give this task, you are
going to be like, okay this is

92
00:06:14,077 --> 00:06:16,757
how I use my manipulation to
grab these things. But when you

93
00:06:16,757 --> 00:06:19,481
pick the pole, it is going to
actually apply force on to your

94
00:06:19,481 --> 00:06:22,340
elbow because the pole is heavy.
So it is going to drag down. Or

95
00:06:22,340 --> 00:06:25,199
when you have the shield and the
person hits on the shield, that

96
00:06:25,199 --> 00:06:27,744
is going to actually have
external force. And your shield

97
00:06:27,744 --> 00:06:30,246
is going to move and the
objective of the humanoid is to

98
00:06:30,246 --> 00:06:32,792
keep it in a certain way. So
that is where the real world

99
00:06:32,792 --> 00:06:35,561
adaption would be needed. Or in
this kind of setting where you

100
00:06:35,561 --> 00:06:38,464
are pushing the door which could
be heavier, lighter. So when the

101
00:06:38,464 --> 00:06:39,000
force comes.

102
00:06:39,000 --> 00:06:59,000
Audrow Nash: Mm-hm. There's a
lot of variation for sure.

103
00:06:59,000 --> 00:07:00,000
Amarjot Singh: Yeah. But also
when you move.

104
00:07:00,000 --> 00:07:02,000
Audrow Nash: Could have stuff
behind it. Could stick.

105
00:07:02,000 --> 00:07:06,142
Amarjot Singh: Yeah, exactly.
And when you move, you have to

106
00:07:06,142 --> 00:07:10,560
move a certain way. So you have
to go and you have to make space

107
00:07:10,560 --> 00:07:14,978
for the other person to come in.
So the way the government wants

108
00:07:14,978 --> 00:07:19,120
to deploy this is in a human
machine tagging fashion. So you

109
00:07:19,120 --> 00:07:23,262
are not working on your own and
learning. You are also doing

110
00:07:23,262 --> 00:07:27,403
group dynamics. So these are
some of the mission sets we are

111
00:07:27,403 --> 00:07:31,890
doing. So as I explained, in all
of these settings, there are new

112
00:07:31,890 --> 00:07:35,963
kind of variations which can
come up. And then your machine

113
00:07:35,963 --> 00:07:39,898
has to first of all estimate
error that, okay, there is a

114
00:07:39,898 --> 00:07:43,626
deviation from what I am
supposed to do. Which kind of

115
00:07:43,626 --> 00:07:47,837
ties to self-awareness because
none of the AI models actually

116
00:07:47,837 --> 00:07:52,117
have that. And once you are able
to figure out that, hey, I am

117
00:07:52,117 --> 00:07:56,327
doing something which I am not
supposed to, then it corrects.

118
00:07:56,327 --> 00:08:00,262
So that is the basic idea. And
the way we do that is this

119
00:08:00,262 --> 00:08:03,990
architecture which you see on
the screen. So the basic

120
00:08:03,990 --> 00:08:08,339
architecture which people use,
that is an end-to-end supervised

121
00:08:08,339 --> 00:08:12,550
architecture or some sort of
transformer. And you train it on

122
00:08:12,550 --> 00:08:16,830
a very large data set and then
you deploy it and then you hope

123
00:08:16,830 --> 00:08:20,212
to god it works. But in our
case, we have a basic

124
00:08:20,212 --> 00:08:24,423
architecture, a foundational
model which is inspired from the

125
00:08:24,423 --> 00:08:28,565
brain. So it is not end-to-end
supervised. It is handcrafted

126
00:08:28,565 --> 00:08:32,431
and semi-supervised and
supervised giving you efficiency

127
00:08:32,431 --> 00:08:36,573
on data and compute. But the
secret sauce beyond that is the

128
00:08:36,573 --> 00:08:40,714
memory. Whatever I am telling
you right now, you are able to

129
00:08:40,714 --> 00:08:44,787
get right away because of this
part of the brain called the

130
00:08:44,787 --> 00:08:48,791
hippocampus. And we have an
equivalent of that, the memory

131
00:08:48,791 --> 00:08:51,000
where all of this magic happens.

132
00:08:51,000 --> 00:08:55,701
Audrow Nash: Okay, tell me more.
So with the, I guess, I'd love

133
00:08:55,701 --> 00:09:00,478
to go through that a bit slower
to understand how that works. So

134
00:09:00,478 --> 00:09:01,000
I, like

135
00:09:01,000 --> 00:09:05,000
Amarjot Singh: Sure. Yes.

136
00:09:05,000 --> 00:09:09,186
Audrow Nash: Tell me about the
foundational models and how

137
00:09:09,186 --> 00:09:13,804
you're approaching this, and how
you're doing some supervised or

138
00:09:13,804 --> 00:09:18,206
semi-supervised. Like I want to
understand the full system in

139
00:09:18,206 --> 00:09:19,000
some sense.

140
00:09:19,000 --> 00:09:23,711
Amarjot Singh: So let's think of
a simple example. You want to

141
00:09:23,711 --> 00:09:28,043
build a system to recognize the
perception side. Everyone

142
00:09:28,043 --> 00:09:32,602
understands more or less these
image data sets where you can

143
00:09:32,602 --> 00:09:37,389
say a dog is a dog or a cat is a
cat. if And generally we would

144
00:09:37,389 --> 00:09:41,796
take that data set, push it
through this black box network

145
00:09:41,796 --> 00:09:44,000
and then it will learn these.

146
00:09:44,000 --> 00:09:45,650
Audrow Nash: And they're all
labeled, so therefore you make

147
00:09:45,650 --> 00:09:47,273
distinctions because it says
this is a cat, this is a dog,

148
00:09:47,273 --> 00:09:48,643
and you propagate it back
through the network and

149
00:09:48,643 --> 00:09:50,406
everything. Okay, so that's like
normal machine learning to me,

150
00:09:50,406 --> 00:09:51,860
like neural networks,
convolutional neural networks.

151
00:09:51,860 --> 00:09:52,000
Okay.

152
00:09:52,000 --> 00:09:56,035
Amarjot Singh: Exactly. The
standard. Yeah. Yes. So what we

153
00:09:56,035 --> 00:10:00,071
do is we would take that whole
data set and we will push it

154
00:10:00,071 --> 00:10:04,448
through this first layer, which
is the battery of Gabor filters,

155
00:10:04,448 --> 00:10:08,142
which is extracting edges at
different frequencies and

156
00:10:08,142 --> 00:10:12,519
scales. And we have transforms.
They're called dual-tree complex

157
00:10:12,519 --> 00:10:16,828
wavelet transform. So you can go
from your input pixel space to

158
00:10:16,828 --> 00:10:21,000
the edge space with no loss of
energy. So they are completely

159
00:10:21,000 --> 00:10:26,000
Audrow Nash: Oh, okay.

160
00:10:26,000 --> 00:10:30,255
Amarjot Singh: Transformable.
That's a lot of my work, which

161
00:10:30,255 --> 00:10:34,582
my professor and I did during my
PhD. Once you have that edge

162
00:10:34,582 --> 00:10:36,000
space that is nicer.

163
00:10:36,000 --> 00:10:38,608
Audrow Nash: And wavelets, I
haven't worked with wavelets

164
00:10:38,608 --> 00:10:41,307
before, but they're just like a
basis set, they're a way of

165
00:10:41,307 --> 00:10:43,000
representing things, is that
correct?

166
00:10:43,000 --> 00:10:45,735
Amarjot Singh: That's pretty
much it. So in the image, when

167
00:10:45,735 --> 00:10:48,748
you have an edge, it's basically
a gradient going from right. And

168
00:10:48,748 --> 00:10:50,000
you just want to pick that.

169
00:10:50,000 --> 00:10:55,000
Audrow Nash: Yep. And is the key
thing of wavelets that they're

170
00:10:55,000 --> 00:11:00,000
smooth and thus gradient descent
works better for optimizing or

171
00:11:00,000 --> 00:11:01,000
Amarjot Singh: You can.

172
00:11:00,000 --> 00:11:03,000
Audrow Nash: Why use wavelets as
opposed to anything else?

173
00:11:03,000 --> 00:11:03,873
Amarjot Singh: So wavelets
allows you to capture these

174
00:11:03,873 --> 00:11:04,875
edges at different orientations,
different scales. But the way

175
00:11:04,875 --> 00:11:05,878
you build the wavelet, you can
try to see when you want to cut

176
00:11:05,878 --> 00:11:06,912
the lobe of your let's say think
of it as a Gaussian and you can

177
00:11:06,912 --> 00:11:07,931
cut it beyond sigma, two sigma.
You can make it fatter. You can

178
00:11:07,931 --> 00:11:08,917
make it smaller. Allows you to
Yeah. So allows you to capture

179
00:11:08,917 --> 00:11:09,935
different kind of edges rather
than just an edge. So it is just

180
00:11:09,935 --> 00:11:10,000
more

181
00:11:10,000 --> 00:11:32,000
Audrow Nash: Mm-hm. Uh-huh.
Change your variance, yeah.

182
00:11:32,000 --> 00:11:33,406
Amarjot Singh: Diverse if you
would. And people design it in

183
00:11:33,406 --> 00:11:34,695
different ways to capture
different kind of things. But

184
00:11:34,695 --> 00:11:36,148
the advantage of the ones we are
using is that you can go from

185
00:11:36,148 --> 00:11:37,625
input to the edge space with no
loss of energy. That's the most

186
00:11:37,625 --> 00:11:38,000
important thing.

187
00:11:38,000 --> 00:11:45,457
Audrow Nash: Okay. What do you
mean by that? I don't fully

188
00:11:45,457 --> 00:11:53,171
understand the loss of energy
and to edge space. Edge space,

189
00:11:53,171 --> 00:11:56,000
do you mean on device?

190
00:11:56,000 --> 00:11:57,525
Amarjot Singh: So you take an
image. You can push it through a

191
00:11:57,525 --> 00:11:58,951
standard edge detector. Let's
say Harris or something like

192
00:11:58,951 --> 00:12:00,426
that. It's going to it's going
to capture some of the edges.

193
00:12:00,426 --> 00:12:02,000
It's going to get rid of all the
smooth areas, things like that.

194
00:12:02,000 --> 00:12:13,000
Audrow Nash: Yep. Yeah.

195
00:12:13,000 --> 00:12:17,417
Amarjot Singh: But when you're
trying to do the Gabor filter or

196
00:12:17,417 --> 00:12:21,694
wavelet filter, whatever that
is. You have your initial input

197
00:12:21,694 --> 00:12:25,900
signal. You can compute the
methods to compute the energy or

198
00:12:25,900 --> 00:12:29,967
how much signal is there. And
then you transform them onto

199
00:12:29,967 --> 00:12:34,454
your edge space or signal space.
The strength of the edges which

200
00:12:34,454 --> 00:12:38,871
are there, you can infuse that
same amount of energy across the

201
00:12:38,871 --> 00:12:43,288
whole set. So when you sum that
signal strength, the signal sum

202
00:12:43,288 --> 00:12:47,775
is same in the input as well as
output. It just gets distributed

203
00:12:47,775 --> 00:12:51,000
to where the edges are rather
than everywhere.

204
00:12:51,000 --> 00:12:54,678
Audrow Nash: Edges are. Okay, so
what I'm hearing with this is

205
00:12:54,678 --> 00:12:58,000
it's an effective compression of
your input data effect.

206
00:12:58,000 --> 00:12:58,979
Amarjot Singh: Yes. Perfect.
Yeah. Yeah. You're

207
00:12:58,979 --> 00:13:00,000
redistributing the energy to
where it's relevant.

208
00:13:00,000 --> 00:13:04,706
Audrow Nash: Okay, so then, yep,
so you're keeping it so you

209
00:13:04,706 --> 00:13:09,569
could, maybe you can't back out
it, but all of the richness of

210
00:13:09,569 --> 00:13:12,000
the information is still there.

211
00:13:12,000 --> 00:13:13,000
Amarjot Singh: Exactly.

212
00:13:13,000 --> 00:13:16,615
Audrow Nash: Okay, and then so
from that, it's effective,

213
00:13:16,615 --> 00:13:19,976
basically very effective
compression of the data that

214
00:13:19,976 --> 00:13:23,717
you're getting. How does that
relate to the whole on-device

215
00:13:23,717 --> 00:13:26,000
learning approach that you're
doing?

216
00:13:26,000 --> 00:13:26,931
Amarjot Singh: So right now we
are building the base model or

217
00:13:26,931 --> 00:13:27,893
the foundational model. So the
first there is how do I get from

218
00:13:27,893 --> 00:13:28,733
my input signal to this edge
space where I am basically

219
00:13:28,733 --> 00:13:29,664
cleaning the input signal. Once
that gets done on top of this

220
00:13:29,664 --> 00:13:30,565
very rich edge space, everything
in the world is made up of

221
00:13:30,565 --> 00:13:31,389
lines. You would put semi
supervised learning. You can

222
00:13:31,389 --> 00:13:32,000
think of as a convolutional
autoencoder.

223
00:13:32,000 --> 00:13:54,000
Audrow Nash: Mm-hm. Yep.

224
00:13:54,000 --> 00:13:56,595
Amarjot Singh: There are many
ways of doing that. But you're

225
00:13:56,595 --> 00:13:58,886
taking these edges and
multilayer without any labels.

226
00:13:58,886 --> 00:14:01,481
You're trying to find common
patterns across all of them. So

227
00:14:01,481 --> 00:14:02,000
a car. Yeah.

228
00:14:02,000 --> 00:14:05,438
Audrow Nash: Patterns and
everything for efficient data

229
00:14:05,438 --> 00:14:07,000
compression again. Mm-hm.

230
00:14:07,000 --> 00:14:08,684
Amarjot Singh: Yeah. So your car
would have bonnet, windscreens,

231
00:14:08,684 --> 00:14:10,316
tires, things like that. And you
find commonalities across the

232
00:14:10,316 --> 00:14:12,000
whole data set. And then you do
a truncated supervised learning.

233
00:14:12,000 --> 00:14:14,967
Audrow Nash: Yep. So, okay, just
to me, I'm thinking of when

234
00:14:14,967 --> 00:14:17,043
we've done stuff with
convolution, or some

235
00:14:17,043 --> 00:14:19,911
convolutional neural network
stuff that I've seen is quite

236
00:14:19,911 --> 00:14:23,075
cool because by hand, initially
doing computer vision, you would

237
00:14:23,075 --> 00:14:25,943
have your edge detectors and
then your feature recognizers

238
00:14:25,943 --> 00:14:28,662
and you'd go and look at
relationships between features

239
00:14:28,662 --> 00:14:31,579
to figure out things. But when
you do it with convolutional

240
00:14:31,579 --> 00:14:34,595
neural networks, the thing that
was cool is you start and you

241
00:14:34,595 --> 00:14:37,660
identify like edges. Like if you
look at the first images that

242
00:14:37,660 --> 00:14:40,627
are made and then eventually
they get less and less abstract

243
00:14:40,627 --> 00:14:43,000
to eventually be like tires and
like full things

244
00:14:43,000 --> 00:15:00,000
Amarjot Singh: Yeah. Yeah. Yeah.
Yes. Yeah.

245
00:15:00,000 --> 00:15:01,082
Audrow Nash: That kind of match
in the image. And then when you

246
00:15:01,082 --> 00:15:02,060
slide these across the image,
you get to see, oh, there's

247
00:15:02,060 --> 00:15:03,124
probably a tire here and then
you look at the relationships of

248
00:15:03,124 --> 00:15:04,000
all the features and that's how
they detect things.

249
00:15:04,000 --> 00:15:11,000
Amarjot Singh: Yeah. Yeah. Yes.

250
00:15:11,000 --> 00:15:14,464
Audrow Nash: What I hear you
saying with this, or what I

251
00:15:14,464 --> 00:15:18,113
believe you're saying, is you
are doing a data transform on

252
00:15:18,113 --> 00:15:21,206
the data to make it more
efficient and then you're

253
00:15:21,206 --> 00:15:24,237
putting it in like a
convolutional neural network

254
00:15:24,237 --> 00:15:27,330
setup to generate features on
your compressed data

255
00:15:27,330 --> 00:15:29,000
effectively. Is that right?

256
00:15:29,000 --> 00:15:32,478
Amarjot Singh: Yes. Yes. So
yeah. So the beauty of

257
00:15:32,478 --> 00:15:36,513
convolutional neural networks is
we don't have to tell the

258
00:15:36,513 --> 00:15:37,000
machine

259
00:15:37,000 --> 00:15:37,678
Audrow Nash: You don't have to
do your old feature detection.

260
00:15:37,678 --> 00:15:38,000
Yeah, yeah. It's much better.

261
00:15:37,000 --> 00:15:40,500
Amarjot Singh: What features.
What. Yeah. Exactly. So yeah. So

262
00:15:40,500 --> 00:15:44,000
the machine can just learn it.
But then you also have to like.

263
00:15:44,000 --> 00:15:46,000
Audrow Nash: Don't need SIFT or
SURF or any of these.

264
00:15:46,000 --> 00:15:49,229
Amarjot Singh: Yeah exactly. But
then you also have to learn a

265
00:15:49,229 --> 00:15:52,042
lot of those additional
features. But I think at least

266
00:15:52,042 --> 00:15:55,323
on the initial piece where you
have low level features. We know

267
00:15:55,323 --> 00:15:58,708
if you are able to get to a good
enough edge space you don't need

268
00:15:58,708 --> 00:16:01,000
to learn them. And then if let's
say you're.

269
00:16:01,000 --> 00:16:02,494
Audrow Nash: That's what you do
with your wavelets and things.

270
00:16:02,494 --> 00:16:03,000
Is that right? Mm-hm.

271
00:16:03,000 --> 00:16:06,278
Amarjot Singh: That's what we
do. Exactly. And then you can

272
00:16:06,278 --> 00:16:09,500
combine those edges together to
form corners and mid-level

273
00:16:09,500 --> 00:16:10,000
features.

274
00:16:10,000 --> 00:16:11,000
Audrow Nash: Everything else.
Yeah, yeah.

275
00:16:11,000 --> 00:16:14,330
Amarjot Singh: Yeah. And then
also this is common across all

276
00:16:14,330 --> 00:16:17,770
of them because if I tie a label
to let's say a car or a truck

277
00:16:17,770 --> 00:16:20,933
you're learning a lot of
redundant features across all of

278
00:16:20,933 --> 00:16:24,152
them which you can later on get
by compressed by doing the

279
00:16:24,152 --> 00:16:27,204
optimization and all. But you
don't need to learn that.

280
00:16:27,204 --> 00:16:30,755
There's a lot of commonality. So
we do unsupervised stuff there.

281
00:16:30,755 --> 00:16:34,307
But then later on where you have
to associate that the bonnet is

282
00:16:34,307 --> 00:16:37,581
next to the tire it's next to
the windscreen and now it's a

283
00:16:37,581 --> 00:16:40,911
car there you need supervised
learning. So we would bring it

284
00:16:40,911 --> 00:16:43,796
in the later layers. But then
that gives you massive

285
00:16:43,796 --> 00:16:46,682
efficiency on the data and
compute. But since we are

286
00:16:46,682 --> 00:16:50,067
building them in parallel I will
build the wavelet thing then

287
00:16:50,067 --> 00:16:53,396
unsupervised learning. I would
also compress it to make sure

288
00:16:53,396 --> 00:16:56,393
that I don't have any
redundancies in my feature maps.

289
00:16:56,393 --> 00:16:59,612
Then I would supervised learning
and that leads to smaller

290
00:16:59,612 --> 00:17:00,000
models.

291
00:17:00,000 --> 00:17:05,738
Audrow Nash: Cool. Okay, so what
I'm hearing with this is it's a

292
00:17:05,738 --> 00:17:10,848
lot of data transforms and
compression for very efficient

293
00:17:10,848 --> 00:17:15,779
data processing effectively.
Okay, so the separation of

294
00:17:15,779 --> 00:17:21,248
things you're doing leads to the
ability to be very efficient

295
00:17:21,248 --> 00:17:26,000
with data and this is kind of,
so that's really cool.

296
00:17:26,000 --> 00:17:28,618
Amarjot Singh: But we require
roughly 1,000x less data and we

297
00:17:28,618 --> 00:17:31,193
have multiple papers on this
when you're building this thing

298
00:17:31,193 --> 00:17:33,811
and this starts to become really
really effective when you're

299
00:17:33,811 --> 00:17:36,000
trying to do temporal stuff like
video or robotics.

300
00:17:36,000 --> 00:17:36,961
Audrow Nash: Mm-hm. Oh, very
cool. Yeah, yeah. Because then

301
00:17:36,961 --> 00:17:37,954
you can, you put your powerful
features into something like a

302
00:17:37,954 --> 00:17:38,932
transformer architecture or
something where it feeds it back

303
00:17:38,932 --> 00:17:39,811
and because you have small
representations for things,

304
00:17:39,811 --> 00:17:40,772
you're able to do a lot and
maybe add a temporal dimension.

305
00:17:40,772 --> 00:17:41,000
Is that right?

306
00:17:41,000 --> 00:17:48,742
Amarjot Singh: Yeah. Exactly.
Yeah. Exactly. But here is the

307
00:17:48,742 --> 00:17:57,000
thing and here's the thing which
is the most important thing. So

308
00:17:57,000 --> 00:18:02,000
Audrow Nash: That's really cool.
Yeah.

309
00:18:02,000 --> 00:18:05,798
Amarjot Singh: When you build
your foundation model it's built

310
00:18:05,798 --> 00:18:09,719
so someone built it using end to
end transformers. I am saying I

311
00:18:09,719 --> 00:18:13,333
don't need to learn the low
level middle level. I only need

312
00:18:13,333 --> 00:18:17,009
to learn associations at the
high level. So whether you like

313
00:18:17,009 --> 00:18:20,746
this process or not we build
this thing. But now when you see

314
00:18:20,746 --> 00:18:24,544
something new in the real world
rather than let's say having a

315
00:18:24,544 --> 00:18:28,342
face which has eyes next to the
nose and nose next to the lips

316
00:18:28,342 --> 00:18:32,018
now you have a weird kind of
face. Somehow someone is trying

317
00:18:32,018 --> 00:18:35,510
to spoof you where you have lips
replaced with one of the

318
00:18:35,510 --> 00:18:36,000
eyelids.

319
00:18:36,000 --> 00:18:38,000
Audrow Nash: On the forehead or
something. Mm-hm.

320
00:18:38,000 --> 00:18:41,000
Amarjot Singh: Yeah. Sorry.

321
00:18:41,000 --> 00:18:43,000
Audrow Nash: It's okay.

322
00:18:43,000 --> 00:18:46,153
Amarjot Singh: If you do that
now if you push that through the

323
00:18:46,153 --> 00:18:49,254
network the relationships are
not there right now. So you can

324
00:18:49,254 --> 00:18:52,407
go through your low level middle
level and when you get to the

325
00:18:52,407 --> 00:18:55,610
high level it's going to be like
well this is not really what I

326
00:18:55,610 --> 00:18:58,407
saw. So this is not really a
face. I am very low on the

327
00:18:58,407 --> 00:19:01,356
confidence score and it will
reject it. And a lot of these

328
00:19:01,356 --> 00:19:04,458
outliers where you have a tank
which is partially camouflaged

329
00:19:04,458 --> 00:19:07,610
or you have a new kind of cyber
signal where someone is trying

330
00:19:07,610 --> 00:19:10,458
to attack you are all of the
things which are beyond two

331
00:19:10,458 --> 00:19:13,610
sigma. So they still have a lot
of the middle level properties

332
00:19:13,610 --> 00:19:16,864
but different relationships. So
we use our middle level layer to

333
00:19:16,864 --> 00:19:19,864
automatically flag a lot of
these outliers which then serve

334
00:19:19,864 --> 00:19:22,000
as a funnel to push stuff into
our memory.

335
00:19:22,000 --> 00:19:28,995
Audrow Nash: Cool. Mm-hm. So if
I understand, so that when

336
00:19:28,995 --> 00:19:35,628
people make a foundational
model, they kind of, they go

337
00:19:35,628 --> 00:19:42,623
from inputs to outputs and you
have nothing, everything is

338
00:19:42,623 --> 00:19:46,000
learned in between. Is that,

339
00:19:46,000 --> 00:19:51,000
Amarjot Singh: Yeah. Exactly.

340
00:19:51,000 --> 00:19:54,527
Audrow Nash: Okay. And then so
what you're doing is you're

341
00:19:54,527 --> 00:19:58,054
having several divisions through
there where you have your

342
00:19:58,054 --> 00:20:01,824
wavelet transform and then you
have your autoencoders and then

343
00:20:01,824 --> 00:20:05,534
you have all these other things.
At each step, you're passing

344
00:20:05,534 --> 00:20:09,122
things to the next stage and
eventually you get to outputs,

345
00:20:09,122 --> 00:20:11,980
but you have all these
intermediate weights and

346
00:20:11,980 --> 00:20:15,568
statistical information and
everything about each layer. So

347
00:20:15,568 --> 00:20:19,338
then that is very good because
you can access it anywhere that

348
00:20:19,338 --> 00:20:22,926
you can put it forward into
another part of your network or

349
00:20:22,926 --> 00:20:26,392
something and then identify this
is a camouflaged tank or

350
00:20:26,392 --> 00:20:27,000
something.

351
00:20:27,000 --> 00:20:29,490
Amarjot Singh: Yeah, exactly.
And you're not going to be able

352
00:20:29,490 --> 00:20:31,939
to do that from the later layers
because in your traditional

353
00:20:31,939 --> 00:20:33,000
methods it's input output.

354
00:20:33,000 --> 00:20:35,000
Audrow Nash: They would already
have failed. Mm-hmm.

355
00:20:35,000 --> 00:20:37,000
Amarjot Singh: Exactly, exactly.
And that is where you...

356
00:20:37,000 --> 00:20:39,000
Audrow Nash: Because you don't
have the data

357
00:20:39,000 --> 00:20:40,000
Amarjot Singh: Exactly.

358
00:20:40,000 --> 00:20:42,000
Audrow Nash: For the new
approach. Okay.

359
00:20:42,000 --> 00:20:46,077
Amarjot Singh: Yeah, and a lot
of our proprietary stuff is how

360
00:20:46,077 --> 00:20:49,694
do you design this base
architecture which is patented?

361
00:20:49,694 --> 00:20:53,245
And then how do you use the
mid-level features to flag

362
00:20:53,245 --> 00:20:57,323
emerging signals? And we do this
all the way from a lot of our

363
00:20:57,323 --> 00:21:01,203
work with the intelligence
organizations where you can have

364
00:21:01,203 --> 00:21:05,082
a new kind of signal which is
online, all the way to drones

365
00:21:05,082 --> 00:21:09,094
picking up new kind of tanks to
a lot of the false positives.

366
00:21:09,094 --> 00:21:12,974
That's where a lot of our value
has come up because when we

367
00:21:12,974 --> 00:21:17,051
started deploying them initially
we were deploying them on the

368
00:21:17,051 --> 00:21:20,602
drones. And one thing is you
want to detect very small

369
00:21:20,602 --> 00:21:24,548
targets on a fast moving asset,
you will always get a lot of

370
00:21:24,548 --> 00:21:28,362
false positives. So what...
Sorry, I'm doing something. So

371
00:21:28,362 --> 00:21:32,637
what big advantage we got on the
memory side initially was how do

372
00:21:32,637 --> 00:21:36,188
we get a reliable AI system for
these mission critical

373
00:21:36,188 --> 00:21:40,397
applications which just doesn't
make mistakes. And this memory's

374
00:21:40,397 --> 00:21:44,277
first value prop was, we have
the lowest false positives in

375
00:21:44,277 --> 00:21:45,000
the market.

376
00:21:45,000 --> 00:21:49,067
Audrow Nash: Yeah. Tell me how
this leads to like, I see sort

377
00:21:49,067 --> 00:21:53,200
of how this could work, but I
don't fully understand it. Let's

378
00:21:53,200 --> 00:21:57,133
go with the tank example. I
think it's a cool one. You have

379
00:21:57,133 --> 00:22:01,000
a camouflage tank. It's not in
your training distribution.

380
00:22:01,000 --> 00:22:04,000
Amarjot Singh: Okay. Yeah. So

381
00:22:04,000 --> 00:22:07,745
Audrow Nash: How do you, how
does this work through the

382
00:22:07,745 --> 00:22:11,898
network? What are we actually
doing when we, oh, there's some

383
00:22:11,898 --> 00:22:16,119
tank like things. Like I see, I
see tracks, but the rest of it

384
00:22:16,119 --> 00:22:20,000
doesn't look right. Or I see
that one, the barrel of the,

385
00:22:20,000 --> 00:22:20,168
Amarjot Singh: Yeah. So yeah, so
let's say you come to me and you

386
00:22:20,168 --> 00:22:20,317
say, "Hey, I want to build a
tank detector. I'm building a

387
00:22:20,317 --> 00:22:20,477
kamikaze drone, real world use
case, and the drone is going to

388
00:22:20,477 --> 00:22:20,642
fly, it's going to find a tank,
it's going to go hit into it. So

389
00:22:20,642 --> 00:22:20,807
I'm like, "Wait, I need to first
see what a tank looks like." So

390
00:22:20,807 --> 00:22:20,956
we get a lot of images of these
tanks. We train our hybrid

391
00:22:20,956 --> 00:22:21,000
model. It's built

392
00:22:20,000 --> 00:22:39,718
Audrow Nash: I don't know what
any of the tank parts are

393
00:22:39,718 --> 00:22:45,000
called. Mm-hmm.

394
00:22:45,000 --> 00:22:49,188
Amarjot Singh: After that, and
there's a memory attached to it.

395
00:22:49,188 --> 00:22:52,644
Once the drone is flying, if I
didn't have my custom

396
00:22:52,644 --> 00:22:56,234
architecture, I would push,
let's say, a tank which is

397
00:22:56,234 --> 00:23:00,288
partially hidden, the AI will
miss it, the drone will not hit

398
00:23:00,288 --> 00:23:04,476
it, story ends. In our case, the
pipeline is two-fold. You take

399
00:23:04,476 --> 00:23:08,531
the image, you push it through
the whole network, and you say

400
00:23:08,531 --> 00:23:12,851
the end of the network, which is
supervised bit, does it say it's

401
00:23:12,851 --> 00:23:17,039
a tank or not? That's great. And
if it says tank, no problem at

402
00:23:17,039 --> 00:23:21,094
all. If it says no tank, we put
it aside. But there is also a

403
00:23:21,094 --> 00:23:24,683
pipeline which is learning from
the middle part of the

404
00:23:24,683 --> 00:23:28,605
architecture. And when you push
an image, you're also doing

405
00:23:28,605 --> 00:23:32,195
matching of the mid-level
features, which are Class A,

406
00:23:32,195 --> 00:23:36,183
Class B, Class C. And you say,
"Hey, it has some features of

407
00:23:36,183 --> 00:23:40,370
Class A, some features of Class
B, or 90% features of Class C."

408
00:23:40,370 --> 00:23:44,558
And based on how you decide, you
can say, "Hey, it looks like a

409
00:23:44,558 --> 00:23:48,679
tank. Let's go and hit it." So
there are validation methods on

410
00:23:48,679 --> 00:23:52,668
top of that. So you would run
these parallel pipelines to do

411
00:23:52,668 --> 00:23:53,000
this.

412
00:23:53,000 --> 00:23:55,133
Audrow Nash: Okay. And that's
where you're saying the two

413
00:23:55,133 --> 00:23:57,452
Sigma approach. Like basically
you are looking outside of your

414
00:23:57,452 --> 00:23:59,548
regular confidence interval. So
if it's, if it's under a

415
00:23:59,548 --> 00:24:01,830
Gaussian distribution, in the
middle, we're very certain it's

416
00:24:01,830 --> 00:24:04,000
a tank. But then when it's two
Sigma out, it's like, we're

417
00:24:04,000 --> 00:24:12,000
Amarjot Singh: Yeah. Yeah. If So

418
00:24:12,000 --> 00:24:14,222
Audrow Nash: 93% sure it's not a
tank or something like, or like,

419
00:24:14,222 --> 00:24:16,000
I don't know, it's over at the
93rd percentile area.

420
00:24:16,000 --> 00:24:16,641
Amarjot Singh: If you were 90%
or above, your supervised

421
00:24:16,641 --> 00:24:17,316
network would have picked it. If
it is partially covered or

422
00:24:17,316 --> 00:24:18,025
something, you don't have a lot
of the features. You have some

423
00:24:18,025 --> 00:24:18,700
of the features. So that would
push it beyond the mean. And

424
00:24:18,700 --> 00:24:19,364
then the only way you can use
that is to have some sort of

425
00:24:19,364 --> 00:24:20,073
network, which is specifically
tuned to pick partial features.

426
00:24:20,073 --> 00:24:20,748
And what are those partial
features, which are good enough?

427
00:24:20,748 --> 00:24:21,492
The mid-level features, which my
autoencoder has learned. So then

428
00:24:21,492 --> 00:24:22,190
I would do basically template
matching. And I would say, "Ah,

429
00:24:22,190 --> 00:24:22,922
here's a tire which is glowing.
Here's a track which is glowing.

430
00:24:22,922 --> 00:24:23,563
Here's a barrel piece." But
nothing else is there. So it

431
00:24:23,563 --> 00:24:24,272
must be useful. So we flag that.
And then you can say, "Out of

432
00:24:24,272 --> 00:24:24,936
all the parts you have learned
for the tank, if there is a

433
00:24:24,936 --> 00:24:25,657
barrel, then you must flag. If
it's a tread, you must flag." So

434
00:24:25,657 --> 00:24:26,000
you can also put more weights.

435
00:24:26,000 --> 00:24:37,355
Audrow Nash: Mm-hmm. Yeah. Cool.
Mm-hmm. Yeah, yeah. So Oops. so

436
00:24:37,355 --> 00:24:47,645
you can, this is where it's
somewhat supervised or you can

437
00:24:47,645 --> 00:24:58,290
tune your own weights in some
sense when you say, for sure a

438
00:24:58,290 --> 00:25:09,468
barrel means tank for this. Like
if it's, if you see wheels and

439
00:25:09,468 --> 00:25:18,871
you see a barrel, it's like,
that's enough. It can be

440
00:25:18,871 --> 00:25:21,000
camouflaged,

441
00:25:21,000 --> 00:25:26,000
Amarjot Singh: Yeah. Yeah. Yeah.

442
00:25:26,000 --> 00:25:28,000
Audrow Nash: But.

443
00:25:28,000 --> 00:25:30,671
Amarjot Singh: And we give this
power to the user through our

444
00:25:30,671 --> 00:25:33,472
platform. So we would say that,
"Hey, this is the category." And

445
00:25:33,472 --> 00:25:36,099
if you want to put more weight
on ABCD, you can do that. You

446
00:25:36,099 --> 00:25:38,638
can also do combinations. You
can say, "Hey, there's a new

447
00:25:38,638 --> 00:25:41,134
kind of thing which has come up,
which is partially tank,

448
00:25:41,134 --> 00:25:43,848
partially whatever. And if you
see two parts of the tank, five

449
00:25:43,848 --> 00:25:46,431
parts of this, definitely flag
that." So you can also build

450
00:25:46,431 --> 00:25:47,000
combinations.

451
00:25:47,000 --> 00:25:51,875
Audrow Nash: Yeah. Okay. What's
the, was this a challenge from a

452
00:25:51,875 --> 00:25:56,598
user interface point of view to
make it so you can expose this

453
00:25:56,598 --> 00:26:00,711
kind of like mid-level
information to users? Or do you

454
00:26:00,711 --> 00:26:05,053
have like, it'll be like a
picture, like a kind of blurry

455
00:26:05,053 --> 00:26:09,928
abstract picture of like a tank
tread and a blurry one of like a

456
00:26:09,928 --> 00:26:14,650
tank barrel. And they say, oh
yeah, this one looks like it and

457
00:26:14,650 --> 00:26:19,449
this one looks like it. But I'm
imagining that is a challenging

458
00:26:19,449 --> 00:26:24,324
thing to get right for users in
the field for this kind of thing

459
00:26:24,324 --> 00:26:26,000
or how does that work?

460
00:26:26,000 --> 00:26:27,266
Amarjot Singh: So the reason why
you want to... So two ways. This

461
00:26:27,266 --> 00:26:28,357
can happen automatically as
well. The reason why we show

462
00:26:28,357 --> 00:26:29,468
this thing to the user is
because explainability is a big

463
00:26:29,468 --> 00:26:30,695
part of our systems. Because we
are protecting kinetic systems,

464
00:26:30,695 --> 00:26:31,844
so we don't want to go and
eliminate something if we cannot

465
00:26:31,844 --> 00:26:32,000
have it.

466
00:26:32,000 --> 00:26:39,063
Audrow Nash: Mm-hmm. What do you
mean kinetic systems? Just

467
00:26:39,063 --> 00:26:46,246
they're moving or do you mean,
is that like a battle term or

468
00:26:46,246 --> 00:26:49,000
anything? I don't know.

469
00:26:49,000 --> 00:26:49,283
Amarjot Singh: So kinetic
systems in defense world would

470
00:26:49,283 --> 00:26:49,556
be you want to go and hit
something. In this case, the

471
00:26:49,556 --> 00:26:49,859
drone going and hitting the
tank. So if you want to actually

472
00:26:49,859 --> 00:26:50,153
want to do kinetic, or if you
want to go and actually blow

473
00:26:50,153 --> 00:26:50,451
something in this simple term,
you want to know what you're

474
00:26:50,451 --> 00:26:50,759
blowing and why. So a lot of our
models later on also go into

475
00:26:50,759 --> 00:26:51,068
decision making, that we are
building systems which are right

476
00:26:51,068 --> 00:26:51,331
now perception only to action
only. Action should be

477
00:26:51,331 --> 00:26:51,649
auditable. How do you actually
then go and take action? So this

478
00:26:51,649 --> 00:26:51,958
decision making chain for us is
very, very important. So flag

479
00:26:51,958 --> 00:26:52,276
something just because you say a
tank or not tank is not useful

480
00:26:52,276 --> 00:26:52,564
enough. So this two layer
pipeline would also get to say,

481
00:26:52,564 --> 00:26:52,842
"Well, I think it's a tank
because it has these kind of

482
00:26:52,842 --> 00:26:53,135
properties and that's why I'm
flagging it." We do a lot of

483
00:26:53,135 --> 00:26:53,464
that for viruses on the internet
as well. This is the ABCD virus.

484
00:26:53,464 --> 00:26:53,783
It has these kind of properties
which are underlying. So that's

485
00:26:53,783 --> 00:26:54,000
why I flagged it. So that's
another reason.

486
00:26:54,000 --> 00:27:09,786
Audrow Nash: Okay. Something up.
Yeah. Mm. Okay, very cool. So

487
00:27:09,786 --> 00:27:25,317
that kind of approach is like
the secret sauce of Skylark. Is

488
00:27:25,317 --> 00:27:41,103
it correct? Or is there like, or
is it one of many approaches?

489
00:27:41,103 --> 00:27:56,125
Because there's the mobility,
it's the first part. What are

490
00:27:56,125 --> 00:28:03,000
the other parts? Where you,

491
00:28:03,000 --> 00:28:08,250
Amarjot Singh: The first part.
It's the first part. Yeah, be

492
00:28:08,250 --> 00:28:10,000
good. Go on, please.

493
00:28:10,000 --> 00:28:11,078
Audrow Nash: Yeah, yeah, no, I'm
just trying to see how it all

494
00:28:11,078 --> 00:28:12,000
fits for all of this. Okay, so
that's the first part.

495
00:28:12,000 --> 00:28:15,006
Amarjot Singh: Yeah, so the
final objective is, okay, if I

496
00:28:15,006 --> 00:28:17,804
know something, I want to flag
it. But if I don't know

497
00:28:17,804 --> 00:28:20,965
something, it is something new.
I want to flag it, but I also

498
00:28:20,965 --> 00:28:23,867
want to label it. I want to
learn it. And I want to make

499
00:28:23,867 --> 00:28:26,873
sure I can do learning on the
device. So most of the other

500
00:28:26,873 --> 00:28:29,723
systems which are doing some
sort of few-shot learning,

501
00:28:29,723 --> 00:28:32,676
someone has to teach them. In
our case, there's no one to

502
00:28:32,676 --> 00:28:35,630
teach. So the first part of how
do you flag something new

503
00:28:35,630 --> 00:28:38,843
becomes very relevant, which is
the bit I just told you about.

504
00:28:38,843 --> 00:28:42,056
The other thing is, how do you
label it? Is it a subclass of a

505
00:28:42,056 --> 00:28:44,906
tank? Is it a new kind of a
tank? Is it a totally novel

506
00:28:44,906 --> 00:28:48,067
category that you decide based
on these mid-level features as

507
00:28:48,067 --> 00:28:51,021
to how many are there in this
new kind of signal you have

508
00:28:51,021 --> 00:28:54,078
flagged? And that leads to the
self-labeling aspect of this

509
00:28:54,078 --> 00:28:56,773
thing. And once you have
identified and labeled, you

510
00:28:56,773 --> 00:28:59,520
will, for the simplest
explanation, extract a feature

511
00:28:59,520 --> 00:29:02,836
vector for this new thing, store
it in the memory, and then next

512
00:29:02,836 --> 00:29:05,686
time you see it, you would know
what it is. And you can

513
00:29:05,686 --> 00:29:09,003
generalize from that. So we have
these generalization functions.

514
00:29:09,003 --> 00:29:12,112
So the best way of thinking
about it is you have a Gaussian.

515
00:29:12,112 --> 00:29:15,273
You are in the center. You know
how to answer that because of

516
00:29:15,273 --> 00:29:18,434
our base foundational model.
Then you flag something which is

517
00:29:18,434 --> 00:29:21,440
a 2 sigma, and then you want to
generalize and extend your

518
00:29:21,440 --> 00:29:24,653
generalization from that. And
you keep on adding these things,

519
00:29:24,653 --> 00:29:26,000
and you keep on extending.

520
00:29:26,000 --> 00:29:49,000
Audrow Nash: Pull that.

521
00:29:49,000 --> 00:29:53,353
Amarjot Singh: But the beautiful
thing here is from that 2 sigma,

522
00:29:53,353 --> 00:29:57,237
you can find another 2 sigma,
another 2 sigma, and you can

523
00:29:57,237 --> 00:30:01,388
keep on doing it. So there is
expansion of capability. But you

524
00:30:01,388 --> 00:30:04,000
don't want to learn noise,
necessarily.

525
00:30:04,000 --> 00:30:06,758
Audrow Nash: Yeah, that was
gonna be what I was gonna ask

526
00:30:06,758 --> 00:30:07,000
next.

527
00:30:07,000 --> 00:30:10,477
Amarjot Singh: Yes. Yes, so we
have mechanisms to flag what is

528
00:30:10,477 --> 00:30:13,000
relevant. That can be done model
level, but--

529
00:30:13,000 --> 00:30:15,402
Audrow Nash: You like a neural
pruning kind of thing where

530
00:30:15,402 --> 00:30:17,888
you're like, oh, I keep seeing
this one and I don't see this

531
00:30:17,888 --> 00:30:20,000
one. Therefore, this one's
relevant. Keep it. What?

532
00:30:20,000 --> 00:30:23,587
Amarjot Singh: Frequency
recency. Yes. Frequency recency.

533
00:30:23,587 --> 00:30:27,175
If you don't see something for a
while, you can prune it.

534
00:30:27,175 --> 00:30:31,140
Whatever you see, you can keep
it. But if you see something for

535
00:30:31,140 --> 00:30:34,727
a while, your memory is still
going to start to clog, and

536
00:30:34,727 --> 00:30:38,692
you're on the edge device. And
then there's another compression

537
00:30:38,692 --> 00:30:42,343
mechanism which happens, which
will say, I have n vectors,

538
00:30:42,343 --> 00:30:45,993
which are being used again. Let
me fit a Gaussian on that.

539
00:30:45,993 --> 00:30:49,958
Replace the mean ones with the
mean, but the outliers which are

540
00:30:49,958 --> 00:30:53,671
beyond 1 sigma. Let's keep all
of them so that I don't lose

541
00:30:53,671 --> 00:30:56,000
information, but I can also
compress.

542
00:30:56,000 --> 00:30:57,591
Audrow Nash: Are you constantly
compressing to a vector of

543
00:30:57,591 --> 00:30:59,237
features for this kind of thing?
Or how do you represent all

544
00:30:59,237 --> 00:31:00,828
this? Like I'm imagining you
have data coming in, you have

545
00:31:00,828 --> 00:31:02,446
your distribution, you're one
sigma away from the mean, and

546
00:31:02,446 --> 00:31:04,092
you say, okay, we keep these
feature labels. And then that's

547
00:31:04,092 --> 00:31:05,656
very efficient because you're
just saying, oh, it's these

548
00:31:05,656 --> 00:31:07,000
features as opposed to keeping
the original data.

549
00:31:07,000 --> 00:31:14,776
Amarjot Singh: Yeah. Yeah. Yeah.
We are not keeping the original

550
00:31:14,776 --> 00:31:21,579
data. We would keep feature
vector only. That's one. And

551
00:31:21,579 --> 00:31:28,991
then when you keep on adding
more and more, it's again in the

552
00:31:28,991 --> 00:31:33,000
feature vector domain. So--
doing

553
00:31:33,000 --> 00:31:34,310
Audrow Nash: Yep, and you change
your distributions around and

554
00:31:34,310 --> 00:31:35,599
you can label them and yeah,
yeah, new center of mass, here's

555
00:31:35,599 --> 00:31:36,000
a new distribution.

556
00:31:36,000 --> 00:31:40,663
Amarjot Singh: Exactly. Exactly.
Exactly. When you're doing,

557
00:31:40,663 --> 00:31:45,249
let's say, robotics, and we are
trying to do this humanoid,

558
00:31:45,249 --> 00:31:49,912
trying to learn a new kind of
graph. Let's say whether it is

559
00:31:49,912 --> 00:31:54,575
trying to pick a pole or trying
to do crawling. Someone will

560
00:31:54,575 --> 00:31:59,083
show it once, or it can also
learn that from its reasoning

561
00:31:59,083 --> 00:32:03,746
model. Then you would add a
single graph, which is temporal,

562
00:32:03,746 --> 00:32:06,000
and you generalize from that.

563
00:32:06,000 --> 00:32:10,537
Audrow Nash: Okay, I'm not fully
all the way there on how this

564
00:32:10,537 --> 00:32:12,000
translates into like

565
00:32:12,000 --> 00:32:13,000
Amarjot Singh: The body.

566
00:32:13,000 --> 00:32:13,622
Audrow Nash: Motion for this
kind of thing for robots. I

567
00:32:13,622 --> 00:32:14,000
understand it from like a
labeling

568
00:32:14,000 --> 00:32:19,000
Amarjot Singh: OK. We'll get to
it.

569
00:32:19,000 --> 00:32:22,709
Audrow Nash: Data, labeling
sensor input data. But so that's

570
00:32:22,709 --> 00:32:26,603
kind of the key. Your efficient
data compression approach where

571
00:32:26,603 --> 00:32:30,249
you have the ability to grab
from the middling label layers

572
00:32:30,249 --> 00:32:34,143
and then kind of redirect them
to reevaluate things because you

573
00:32:34,143 --> 00:32:37,790
may have only, you may have a
camouflage tank. And then the

574
00:32:37,790 --> 00:32:41,437
second part is compressing
everything down to kind of those

575
00:32:41,437 --> 00:32:45,022
mid-level features so you can
learn a distribution of what

576
00:32:45,022 --> 00:32:47,000
you're seeing in the real world.

577
00:32:47,000 --> 00:32:51,000
Amarjot Singh: Yeah. Yeah. All
right.

578
00:32:51,000 --> 00:32:52,012
Audrow Nash: Does that relate to
motion? Or is that kind of, is

579
00:32:52,012 --> 00:32:52,960
that the key part that you would
make sense that's like the

580
00:32:52,960 --> 00:32:53,924
largest problem you're solving?
I know that there's a ton of

581
00:32:53,924 --> 00:32:54,888
other subdomains that you're
interested in, but is that kind

582
00:32:54,888 --> 00:32:55,000
of like

583
00:32:55,000 --> 00:33:09,000
Amarjot Singh: Look at this.

584
00:33:09,000 --> 00:33:10,506
Audrow Nash: The core of most of
the things we're going to be

585
00:33:10,506 --> 00:33:11,000
talking about today?

586
00:33:11,000 --> 00:33:12,345
Amarjot Singh: Yeah. So the same
architecture applies to the

587
00:33:12,345 --> 00:33:13,691
whole stack. So after this, I
can take you to the stack. But

588
00:33:13,691 --> 00:33:15,013
if you see the diagram on the
screen, so you can see as the

589
00:33:15,013 --> 00:33:16,000
architecture builds, it does
your-- wavelet.

590
00:33:16,000 --> 00:33:29,000
Audrow Nash: Yep. Mm-hm. You get
your edge detection

591
00:33:29,000 --> 00:33:30,000
Amarjot Singh: Yeah. Wavelets.

592
00:33:29,000 --> 00:33:31,027
Audrow Nash: And you have your
like middle feature detection

593
00:33:31,027 --> 00:33:32,986
and then you have your like
object detection from that. So

594
00:33:32,986 --> 00:33:34,000
yeah, I see your three layers.

595
00:33:34,000 --> 00:33:39,000
Amarjot Singh: And then-- Yeah.
And then you have the--

596
00:33:39,000 --> 00:33:41,197
Audrow Nash: And I know that
that's just a representation.

597
00:33:41,197 --> 00:33:43,394
You probably have all sorts of
intermediate layers between

598
00:33:43,394 --> 00:33:44,000
everything, but.

599
00:33:44,000 --> 00:33:45,197
Amarjot Singh: Yeah. Yeah. So
coarser level, if you want to

600
00:33:45,197 --> 00:33:46,313
see, it's handcrafted,
unsupervised, and supervised. So

601
00:33:46,313 --> 00:33:47,551
you can see three blocks. And
then you can see the memory bit

602
00:33:47,551 --> 00:33:48,809
below, where you have a bunch of
vectors you add. And then the

603
00:33:48,809 --> 00:33:50,046
compression is, let's keep the
mean with the outliers so that

604
00:33:50,046 --> 00:33:51,000
the memory gets compressed. So
that's the idea.

605
00:33:51,000 --> 00:34:07,000
Audrow Nash: Mm-hm. Yep. It's a
very cool approach.

606
00:34:07,000 --> 00:34:07,249
Amarjot Singh: Yeah. But when
you get to the stack part of it,

607
00:34:07,249 --> 00:34:07,498
so the stack would be, let's say
you have this kind of data on

608
00:34:07,498 --> 00:34:07,739
which you want to build the
model, which is how do you learn

609
00:34:07,739 --> 00:34:07,964
motion, how do you learn
manipulation. There are a bunch

610
00:34:07,964 --> 00:34:08,000
of things

611
00:34:07,000 --> 00:34:22,000
Audrow Nash: Mm-hm. Yeah.

612
00:34:22,000 --> 00:34:23,000
Amarjot Singh: Happening here.

613
00:34:22,000 --> 00:34:23,738
Audrow Nash: Yeah, so tell me
how that works. How do you, how

614
00:34:23,738 --> 00:34:25,533
do you relate it? How do you do
a, like, if you watch a person,

615
00:34:25,533 --> 00:34:27,157
maybe you do like a skeleton
detection kind of thing. And

616
00:34:27,157 --> 00:34:28,895
then you use that as a feature
representation and you look at

617
00:34:28,895 --> 00:34:30,661
like several frames of that put
into one vector and that's how

618
00:34:30,661 --> 00:34:32,000
you start to learn a skill or
how does it work?

619
00:34:32,000 --> 00:34:45,000
Amarjot Singh: Yeah. Yeah. Yeah.
Yeah, let Yeah.

620
00:34:45,000 --> 00:34:45,690
Audrow Nash: I guess you're
saying, yeah, so. Yeah, but go

621
00:34:45,690 --> 00:34:46,000
ahead. Tell me about this.

622
00:34:46,000 --> 00:34:47,901
Amarjot Singh: Let's get into
the-- Let's get into the

623
00:34:47,901 --> 00:34:49,979
robotics stack. Yeah. So let's
get into the robotics stack.

624
00:34:49,979 --> 00:34:52,197
Right? So what we are doing in
the robotics stack is you give a

625
00:34:52,197 --> 00:34:54,310
task, which can be all the way
from scanning for the bomb or

626
00:34:54,310 --> 00:34:56,000
firing a gun or picking a box,
whatever that is.

627
00:34:56,000 --> 00:35:04,000
Audrow Nash: Mm-hm. Which is so
crazy. Yeah, what a thing.

628
00:35:04,000 --> 00:35:05,577
Amarjot Singh: Yeah. So when you
give a task, there are a couple

629
00:35:05,577 --> 00:35:07,056
of things happening. Right? So
it goes to a reasoning model,

630
00:35:07,056 --> 00:35:08,461
which could be part of the
foundational model or it could

631
00:35:08,461 --> 00:35:09,965
be a separate stack, where you
say, hey, this is a task. Tell

632
00:35:09,965 --> 00:35:11,000
me what to do. What is my high
level plan?

633
00:35:11,000 --> 00:35:20,000
Audrow Nash: Mm-hm.

634
00:35:20,000 --> 00:35:24,663
Amarjot Singh: Now, in our case,
generally people will use a

635
00:35:24,663 --> 00:35:29,403
standard VLM, a ChatGPT, or a
Claude to do that. And they are

636
00:35:29,403 --> 00:35:34,221
mainly reasoning over vision.
You give a language. You'll give

637
00:35:34,221 --> 00:35:39,039
a picture or a small video of
what is in front of you, usually

638
00:35:39,039 --> 00:35:44,012
a picture. And then ChatGPT will
say, do this, do that. And then

639
00:35:44,012 --> 00:35:48,675
eventually you'll be able to
complete the task. In our case,

640
00:35:48,675 --> 00:35:53,337
we are not reasoning just over
vision. We are reasoning over

641
00:35:53,337 --> 00:35:58,000
vision, audio, gestures, methane
gas detectors, gas sensors.

642
00:35:58,000 --> 00:35:59,000
Audrow Nash: So various sensors,
yeah.

643
00:35:59,000 --> 00:36:00,155
Amarjot Singh: Exactly. So we
have our custom sensor language

644
00:36:00,155 --> 00:36:01,309
model. Let's say that is the
first thing we want to build. So

645
00:36:01,309 --> 00:36:02,483
we will take the architecture,
which I have been explaining to

646
00:36:02,483 --> 00:36:03,637
you, which is here. We will take
all of that data and we will

647
00:36:03,637 --> 00:36:04,584
build this very nice
foundational model so that if

648
00:36:04,584 --> 00:36:05,700
you give me a language along
with a bunch of these sensors,

649
00:36:05,700 --> 00:36:06,817
it should produce at the end
what are the things you should

650
00:36:06,817 --> 00:36:07,896
do. That's pretty much it. Now,
if you have a new kind of

651
00:36:07,896 --> 00:36:09,107
reasoning, which I have not seen
or a new kind of condition, you

652
00:36:09,107 --> 00:36:10,167
want to learn that in your
memory. And that is where you

653
00:36:10,167 --> 00:36:11,000
add vectors corresponding to it.
Make sense?

654
00:36:11,000 --> 00:36:45,000
Audrow Nash: Mm-hm. Yes.

655
00:36:45,000 --> 00:36:47,443
Amarjot Singh: Right. And in
case of action, once you have

656
00:36:47,443 --> 00:36:50,139
understood that, you want to see
and you want to act. So the see

657
00:36:50,139 --> 00:36:52,708
part I told you, the act part is
you're trying to learn these

658
00:36:52,708 --> 00:36:55,446
kind of models. You're trying to
learn the whole body control and

659
00:36:55,446 --> 00:36:57,889
then you're trying to learn
manipulation. So let's say the

660
00:36:57,889 --> 00:37:00,122
whole body control, we would
gather a lot of data for

661
00:37:00,122 --> 00:37:02,649
different kind of graphs we
have. And then we will train our

662
00:37:02,649 --> 00:37:05,176
base model, which is there. And
if you ask it to do that, it

663
00:37:05,176 --> 00:37:07,704
will do it. But if let's say you
tell it to do a new kind of

664
00:37:07,704 --> 00:37:10,105
graph where this person is,
let's say, doing this kind of

665
00:37:10,105 --> 00:37:12,505
action, we will record the
skeleton and we will condition

666
00:37:12,505 --> 00:37:15,075
the skeleton on our base model
so that if you're doing like a

667
00:37:15,075 --> 00:37:17,729
loose kind of recording, you can
kind of adjust that so that it

668
00:37:17,729 --> 00:37:20,087
is sensible. Then you store that
to the memory, which is

669
00:37:20,087 --> 00:37:22,320
basically a bunch of graphs
recorded at 30 frames per

670
00:37:22,320 --> 00:37:24,721
second. And let's say your
action lasts for five seconds.

671
00:37:24,721 --> 00:37:27,122
So you have 30 into 5, 150
graphs. Now, if you say do the

672
00:37:27,122 --> 00:37:29,607
same thing, it will just call
that graph and it will do it.

673
00:37:29,607 --> 00:37:32,050
But let's say now if you put an
obstacle in front of it or

674
00:37:32,050 --> 00:37:34,619
something like that, then it
would pull that graph, the model

675
00:37:34,619 --> 00:37:37,231
will predict what actions you
should take given this obstacle.

676
00:37:37,231 --> 00:37:39,000
It will modify the graph and
then execute.

677
00:37:39,000 --> 00:37:57,860
Audrow Nash: Mm-hm. Cool. Okay,
tell me about the graph and how

678
00:37:57,860 --> 00:38:15,822
that relates to the temporal
nature of actions for this kind

679
00:38:15,822 --> 00:38:26,000
of thing. Like, what is the
graph?

680
00:38:26,000 --> 00:38:28,028
Amarjot Singh: Yeah. Yeah, so
the graph is basically whatever

681
00:38:28,028 --> 00:38:29,957
degree of freedoms you have. So
in your humanoid, you have

682
00:38:29,957 --> 00:38:31,918
whatever 20 plus or 25 plus
degrees of freedom. If you look

683
00:38:31,918 --> 00:38:33,813
at the example of the hand
below, it's a five-finger hand

684
00:38:33,813 --> 00:38:35,808
for us. So you have whatever
degrees of freedom. So when I'm

685
00:38:35,808 --> 00:38:37,803
recording, initially I'm more or
less recording what are the

686
00:38:37,803 --> 00:38:39,000
motor values for each of them?
Yeah.

687
00:38:39,000 --> 00:38:44,299
Audrow Nash: Awesome. For
everything. So you take the pose

688
00:38:44,299 --> 00:38:49,690
tracking, the skeleton, and you
infer the joint angles from

689
00:38:49,690 --> 00:38:55,081
everything. Some coordinate
system that you've established.

690
00:38:55,081 --> 00:38:57,000
You map it into that.

691
00:38:57,000 --> 00:39:04,000
Amarjot Singh: You're right.
Yeah. Yes.

692
00:39:04,000 --> 00:39:06,000
Audrow Nash: And then, okay.

693
00:39:06,000 --> 00:39:07,388
Amarjot Singh: And map from
there to the motor functions so

694
00:39:07,388 --> 00:39:08,000
that if I have to do this.

695
00:39:08,000 --> 00:39:10,100
Audrow Nash: And map from there
to the motor functions. How imp

696
00:39:10,100 --> 00:39:12,033
how difficult is that, I guess?
Because humanoids, they're

697
00:39:12,033 --> 00:39:14,067
shaped kind of like people, but
they might be able to move in

698
00:39:14,067 --> 00:39:16,233
different ways. Also, if the one
is four and a half feet tall, it

699
00:39:16,233 --> 00:39:18,267
may have to do the task in a
different way than someone who's

700
00:39:18,267 --> 00:39:20,133
six foot six or something. how
tell me Tell me about how

701
00:39:20,133 --> 00:39:21,000
difficult that mapping is.

702
00:39:21,000 --> 00:39:29,487
Amarjot Singh: Yeah. Yeah. It's
a very good point. So you can

703
00:39:29,487 --> 00:39:37,000
learn that mapping. That's what
we do. So for example,

704
00:39:37,000 --> 00:39:40,000
Audrow Nash: Yeah. That's such a
cool point.

705
00:39:40,000 --> 00:39:42,557
Amarjot Singh: Yeah, because
like, let's say you're trying to

706
00:39:42,557 --> 00:39:45,030
do certain action and we will
record your graph and then we

707
00:39:45,030 --> 00:39:47,629
will record a machine doing the
same thing. And they have that

708
00:39:47,629 --> 00:39:50,353
graph and you want to learn this
model in between. Yeah, exactly.

709
00:39:50,353 --> 00:39:52,826
Minimizes the distance between
the graphs and we would do a

710
00:39:52,826 --> 00:39:55,467
bunch of that. And that more or
less acts as a mapping function

711
00:39:55,467 --> 00:39:57,898
so that if you do it, we will
push it through that mapping

712
00:39:57,898 --> 00:40:00,581
function so that when it goes on
to the motor side, it is how it

713
00:40:00,581 --> 00:40:01,000
should be.

714
00:40:01,000 --> 00:40:07,613
Audrow Nash: Relationship.
Mm-hm. How do you... I guess,

715
00:40:07,613 --> 00:40:14,934
how does it work if something is
I guess, how does it work for

716
00:40:14,934 --> 00:40:22,137
slight variations or variations
in like, say you've trained a

717
00:40:22,137 --> 00:40:28,985
robot how to open a door. Now
you get a slightly different

718
00:40:28,985 --> 00:40:33,000
door handle than you've ever
seen.

719
00:40:33,000 --> 00:40:36,000
Amarjot Singh: Yeah. Yeah.

720
00:40:36,000 --> 00:40:37,210
Audrow Nash: Does it just throw
its hand up and doesn't know how

721
00:40:37,210 --> 00:40:38,326
to do it? Or how do you deal...
Because you're talking also

722
00:40:38,326 --> 00:40:39,499
about learning on the job. Does
it always need to see a person

723
00:40:39,499 --> 00:40:40,652
do it the first time? Or what
would happen? Like, okay, we've

724
00:40:40,652 --> 00:40:41,655
only trained on round door
knobs. Now it's one with a

725
00:40:41,655 --> 00:40:42,827
lever. How does it work? Or that
it's like a ship door to open

726
00:40:42,827 --> 00:40:43,887
where you have to spin the
thing. How would we deal with

727
00:40:43,887 --> 00:40:44,000
those?

728
00:40:44,000 --> 00:40:48,319
Amarjot Singh: Yeah. Yeah. Yeah.
Yeah. Yeah. Yeah. Yeah, that's

729
00:40:48,319 --> 00:40:52,431
great. That's great. So let's
say you teach it how to grab a

730
00:40:52,431 --> 00:40:56,202
specific bottle. Here is an
example at the bottom right

731
00:40:56,202 --> 00:41:00,383
where you are telling it to more
or less grab a bottle. So we

732
00:41:00,383 --> 00:41:04,702
would tell it grab a bottle and
you can either put an object in

733
00:41:04,702 --> 00:41:08,883
front of it. So it has to go
around it or you can put it at a

734
00:41:08,883 --> 00:41:12,585
height. And now there is
variation in height. So it is

735
00:41:12,585 --> 00:41:16,835
able to do a lot of cool things
because we don't learn a pixel

736
00:41:16,835 --> 00:41:18,000
to action policy.

737
00:41:18,000 --> 00:41:37,000
Audrow Nash: Yep.

738
00:41:37,000 --> 00:41:41,292
Amarjot Singh: We learn a
geometry to action policy. So we

739
00:41:41,292 --> 00:41:45,066
use 3D data, whether it is
coming from a LiDAR or a

740
00:41:45,066 --> 00:41:49,654
RealSense camera. So if you put
a bottle somewhere else and if

741
00:41:49,654 --> 00:41:53,798
you recorded one trajectory, I
can add and subtract that

742
00:41:53,798 --> 00:41:58,387
original trajectory to say, hey,
this is where the bottle was.

743
00:41:58,387 --> 00:42:02,827
This is where the bottle has
moved. So I should do this. And

744
00:42:02,827 --> 00:42:07,045
that is what you can see that
curve getting formed at the

745
00:42:07,045 --> 00:42:11,559
bottom. And since we have this
also skeleton, which is in the

746
00:42:11,559 --> 00:42:16,073
3D space, you can also add and
subtract to that to adjust it.

747
00:42:16,073 --> 00:42:20,587
Now, how much can you do? It's a
function of what you have in

748
00:42:20,587 --> 00:42:25,249
your base skill. If you have a
certain way, the hand is and you

749
00:42:25,249 --> 00:42:29,763
want to add and subtract little
things to that, the algorithm

750
00:42:29,763 --> 00:42:33,833
can do that. It's basically a
key point transform. It's

751
00:42:33,833 --> 00:42:37,904
basically a graph neural
network. So it will modify the

752
00:42:37,904 --> 00:42:42,492
original graph. But if you have
the need to have a totally new

753
00:42:42,492 --> 00:42:46,710
graph, we call our sensor
language model, which will then

754
00:42:46,710 --> 00:42:51,002
generate a graph which you can
add to your memory and then

755
00:42:51,002 --> 00:42:53,000
generalize from that point.

756
00:42:53,000 --> 00:42:56,974
Audrow Nash: Cool. Nice. Okay. I
really like the approach of

757
00:42:56,974 --> 00:43:01,013
splitting it up. So you have the
trained version and then you

758
00:43:01,013 --> 00:43:04,722
have a memory, which is like a
local way of executing or

759
00:43:04,722 --> 00:43:08,563
perceiving things. I think
that's a very clever split. And

760
00:43:08,563 --> 00:43:12,669
then you can take the data from
the memory and aggregate it on

761
00:43:12,669 --> 00:43:16,709
there and then incorporate it
into your model, which is quite

762
00:43:16,709 --> 00:43:20,483
cool. How do you, for getting
the pose of people, are you

763
00:43:20,483 --> 00:43:24,788
mentioned needing 3D information
for this. I'm wondering, can you

764
00:43:24,788 --> 00:43:28,695
use a pose detector on a 2D
image and try to, like, are you

765
00:43:28,695 --> 00:43:32,404
able to scrape YouTube, for
example, for learning how to

766
00:43:32,404 --> 00:43:33,000
locomote?

767
00:43:33,000 --> 00:43:40,000
Amarjot Singh: Yes. Yes. Yes.
Yes. So. the Yeah.

768
00:43:40,000 --> 00:43:41,476
Audrow Nash: Because the pose
detectors are quite good lately,

769
00:43:41,476 --> 00:43:42,881
I think. And so you can learn
from... Mm-hmm. Tell me about

770
00:43:42,881 --> 00:43:43,000
that.

771
00:43:43,000 --> 00:43:44,604
Amarjot Singh: Yeah. So yeah. So
I think the way I mean, one

772
00:43:44,604 --> 00:43:46,314
thing is that we want to cut the
data bottlenecks. So we want to

773
00:43:46,314 --> 00:43:47,998
do few-shot learning. So we want
to have a very big base model.

774
00:43:47,998 --> 00:43:49,682
We want to have a memory so you
can add new graphs to that. And

775
00:43:49,682 --> 00:43:51,258
then we want to either do it
automatically using our Oracle

776
00:43:51,258 --> 00:43:52,969
so it can estimate error, small
or large, and produce new graphs

777
00:43:52,969 --> 00:43:54,626
automatically. Or you can teach
it. Now when you teach it, you

778
00:43:54,626 --> 00:43:55,000
can either use

779
00:43:55,000 --> 00:44:17,000
Audrow Nash: Mm-hmm.

780
00:44:17,000 --> 00:44:21,000
Amarjot Singh: The Apple Vision
Pro, wear those glasses.

781
00:44:21,000 --> 00:44:22,909
Audrow Nash: That's amazing. I
haven't, when I saw it come out,

782
00:44:22,909 --> 00:44:24,697
I thought it was such a good
product for robotics. Like, it

783
00:44:24,697 --> 00:44:26,576
was just like, this, Apple is
getting into robotics was my big

784
00:44:26,576 --> 00:44:28,455
thought. And then I haven't seen
any companies using it really

785
00:44:28,455 --> 00:44:30,242
well. And that's really cool
that you guys are using it. So

786
00:44:30,242 --> 00:44:31,000
yeah, tell me about that.

787
00:44:31,000 --> 00:44:33,041
Amarjot Singh: Yes. Yeah.
Because then you can have a one

788
00:44:33,041 --> 00:44:35,189
to one mapping of how you're
looking to the machine. So it's

789
00:44:35,189 --> 00:44:37,481
more accurate. But if let's say
you deploy this thing in my home

790
00:44:37,481 --> 00:44:39,629
and my mom would say, hey, can
you do this? Let me show you.

791
00:44:39,629 --> 00:44:41,670
She's not going to wear the
Apple Vision Pro. So we use a

792
00:44:41,670 --> 00:44:43,747
lot of egocentric data, which
you can see on the top right

793
00:44:43,747 --> 00:44:45,000
where this person is
demonstrating.

794
00:44:45,000 --> 00:45:07,000
Audrow Nash: Yep. Mm-hmm.

795
00:45:07,000 --> 00:45:08,327
Amarjot Singh: And the machine
can do the same thing, but also

796
00:45:08,327 --> 00:45:09,697
it can adapt to a new graph. It
is all one shot. So yes, you can

797
00:45:09,697 --> 00:45:11,003
show it. We can get a 2D pose
and then we can map the 2D pose

798
00:45:11,003 --> 00:45:12,352
to the machine pose and do that.
But you need to have a mapping

799
00:45:12,352 --> 00:45:13,679
function in between, which is
trained. So we have a lot of the

800
00:45:13,679 --> 00:45:14,000
demonstrations.

801
00:45:14,000 --> 00:45:29,000
Audrow Nash: Hmm. Mm-hmm. Yes.

802
00:45:29,000 --> 00:45:32,000
Amarjot Singh: We have a lot of
the similar demonstrations the

803
00:45:32,000 --> 00:45:35,000
machine is doing. And then we
try to minimize that difference.

804
00:45:35,000 --> 00:45:39,026
Audrow Nash: That's clever. It's
clever to me. I mean, it's

805
00:45:39,026 --> 00:45:43,121
something, I guess, you would
think about if you were trying

806
00:45:43,121 --> 00:45:47,010
to do this, that you need a
mapping function of how to go

807
00:45:47,010 --> 00:45:50,832
from a human to whatever
humanoid or robot you're doing.

808
00:45:50,832 --> 00:45:55,268
Does this, so I know humanoid is
becoming a bigger focus. Is this

809
00:45:55,268 --> 00:45:59,499
also true for like, I saw some
robot dogs when you were giving

810
00:45:59,499 --> 00:46:01,000
me a tour earlier. Are

811
00:46:01,000 --> 00:46:03,000
Amarjot Singh: Yes. Yes.

812
00:46:03,000 --> 00:46:03,797
Audrow Nash: You able to map
onto different morphologies for

813
00:46:03,797 --> 00:46:04,608
everything? Like, can you map
from a human to a robot dog? Or

814
00:46:04,608 --> 00:46:05,458
how does all that work? I guess
you can just map a part of them.

815
00:46:05,458 --> 00:46:06,269
So if you have an arm strapped
to a robot dog, you can map to

816
00:46:06,269 --> 00:46:07,000
the arm of a human or something.
Or how does that work?

817
00:46:07,000 --> 00:46:09,451
Amarjot Singh: Yes. Yes. Yes. So
two parts, right? One is the

818
00:46:09,451 --> 00:46:11,741
whole body control and the
second is the manipulation. So

819
00:46:11,741 --> 00:46:14,353
on a lot of our robotic dogs, we
would use a similar kind of arm,

820
00:46:14,353 --> 00:46:16,844
maybe shorter, longer. Try to
use a five finger manipulator in

821
00:46:16,844 --> 00:46:19,375
all of them. So easy to use that
same manipulation model. Yeah,

822
00:46:19,375 --> 00:46:21,786
and different kind of things.
The whole body control differs

823
00:46:21,786 --> 00:46:24,196
because humanoids are different
in shape, obviously, bipedal

824
00:46:24,196 --> 00:46:25,000
versus quadruped. So

825
00:46:25,000 --> 00:46:49,000
Audrow Nash: Mm-hmm. So cool.

826
00:46:49,000 --> 00:46:52,761
Amarjot Singh: That is where the
change comes in the locomotion

827
00:46:52,761 --> 00:46:56,164
policy. But you can do the
manipulation policy with small

828
00:46:56,164 --> 00:46:57,000
modifications.

829
00:46:57,000 --> 00:46:57,642
Audrow Nash: Mm-hmm. Super cool.
Tell me how much of a role does

830
00:46:57,642 --> 00:46:58,274
simulation play in what you're
doing? Like, are you doing a lot

831
00:46:58,274 --> 00:46:58,866
of like whole body control
tuning and like control figuring

832
00:46:58,866 --> 00:46:59,388
out in simulation? How do you
leverage real data and

833
00:46:59,388 --> 00:47:00,000
simulation? Like, what's the,
what's your pipeline look like?

834
00:47:00,000 --> 00:47:02,673
Amarjot Singh: Yes. Yes. Yes. So
we do quite a bit of simulation.

835
00:47:02,673 --> 00:47:05,263
We mainly use NVIDIA's Isaac Sim
and Isaac Gym. Sim has more or

836
00:47:05,263 --> 00:47:07,895
less the Gym part as well, where
we are modeling forces. So if I

837
00:47:07,895 --> 00:47:10,403
want to push my humanoid or I
want to teach it that if I push

838
00:47:10,403 --> 00:47:12,993
you and you fall down, and you
should learn how to get up, that

839
00:47:12,993 --> 00:47:15,625
is very hard to do all the time
in the real world because you're

840
00:47:15,625 --> 00:47:17,969
going to damage that humanoid
and it's a $70,000, $80,000

841
00:47:17,969 --> 00:47:20,641
machine. So we would build a lot
of jumping from one point to the

842
00:47:20,641 --> 00:47:22,944
other. How big of a ditch can
you jump? Can you actually

843
00:47:22,944 --> 00:47:25,000
recover if you fall? So all of
that is simulation.

844
00:47:25,000 --> 00:47:32,947
Audrow Nash: Mm-hmm. Okay.
Mm-hmm. It's awesome. Ah. That

845
00:47:32,947 --> 00:47:41,870
makes sense. How, so when you go
from simulation to real robots,

846
00:47:41,870 --> 00:47:50,375
is there any gap there? Have you
guys figured it out? The way

847
00:47:50,375 --> 00:47:58,322
that I've heard that people
figured it out pretty well is

848
00:47:58,322 --> 00:48:06,827
they just vary things in physics
wildly. And then the robot's

849
00:48:06,827 --> 00:48:15,750
able to generalize fairly well.
But, and for whole body control,

850
00:48:15,750 --> 00:48:23,000
say, but tell me about that.
Tell me about Sim2Real.

851
00:48:23,000 --> 00:48:26,145
Amarjot Singh: Yes. I think most
of the things which people are

852
00:48:26,145 --> 00:48:29,290
trying to do is in manufacturing
and home. So your objective is

853
00:48:29,290 --> 00:48:32,434
to maybe stand straight or bend
down to pick something. You are

854
00:48:32,434 --> 00:48:35,529
not making the robot jump from
stair to stair or room to room,

855
00:48:35,529 --> 00:48:38,324
right? So the space of whole
body control in most of the

856
00:48:38,324 --> 00:48:41,170
other applications is quite
limited. So the idea that you

857
00:48:41,170 --> 00:48:44,115
can more or less vary all the
physics and hopefully get it.

858
00:48:44,115 --> 00:48:47,010
Also maybe you pick something
which is a certain amount of

859
00:48:47,010 --> 00:48:49,955
weight and then that can pull
the center of gravity so that

860
00:48:49,955 --> 00:48:52,950
you're going to do something
there. I think it's possible to

861
00:48:52,950 --> 00:48:56,045
do, we try to do that as well.
But still I think the number of

862
00:48:56,045 --> 00:48:58,990
variations which can come up are
quite a bit because you're

863
00:48:58,990 --> 00:49:01,885
moving on uneven surfaces,
you're carrying heavy stuff. So

864
00:49:01,885 --> 00:49:04,730
that kind of works for us but
you still need that adapter

865
00:49:04,730 --> 00:49:07,775
piece. And I think without the
adapter piece you would not be

866
00:49:07,775 --> 00:49:10,870
able to have reliable work in
our kind of field. I don't think

867
00:49:10,870 --> 00:49:13,815
you can even do that in the
homes and manufacturing side as

868
00:49:13,815 --> 00:49:16,661
well. I still feel like you
would need on-device adaption

869
00:49:16,661 --> 00:49:19,755
and people have been doing that
like Physical Intelligence but

870
00:49:19,755 --> 00:49:22,651
they would more or less do it in
an iterative way on a big

871
00:49:22,651 --> 00:49:23,000
server.

872
00:49:23,000 --> 00:49:53,000
Audrow Nash: Yep. Mm-hmm.

873
00:49:53,000 --> 00:49:57,377
Amarjot Singh: Skild has been
trying to do few shots but then

874
00:49:57,377 --> 00:50:02,040
they go and teach it or show it.
While in our case the part about

875
00:50:02,040 --> 00:50:06,058
where you can find something
new, learn it, that kind of

876
00:50:06,058 --> 00:50:09,000
makes it unattended and fully
autonomous.

877
00:50:09,000 --> 00:50:13,179
Audrow Nash: Mm-hmm. Which is
super cool. Okay. let's see.

878
00:50:13,179 --> 00:50:17,646
Tell me, so just to understand
better where the limits of this

879
00:50:17,646 --> 00:50:21,609
technology, like the approach
you're taking are at this

880
00:50:21,609 --> 00:50:25,788
moment, tell me some of the,
like, hard tasks, easy tasks,

881
00:50:25,788 --> 00:50:29,751
unattainable tasks, for
automating given this approach.

882
00:50:29,751 --> 00:50:33,930
Like, you can move a water
bottle, you can army crawl, You

883
00:50:33,930 --> 00:50:38,469
can bash down a door, but like,
so tell me what's easy, tell me

884
00:50:38,469 --> 00:50:42,000
what's hard, tell me what's
infeasible currently.

885
00:50:42,000 --> 00:50:46,571
Amarjot Singh: Yeah I think like
so if you look at from a user's

886
00:50:46,571 --> 00:50:50,786
perspective and a lot of our
users would be the US military

887
00:50:50,786 --> 00:50:55,214
or let's say oil and gas. Oil
and gas would say would love for

888
00:50:55,214 --> 00:50:59,357
your humanoid to come, go into
my mine and do some sort of

889
00:50:59,357 --> 00:51:03,286
blast so I don't have to send
someone in. Great. You're

890
00:51:03,286 --> 00:51:07,143
working in very dusty
environments. Sometimes there is

891
00:51:07,143 --> 00:51:11,643
oil, sometimes there is water.
So your humanoid should be water

892
00:51:11,643 --> 00:51:15,929
resistant. It'll be great if
it's waterproof. So none of the

893
00:51:15,929 --> 00:51:20,143
humanoids have that. So that's
one hardware bottleneck. The

894
00:51:20,143 --> 00:51:24,214
battery should not catch fire
because if you're operating

895
00:51:24,214 --> 00:51:28,643
around an oil mine that's a bad
idea. So none of the humanoids

896
00:51:28,643 --> 00:51:30,000
have that property.

897
00:51:30,000 --> 00:51:33,043
Audrow Nash: I was meaning more
from the learning approach, but

898
00:51:33,043 --> 00:51:35,795
this is interesting too. Like,
from the, cause, like, I'm

899
00:51:35,795 --> 00:51:38,886
saying what kinds of tasks, but
yeah, it is very funny, like, of

900
00:51:38,886 --> 00:51:41,736
course, you shouldn't have a
robot that lights on fire with

901
00:51:41,736 --> 00:51:44,634
its battery in an oil mine, that
would be like, the customer

902
00:51:44,634 --> 00:51:47,000
doesn't want you back at all,
this kind of thing.

903
00:51:47,000 --> 00:51:47,891
Amarjot Singh: Yeah fair. I
think on the software side of

904
00:51:47,891 --> 00:51:48,000
things,

905
00:51:47,000 --> 00:51:51,000
Audrow Nash: But.

906
00:51:51,000 --> 00:51:55,044
Amarjot Singh: The biggest
challenge we have is primarily

907
00:51:55,044 --> 00:51:59,017
the simple tasks we are doing
are picking things, moving

908
00:51:59,017 --> 00:52:03,133
things around which could be a
metal detector or let's say

909
00:52:03,133 --> 00:52:07,531
pushing the door in. Those are
things which are fairly easy to

910
00:52:07,531 --> 00:52:11,930
do, how reliable they are, how
many times can you run it again

911
00:52:11,930 --> 00:52:16,400
as a function of hardware. But
from a software point of view, I

912
00:52:16,400 --> 00:52:20,373
think you can manage that
because there isn't actually a

913
00:52:20,373 --> 00:52:24,985
lot of variability. You're doing
the same or very similar kind of

914
00:52:24,985 --> 00:52:29,171
task with a similar kind of
metal detector or you're trying

915
00:52:29,171 --> 00:52:33,286
to punch down the door. Very
minimal variability there but

916
00:52:33,286 --> 00:52:37,898
massive ROI, even if it works 70
or 80% of the time that is still

917
00:52:37,898 --> 00:52:42,368
okay because you're not getting
shot, someone else is. So yeah,

918
00:52:42,368 --> 00:52:44,000
so that has been great.

919
00:52:44,000 --> 00:52:46,936
Audrow Nash: What an interesting
thing. I hadn't thought about

920
00:52:46,936 --> 00:52:49,729
this until you just said it.
because one of the things with

921
00:52:49,729 --> 00:52:52,665
like, imitation learning is that
it's like, a lot of the demos

922
00:52:52,665 --> 00:52:55,459
you see, it's like, it's only
working 90% of the time, this

923
00:52:55,459 --> 00:52:58,489
kind of thing. but the thing is,
depending on the industry, that

924
00:52:58,489 --> 00:53:01,472
might be okay, because it's not
you getting shot or blown up or

925
00:53:01,472 --> 00:53:04,455
whatever it is. it's the robot,
which is expensive and a bummer

926
00:53:04,455 --> 00:53:07,249
if it, whatever, but it can,
it's still lower risk and less

927
00:53:07,249 --> 00:53:10,090
expensive than a human, I would
think. very interesting that

928
00:53:10,090 --> 00:53:12,600
because it's like, different
interest industries have

929
00:53:12,600 --> 00:53:15,157
different standards of
reliability that they need. And

930
00:53:15,157 --> 00:53:17,950
if this can do a lot of harm
reduction, even if it doesn't,

931
00:53:17,950 --> 00:53:20,791
or even if it works 70, 80% of
the time, it's still fine. Or

932
00:53:20,791 --> 00:53:23,538
you can just retry and then
it'll get it the next time for

933
00:53:23,538 --> 00:53:26,000
this kind of thing. That is
super interesting to me.

934
00:53:26,000 --> 00:53:31,381
Amarjot Singh: Yeah. Yeah. The
other thing which is very

935
00:53:31,381 --> 00:53:37,147
interesting and again goes to
the software is there are many

936
00:53:37,147 --> 00:53:43,009
robotic platforms in the mission
critical side but they would

937
00:53:43,009 --> 00:53:48,775
just do it one way. You can have
like a big robot with a big

938
00:53:48,775 --> 00:53:54,733
barrel and it will go and hit
the door and everyone is waiting

939
00:53:54,733 --> 00:53:58,000
for that to happen. It will go
in.

940
00:53:58,000 --> 00:54:00,000
Audrow Nash: So specialized.
Yeah, yeah.

941
00:54:00,000 --> 00:54:03,000
Amarjot Singh: Yeah, but you
want to have a machine

942
00:54:03,000 --> 00:54:04,000
Audrow Nash: It's a general, do
a lot of things.

943
00:54:03,000 --> 00:54:07,119
Amarjot Singh: Or a software
package. Yeah, mission set on

944
00:54:07,119 --> 00:54:11,451
mission set on mission sets. So
the foundational models right

945
00:54:11,451 --> 00:54:15,144
now we have, they can do
breaching, they can do your

946
00:54:15,144 --> 00:54:19,405
weapon engagement, IED. But as
you start to go up, data is a

947
00:54:19,405 --> 00:54:23,666
bottleneck and then software is
a bottleneck and then it can

948
00:54:23,666 --> 00:54:27,785
start to balloon up. So we
haven't tested beyond that. But

949
00:54:27,785 --> 00:54:32,188
also the way you move, right, a
lot of the temporalness starts

950
00:54:32,188 --> 00:54:36,591
to become very, very unique. So
that is something which we are

951
00:54:36,591 --> 00:54:40,000
working through but right now
the biggest point.

952
00:54:40,000 --> 00:54:41,280
Audrow Nash: I don't understand.
What do you mean? When you move

953
00:54:41,280 --> 00:54:42,440
temporal, things become unique.
What do you mean? Just the

954
00:54:42,440 --> 00:54:43,600
ground being varied or the
robot's dynamics changing as it

955
00:54:43,600 --> 00:54:44,000
runs or what do you,

956
00:54:44,000 --> 00:54:51,000
Amarjot Singh: Oh yeah, yeah.
Robot dynamics. Robot dynamics.

957
00:54:51,000 --> 00:54:54,861
Audrow Nash: Things heat up. You
get mud in the area. There's

958
00:54:54,861 --> 00:54:58,405
friction to go through, like all
sorts of variety by the

959
00:54:58,405 --> 00:55:01,000
environment and the parts
wearing, maybe.

960
00:55:01,000 --> 00:55:03,542
Amarjot Singh: If you just take
the example of the human just

961
00:55:03,542 --> 00:55:06,083
kicking the door, the way the
government wants to deploy this

962
00:55:06,083 --> 00:55:08,583
is that it's a group moving and
they want the humanoid to be

963
00:55:08,583 --> 00:55:10,833
part of the group. So the
humanoid should move how the

964
00:55:10,833 --> 00:55:13,458
humans move. So you are part of
the group. So when you kick the

965
00:55:13,458 --> 00:55:15,875
door in, you make space, the
other humans move in. So it's

966
00:55:15,875 --> 00:55:18,250
formation dynamics and that is
basically another temporal

967
00:55:18,250 --> 00:55:20,667
dimension on top of just your
human. And that is currently

968
00:55:20,667 --> 00:55:23,000
difficult and we are working on
that side. And they call

969
00:55:23,000 --> 00:55:40,000
Audrow Nash: Oh. Mm-hmm. Huh.
That's really cool.

970
00:55:40,000 --> 00:55:43,879
Amarjot Singh: It the humanoid
Turing test. So we haven't

971
00:55:43,879 --> 00:55:48,166
pushed that out yet. That if you
totally cover it and you are a

972
00:55:48,166 --> 00:55:52,180
serviceman and you are saying,
"Hey, this is your buddy and

973
00:55:52,180 --> 00:55:56,399
he's trained the same way." Go
and do breaching and they do it

974
00:55:56,399 --> 00:56:00,618
and then you replace the human
with the humanoid. Can you tell

975
00:56:00,618 --> 00:56:03,000
it's an actual human or a
humanoid?

976
00:56:03,000 --> 00:56:05,961
Audrow Nash: It's awesome. What
a world we live in. It's so

977
00:56:05,961 --> 00:56:09,073
freaking cool. okay, I love it.
What are, what are some tasks?

978
00:56:09,073 --> 00:56:12,185
So moving with the group is very
interesting because you don't

979
00:56:12,185 --> 00:56:15,347
want to be friction. You don't
want these robots to be awkward.

980
00:56:15,347 --> 00:56:16,000
To work with.

981
00:56:16,000 --> 00:56:20,000
Amarjot Singh: Exactly.

982
00:56:20,000 --> 00:56:23,447
Audrow Nash: So you have to
realize, and that might vary

983
00:56:23,447 --> 00:56:27,325
between groups. Like some groups
move faster, some move slower,

984
00:56:27,325 --> 00:56:31,018
some whatever. So you have to
vary how they do that. So that

985
00:56:31,018 --> 00:56:34,650
sounds like a very interesting
problem to me. but what are,

986
00:56:34,650 --> 00:56:38,466
what are some tasks? That's, so
I'm hearing moving stuff, like

987
00:56:38,466 --> 00:56:42,467
picking up objects, manipulating
them and getting around are kind

988
00:56:42,467 --> 00:56:46,284
of like the relatively easy ones
that are relatively reliable.

989
00:56:46,284 --> 00:56:49,792
Then you have tasks like
operating well in the group that

990
00:56:49,792 --> 00:56:53,424
are hard and not quite solved
yet. what are some other ones

991
00:56:53,424 --> 00:56:56,994
that are hard and not quite
solved? And then what are some

992
00:56:56,994 --> 00:57:00,503
ones that you think are just
like our current approach in

993
00:57:00,503 --> 00:57:04,380
kind of this imitation learning
setup is not capable of solving

994
00:57:04,380 --> 00:57:08,135
yet. And then I'd love to hear
how you think those will fall.

995
00:57:08,135 --> 00:57:12,013
Like when will we get to those?
Which ones will be the absolute

996
00:57:12,013 --> 00:57:15,768
hardest problems for us for this
kind of thing? Moving in the

997
00:57:15,768 --> 00:57:19,584
group seems wonderful as a task.
That seems like an incredible

998
00:57:19,584 --> 00:57:21,000
thing once that's done.

999
00:57:21,000 --> 00:57:24,624
Amarjot Singh: Yeah. I think the
biggest challenge in all of this

1000
00:57:24,624 --> 00:57:27,690
is generalization. And that's
the biggest bottleneck of

1001
00:57:27,690 --> 00:57:31,202
imitation learning, right? How
general is it? And can it extend

1002
00:57:31,202 --> 00:57:34,491
on its own, few shots, one shot?
I think that's the biggest

1003
00:57:34,491 --> 00:57:37,836
challenge. And I think whatever
we have deployed, it kind of

1004
00:57:37,836 --> 00:57:41,181
works. And then we have this VLM
model. If your graph is not

1005
00:57:41,181 --> 00:57:44,303
there, it can generate that
graph. But how good it is at

1006
00:57:44,303 --> 00:57:47,537
generating that graph. Can you
really generalize from that

1007
00:57:47,537 --> 00:57:50,993
graph? Can the modifications be
there? And all of that I would

1008
00:57:50,993 --> 00:57:53,000
put under the generalization
bucket.

1009
00:57:53,000 --> 00:58:05,000
Audrow Nash: Mm hmm.

1010
00:58:05,000 --> 00:58:08,956
Amarjot Singh: Because we don't
want to go down the path of keep

1011
00:58:08,956 --> 00:58:12,602
on retuning the foundational
model. Because if you do that,

1012
00:58:12,602 --> 00:58:16,620
you will never actually get to a
reliable system. Adaption is the

1013
00:58:16,620 --> 00:58:20,575
key. But then the adaption side,
you can say, well, I'm using 15

1014
00:58:20,575 --> 00:58:24,098
minutes of data rather than
10,000. But how good is that?

1015
00:58:24,098 --> 00:58:27,745
And that will all come from
testing in the field. Because a

1016
00:58:27,745 --> 00:58:31,330
lot of the testing we do, we
capture these failure points,

1017
00:58:31,330 --> 00:58:35,223
which the system automatically
picks. And then we add it and it

1018
00:58:35,223 --> 00:58:38,994
generalize from there. But can
your mid-level detector or the

1019
00:58:38,994 --> 00:58:42,702
thing which is looking for an
emerging signal, is it able to

1020
00:58:42,702 --> 00:58:46,349
capture all of that? Do your
things fall only on 2 sigma or

1021
00:58:46,349 --> 00:58:50,242
are they at 10 sigma? So there
are bottlenecks on, can you find

1022
00:58:50,242 --> 00:58:54,198
something new? Can you reliably
label it? Can you reliably learn

1023
00:58:54,198 --> 00:58:57,845
it, which is a function of
generalization? And if you don't

1024
00:58:57,845 --> 00:59:01,615
have that, then do you have to
go and retrain your base model

1025
00:59:01,615 --> 00:59:05,076
because the base model wasn't
good enough to do all that

1026
00:59:05,076 --> 00:59:08,970
stuff? So that we will discover
as we go and deploy more. But I

1027
00:59:08,970 --> 00:59:12,307
think that is the biggest
bottleneck in robotics, that

1028
00:59:12,307 --> 00:59:16,325
generalization is not there. And
for our two mission sets, we can

1029
00:59:16,325 --> 00:59:20,219
more or less still be okay with
that because the variability is

1030
00:59:20,219 --> 00:59:23,000
less. And the other cases,
that's quite high.

1031
00:59:23,000 --> 00:59:27,536
Audrow Nash: Mm hmm. Well,
that's really cool that you can

1032
00:59:27,536 --> 00:59:32,306
find markets that you can that
have lower variability in some

1033
00:59:32,306 --> 00:59:36,998
sense. And then, from that you
can go and you can add value,

1034
00:59:36,998 --> 00:59:41,142
you can improve your
foundational model, and then you

1035
00:59:41,142 --> 00:59:45,834
can go to less structured ones
that have higher variability.

1036
00:59:45,834 --> 00:59:50,448
And you can kind of keep going
until you can eat up all the

1037
00:59:50,448 --> 00:59:55,140
different tasks that have a lot
of variability. This kind of

1038
00:59:55,140 --> 00:59:56,000
thing. What

1039
00:59:56,000 --> 01:00:02,000
Amarjot Singh: Yeah, absolutely.

1040
01:00:02,000 --> 01:00:05,991
Audrow Nash: A thing. Okay, very
interesting. Now tell me about

1041
01:00:05,991 --> 01:00:09,982
kind of your hardware path for
everything because you guys, you

1042
01:00:09,982 --> 01:00:13,719
do drones, you do humanoids, you
do robot dogs, you do some

1043
01:00:13,719 --> 01:00:17,520
surveillance. Like it's quite
about quite a bit. Tell me how

1044
01:00:17,520 --> 01:00:21,385
it all fits together and kind of
how you went from one to the

1045
01:00:21,385 --> 01:00:25,249
next. and then I think humanoids
is probably your most recent

1046
01:00:25,249 --> 01:00:29,113
one. And I'd love to see how
that fits in kind of the company

1047
01:00:29,113 --> 01:00:30,000
vision to you.

1048
01:00:30,000 --> 01:00:33,573
Amarjot Singh: Yeah, so
initially we started with

1049
01:00:33,573 --> 01:00:37,875
markets where there was an ROI,
so that if we could deploy,

1050
01:00:37,875 --> 01:00:41,813
someone would pay for our
self-learning model. And the

1051
01:00:41,813 --> 01:00:46,260
sectors were any asset which is
moving very fast. You want to

1052
01:00:46,260 --> 01:00:50,198
pick targets which are very
small. And if you miss the

1053
01:00:50,198 --> 01:00:54,208
target or if you have a false
positive, you're going to

1054
01:00:54,208 --> 01:00:58,583
actually have a challenge. So we
deployed it to detect small

1055
01:00:58,583 --> 01:01:03,177
debris on tarmac, which can get
into fighter jets and can cause

1056
01:01:03,177 --> 01:01:05,000
damage to F-22 and F-35s.

1057
01:01:05,000 --> 01:01:08,000
Audrow Nash: Oh, hmm.

1058
01:01:08,000 --> 01:01:13,133
Amarjot Singh: So we have
actually this ground vehicle,

1059
01:01:13,133 --> 01:01:18,453
which you can see here, which
was running up and down the

1060
01:01:18,453 --> 01:01:22,000
tarmac. And it was, you can see
on the

1061
01:01:22,000 --> 01:01:23,000
Audrow Nash: Yeah, it's so cool.

1062
01:01:22,000 --> 01:01:24,824
Amarjot Singh: Left side. Yeah,
Yeah, that's what we started

1063
01:01:24,824 --> 01:01:26,000
with. So it's going to...

1064
01:01:26,000 --> 01:01:29,444
Audrow Nash: So just to describe
for listeners, there's a car,

1065
01:01:29,444 --> 01:01:32,778
it's kind of like a Waymo in
some sense where it has tons of

1066
01:01:32,778 --> 01:01:36,222
sensors on it, a bunch of like
security cameras looking things

1067
01:01:36,222 --> 01:01:39,333
on its front bumper. And then
there's a robot dog that's

1068
01:01:39,333 --> 01:01:42,556
running around with an arm
picking up the things that it's

1069
01:01:42,556 --> 01:01:44,000
identified. Is that right?

1070
01:01:44,000 --> 01:01:44,685
Amarjot Singh: Yeah, so when we
started, we were... d Yeah, the

1071
01:01:44,685 --> 01:01:45,380
cars were there, right? But the
intelligence wasn't. Because the

1072
01:01:45,380 --> 01:01:46,000
wind would blow up new debris on
a tarmac, that's FOD, or

1073
01:01:46,000 --> 01:02:00,000
Audrow Nash: It's awesome. Mm
hmm.

1074
01:02:00,000 --> 01:02:01,000
Amarjot Singh: You get a lot of
oil patches.

1075
01:02:01,000 --> 01:02:02,837
Audrow Nash: And tarmacs, are
they just runways? Like it's a

1076
01:02:02,837 --> 01:02:04,000
runway for planes? Is that what
it is?

1077
01:02:04,000 --> 01:02:06,089
Amarjot Singh: Yeah, absolutely.
It's a tarmac. It's wherever you

1078
01:02:06,089 --> 01:02:08,018
ha So this is on an aircraft
carrier. It's on a military air

1079
01:02:08,018 --> 01:02:10,043
tarmac, same stuff. And you have
these fighter jets, the engine

1080
01:02:10,043 --> 01:02:12,100
is close to the ground. So they
would suck in the debris. So the

1081
01:02:12,100 --> 01:02:13,900
first thing we did was, can we
give you AI, which is not

1082
01:02:13,900 --> 01:02:15,732
expensive, first of all, because
if you use a radar-based

1083
01:02:15,732 --> 01:02:17,629
solution, that is $2 million,
and not everyone wants to buy

1084
01:02:17,629 --> 01:02:19,621
that. So we were using cameras,
and the downside is it'll give

1085
01:02:19,621 --> 01:02:21,582
you false positive, so it's not
reliable. So that is what the

1086
01:02:21,582 --> 01:02:22,000
self-learning

1087
01:02:22,000 --> 01:02:40,000
Audrow Nash: Mm hmm. Oh, yeah.

1088
01:02:40,000 --> 01:02:42,000
Amarjot Singh: Initially
corrected.

1089
01:02:42,000 --> 01:02:43,000
Audrow Nash: Cool. Yeah.

1090
01:02:43,000 --> 01:02:46,954
Amarjot Singh: Yeah, and that's
what we started with. And now

1091
01:02:46,954 --> 01:02:50,972
some... But initially someone
had to drive that car. They will

1092
01:02:50,972 --> 01:02:54,926
get an alert, get out, and pick
that debris. But we wanted to

1093
01:02:54,926 --> 01:02:59,074
automate that, so now we have a
robotic dog, which uses the same

1094
01:02:59,074 --> 01:03:03,093
perception model, but also has
an arm, using which it can pick

1095
01:03:03,093 --> 01:03:04,000
the debris. So

1096
01:03:04,000 --> 01:03:05,000
Audrow Nash: Sweet.

1097
01:03:05,000 --> 01:03:06,148
Amarjot Singh: Automation. But
now when it is not doing... the

1098
01:03:06,148 --> 01:03:07,185
So now when it's not doing the
FOD, you can use it to do

1099
01:03:07,185 --> 01:03:08,296
perimeter security. When it is
not doing perimeter security,

1100
01:03:08,296 --> 01:03:09,370
yeah, it can go to other areas
where there is some sort of

1101
01:03:09,370 --> 01:03:10,537
critical infrastructure. There
are gauges. They're not digital.

1102
01:03:10,537 --> 01:03:11,574
You can go and read the gauge.
But even on the FOD side,

1103
01:03:11,574 --> 01:03:12,667
previously on the car you're
just doing the tarmac, but now

1104
01:03:12,667 --> 01:03:13,759
it can also go to the hangar
areas. It can go to the engine

1105
01:03:13,759 --> 01:03:14,852
testing bay. So whatever you're
doing on the FOD side, this

1106
01:03:14,852 --> 01:03:15,000
expands.

1107
01:03:15,000 --> 01:03:39,000
Audrow Nash: Nice.

1108
01:03:39,000 --> 01:03:44,078
Amarjot Singh: And whatever
you're doing on the mission set,

1109
01:03:44,078 --> 01:03:49,157
a single machine can do that.
And the reason why we are very

1110
01:03:49,157 --> 01:03:54,489
excited about the humanoid is,
humanoid is one machine based on

1111
01:03:54,489 --> 01:03:59,906
its degrees of freedom, has the
capacity to do many mission sets

1112
01:03:59,906 --> 01:04:05,154
together. It can go from riot
gear to defusing bombs to firing

1113
01:04:05,154 --> 01:04:06,000
weapons to

1114
01:04:06,000 --> 01:04:07,000
Audrow Nash: Anything people do.
Mm

1115
01:04:07,000 --> 01:04:10,210
Amarjot Singh: Everything people
do. So our bet is that once the

1116
01:04:10,210 --> 01:04:13,019
hardware matures, we would be
able to basically have one

1117
01:04:13,019 --> 01:04:16,079
machine which can do everything.
Now, on the hardware side of

1118
01:04:16,079 --> 01:04:19,039
things, wherever there is an
established player, whether it

1119
01:04:19,039 --> 01:04:21,998
is drones and dogs, and they
have significant market share,

1120
01:04:21,998 --> 01:04:25,008
we don't want to build that
hardware because we can't be the

1121
01:04:25,008 --> 01:04:28,018
category leader there. So we'll
give you our software. We'll

1122
01:04:28,018 --> 01:04:31,228
partner with you, increase your
revenue in the process also, get

1123
01:04:31,228 --> 01:04:33,987
money. But on humanoid side,
since there is no category

1124
01:04:33,987 --> 01:04:36,896
leader for the mission critical
side, that is where we are

1125
01:04:36,896 --> 01:04:38,000
building our humanoid,

1126
01:04:38,000 --> 01:04:50,000
Audrow Nash: Mm hmm.

1127
01:04:50,000 --> 01:04:54,411
Amarjot Singh: Which you can see
on the screen. That's a graphic.

1128
01:04:54,411 --> 01:04:58,346
But ideally, once it is built,
it should be waterproof. It

1129
01:04:58,346 --> 01:05:02,621
should have battery which does
not catch fire. But also you can

1130
01:05:02,621 --> 01:05:06,761
go from bipedal to quadruped
because we don't feel like biped

1131
01:05:06,761 --> 01:05:09,000
is the right design all the
time.

1132
01:05:09,000 --> 01:05:11,000
Audrow Nash: Warm for distance.
Mm hmm.

1133
01:05:11,000 --> 01:05:13,202
Amarjot Singh: Exactly. So you
want to put it on quadruped. So

1134
01:05:13,202 --> 01:05:15,475
a lot of our work is on building
this LEGO kind of hardware, but

1135
01:05:15,475 --> 01:05:17,713
then building control policies
so that if you take the top part

1136
01:05:17,713 --> 01:05:19,986
and put it onto the legs or legs
with wheels or treads, the same

1137
01:05:19,986 --> 01:05:22,260
control policy will work across
these humanoid form factors. And

1138
01:05:22,260 --> 01:05:24,000
that's the work we're doing on
the software side.

1139
01:05:24,000 --> 01:05:29,870
Audrow Nash: Yeah. Damn. That
the graphic just to describe for

1140
01:05:29,870 --> 01:05:35,456
people listening is it almost
looks like something out of a

1141
01:05:35,456 --> 01:05:41,231
sci-fi movie. It's a big black
robot humanoid that jumps over

1142
01:05:41,231 --> 01:05:46,817
something, goes into running
like a dog, and then shoots at

1143
01:05:46,817 --> 01:05:52,497
targets, which is just bonkers.
Super cool looking. no. It's

1144
01:05:52,497 --> 01:05:56,000
like it's even scarier than
those the

1145
01:05:56,000 --> 01:06:14,843
Amarjot Singh: You know, You
don't want this thing coming on

1146
01:06:14,843 --> 01:06:31,488
to you when you're in the
battlefield. Yes. Yes. Yes.

1147
01:06:31,488 --> 01:06:34,000
Exactly.

1148
01:06:34,000 --> 01:06:43,935
Audrow Nash: Mm hmm. Wow. So
what do you think? Yeah, go

1149
01:06:43,935 --> 01:06:45,000
ahead.

1150
01:06:45,000 --> 01:06:49,180
Amarjot Singh: We get it. No,
no, no. I'm sorry. We're getting

1151
01:06:49,180 --> 01:06:51,000
there. We're getting there.

1152
01:06:51,000 --> 01:06:55,794
Audrow Nash: That's bonkers.
wow. So where do you, I guess.

1153
01:06:55,794 --> 01:07:00,670
So what's your read on the rest
of the humanoid n companies?

1154
01:07:00,670 --> 01:07:05,708
Like you have, there's a whole
bunch of them. And I guess none

1155
01:07:05,708 --> 01:07:10,746
of them are really going after
this. There are a few, I think,

1156
01:07:10,746 --> 01:07:15,947
making military humanoids. but,
and I don't know, like, I assume

1157
01:07:15,947 --> 01:07:20,091
Optimus will try to be
weatherproof and things like

1158
01:07:20,091 --> 01:07:25,211
this. But, what's your, it's or
it's just too early of a market

1159
01:07:25,211 --> 01:07:30,167
to say that there will be any
leaders. So that's why you guys

1160
01:07:30,167 --> 01:07:34,881
are betting there as opposed to
like competing with DJI on

1161
01:07:34,881 --> 01:07:40,000
drones or whatever it might be.
Where they're well established.

1162
01:07:40,000 --> 01:07:41,694
Amarjot Singh: So I think so I
think the hardware market will

1163
01:07:41,694 --> 01:07:43,333
become commoditized, similar to
the drone market. You go on

1164
01:07:43,333 --> 01:07:45,083
LinkedIn, there's a new humanoid
coming out every other day. So

1165
01:07:45,083 --> 01:07:46,778
value will move to the brain.
And I think on our side, we are

1166
01:07:46,778 --> 01:07:47,000
building

1167
01:07:47,000 --> 01:08:02,000
Audrow Nash: Mm hmm. Yeah.

1168
01:08:02,000 --> 01:08:02,503
Amarjot Singh: The brain for
specific mission sets on

1169
01:08:02,503 --> 01:08:03,044
proprietary data on this
adaptable brain. So I think that

1170
01:08:03,044 --> 01:08:03,642
positions us well in the mission
critical space. And I think as

1171
01:08:03,642 --> 01:08:04,184
compar as far as the other
players are concerned, whether

1172
01:08:04,184 --> 01:08:04,715
it is Figure or Agility or
Tesla, I think they are doing

1173
01:08:04,715 --> 01:08:05,000
mainly homes or manufacturing.

1174
01:08:05,000 --> 01:08:25,000
Audrow Nash: Mm hmm. Mm

1175
01:08:25,000 --> 01:08:25,184
Amarjot Singh: I think it's
great, but I think the

1176
01:08:25,184 --> 01:08:25,415
variability in that space is so
much that it's very hard to get

1177
01:08:25,415 --> 01:08:25,651
reliability. That's one and then
get reliability to the standard

1178
01:08:25,651 --> 01:08:25,857
of your user. If I put a
humanoid in your home and it is

1179
01:08:25,857 --> 01:08:26,000
breaking the plate one out of 10
times.

1180
01:08:25,000 --> 01:08:44,000
Audrow Nash: Hmm. Or just falls
down.

1181
01:08:44,000 --> 01:08:45,025
Amarjot Singh: Yeah. And you're
pissed off, right? Because you

1182
01:08:45,025 --> 01:08:46,000
don't even at that point, 90
percent of liability. Exactly.

1183
01:08:46,000 --> 01:08:48,553
Audrow Nash: Mm hmm. I mean,
they're heavy. Yeah. Like a big

1184
01:08:48,553 --> 01:08:51,106
heavy heart or a metal robot
falling over in the house could

1185
01:08:51,106 --> 01:08:53,489
damage a lot. Like there's real
risk in that. Or step on

1186
01:08:53,489 --> 01:08:56,000
someone's toes or I don't know,
squish our dog or whatever.

1187
01:08:56,000 --> 01:08:59,695
Amarjot Singh: Exactly. Yeah,
exactly. And also depends on how

1188
01:08:59,695 --> 01:09:03,152
much you pay, right? Because in
our case, if we get like a

1189
01:09:03,152 --> 01:09:05,000
Boston Dynamics dog, it's 200K.

1190
01:09:05,000 --> 01:09:11,000
Audrow Nash: Mm hmm. Mm hmm.

1191
01:09:11,000 --> 01:09:13,932
Amarjot Singh: If you get a
humanoid, we want to sell it

1192
01:09:13,932 --> 01:09:17,283
for, let's say, maybe double of
that or 300 or 350. But then you

1193
01:09:17,283 --> 01:09:20,319
can justify the ROI, right?
Because it can do many mission

1194
01:09:20,319 --> 01:09:23,408
sets. And you have to train
people for all of those mission

1195
01:09:23,408 --> 01:09:26,654
sets. You can justify the ROI.
What are the kind of things you

1196
01:09:26,654 --> 01:09:29,953
are making a humanoid do in the
house, which is really going to

1197
01:09:29,953 --> 01:09:31,000
get to the ROI? Even

1198
01:09:31,000 --> 01:09:33,000
Audrow Nash: Yeah. I mean,

1199
01:09:33,000 --> 01:09:35,000
Amarjot Singh: Manufacturing.

1200
01:09:35,000 --> 01:09:37,524
Audrow Nash: Manufacturing maybe
because bottlenecks in

1201
01:09:37,524 --> 01:09:40,278
manufacturing really like, I
don't know, have you, I read an

1202
01:09:40,278 --> 01:09:43,216
excellent book that talks about
how bottlenecks are unbelievably

1203
01:09:43,216 --> 01:09:45,786
like they dictate the pace that
you can actually produce

1204
01:09:45,786 --> 01:09:48,264
products. And so if you can
automate a bottleneck in a

1205
01:09:48,264 --> 01:09:51,156
manufacturing process, it could
be incredibly valuable. or just

1206
01:09:51,156 --> 01:09:54,001
make it so you can go. Or data
centers. Yeah, maybe. But yeah,

1207
01:09:54,001 --> 01:09:56,801
in the house, I mean, you're
competing to the bottom for this

1208
01:09:56,801 --> 01:09:59,555
kind of thing for the price
because like I would if we could

1209
01:09:59,555 --> 01:10:02,263
buy a 20K humanoid that would
just clean up our house. It's

1210
01:10:02,263 --> 01:10:05,246
like that's an easy decision. We
would for sure do it. But that's

1211
01:10:05,246 --> 01:10:08,000
a much lower thing than a 400K
robot for this kind of thing.

1212
01:10:08,000 --> 01:10:10,799
Amarjot Singh: Yeah. Yes. I
think 20K is okay, but process a

1213
01:10:10,799 --> 01:10:13,690
function of your acceptance. I
think there are some people who

1214
01:10:13,690 --> 01:10:16,722
are early technology adopters. I
buy a lot of things even if they

1215
01:10:16,722 --> 01:10:19,474
work 50 percent, right? So I
like that. But I think for the

1216
01:10:19,474 --> 01:10:22,179
average user, the market is,
look, everyone is going to be

1217
01:10:22,179 --> 01:10:24,978
using a humanoid. And I think
that's good. But then you need

1218
01:10:24,978 --> 01:10:27,776
to get the product to a point
where if you say it's going to

1219
01:10:27,776 --> 01:10:30,668
do A, it's going to do A. And I
think that is where it's going

1220
01:10:30,668 --> 01:10:33,000
to be a little bit of a time
before it gets there.

1221
01:10:33,000 --> 01:10:40,274
Audrow Nash: Yeah, you do. Yep.
Mm hmm. So it seems to me like

1222
01:10:40,274 --> 01:10:47,314
military because they're so
mission critical and the cost is

1223
01:10:47,314 --> 01:10:54,236
so high of putting humans at
risk that it's actually a very

1224
01:10:54,236 --> 01:11:01,511
good place for humanoids because
even if they're not perfectly

1225
01:11:01,511 --> 01:11:08,903
reliable, you have the ability
for them to do a dangerous task.

1226
01:11:08,903 --> 01:11:15,942
And so it seems to me like this
is a very good bootstrapping

1227
01:11:15,942 --> 01:11:23,451
humanoid approach. And then once
you do this, then you move into

1228
01:11:23,451 --> 01:11:30,726
manufacturing or like even home
servicing company like service

1229
01:11:30,726 --> 01:11:36,475
companies like HVAC and stuff
like this, like air

1230
01:11:36,475 --> 01:11:38,000
conditioning.

1231
01:11:38,000 --> 01:11:42,631
Amarjot Singh: I think for us,
we try to stay, whether when we

1232
01:11:42,631 --> 01:11:47,112
did drones or dogs or cars, our
objective has always been in

1233
01:11:47,112 --> 01:11:51,369
places where they are out,
remote, expensive to maintain.

1234
01:11:51,369 --> 01:11:56,000
And if you don't do that, the
outcome can be very significant.

1235
01:11:56,000 --> 01:11:57,000
Audrow Nash: Bad.

1236
01:11:57,000 --> 01:12:00,249
Amarjot Singh: Yeah. So I think
like HVAC and all that stuff are

1237
01:12:00,249 --> 01:12:03,447
there. But I think as the market
evolves, there will be company

1238
01:12:03,447 --> 01:12:06,645
who would start to capture this
thing from the top side and the

1239
01:12:06,645 --> 01:12:09,893
bottom side, whether it is your
home or manufacturing or HVAC or

1240
01:12:09,893 --> 01:12:12,838
all those kind of things. And
one more thing which is very

1241
01:12:12,838 --> 01:12:15,985
important for us is barriers to
entry. So you cannot use a lot

1242
01:12:15,985 --> 01:12:17,000
of anything Chinese.

1243
01:12:17,000 --> 01:12:23,000
Audrow Nash: Mm hmm.

1244
01:12:23,000 --> 01:12:24,818
Amarjot Singh: So and a lot of
the humanoid market sits over

1245
01:12:24,818 --> 01:12:25,000
there.

1246
01:12:25,000 --> 01:12:27,000
Audrow Nash: Oh, yeah.

1247
01:12:27,000 --> 01:12:28,000
Amarjot Singh: Yeah. So for us,

1248
01:12:28,000 --> 01:12:29,891
Audrow Nash: So that's a big
barrier to entry because like

1249
01:12:29,891 --> 01:12:31,848
the what's the company Unitree
making humanoids I guess like

1250
01:12:31,848 --> 01:12:33,804
their spot equivalent is way
cheaper. I think their humanoid

1251
01:12:33,804 --> 01:12:35,696
can be pretty expensive still
but like if you get it fully

1252
01:12:35,696 --> 01:12:37,000
featured with good motors and
stuff. But

1253
01:12:37,000 --> 01:12:39,820
Amarjot Singh: Lot of our. Yeah,
exactly. We. Yeah, so we have a

1254
01:12:39,820 --> 01:12:42,597
lot of Unitrees. We have a lot
of Boston Dynamics. So a Unitree

1255
01:12:42,597 --> 01:12:45,329
would be like 80 K versus 300 K.
But I think even if you don't

1256
01:12:45,329 --> 01:12:48,061
worry about the price, let's say
the military can pay whatever

1257
01:12:48,061 --> 01:12:50,000
they don't want to use Chinese
equipment. So

1258
01:12:50,000 --> 01:13:00,633
Audrow Nash: Yeah. Mm hmm. Mm
hmm. Chinese. Yeah, it's a

1259
01:13:00,633 --> 01:13:05,000
security risk for them.

1260
01:13:05,000 --> 01:13:08,830
Amarjot Singh: Exactly. But that
barrier is not in other sectors.

1261
01:13:08,830 --> 01:13:12,365
So anyone eventually and the
Chinese machines are getting so

1262
01:13:12,365 --> 01:13:15,901
good. What will happen is in the
next one year or so, you'll

1263
01:13:15,901 --> 01:13:19,259
start to get these Chinese
machines, humanoids with these

1264
01:13:19,259 --> 01:13:22,971
kind of mission sets. And then
people who really like the field

1265
01:13:22,971 --> 01:13:26,801
will start to import from there,
modify it, and start to service.

1266
01:13:26,801 --> 01:13:30,631
So the idea, if you have built a
foundational model with a lot of

1267
01:13:30,631 --> 01:13:33,872
costs, a lot of money, it's
really nice. And it may not

1268
01:13:33,872 --> 01:13:37,584
really be that attractive to a
user because at that point price

1269
01:13:37,584 --> 01:13:40,000
will start to play a role. And
if you're.

1270
01:13:40,000 --> 01:13:42,000
Audrow Nash: The deciding role.

1271
01:13:42,000 --> 01:13:44,634
Amarjot Singh: Yeah, and if the
reliability is more or less

1272
01:13:44,634 --> 01:13:47,000
similar for both of them. And
that's another problem.

1273
01:13:47,000 --> 01:13:51,253
Audrow Nash: Interesting. So
you're you've positioned

1274
01:13:51,253 --> 01:13:56,227
yourself very well, where if say
China produces a ton of cheap

1275
01:13:56,227 --> 01:14:01,363
humanoids that are high quality,
they're still not going to want

1276
01:14:01,363 --> 01:14:06,257
to be used by the government and
therefore you're still in an

1277
01:14:06,257 --> 01:14:10,911
excellent place for this. Mm
hmm. Seems very smart. I feel

1278
01:14:10,911 --> 01:14:15,805
like just from talking through
and kind of seeing some of the

1279
01:14:15,805 --> 01:14:20,299
progression of your business
that it has been very smart

1280
01:14:20,299 --> 01:14:24,952
strategic like running like a
startup into different areas

1281
01:14:24,952 --> 01:14:29,285
that make good sense and kind of
growing your robotics

1282
01:14:29,285 --> 01:14:33,939
capabilities and positioning
yourself well just in general

1283
01:14:33,939 --> 01:14:38,753
for like long term outcomes.
huh. Hell yeah. Can you tell me

1284
01:14:38,753 --> 01:14:43,889
actually a bit about the journey
because I mean doing the tarmac

1285
01:14:43,889 --> 01:14:48,863
and things like that's wildly
different than I guess you still

1286
01:14:48,863 --> 01:14:53,838
had to do perception on that so
you could identify objects and

1287
01:14:53,838 --> 01:14:58,733
maybe so there is kind of like a
common core to that but also

1288
01:14:58,733 --> 01:15:03,788
doing drones to help military so
if they do a kamikaze drone or

1289
01:15:03,788 --> 01:15:07,960
something like this but then
doing humanoids for the

1290
01:15:07,960 --> 01:15:12,774
military. Application and then
it seems like from that video

1291
01:15:12,774 --> 01:15:16,786
that you shared with the
humanoid running over and

1292
01:15:16,786 --> 01:15:21,199
jumping the thing and then
shooting the gun and running

1293
01:15:21,199 --> 01:15:26,174
away like a dog and just like
bonkers. Seems like it's is kind

1294
01:15:26,174 --> 01:15:30,828
of the goal you think to make
like a super soldier or just

1295
01:15:30,828 --> 01:15:35,000
like a very capable robot for
military applications.

1296
01:15:35,000 --> 01:15:37,262
Amarjot Singh: So I think like
the evolution from the drones to

1297
01:15:37,262 --> 01:15:39,201
the cars is relatively easy
because we trained on very

1298
01:15:39,201 --> 01:15:41,499
minimal base data. you're gonna
get error anyways and the memory

1299
01:15:41,499 --> 01:15:43,402
can learn that. So that was
relatively easier because

1300
01:15:43,402 --> 01:15:45,557
otherwise I would have to go
through many cycles of updating

1301
01:15:45,557 --> 01:15:47,783
again and again. And the reason
why we did different use cases

1302
01:15:47,783 --> 01:15:49,937
was we didn't really know what
would be useful and where the

1303
01:15:49,937 --> 01:15:52,199
value chain would be. Everything
from finding suspicious people

1304
01:15:52,199 --> 01:15:54,282
moving in front of a jewelry
store so that eventually they

1305
01:15:54,282 --> 01:15:55,000
will break in. Where

1306
01:15:55,000 --> 01:16:08,000
Audrow Nash: Mm hmm. Wow. Yeah.

1307
01:16:08,000 --> 01:16:13,870
Amarjot Singh: No one will pay
us, to the tarmacs. So it was

1308
01:16:13,870 --> 01:16:17,000
just also exploration for us and

1309
01:16:17,000 --> 01:16:18,000
Audrow Nash: Cool.

1310
01:16:18,000 --> 01:16:19,000
Amarjot Singh: And the sector.

1311
01:16:19,000 --> 01:16:21,344
Audrow Nash: You basically
you've been trying to make value

1312
01:16:21,344 --> 01:16:23,291
in all these different
industries and seeing what

1313
01:16:23,291 --> 01:16:25,000
sticks and trying lots of
different things.

1314
01:16:25,000 --> 01:16:26,487
Amarjot Singh: Who would pay us?
Where is the actual life?

1315
01:16:26,487 --> 01:16:27,000
Exactly. And pay us.

1316
01:16:27,000 --> 01:16:29,294
Audrow Nash: Yeah who would pay
you? And also because like

1317
01:16:29,294 --> 01:16:31,706
because you also need to figure
out the go to market strategy

1318
01:16:31,706 --> 01:16:34,079
and everything and then once
you're in with the military and

1319
01:16:34,079 --> 01:16:36,373
they're using you I assume
that's a huge moat because it's

1320
01:16:36,373 --> 01:16:38,706
probably slow sales cycles and
you have to know someone and

1321
01:16:38,706 --> 01:16:41,000
it's just like it's a hard thing
for other startups to do.

1322
01:16:41,000 --> 01:16:43,316
Amarjot Singh: Yeah. But also a
hard thing, not from just the

1323
01:16:43,316 --> 01:16:45,519
GTM, but also the technology
side. Not anyone can come and

1324
01:16:45,519 --> 01:16:47,911
say, hey, I can also detect ABCD
on a camera. If you have to go

1325
01:16:47,911 --> 01:16:50,152
to the stationary asset, which
is the CCTV versus drones or

1326
01:16:50,152 --> 01:16:52,392
cars, you will get probably
hundreds of companies which can

1327
01:16:52,392 --> 01:16:53,000
make that claim.

1328
01:16:53,000 --> 01:17:09,000
Audrow Nash: Mm hmm. Mm hmm.

1329
01:17:09,000 --> 01:17:12,942
Amarjot Singh: When it comes to
putting AI, which can work

1330
01:17:12,942 --> 01:17:16,951
sensibly on a drone or a car,
there would be maybe handful.

1331
01:17:16,951 --> 01:17:21,233
And that also is a big decision
because I want to be in a space

1332
01:17:21,233 --> 01:17:24,903
which is growing. There is
active users. There is very

1333
01:17:24,903 --> 01:17:28,981
clear defined ROI, but less
competition so you don't have to

1334
01:17:28,981 --> 01:17:30,000
fight on price.

1335
01:17:30,000 --> 01:17:32,591
Audrow Nash: Hell yeah. Yeah and
then you can invest and then you

1336
01:17:32,591 --> 01:17:34,904
can keep doing awesome things
like you can keep playing it

1337
01:17:34,904 --> 01:17:37,375
forward for the future which I
think is what you're doing with

1338
01:17:37,375 --> 01:17:39,887
humanoids now like this is this
is a new process for that where

1339
01:17:39,887 --> 01:17:42,000
that's a big future bet because
the market's growing.

1340
01:17:42,000 --> 01:17:47,000
Amarjot Singh: Yeah. Yeah.

1341
01:17:47,000 --> 01:17:48,477
Audrow Nash: The world's chaotic
and the technology is just about

1342
01:17:48,477 --> 01:17:49,000
there is what it seems.

1343
01:17:49,000 --> 01:17:51,916
Amarjot Singh: Yeah. Yeah. Yeah.
So my approach is we were doing

1344
01:17:51,916 --> 01:17:54,787
it on the car. How do we fully
automate that? So we went to the

1345
01:17:54,787 --> 01:17:57,429
dog and then when we are going
to the dog, we also want to

1346
01:17:57,429 --> 01:17:59,571
improve its mission set
capability. So it's not

1347
01:17:59,571 --> 01:18:02,305
specialized. It can do multiple
things. And I feel like most

1348
01:18:02,305 --> 01:18:05,084
capable machine which can do
multiple things is the humanoid.

1349
01:18:05,084 --> 01:18:07,681
And that's the reason why we're
pushing more towards that

1350
01:18:07,681 --> 01:18:08,000
domain.

1351
01:18:08,000 --> 01:18:11,565
Audrow Nash: Mm hmm. Mm hmm. Mm
hmm. Yeah because then you can

1352
01:18:11,565 --> 01:18:14,901
go adjacent and that helps
justify the cost for everything

1353
01:18:14,901 --> 01:18:18,639
because it can be a soldier it's
like your robot dog which can do

1354
01:18:18,639 --> 01:18:22,319
picking up stuff but it can also
do surveillance and then it can

1355
01:18:22,319 --> 01:18:26,000
also do some other tasks but the
humanoid will be a more extreme

1356
01:18:26,000 --> 01:18:37,000
Amarjot Singh: Is that. Yeah.

1357
01:18:37,000 --> 01:18:42,112
Audrow Nash: Version of that.
And so, hell yeah. What do you

1358
01:18:42,112 --> 01:18:47,224
think you showed me that you're
building a humanoid? Tell me

1359
01:18:47,224 --> 01:18:52,507
about where you are for that and
how that turns into the crazy

1360
01:18:52,507 --> 01:18:56,000
machine that you showed me in
that video.

1361
01:18:56,000 --> 01:18:59,482
Amarjot Singh: Yeah. I think the
two dimensions to the building

1362
01:18:59,482 --> 01:19:02,632
one is the body and the second
are the hands. Well, power

1363
01:19:02,632 --> 01:19:06,224
system is another one. So we are
working with a couple of vendors

1364
01:19:06,224 --> 01:19:09,484
in Korea. One who can build us
the battery packs the way we

1365
01:19:09,484 --> 01:19:12,855
want. One specializes in hands
which are not motor-driven but

1366
01:19:12,855 --> 01:19:16,447
tendon-driven so you can put the
motors on the exactly so you can

1367
01:19:16,447 --> 01:19:17,000
pull more.

1368
01:19:17,000 --> 01:19:20,042
Audrow Nash: Cool. Motors are
somewhere away. And that

1369
01:19:20,042 --> 01:19:23,592
probably is better for running
like your bipedal running so you

1370
01:19:23,592 --> 01:19:27,197
don't have heavy weights at the
end of the end effectors or that

1371
01:19:27,197 --> 01:19:29,000
end of the arms I guess. Mm hmm.

1372
01:19:29,000 --> 01:19:33,340
Amarjot Singh: Exactly. Exactly.
So we are working on all of

1373
01:19:33,340 --> 01:19:37,753
this. I think it will take us
probably six to eight months to

1374
01:19:37,753 --> 01:19:42,383
get the V1 out. But I think once
we have that then we are off to

1375
01:19:42,383 --> 01:19:46,000
the races because that will make
a big difference.

1376
01:19:46,000 --> 01:19:49,257
Audrow Nash: Mm hmm. What are
some challenges that you like

1377
01:19:49,257 --> 01:19:52,349
what are some of the most
challenging things with this I

1378
01:19:52,349 --> 01:19:55,440
guess I my suspicion is
manufacturing for any scale will

1379
01:19:55,440 --> 01:19:58,532
be a huge challenge for this.
But like are parts is part

1380
01:19:58,532 --> 01:20:01,678
sourcing difficult are is
designing for manufacturability

1381
01:20:01,678 --> 01:20:04,825
a big challenge I guess because
you're going for high end

1382
01:20:04,825 --> 01:20:08,192
applications with cross usage
you don't have to push down the

1383
01:20:08,192 --> 01:20:11,229
price as much as like people
trying to put humanoids in

1384
01:20:11,229 --> 01:20:14,265
people's homes. But and then I
wonder also from like an

1385
01:20:14,265 --> 01:20:17,577
imitation learning and control
and I don't know just general

1386
01:20:17,577 --> 01:20:21,000
what are what are some of the
challenges of making a humanoid.

1387
01:20:21,000 --> 01:20:27,099
Amarjot Singh: Yeah. Well the
first thing is as you said

1388
01:20:27,099 --> 01:20:33,960
manufacturing. You don't have a
lot of manufacturing capability

1389
01:20:33,960 --> 01:20:40,604
in the U.S. for humanoids right
now. The ones which is there.

1390
01:20:40,604 --> 01:20:43,000
The one that is there.

1391
01:20:43,000 --> 01:20:45,000
Audrow Nash: What's missing?
Yeah, yeah.

1392
01:20:45,000 --> 01:20:47,775
Amarjot Singh: So I think like
for example you have a humanoid,

1393
01:20:47,775 --> 01:20:50,507
you run, you run wires through
the humanoids pin connectors. A

1394
01:20:50,507 --> 01:20:53,106
lot of that knowledge as to how
do you do that is very very

1395
01:20:53,106 --> 01:20:55,000
specialized. it Most of that
sits in China.

1396
01:20:55,000 --> 01:21:06,000
Audrow Nash: Mm Hmm. Mm hmm.

1397
01:21:06,000 --> 01:21:07,612
Amarjot Singh: Milling. You want
to mill the chassis. The

1398
01:21:07,612 --> 01:21:09,308
actuators more or less. They
have rare earth metals in that.

1399
01:21:09,308 --> 01:21:11,062
A lot of that is controlled by
China which comes from a lot of

1400
01:21:11,062 --> 01:21:12,730
that southern America. So if you
and everything is built on

1401
01:21:12,730 --> 01:21:14,427
actuators if you don't have
actuators that doesn't work. And

1402
01:21:14,427 --> 01:21:16,123
we are starting to build some
but most of them is in Asia or

1403
01:21:16,123 --> 01:21:17,000
China. So that's the first one.

1404
01:21:17,000 --> 01:21:23,548
Audrow Nash: Yeah. Mm hmm. Yep.
And China does all those rare

1405
01:21:23,548 --> 01:21:30,311
earth processing to for I think
magnets in general like I think

1406
01:21:30,311 --> 01:21:36,000
they have a huge amount of the
market share for that.

1407
01:21:36,000 --> 01:21:37,237
Amarjot Singh: Exactly. So even
if you don't want to buy the

1408
01:21:37,237 --> 01:21:38,474
whole Chinese humanoid you are
more or less somewhere buying

1409
01:21:38,474 --> 01:21:39,753
Chinese. Exactly. Then the other
thing would be the cost. If I

1410
01:21:39,753 --> 01:21:40,928
want to build something over
here it'll cost me X. If I'm

1411
01:21:40,928 --> 01:21:42,227
doing prototyping I have to do
five prototypes which eventually

1412
01:21:42,227 --> 01:21:43,443
the last one's going to work. So
my cost of doing that here

1413
01:21:43,443 --> 01:21:44,000
versus offshore is extreme.

1414
01:21:44,000 --> 01:22:04,000
Audrow Nash: Sourcing through
them. Mm hmm.

1415
01:22:04,000 --> 01:22:06,161
Amarjot Singh: And that's the
thing. And I don't even know if

1416
01:22:06,161 --> 01:22:08,323
I put that money if I would be
able to actually get that done

1417
01:22:08,323 --> 01:22:10,520
because that scale is not tried
and tested at scale while they

1418
01:22:10,520 --> 01:22:12,504
have that. So a bunch of those
kind of challenges on the

1419
01:22:12,504 --> 01:22:13,000
hardware side.

1420
01:22:13,000 --> 01:22:17,900
Audrow Nash: Mm hmm. Mm hmm.
What do you think so you have 20

1421
01:22:17,900 --> 01:22:22,640
people in the US and then you
have seven 50 people or so in

1422
01:22:22,640 --> 01:22:26,817
India. How do you leverage that?
I think in a lot of

1423
01:22:26,817 --> 01:22:31,396
manufacturing seemed to be
moving from China to India. Do

1424
01:22:31,396 --> 01:22:35,814
you think that will play a
strategic role for you where

1425
01:22:35,814 --> 01:22:40,313
you'll do some I don't I don't
know much about India for

1426
01:22:40,313 --> 01:22:42,000
manufacturing things.

1427
01:22:42,000 --> 01:22:43,000
Amarjot Singh: So India

1428
01:22:43,000 --> 01:22:45,000
Audrow Nash: But.

1429
01:22:45,000 --> 01:22:48,953
Amarjot Singh: India is very
strategic for us because of two

1430
01:22:48,953 --> 01:22:53,105
things. One data for robotics.
So a lot of this real world data

1431
01:22:53,105 --> 01:22:56,926
real world testing which we do
deploying teams operations,

1432
01:22:56,926 --> 01:23:00,946
that's way more economical there
as compared to India. That's

1433
01:23:00,946 --> 01:23:04,965
one. The second thing which is
very important which is a very

1434
01:23:04,965 --> 01:23:08,523
very very big bottleneck is
talent. If I have to get a

1435
01:23:08,523 --> 01:23:12,543
robotics engineer in the Bay
Area very hard to do and we have

1436
01:23:12,543 --> 01:23:16,364
people but then they also want
to talk to Figure they also

1437
01:23:16,364 --> 01:23:19,000
talking to Tesla. A lot of
options here.

1438
01:23:19,000 --> 01:23:22,000
Audrow Nash: Oh yeah there's so
much competition. Mm hmm.

1439
01:23:22,000 --> 01:23:24,218
Amarjot Singh: Yeah if I go to
New York relatively less if I go

1440
01:23:24,218 --> 01:23:26,190
to India a lot of very talented
people but not that much

1441
01:23:26,190 --> 01:23:28,092
competition. So we also use
India as our talent hub to

1442
01:23:28,092 --> 01:23:30,239
retrain the policies test these
things do breaching there all

1443
01:23:30,239 --> 01:23:32,000
that kind of stuff. Get data and
also manufacture.

1444
01:23:32,000 --> 01:23:36,920
Audrow Nash: Hmm. Cool. Huh. Do
you think that how from your

1445
01:23:36,920 --> 01:23:41,923
perspective how is AI changing
all this like I wonder I don't

1446
01:23:41,923 --> 01:23:46,843
know like if you have say you
say if you go hire an engineer

1447
01:23:46,843 --> 01:23:51,927
in Silicon Valley. They're great
and bring a lot of knowledge.

1448
01:23:51,927 --> 01:23:56,683
But if you go to India and you
hire someone who's probably

1449
01:23:56,683 --> 01:24:01,604
quite good too but maybe doesn't
have quite the education or

1450
01:24:01,604 --> 01:24:06,770
maybe does. But if they use AI
it's probably a bit of a leveler

1451
01:24:06,770 --> 01:24:08,000
would you think

1452
01:24:08,000 --> 01:24:13,000
Amarjot Singh: Sure.

1453
01:24:13,000 --> 01:24:16,026
Audrow Nash: Like I think you
could probably leverage it to

1454
01:24:16,026 --> 01:24:19,000
make your teams even smarter or
how do you think about it.

1455
01:24:19,000 --> 01:24:20,671
Amarjot Singh: We are some most
of the specialized end to end

1456
01:24:20,671 --> 01:24:22,370
autonomy knowledge, VLAs and all
that stuff is in the US Menlo

1457
01:24:22,370 --> 01:24:23,767
Park. We fund university
professors at UC Berkeley,

1458
01:24:23,767 --> 01:24:25,274
Carnegie Mellon. So a lot of our
whole body control and

1459
01:24:25,274 --> 01:24:27,000
manipulation work happens with
them through their PhD students.

1460
01:24:27,000 --> 01:24:28,644
But once that we get the base
structure and you want to test

1461
01:24:28,644 --> 01:24:30,315
it for n kind of applications
you want to tweak the different

1462
01:24:30,315 --> 01:24:32,068
kind of policies. We send all of
that back to India and they can

1463
01:24:32,068 --> 01:24:33,712
then pick that up and they can
improve on that. A lot of the

1464
01:24:33,712 --> 01:24:35,137
people there we have used to
work for companies like

1465
01:24:35,137 --> 01:24:36,918
Dexterity or something like that
offshore and they were doing the

1466
01:24:36,918 --> 01:24:38,534
same thing for them. So we have
brought those people on. So

1467
01:24:38,534 --> 01:24:39,000
refining updating

1468
01:24:39,000 --> 01:25:05,000
Audrow Nash: Nice. Mm hmm. Cool.

1469
01:25:05,000 --> 01:25:08,743
Amarjot Singh: All of that we do
there and obviously Claude helps

1470
01:25:08,743 --> 01:25:11,853
because you don't have to
reinvent the wheel. You just

1471
01:25:11,853 --> 01:25:15,309
need to know how to think and
that we can either do directly

1472
01:25:15,309 --> 01:25:16,000
through the.

1473
01:25:16,000 --> 01:25:17,701
Audrow Nash: You just need to
know half the what. Sorry I

1474
01:25:17,701 --> 01:25:18,000
missed it.

1475
01:25:18,000 --> 01:25:19,338
Amarjot Singh: Just they just
have to think they don't need to

1476
01:25:19,338 --> 01:25:20,612
know coding and all that stuff
which Claude Code has helped

1477
01:25:20,612 --> 01:25:21,000
with. quite a bit.

1478
01:25:21,000 --> 01:25:26,652
Audrow Nash: Oh yeah. Yeah it's
such a wild world for everything

1479
01:25:26,652 --> 01:25:32,126
with all the AI going where it's
going. What do you I guess so

1480
01:25:32,126 --> 01:25:37,778
we're starting to wrap up where
do you see Skylark going like so

1481
01:25:37,778 --> 01:25:43,341
I think you were thinking that
humanoid robot in three years or

1482
01:25:43,341 --> 01:25:48,463
so for everything but what do
you see kind of as the next.

1483
01:25:48,463 --> 01:25:54,026
Three to five years for you all
like what's where were you guys

1484
01:25:54,026 --> 01:25:58,000
headed what's the robotics
climate look like.

1485
01:25:58,000 --> 01:26:02,158
Amarjot Singh: Yeah. Yeah. So
short term we are focused

1486
01:26:02,158 --> 01:26:06,618
heavily on where we get the
money where the business is. So

1487
01:26:06,618 --> 01:26:11,003
we have drones which are doing
primarily detection. So for

1488
01:26:11,003 --> 01:26:15,841
example we are detecting cracks
on a solar panel to cracks on an

1489
01:26:15,841 --> 01:26:20,151
insulator. And now we have
drones which have arms. So the

1490
01:26:20,151 --> 01:26:24,687
drone is also going and cleaning
the solar panel so that the

1491
01:26:24,687 --> 01:26:29,449
efficiency of the solar panel is
there to going and sealing the

1492
01:26:29,449 --> 01:26:34,212
crack on the insulator. So a lot
of our customers are ingesting

1493
01:26:34,212 --> 01:26:37,538
this aerial robotics
application. So that is

1494
01:26:37,538 --> 01:26:42,150
immediate extension for us. Then
same goes on the ground side

1495
01:26:42,150 --> 01:26:46,837
where we were using cars and now
using robotic dogs. So that's

1496
01:26:46,837 --> 01:26:51,524
the other dimension which we are
scaling through our own sales

1497
01:26:51,524 --> 01:26:56,363
through our new engineering new
model class that's that. Then on

1498
01:26:56,363 --> 01:27:00,445
the humanoid side we are
building the brain because we

1499
01:27:00,445 --> 01:27:05,208
feel like there is there will be
significant commoditization on

1500
01:27:05,208 --> 01:27:09,517
the hardware side and we are
starting to build our own as

1501
01:27:09,517 --> 01:27:14,204
well which we're going to push
out in the next eight months or

1502
01:27:14,204 --> 01:27:19,042
10 months. But currently we are
working with partners from Korea

1503
01:27:19,042 --> 01:27:23,503
like Robotis, using and through
whom we are going and still

1504
01:27:23,503 --> 01:27:28,114
showing this brain to the end
customer. So we are still doing

1505
01:27:28,114 --> 01:27:32,877
the data flywheel still showing
still improving on the humanoid

1506
01:27:32,877 --> 01:27:37,715
side while we build our own with
hardware which is acceptable so

1507
01:27:37,715 --> 01:27:42,251
that we don't have to bear the
cost. That's one. And then on

1508
01:27:42,251 --> 01:27:46,939
the side where the business is
there air and ground over there

1509
01:27:46,939 --> 01:27:51,550
we are expanding our current
business expanding those product

1510
01:27:51,550 --> 01:27:56,313
suites with aerial robotics or
ground robotics and then scaling

1511
01:27:56,313 --> 01:28:01,000
that. So that's how we see this
short term and then long term.

1512
01:28:01,000 --> 01:28:05,089
Audrow Nash: Hell yeah very
exciting. What do you think to

1513
01:28:05,089 --> 01:28:09,389
me it seems like you're running
effectively a few startups at

1514
01:28:09,389 --> 01:28:13,548
once like all these different
areas and I know that there's

1515
01:28:13,548 --> 01:28:17,918
kind of a common core to it but
it still seems like completely

1516
01:28:17,918 --> 01:28:22,148
different markets completely
different products. These kinds

1517
01:28:22,148 --> 01:28:26,448
of things what advice do you
have for I don't know people who

1518
01:28:26,448 --> 01:28:29,690
want to be robotics
entrepreneurs want to do a

1519
01:28:29,690 --> 01:28:33,779
robotic startup seems like
you're chasing the money really

1520
01:28:33,779 --> 01:28:38,220
good finding good market fit and
then deploying but what advice

1521
01:28:38,220 --> 01:28:42,661
what things would have saved you
a bit of learning. If you were

1522
01:28:42,661 --> 01:28:44,000
to do it all again.

1523
01:28:44,000 --> 01:28:48,250
Amarjot Singh: Yeah. Yeah I
think it's a tricky market

1524
01:28:48,250 --> 01:28:53,051
because if you're trying to I
mean the reason why we are able

1525
01:28:53,051 --> 01:28:57,302
to do multiple markets is
because we don't really sell

1526
01:28:57,302 --> 01:29:01,867
anything directly. The go to
market motion is with the end

1527
01:29:01,867 --> 01:29:06,668
customer on whose hardware we
are overlaying. So partnerships

1528
01:29:06,668 --> 01:29:11,233
I think big play a big role in
offsetting the price of the

1529
01:29:11,233 --> 01:29:15,247
hardware or the data or
operations. I think that is

1530
01:29:15,247 --> 01:29:19,970
something which everyone should
try to leverage. But I think

1531
01:29:19,970 --> 01:29:25,007
then finding sectors where your
solution no matter how effective

1532
01:29:25,007 --> 01:29:29,887
it is able to solve a very big
pain problem. Which they cannot

1533
01:29:29,887 --> 01:29:34,609
live without. And then they will
buy things even if it is 70

1534
01:29:34,609 --> 01:29:39,411
percent efficient or 80 percent
efficient. So I think finding

1535
01:29:39,411 --> 01:29:43,661
something which really where you
can get to delivering

1536
01:29:43,661 --> 01:29:48,226
reliability whatever their
definition is as quickly as you

1537
01:29:48,226 --> 01:29:53,185
can. I think that's the next the
next big thing because I think

1538
01:29:53,185 --> 01:29:57,986
robotics kind of works and kind
of is dependent on who you're

1539
01:29:57,986 --> 01:30:02,000
showing it to. So you get a
selective over there. I

1540
01:30:02,000 --> 01:30:06,889
Audrow Nash: Hell yeah. Where do
you think robotics is going like

1541
01:30:06,889 --> 01:30:11,553
if anyone wanted to skate to the
puck in some sense like where

1542
01:30:11,553 --> 01:30:15,991
would you you're obviously
placing your bet on humanoid and

1543
01:30:15,991 --> 01:30:19,000
military. Where do you think
it's going.

1544
01:30:19,000 --> 01:30:23,596
Amarjot Singh: Think it's a very
very exciting field. I can

1545
01:30:23,596 --> 01:30:28,037
easily imagine whatever has
happened on the language side

1546
01:30:28,037 --> 01:30:32,789
and the reasoning side with LLMs
would happen on the robotics

1547
01:30:32,789 --> 01:30:36,917
side as well. Once the
foundational models either get

1548
01:30:36,917 --> 01:30:41,436
adaptability or they get good
enough and I think that will

1549
01:30:41,436 --> 01:30:46,266
happen. I think there will be
massive disruption on everything

1550
01:30:46,266 --> 01:30:51,251
we do which is repeatable which
goes all the way from warehouses

1551
01:30:51,251 --> 01:30:56,081
to manufacturing to hopefully
homes. And I think there will be

1552
01:30:56,081 --> 01:31:00,599
a big shift. We'll see this
dystopian world where you have

1553
01:31:00,599 --> 01:31:05,351
humanoids walking around all the
time. So I think that in the

1554
01:31:05,351 --> 01:31:09,792
next five years. But just
imagine right if you're able to

1555
01:31:09,792 --> 01:31:14,544
build a brain which is able to
learn alongside you which is a

1556
01:31:14,544 --> 01:31:19,296
lot of your personalized agents
right. And you can embed them

1557
01:31:19,296 --> 01:31:24,281
and now we're getting humanoids
which also have human like skin.

1558
01:31:24,281 --> 01:31:28,800
Do you actually get to a world
where you can transfer your

1559
01:31:28,800 --> 01:31:33,630
consciousness. Do you even get
if you're talking to a human or

1560
01:31:33,630 --> 01:31:38,304
a humanoid. I don't even know
where the world is going to go

1561
01:31:38,304 --> 01:31:42,666
right. If you are able to
actually build the brain which

1562
01:31:42,666 --> 01:31:47,340
is adaptable and sensible and
the form factor is there which

1563
01:31:47,340 --> 01:31:52,092
is a lot of the work coming from
Japan and China. It's a very

1564
01:31:52,092 --> 01:31:57,000
it's a it the world's going to
look very different in 10 years.

1565
01:31:57,000 --> 01:32:02,302
Audrow Nash: Yeah I think that's
for sure. Okay hell yeah if

1566
01:32:02,302 --> 01:32:07,428
there's anything that someone
takes away from this episode

1567
01:32:07,428 --> 01:32:12,553
like one thing you want our
listeners and watchers to come

1568
01:32:12,553 --> 01:32:16,000
away from this episode with what
is it.

1569
01:32:16,000 --> 01:32:21,262
Amarjot Singh: I think the key
take away is physical AI is a

1570
01:32:21,262 --> 01:32:26,085
very exciting field. It's
coming. It's coming fast. The

1571
01:32:26,085 --> 01:32:30,733
core the core thing which is
going to get it there is

1572
01:32:30,733 --> 01:32:35,994
improvements in hardware which
is happening very rapidly but

1573
01:32:35,994 --> 01:32:41,081
also improvements in software
which is adaptable. And it's

1574
01:32:41,081 --> 01:32:46,430
going to affect many sectors.
And I think the focus should be

1575
01:32:46,430 --> 01:32:51,165
on finding sectors where
robotics can get value to the

1576
01:32:51,165 --> 01:32:55,638
end customer quickly and data
requirements are less

1577
01:32:55,638 --> 01:33:00,900
variability is less so that
whatever foundation model you're

1578
01:33:00,900 --> 01:33:06,512
using can be useful and that is
for the builder. And I think for

1579
01:33:06,512 --> 01:33:12,212
the user I would say it's a very
exciting world and it's going to

1580
01:33:12,212 --> 01:33:17,562
solve a lot of people's problems
and make their life easy. So

1581
01:33:17,562 --> 01:33:18,000
yeah.

1582
01:33:18,000 --> 01:33:21,717
Audrow Nash: Hell yeah. Alright
well it's been great talking

1583
01:33:21,717 --> 01:33:25,000
with you and you guys are doing
super exciting stuff.

1584
01:33:25,000 --> 01:33:28,420
Amarjot Singh: Thank you very
much for having us. Thank you

1585
01:33:28,420 --> 01:33:29,000
very much.

1586
01:33:29,000 --> 01:33:30,000
Audrow Nash: Bye everyone.
