1
00:00:00,000 --> 00:00:03,918
Pras Velagapudi: We do paid
labor. We have customers that

2
00:00:03,918 --> 00:00:07,699
deploy our robots in their
manufacturing facilities and

3
00:00:07,699 --> 00:00:11,961
their logistics facilities and
Digit works in a full shift and

4
00:00:11,961 --> 00:00:16,223
gets paid for it. So that's the
real value of creating a human

5
00:00:16,223 --> 00:00:20,347
centric robot like Digit. It's
to be able to be flexible and

6
00:00:20,347 --> 00:00:24,403
easy to deploy, and all of these
different places in a in a

7
00:00:24,403 --> 00:00:28,183
process flow that were
traditionally very difficult and

8
00:00:28,183 --> 00:00:32,376
specific. I would say if you are
a humanoid robotics provider

9
00:00:32,376 --> 00:00:36,707
yourself, I would encourage you
to understand safety and start,

10
00:00:36,707 --> 00:00:40,625
you know, working with the
standards bodies that are that

11
00:00:40,625 --> 00:00:42,000
are working on that.

12
00:00:42,000 --> 00:00:44,000
Audrow Nash: Hi Pras, would you
introduce yourself?

13
00:00:44,000 --> 00:00:47,480
Pras Velagapudi: Hey, I'm Pras
Velagapudi. I'm the chief

14
00:00:47,480 --> 00:00:51,209
technology officer at Agility
Robotics and at Agility we are

15
00:00:51,209 --> 00:00:55,000
working on building Digit, the
humanoid that's made for work.

16
00:00:55,000 --> 00:00:57,462
Audrow Nash: Hell yeah. Tell me
about Digit and tell me how it's

17
00:00:57,462 --> 00:00:58,000
made for work.

18
00:00:58,000 --> 00:01:01,669
Pras Velagapudi: Yeah. So Digit
is our human centric robotic

19
00:01:01,669 --> 00:01:05,522
platform. So it's a robot with
two arms and two legs. It's spun

20
00:01:05,522 --> 00:01:09,314
out of technology coming out of
Oregon State University. And a

21
00:01:09,314 --> 00:01:12,861
bunch of related academic
research around how to create. A

22
00:01:12,861 --> 00:01:16,530
dynamically stable robot. And
that evolved into the platform

23
00:01:16,530 --> 00:01:20,383
that we created, which is doing
right now material handling. So

24
00:01:20,383 --> 00:01:24,236
moving around totes and bins and
things like that in warehouses

25
00:01:24,236 --> 00:01:27,783
and logistics and manufacturing
applications. And so we're

26
00:01:27,783 --> 00:01:31,391
building up a robot that's
really intended not just to be a

27
00:01:31,391 --> 00:01:34,754
thing that looks kind of human
like, but something that

28
00:01:34,754 --> 00:01:37,873
actually can go into
environments that are designed

29
00:01:37,873 --> 00:01:41,420
for humans, those parts of
facilities, and do useful work.

30
00:01:41,420 --> 00:01:44,967
We do paid labor. We have
customers that deploy our robots

31
00:01:44,967 --> 00:01:48,208
in their manufacturing
facilities, in their logistics

32
00:01:48,208 --> 00:01:52,000
facilities, and it works in a
full shift and gets paid for it.

33
00:01:52,000 --> 00:01:54,932
Audrow Nash: Awesome. And yeah,
you just celery celebrated, an

34
00:01:54,932 --> 00:01:57,818
anniversary of Digit working
somewhere for a while. Would you

35
00:01:57,818 --> 00:01:59,000
tell me a bit about that?

36
00:01:59,000 --> 00:02:03,743
Pras Velagapudi: Yeah. So we
recently celebrated our one year

37
00:02:03,743 --> 00:02:08,252
anniversary being deployed at
GXO. So GXO is a third party

38
00:02:08,252 --> 00:02:12,995
logistics provider. So they do
order fulfillment for a number

39
00:02:12,995 --> 00:02:17,971
of different customers. And, one
of them is Spanx. And Digit has

40
00:02:17,971 --> 00:02:22,324
been helping to do order
fulfillment for Spanx orders at

41
00:02:22,324 --> 00:02:27,300
a facility in, Georgia for just
over a year now. Just cross that

42
00:02:27,300 --> 00:02:28,000
boundary.

43
00:02:28,000 --> 00:02:31,424
Audrow Nash: So what does that
look like? Does it mean that

44
00:02:31,424 --> 00:02:34,964
Digit is working an eight hour
day, then taking a rest? Or is

45
00:02:34,964 --> 00:02:38,621
it working continuously during
the day? Or how does it how does

46
00:02:38,621 --> 00:02:41,000
it look for Digit to be deployed
in this.

47
00:02:41,000 --> 00:02:44,783
Pras Velagapudi: A deployment
looks a lot like Digit basically

48
00:02:44,783 --> 00:02:48,200
matching what the rest of the
process can do. So in this

49
00:02:48,200 --> 00:02:51,862
particular facility, Digits
working a full shift because the

50
00:02:51,862 --> 00:02:55,706
things upstream and downstream
of Digit, the other stations are

51
00:02:55,706 --> 00:02:59,611
also working that same shift. So
that's the limiting factor in a

52
00:02:59,611 --> 00:03:03,212
facility where there were two
shifts, Digit could also work

53
00:03:03,212 --> 00:03:06,568
two shifts. So in this
particular facility I think it's

54
00:03:06,568 --> 00:03:10,168
basically a single shift. I
think it's roughly eight hours.

55
00:03:10,168 --> 00:03:13,951
Digit basically starts up in the
morning and starts doing work

56
00:03:13,951 --> 00:03:17,734
where essentially it's a part of
the process where orders have

57
00:03:17,734 --> 00:03:21,579
been picked from some, inventory
upstream, and they arrive at a

58
00:03:21,579 --> 00:03:25,423
station on an Amr, an autonomous
mobile robot that's carrying a

59
00:03:25,423 --> 00:03:28,962
bunch of different bins for the
different orders. So Digit

60
00:03:28,962 --> 00:03:32,807
actually unloads those bins off
of the Amr and puts them onto a

61
00:03:32,807 --> 00:03:36,651
conveyor so they can go to the
next station, which is a packing

62
00:03:36,651 --> 00:03:40,617
station where they get packed to
go out to customers. So it's one

63
00:03:40,617 --> 00:03:44,462
part of a process that interacts
with automation on both sides.

64
00:03:44,462 --> 00:03:48,001
And I think that's really
important is that we're building

65
00:03:48,001 --> 00:03:51,418
Digit not to replace all
automation in the facility, but

66
00:03:51,418 --> 00:03:55,140
really to work with the other
types of things that are there,

67
00:03:55,140 --> 00:03:58,862
like we have great automation
for things like conveyance, and

68
00:03:58,862 --> 00:04:02,463
we have great automation for
driving around a facility with

69
00:04:02,463 --> 00:04:06,307
an autonomous mobile robot. And
then maybe not as great options

70
00:04:06,307 --> 00:04:10,029
when we're getting to a place
where those two things interact

71
00:04:10,029 --> 00:04:13,385
and you're trying to get
something from place to place,

72
00:04:13,385 --> 00:04:17,047
or you're trying to go from a
shelf to a conveyor or a shelf

73
00:04:17,047 --> 00:04:20,769
to a mobile robot, or a mobile
robot to a pallet that's going

74
00:04:20,769 --> 00:04:24,613
to go on a truck. Those are all
places where you typically have

75
00:04:24,613 --> 00:04:27,908
to have higher cost, more
customized automation that's

76
00:04:27,908 --> 00:04:31,631
built around what those other
pieces are doing. So instead of

77
00:04:31,631 --> 00:04:35,170
doing that, we can have this
single flexible platform that

78
00:04:35,170 --> 00:04:38,709
can go in and glue these other
process flows together. And

79
00:04:38,709 --> 00:04:42,553
that's really the utility of a
platform like Digit. We can come

80
00:04:42,553 --> 00:04:45,665
into a use case like that and
without reconfiguring

81
00:04:45,665 --> 00:04:49,205
everything, without bolting a
bunch of additional stuff to

82
00:04:49,205 --> 00:04:52,683
your armor, or completely
reconfiguring your conveyor and

83
00:04:52,683 --> 00:04:56,161
adding a big gantry robot or
something like that, you can

84
00:04:56,161 --> 00:04:59,395
just connect these pieces
together and if you need to

85
00:04:59,395 --> 00:05:02,995
usage at somewhere else in the
facility, the same platform,

86
00:05:02,995 --> 00:05:06,656
same Digit unit can go over
there and basically with more or

87
00:05:06,656 --> 00:05:09,646
less mostly software
reconfiguration, go do other

88
00:05:09,646 --> 00:05:13,369
tasks and that other part of the
facility. So that's the real

89
00:05:13,369 --> 00:05:16,786
value of creating a human
centric robot like Digit. It's

90
00:05:16,786 --> 00:05:20,447
to be able to be flexible and
easy to deploy in all of these

91
00:05:20,447 --> 00:05:23,437
different places in a in a
process flow that were

92
00:05:23,437 --> 00:05:26,000
traditionally very difficult and
specific.

93
00:05:26,000 --> 00:05:30,209
Audrow Nash: Very interesting.
Yeah. So, I mean, there's a lot

94
00:05:30,209 --> 00:05:33,874
of questions around why
humanoids and I think that's a

95
00:05:33,874 --> 00:05:38,015
big answer to it in some sense,
because what you're saying is

96
00:05:38,015 --> 00:05:42,360
that we have these systems that
are highly optimized, but in the

97
00:05:42,360 --> 00:05:46,093
intersection or like in the
boundary between them, it's

98
00:05:46,093 --> 00:05:50,302
often it's harder. And so like
people would do it or you could

99
00:05:50,302 --> 00:05:54,239
have really complex, costly
integrations that would be set

100
00:05:54,239 --> 00:05:58,108
up to make the systems map from
one to the other. But the

101
00:05:58,108 --> 00:06:02,385
flexibility of the form factor
makes it so you can kind of drop

102
00:06:02,385 --> 00:06:05,983
them in and they can do the
task, and it's a software

103
00:06:05,983 --> 00:06:09,852
change. But I think it'd be
worth talking more about this

104
00:06:09,852 --> 00:06:13,925
explicitly. Why humanoid for
this? Like, what do you what do

105
00:06:13,925 --> 00:06:18,270
you say to people when they say,
why not just a mobile base with

106
00:06:18,270 --> 00:06:22,614
an arm on it, like a single arm?
Why do we need it to have legs?

107
00:06:22,614 --> 00:06:26,687
Why do we need it? To have the
head? How it is like, what do

108
00:06:26,687 --> 00:06:30,964
you think is special about the
human form factor? And why would

109
00:06:30,964 --> 00:06:33,000
this make good business sense?

110
00:06:33,000 --> 00:06:35,996
Pras Velagapudi: Yeah. So we
often get asked why. Humanoid,

111
00:06:35,996 --> 00:06:38,941
right? I'm sure it's a very
common question. And I think a

112
00:06:38,941 --> 00:06:41,734
lot of people are coming from
the perspective of seeing

113
00:06:41,734 --> 00:06:44,629
companies that just arrive on
the scene with a robot that

114
00:06:44,629 --> 00:06:46,000
looks exactly like a human.

115
00:06:46,000 --> 00:06:47,000
Audrow Nash: Yeah, you guys have
been doing it for a while.

116
00:06:47,000 --> 00:06:50,564
Pras Velagapudi: Yeah, well, not
only have we've been doing it

117
00:06:50,564 --> 00:06:53,898
for a while at Agility, you
know, the company was founded.

118
00:06:53,898 --> 00:06:57,519
Yeah, back in 20, I think. About
ten years ago. The company was

119
00:06:57,519 --> 00:07:01,140
founded a little over ten years
ago. The design that we started

120
00:07:01,140 --> 00:07:04,531
with was not very lacking at
all. Cassie. Prototype, if you

121
00:07:04,531 --> 00:07:07,865
go back to those days, was
essentially just a set of legs.

122
00:07:07,865 --> 00:07:11,486
And the legs look more birdlike
than human, and it was designed

123
00:07:11,486 --> 00:07:14,303
to basically emulate a certain
set of physics. So

124
00:07:14,303 --> 00:07:17,809
interestingly, with Agility,
we've almost gone out of our way

125
00:07:17,809 --> 00:07:21,258
not to start out trying to aim
at a human form factor. We've

126
00:07:21,258 --> 00:07:24,477
sort of evolved in that
direction by trying to solve for

127
00:07:24,477 --> 00:07:27,926
the physics problems around
interacting with things in human

128
00:07:27,926 --> 00:07:31,259
environments, and ending up with
something that happens to

129
00:07:31,259 --> 00:07:34,133
analogously look relatively
humanoid with some key

130
00:07:34,133 --> 00:07:37,410
differences. So you can see
that, you know, our legs, for

131
00:07:37,410 --> 00:07:40,973
example, don't have the knees
bending in the same direction as

132
00:07:40,973 --> 00:07:44,192
a human knee. Now, will that
change in the future? Maybe

133
00:07:44,192 --> 00:07:47,641
because for our perspective,
it's not about whether it looks

134
00:07:47,641 --> 00:07:50,860
like a human or not. We can we
might not. It's about the

135
00:07:50,860 --> 00:07:54,539
physics of interacting with the
world and how we do things like,

136
00:07:54,539 --> 00:07:58,045
the linkage transmissions to the
actuators in our leg. That's

137
00:07:58,045 --> 00:08:01,609
what's driving the design. It
just so happens that when you're

138
00:08:01,609 --> 00:08:04,943
dealing with things that are
weights and payloads in human

139
00:08:04,943 --> 00:08:08,449
scale and around eyelets and
things like that, that are again

140
00:08:08,449 --> 00:08:11,725
sized for humans, you start to
up, you start to find that

141
00:08:11,725 --> 00:08:15,116
you're optimizing towards a
human form factor. The legs end

142
00:08:15,116 --> 00:08:18,450
up being similar in size. In
order to move human payloads,

143
00:08:18,450 --> 00:08:21,727
you need to be about human
weight and human height. A lot

144
00:08:21,727 --> 00:08:25,348
of these things are converging,
even though we start out saying

145
00:08:25,348 --> 00:08:29,027
like, look, make any arm of any
length. And then it sort of ends

146
00:08:29,027 --> 00:08:32,303
up being kind of a similar
length to a human arm to reach

147
00:08:32,303 --> 00:08:35,522
these types of payloads. You
know why? Two arms? Well, a

148
00:08:35,522 --> 00:08:38,971
single arm can contact a payload
at a certain point. But the

149
00:08:38,971 --> 00:08:42,707
interesting thing about it there
is that you have to exert all of

150
00:08:42,707 --> 00:08:46,098
your force around that one
contact point with your payload.

151
00:08:46,098 --> 00:08:49,719
So if your pillow is really big,
then what's going to happen is

152
00:08:49,719 --> 00:08:53,225
oftentimes you're going to grab
it at the corner and then the

153
00:08:53,225 --> 00:08:56,904
rest of the payload is going to
break off. Like, I don't know if

154
00:08:56,904 --> 00:09:00,296
you've ever tried to grab a
really heavy cardboard box from

155
00:09:00,296 --> 00:09:03,457
just one corner. Not super
great. And so you can either

156
00:09:03,457 --> 00:09:06,963
make a specialized gripper to
try and exert torques and large

157
00:09:06,963 --> 00:09:10,297
forces across that, which is
what a lot of special purpose

158
00:09:10,297 --> 00:09:13,746
systems do that are handling
just a handful of boxes. Or you

159
00:09:13,746 --> 00:09:17,137
can add a second arm to hold
another corner of the box. And

160
00:09:17,137 --> 00:09:20,758
now you're exerting really small
torques comparatively, because

161
00:09:20,758 --> 00:09:23,000
you can do it at two points on
the box.

162
00:09:23,000 --> 00:09:25,000
Audrow Nash: Now, could you do
it with three?

163
00:09:25,000 --> 00:09:28,263
Pras Velagapudi: You could. In
fact, some researchers have

164
00:09:28,263 --> 00:09:31,077
taken one of our research
Digits, the, these three

165
00:09:31,077 --> 00:09:34,284
platforms that we produced
several years ago. And they've

166
00:09:34,284 --> 00:09:37,435
put a third arm on it in a
research application. But the

167
00:09:37,435 --> 00:09:40,642
more arms you put, the more
expense you're adding. And so

168
00:09:40,642 --> 00:09:44,187
two is sort of our minimum. It's
the least arms that can handle

169
00:09:44,187 --> 00:09:47,000
the range of payload sizes that
we think is, yeah.

170
00:09:47,000 --> 00:09:49,714
Audrow Nash: Probably hit
diminishing returns in terms of

171
00:09:49,714 --> 00:09:50,000
value.

172
00:09:50,000 --> 00:09:51,000
Pras Velagapudi: Right. And so
you end up.

173
00:09:50,000 --> 00:09:52,000
Audrow Nash: With more arms. You
go.

174
00:09:52,000 --> 00:09:55,379
Pras Velagapudi: Yeah you can
make an octopus robot. But then

175
00:09:55,379 --> 00:09:58,315
that's a lot of complexity.
You're trying to find the

176
00:09:58,315 --> 00:10:01,695
minimum complexity to solve the
problem. And so similarly you

177
00:10:01,695 --> 00:10:05,074
know now we have now we have two
arms because of the way that

178
00:10:05,074 --> 00:10:08,232
we're manipulating payloads. So
why legs. Why two legs in

179
00:10:08,232 --> 00:10:11,500
particular. You end up with a
similar sort of thing. Right.

180
00:10:11,500 --> 00:10:14,437
You want to be able to
dynamically balance loads in a

181
00:10:14,437 --> 00:10:17,650
really tight form factor. So you
want to have the smallest

182
00:10:17,650 --> 00:10:20,808
footprint on the ground so you
can navigate through these

183
00:10:20,808 --> 00:10:23,910
really small aisles, but be able
to access a really wide

184
00:10:23,910 --> 00:10:27,068
workspace of the robot really
high up, really low without

185
00:10:27,068 --> 00:10:30,059
expanding that footprint. So if
you're doing that in a

186
00:10:30,059 --> 00:10:33,438
statically stable way, which is
what a lot of the robot basis

187
00:10:33,438 --> 00:10:36,652
with arms on them are doing,
then in order to do that, you

188
00:10:36,652 --> 00:10:40,086
have to make sure that all of
the dynamic loads with your arms

189
00:10:40,086 --> 00:10:43,244
essentially keep you within the
stability polygon. So the

190
00:10:43,244 --> 00:10:46,457
dynamic loads are like when
you're moving a force back and

191
00:10:46,457 --> 00:10:50,003
forth, you have some weight and
you're moving it back and forth.

192
00:10:50,003 --> 00:10:51,000
Yeah, dynamically.

193
00:10:51,000 --> 00:10:53,000
Audrow Nash: You want to get out
of your support polygon.

194
00:10:53,000 --> 00:10:56,380
Pras Velagapudi: Moving. Yeah.
You're moving your center of

195
00:10:56,380 --> 00:10:59,817
gravity. You're moving your
center of pressure around in the

196
00:10:59,817 --> 00:11:03,082
polygon of where you make
contact with the ground. And if

197
00:11:03,082 --> 00:11:06,404
you go past that then you start
to tip, which is bad news,

198
00:11:06,404 --> 00:11:10,013
right? So if you want to do that
well with the static base, you

199
00:11:10,013 --> 00:11:12,934
have two options, right? You
either as your payload

200
00:11:12,934 --> 00:11:16,371
increases, you increase the
weight of your base, which moves

201
00:11:16,371 --> 00:11:19,980
your center of gravity down and
means that you move that center

202
00:11:19,980 --> 00:11:23,245
of pressure around less, less.
Or you make your footprint

203
00:11:23,245 --> 00:11:26,854
really large, which makes your
polygon big. And so you can move

204
00:11:26,854 --> 00:11:30,348
around more before you get to
the edge of it. Right. But both

205
00:11:30,348 --> 00:11:33,899
of those are compromises that
you have to make in terms of how

206
00:11:33,899 --> 00:11:37,393
much your robot weighs and how
big it is physically, how wide

207
00:11:37,393 --> 00:11:40,945
it is now, with a dynamically
stable base, you can have a very

208
00:11:40,945 --> 00:11:44,325
small footprint and dynamically
change where your stability

209
00:11:44,325 --> 00:11:47,361
polygon is while you're
manipulating stuff. So we can

210
00:11:47,361 --> 00:11:50,626
put our feet in very specific
positions and push reaction

211
00:11:50,626 --> 00:11:54,120
forces through them to be able
to maintain this really narrow

212
00:11:54,120 --> 00:11:57,213
footprint, even as we're moving
through a pretty large

213
00:11:57,213 --> 00:12:00,535
workspace. So that's what we've
found is a really valuable

214
00:12:00,535 --> 00:12:04,259
thing. It means that in a lot of
different scenarios and a lot of

215
00:12:04,259 --> 00:12:07,466
different environments, whether
it's narrow aisles or in

216
00:12:07,466 --> 00:12:11,018
situations in which there might
be something on the floor that

217
00:12:11,018 --> 00:12:14,398
you're stepping around like a
pallet, having the ability to

218
00:12:14,398 --> 00:12:17,720
really control that stability
polygon and place where it's

219
00:12:17,720 --> 00:12:21,214
being, you know, where you're
going to construct it and exert

220
00:12:21,214 --> 00:12:24,766
forces all the time to be able
to basically control your whole

221
00:12:24,766 --> 00:12:28,260
body rather than just, react to
the disturbances of the upper

222
00:12:28,260 --> 00:12:31,639
body, let's us do a lot of
motion of payloads with a lot of

223
00:12:31,639 --> 00:12:34,905
dynamic load and a really
compact footprint. And we think

224
00:12:34,905 --> 00:12:38,341
that's something that provides
value. We think it's a reason

225
00:12:38,341 --> 00:12:41,721
that we're able to go into a lot
of spaces where we can get

226
00:12:41,721 --> 00:12:45,215
really up close to things and we
can access, you know, really

227
00:12:45,215 --> 00:12:48,308
large swaths of area. Now,
there's still going to be a

228
00:12:48,308 --> 00:12:51,860
place for those static robots.
And I would expect that we will

229
00:12:51,860 --> 00:12:55,526
see them coming on the market as
prices decrease and things like

230
00:12:55,526 --> 00:12:58,963
that. But we were talking about
what scales really well, and

231
00:12:58,963 --> 00:13:02,343
what can move from environment
to environment. We're pretty

232
00:13:02,343 --> 00:13:05,207
confident that by starting with
this base level of

233
00:13:05,207 --> 00:13:08,873
functionality, we won't have to
adapt a whole lot as we get into

234
00:13:08,873 --> 00:13:12,424
more complex environments where
there's more occlusions, where

235
00:13:12,424 --> 00:13:15,804
there's more narrow spaces or
things that we have to fit in

236
00:13:15,804 --> 00:13:19,355
and around. Because again, yeah,
we're human centric. So we're

237
00:13:19,355 --> 00:13:22,792
roughly going to be able to map
to the same ground footprint

238
00:13:22,792 --> 00:13:26,000
that a human would need to move
those same loads around.

239
00:13:26,000 --> 00:13:29,370
Audrow Nash: Yeah. Okay. Another
one with this. And that's very

240
00:13:29,370 --> 00:13:32,579
interesting. All of that. And
that makes a lot of sense. One

241
00:13:32,579 --> 00:13:35,895
other question that I hear often
is why not have wheels on the

242
00:13:35,895 --> 00:13:38,784
feet? So you have legs. And so
kind of like the Boston

243
00:13:38,784 --> 00:13:41,940
Dynamics, I forget what it's
called, but has legs. And then

244
00:13:41,940 --> 00:13:45,202
it also has wheels at the feet.
And then I think, I think you

245
00:13:45,202 --> 00:13:48,358
Unitree have some of these.
Yeah. What do you think of this

246
00:13:48,358 --> 00:13:49,000
form factor.

247
00:13:49,000 --> 00:13:52,971
Pras Velagapudi: So we have seen
it. So Boston Dynamics had the

248
00:13:52,971 --> 00:13:56,689
handle platform, which was kind
of a balancing robot on two

249
00:13:56,689 --> 00:14:00,345
wheels. Super cool. Yeah. Very
cool. And there's, I think,

250
00:14:00,345 --> 00:14:04,063
hexagon is trying to make a
similar platform with two legs.

251
00:14:04,063 --> 00:14:07,845
We'll see how that goes. And
then the Unitree, they have the

252
00:14:07,845 --> 00:14:11,311
go w which has the wheels at
each edge at each foot. So

253
00:14:11,311 --> 00:14:14,903
combining wheels with legs,
it's, it's an option that you

254
00:14:14,903 --> 00:14:18,685
can do. There's a couple of
caveats to, to be aware of. And,

255
00:14:18,685 --> 00:14:22,592
you know, I won't say that, you
know, Agility wheel will never

256
00:14:22,592 --> 00:14:26,437
explore having wheels at the, at
the bottom. But there's some

257
00:14:26,437 --> 00:14:30,282
trade offs that you have to kind
of consider. The first is if

258
00:14:30,282 --> 00:14:34,000
wheels are part of the dynamic
stability, which is the case

259
00:14:34,000 --> 00:14:37,277
with handle or a Segway, or
there's been a couple of

260
00:14:37,277 --> 00:14:41,185
companies at various points that
have made, balancing, bots on

261
00:14:41,185 --> 00:14:45,029
top of balls. So, bossa nova at
one point, for example, bossa

262
00:14:45,029 --> 00:14:48,811
nova robotics, and there's a
couple of others. The challenge

263
00:14:48,811 --> 00:14:52,655
there is if you start to lose
stability. So say you go over a

264
00:14:52,655 --> 00:14:56,437
slippery patch, right, and you
lose contact and you start to

265
00:14:56,437 --> 00:15:00,345
lean the way that you have to
correct a lean in a robot that's

266
00:15:00,345 --> 00:15:04,315
dynamically stable via, a rotary
joint, like a wheel or a ball,

267
00:15:04,315 --> 00:15:07,718
is that you have to accelerate
under the robot and get

268
00:15:07,718 --> 00:15:11,626
essentially catch up. The wheel
inverted pendulum is where the

269
00:15:11,626 --> 00:15:15,597
inverted pendulum is. Exactly.
And if you fail to do that, your

270
00:15:15,597 --> 00:15:19,315
only option is to accelerate
more. And so, yeah, you end up

271
00:15:19,315 --> 00:15:23,223
in this sort of degraded state
where the only thing you can do

272
00:15:23,223 --> 00:15:26,941
is to try and speed up the
motion of the base. And the more

273
00:15:26,941 --> 00:15:30,597
that you're doing that, you're
spinning up, obviously, the

274
00:15:30,597 --> 00:15:34,252
inertia there. And it's easy to
get into a sort of unsafe,

275
00:15:34,252 --> 00:15:37,845
dangerous cycle where you're
essentially accelerating the

276
00:15:37,845 --> 00:15:41,059
robot because you're
destabilizing. And so the more

277
00:15:41,059 --> 00:15:45,029
you're falling over, the faster
you have to go. And that is not

278
00:15:45,029 --> 00:15:49,000
a good loop when you're trying
to get a safety case worked out.

279
00:15:49,000 --> 00:15:50,000
Audrow Nash: Yeah, that's very
interesting.

280
00:15:50,000 --> 00:15:52,857
Pras Velagapudi: As someone
who's face planted a Segway

281
00:15:52,857 --> 00:15:55,818
before, I can tell you that
that's not a loop gets really

282
00:15:55,818 --> 00:15:59,039
fast sometimes. You were driving
it. Yeah. Yeah. If you take a

283
00:15:59,039 --> 00:16:02,000
Segway sufficiently off road,
you can get this to happen.

284
00:16:02,000 --> 00:16:05,482
Audrow Nash: To face planet. I
had one where I was riding a

285
00:16:05,482 --> 00:16:09,200
segue, and, You reading a lot to
keep it accelerating, and then

286
00:16:09,200 --> 00:16:12,446
it, like, kicks forward or
something. That's it. Do you

287
00:16:12,446 --> 00:16:16,105
want to go faster? Yeah. Yes. So
stabilize itself and keep it,

288
00:16:16,105 --> 00:16:19,528
keep it. It's, center of mass
under its support polygon, I

289
00:16:19,528 --> 00:16:20,000
suppose.

290
00:16:20,000 --> 00:16:23,425
Pras Velagapudi: Right. And so
if you lean really far forward,

291
00:16:23,425 --> 00:16:26,796
right, it loses the ability to
go fast enough to keep up with

292
00:16:26,796 --> 00:16:30,000
that. And that's when it starts
to get a little bit hairy.

293
00:16:30,000 --> 00:16:33,302
Audrow Nash: So wheels basically
make it so it's hard to gear.

294
00:16:33,302 --> 00:16:36,657
It's harder to can think about
safety for this because you have

295
00:16:36,657 --> 00:16:39,000
to outrun. It's falling over if
that occurs.

296
00:16:39,000 --> 00:16:42,152
Pras Velagapudi: That's
specifically in the case of when

297
00:16:42,152 --> 00:16:45,472
they're a part of the dynamic
stability. Now you have other

298
00:16:45,472 --> 00:16:48,567
robot platforms. And a good
example of this is the, the

299
00:16:48,567 --> 00:16:52,056
Unitree, go where they have the
wheels just on every foot of a

300
00:16:52,056 --> 00:16:55,152
quadruped. So, so they're,
you're really just if you're

301
00:16:55,152 --> 00:16:58,584
just using wheels to create,
fixed velocity on your otherwise

302
00:16:58,584 --> 00:17:02,186
dynamically stable base. Right.
They're using the legs to do all

303
00:17:02,186 --> 00:17:05,732
the dynamic stability there. And
that's a very different thing.

304
00:17:05,732 --> 00:17:08,883
That would be the same as if
you, for example, made like

305
00:17:08,883 --> 00:17:12,203
roller skates on each foot of a,
of a biped because they're

306
00:17:12,203 --> 00:17:15,411
you're still using the fetus
feet. It's just that you can

307
00:17:15,411 --> 00:17:19,013
also induce a velocity on them.
So that doesn't necessarily have

308
00:17:19,013 --> 00:17:22,502
the same constraints because it
can degrade by breaking all of

309
00:17:22,502 --> 00:17:25,879
the wheels together. You're
basically just turning it into a

310
00:17:25,879 --> 00:17:29,312
foot. But it's a lot of extra
joints and a lot of extra power

311
00:17:29,312 --> 00:17:32,745
and mass that you're putting at
your feet, which ideally, the

312
00:17:32,745 --> 00:17:36,065
less mass that you can put in
your foot, the faster you can

313
00:17:36,065 --> 00:17:39,104
move the foot to do the
stepping. So in some sense, if

314
00:17:39,104 --> 00:17:42,649
you make a really heavy roller
skate or like a skateboard, that

315
00:17:42,649 --> 00:17:46,251
you're riding all the time now,
if you do need to move your foot

316
00:17:46,251 --> 00:17:49,290
to do, you know, stepping
things, you've put that much

317
00:17:49,290 --> 00:17:52,779
more mass on it. So it's a bit
of a trade off that you usually

318
00:17:52,779 --> 00:17:55,818
don't want to take on until
everything else is sort of

319
00:17:55,818 --> 00:17:57,000
solved for right now.

320
00:17:57,000 --> 00:18:00,218
Audrow Nash: So in the
environments that you're in,

321
00:18:00,218 --> 00:18:03,751
that Agility and Digit, that
Digit is and are the floors

322
00:18:03,751 --> 00:18:07,726
rough or anything like this? Is
it like I'm wondering about how

323
00:18:07,726 --> 00:18:11,637
that trade off would be, because
I get that. So if you, if you

324
00:18:11,637 --> 00:18:15,549
treat if you have wheels that
act as roller skates and you are

325
00:18:15,549 --> 00:18:19,397
just kind of using them to move
faster, but you can turn them

326
00:18:19,397 --> 00:18:23,498
off to have them brakes, so then
they're just regular feet. I get

327
00:18:23,498 --> 00:18:27,410
that they are heavier to pick up
and thus you lose more energy

328
00:18:27,410 --> 00:18:31,069
when you step. But, would you
lose less energy if you were

329
00:18:31,069 --> 00:18:34,855
just traversing on a flat floor
with thumb, or how would you

330
00:18:34,855 --> 00:18:37,000
think of that kind of thing?
Yeah.

331
00:18:37,000 --> 00:18:40,185
Pras Velagapudi: If you're if
you're just comparing wheels to

332
00:18:40,185 --> 00:18:43,369
legs, when you're moving around
in a flat environment, wheels

333
00:18:43,369 --> 00:18:46,398
will be more efficient. They,
they are for humans as well.

334
00:18:46,398 --> 00:18:49,687
That's why we have things like
bicycles and so forth ourselves.

335
00:18:49,687 --> 00:18:50,000
Right?

336
00:18:50,000 --> 00:18:51,000
Audrow Nash: Hoverboard and
whatever else. Yeah, it.

337
00:18:51,000 --> 00:18:54,386
Pras Velagapudi: Shouldn't
necessarily be that much of a

338
00:18:54,386 --> 00:18:58,256
difference in, short distances.
And when there's a lot of things

339
00:18:58,256 --> 00:19:01,884
like turning and reorienting
involved there, if you're doing

340
00:19:01,884 --> 00:19:05,633
an efficient walking gait and
again, you can do an inefficient

341
00:19:05,633 --> 00:19:09,502
walking gait. This is something
that you have to actually design

342
00:19:09,502 --> 00:19:13,130
for. You want low mass in your
legs. You want to have a good

343
00:19:13,130 --> 00:19:16,577
control algorithm that's not
making huge static forces in

344
00:19:16,577 --> 00:19:20,205
your toes and things like that.
But once you get all of that

345
00:19:20,205 --> 00:19:23,772
right, it's not too much worse
to be doing walking in these

346
00:19:23,772 --> 00:19:27,521
small areas like turning around
in small, you know, work cells

347
00:19:27,521 --> 00:19:30,967
and in between hotels and things
like that. Over a longer

348
00:19:30,967 --> 00:19:34,656
distances, a wheel is a little
bit better now where it's more

349
00:19:34,656 --> 00:19:38,405
interesting is actually not in
the hey, I need to step versus,

350
00:19:38,405 --> 00:19:41,549
drive and I am using a little
bit more power. It's I

351
00:19:41,549 --> 00:19:45,358
unexpectedly need to change my
stability polygon because I, you

352
00:19:45,358 --> 00:19:48,986
know, maybe misstep or I clipped
some object, right. Like, I

353
00:19:48,986 --> 00:19:52,614
maybe I'm, I'm, walking along or
you're driving along, and I

354
00:19:52,614 --> 00:19:55,879
encountered the edge of
something like an anti fatigue

355
00:19:55,879 --> 00:19:59,205
mat, or someone has left
something on the floor. Right.

356
00:19:59,205 --> 00:20:03,014
Yeah. With Digit right now will
very quickly be able to do what

357
00:20:03,014 --> 00:20:06,581
I call step recovery, right
where we essentially sense that

358
00:20:06,581 --> 00:20:09,967
this is the disturbance and
start moving our foot really

359
00:20:09,967 --> 00:20:13,293
quickly to a new point, you
know, go over or around the

360
00:20:13,293 --> 00:20:17,102
obstacle to a new point where we
can reestablish our stability.

361
00:20:17,102 --> 00:20:20,972
Polygon. If you create a lot of
mass on the end of your foot, it

362
00:20:20,972 --> 00:20:22,000
actually changes.

363
00:20:22,000 --> 00:20:24,000
Audrow Nash: How slows it down,
move it right.

364
00:20:24,000 --> 00:20:27,114
Pras Velagapudi: Do you have a
lot of inertia to move? So now

365
00:20:27,114 --> 00:20:30,177
you need bigger action to do
that response, a lot more power

366
00:20:30,177 --> 00:20:33,342
to do that response. And that
makes it harder to do so. Having

367
00:20:33,342 --> 00:20:36,508
faster feet really lets you be
very responsive. And that's the

368
00:20:36,508 --> 00:20:39,622
bigger thing that you're losing
when you put a lot of mass in

369
00:20:39,622 --> 00:20:42,838
the feet is actually that things
like step recovery, you're now

370
00:20:42,838 --> 00:20:45,952
going to use a lot more power
and you're going to be sluggish

371
00:20:45,952 --> 00:20:49,066
to respond. In the same way that
if you're a human and you're

372
00:20:49,066 --> 00:20:52,384
wearing like really bulky roller
skates, it's definitely going to

373
00:20:52,384 --> 00:20:55,550
be easier to trip you up than a
person who's just going around

374
00:20:55,550 --> 00:20:58,000
and, you know, like cross
trainers or something.

375
00:20:58,000 --> 00:21:01,320
Audrow Nash: Yeah, for sure
makes sense. Yeah. Because if I

376
00:21:01,320 --> 00:21:04,920
think of a lot of times like if
I trip, I'll like run out of it,

377
00:21:04,920 --> 00:21:08,465
almost this kind of thing. And
that's because of that fast step

378
00:21:08,465 --> 00:21:11,953
that you can do to recover. And
that's what the robot needs to

379
00:21:11,953 --> 00:21:15,329
do. And if you have super heavy
feet, that doesn't work very

380
00:21:15,329 --> 00:21:18,649
good. What do you think about
the idea of, like, a humanoid

381
00:21:18,649 --> 00:21:22,193
torso on a centaur, kind of or
on a, on a quadruped base? Yeah,

382
00:21:22,193 --> 00:21:25,625
it's kind of thing because it
seems like you could have a lot

383
00:21:25,625 --> 00:21:28,832
of. You could have a lot of the
benefits of legs and even

384
00:21:28,832 --> 00:21:32,152
wheels. But then you don't have
things like these, the, the

385
00:21:32,152 --> 00:21:35,528
tripping issue as much because
it would be better if you had

386
00:21:35,528 --> 00:21:39,016
four legs if you were to trip.
But I guess it's more parts and

387
00:21:39,016 --> 00:21:42,673
it probably can't turn around as
easily in narrow spaces. But how

388
00:21:42,673 --> 00:21:46,161
do you think about that kind of
thing? Because a lot of people

389
00:21:46,161 --> 00:21:49,818
have thought have mentioned that
like, a centaur shape seems like

390
00:21:49,818 --> 00:21:51,000
a very optimal shape.

391
00:21:51,000 --> 00:21:54,478
Pras Velagapudi: Yeah. I think
you're more or less nailed it

392
00:21:54,478 --> 00:21:57,957
with the footprint really being
the thing that becomes quite

393
00:21:57,957 --> 00:22:01,493
inconvenient. So if you start to
build a four legged platform

394
00:22:01,493 --> 00:22:04,797
with a with a torso on it, you
start to get into the same

395
00:22:04,797 --> 00:22:08,217
degradation of, you're starting
to really take up a lot of,

396
00:22:08,217 --> 00:22:11,696
footprint on the ground in order
to do that. And that really

397
00:22:11,696 --> 00:22:15,290
impedes you being able to get
into some of the narrow passages

398
00:22:15,290 --> 00:22:18,420
that humans are expected to
operate in. And therefore,

399
00:22:18,420 --> 00:22:22,072
robots like Digit are expected
to operate in. When you're doing

400
00:22:22,072 --> 00:22:25,319
things like going between a
conveyor and a shelf down in

401
00:22:25,319 --> 00:22:29,029
Iowa or something like that, you
wouldn't really be able to turn

402
00:22:29,029 --> 00:22:32,449
around a centaur robot, as
effectively in those spaces. And

403
00:22:32,449 --> 00:22:35,986
it's not too different once you
do all of that of versus what

404
00:22:35,986 --> 00:22:39,406
you might need to do, have a
statically stable wheeled base

405
00:22:39,406 --> 00:22:42,652
in kind of that same form
factor, you're also taking up,

406
00:22:42,652 --> 00:22:46,072
you know, similar expansion of
the ground footprint. So the

407
00:22:46,072 --> 00:22:49,667
sense of a robot is sort of an
interesting point in the space.

408
00:22:49,667 --> 00:22:53,319
It's maybe not clearly so much
better than just doing that with

409
00:22:53,319 --> 00:22:56,913
a statically stable base that's
that large, if you're going to

410
00:22:56,913 --> 00:23:00,333
do it anyway. Folks will explore
it. But really, that isn't

411
00:23:00,333 --> 00:23:03,754
necessarily a great trade off if
you're trying to get into,

412
00:23:03,754 --> 00:23:06,768
getting up close to different
types of machinery and

413
00:23:06,768 --> 00:23:10,188
applications. That for the same
reason that it's not super,

414
00:23:10,188 --> 00:23:13,725
convenient to, you know, take a
small horse through an office

415
00:23:13,725 --> 00:23:15,000
building. Right? Yeah.

416
00:23:15,000 --> 00:23:18,530
Audrow Nash: No, but it'd be
great to see. That would be

417
00:23:18,530 --> 00:23:22,312
funny. Okay. Hell, yeah. So tell
me more about Digit and the

418
00:23:22,312 --> 00:23:26,157
applications that you guys are
using it for. So it's a lot of

419
00:23:26,157 --> 00:23:29,813
code manipulation and moving
objects. That the last time I

420
00:23:29,813 --> 00:23:33,784
saw I thought Digits hands were
flippers. Like, they, they only

421
00:23:33,784 --> 00:23:37,629
had one degree of freedom. Maybe
they have a wrist actuation,

422
00:23:37,629 --> 00:23:41,537
too. So two degrees of freedom
on the hand. But you have their

423
00:23:41,537 --> 00:23:45,382
Digits. Design is modular so
that you can swap out parts. But

424
00:23:45,382 --> 00:23:49,227
tell me about the tasks that
Digits doing and I guess the end

425
00:23:49,227 --> 00:23:52,000
effectors and how it's equipped
to do those.

426
00:23:52,000 --> 00:23:55,922
Pras Velagapudi: Yeah. So right
now Digits primary use cases are

427
00:23:55,922 --> 00:23:59,354
all material handling. It's
where we wanted to enter the

428
00:23:59,354 --> 00:24:03,214
market. So Digit right now has
for user field replaceable units

429
00:24:03,214 --> 00:24:06,952
that are essentially the limbs
of its body. Its arms and legs

430
00:24:06,952 --> 00:24:10,813
are the field replaceable units.
So we can swap out an arm. And

431
00:24:10,813 --> 00:24:14,551
usually we just swap it out with
another arm. It's, you know,

432
00:24:14,551 --> 00:24:18,350
service operation. But when we
want to change functionality of

433
00:24:18,350 --> 00:24:21,721
an ARM, we can swap out an arm
that has a different end

434
00:24:21,721 --> 00:24:25,581
effector on it, for example. And
so the, the, the end effectors

435
00:24:25,581 --> 00:24:29,258
that you're talking about, that
sort of flipper style of end

436
00:24:29,258 --> 00:24:33,119
effector, we're actually from an
older generation of Digit that

437
00:24:33,119 --> 00:24:36,489
was, moving around types of
cardboard boxes and certain

438
00:24:36,489 --> 00:24:39,921
types of very, friction based,
material handling contact

439
00:24:39,921 --> 00:24:43,659
points. So it kind of clamped
down on the sides of that, what

440
00:24:43,659 --> 00:24:47,581
we've switched to and have been
using in most of our deployments

441
00:24:47,581 --> 00:24:51,503
now is actually what looks like
almost like a pincer. So it does

442
00:24:51,503 --> 00:24:55,057
have it has essentially four
fingers, two pairs of fingers

443
00:24:55,057 --> 00:24:58,857
and thumbs that grab something
like a tote, at sort of a front

444
00:24:58,857 --> 00:25:02,288
and rear, pinch foods. And so
those work really well for

445
00:25:02,288 --> 00:25:05,904
things like bins and totes and
other places, other types of

446
00:25:05,904 --> 00:25:09,642
containers that have an overhand
grasp position. And so we've

447
00:25:09,642 --> 00:25:13,135
been doing a lot of that type of
material handling. Those

448
00:25:13,135 --> 00:25:17,118
grippers have been the ones that
we pretty much most concern most

449
00:25:17,118 --> 00:25:20,427
things. And those are the ones
that you can see in our

450
00:25:20,427 --> 00:25:24,288
deployments to conferences like,
Pro at this year where we were

451
00:25:24,288 --> 00:25:27,291
showing off our new Z4
production units with, the

452
00:25:27,291 --> 00:25:30,967
extended battery packs and a new
version of that, a poseable

453
00:25:30,967 --> 00:25:34,644
grasping gripper with the four
fingers. So that's what we've

454
00:25:34,644 --> 00:25:38,321
been deploying on, on those
robots, primarily as we're going

455
00:25:38,321 --> 00:25:41,875
forward. What we're planning to
be designing into the next

456
00:25:41,875 --> 00:25:45,613
version of Digit is actually
just a tool flange. So Digit can

457
00:25:45,613 --> 00:25:49,290
just change out its own end
effector super as it sees fit. I

458
00:25:49,290 --> 00:25:52,660
mean, that's where the industry
has been for industrial

459
00:25:52,660 --> 00:25:55,970
manipulators for some time.
Yeah. So it's not really a

460
00:25:55,970 --> 00:25:59,892
surprise that it makes sense for
us to pursue that basic awesome

461
00:25:59,892 --> 00:26:03,568
subdivide our field replaceable
unit into an arm and a hand.

462
00:26:03,568 --> 00:26:07,000
Essentially, although awesome
hands might not be a hand.

463
00:26:07,000 --> 00:26:08,000
Audrow Nash: It might be a.

464
00:26:08,000 --> 00:26:10,000
Pras Velagapudi: Claw or a hook.
Or maybe.

465
00:26:10,000 --> 00:26:12,997
Audrow Nash: A hand. Could be
anything. Yeah, could be a saw.

466
00:26:12,997 --> 00:26:15,945
Could be a solder. They could be
anything. Yeah. Yeah. Super

467
00:26:15,945 --> 00:26:18,844
cool. Yeah. I love the idea of,
like, tool changing robotic

468
00:26:18,844 --> 00:26:21,792
parts. Well, the robot's doing
that. Like, I don't need that

469
00:26:21,792 --> 00:26:24,789
anymore. It switches its hands,
grabs a more optimal hand. So

470
00:26:24,789 --> 00:26:27,000
cool. We're in the future. It's
unbelievable.

471
00:26:27,000 --> 00:26:29,528
Pras Velagapudi: It's really
funny because a lot of, a lot of

472
00:26:29,528 --> 00:26:31,974
times you see people really
thinking about, you know, these

473
00:26:31,974 --> 00:26:34,461
five fingered hands and they're
trying to jam just like, you

474
00:26:34,461 --> 00:26:35,000
know, 20, 30.

475
00:26:35,000 --> 00:26:37,000
Audrow Nash: Did incredibly
complex that.

476
00:26:37,000 --> 00:26:39,873
Pras Velagapudi: You know,
exactly replicates what our

477
00:26:39,873 --> 00:26:43,066
hands do. But if you really
think about it, there's two kind

478
00:26:43,066 --> 00:26:46,152
of things there. One is if you
actually look at how people

479
00:26:46,152 --> 00:26:49,398
grasp things, it's not like they
necessarily use all of their

480
00:26:49,398 --> 00:26:52,006
degrees of freedom for
everything. And you have a

481
00:26:52,006 --> 00:26:55,251
number of folks who are highly
capable, who don't have all of

482
00:26:55,251 --> 00:26:58,391
those degrees of freedom for
various, disabilities, medical

483
00:26:58,391 --> 00:27:01,477
conditions or things like that.
Right? Whatever. Good. But

484
00:27:01,477 --> 00:27:04,776
they're highly capable, and even
people using various types of

485
00:27:04,776 --> 00:27:07,862
prosthetics can be as dexterous
as any robot on the market

486
00:27:07,862 --> 00:27:10,949
today. And so it's clearly not
that you have to exactly be

487
00:27:10,949 --> 00:27:13,982
replicating a human form factor
to get to dexterity. It's

488
00:27:13,982 --> 00:27:16,749
there's some notion of
dexterity, which is helped by

489
00:27:16,749 --> 00:27:19,622
having some part of those
degrees of freedom, but it's

490
00:27:19,622 --> 00:27:22,761
sort of an open question of
what's the important bit there.

491
00:27:22,761 --> 00:27:26,114
And certainly now that we have
AI based models for training it,

492
00:27:26,114 --> 00:27:29,413
we can explore that space in a
lot more, effective way than we

493
00:27:29,413 --> 00:27:32,765
could before by learning about
how to grasp things with various

494
00:27:32,765 --> 00:27:35,000
types of gripper configurations.
And then.

495
00:27:35,000 --> 00:27:37,426
Audrow Nash: The oh, you mean
you mean making a simple gripper

496
00:27:37,426 --> 00:27:39,930
go a longer way through, I don't
know, reinforcement learning or

497
00:27:39,930 --> 00:27:42,161
something like that, where it
goes and iterates and tries

498
00:27:42,161 --> 00:27:44,000
everything and you see what you
can do with it.

499
00:27:44,000 --> 00:27:47,272
Pras Velagapudi: Right? If you
look at a lot of the work

500
00:27:47,272 --> 00:27:50,660
that's, coming out of groups
that are that kind of started

501
00:27:50,660 --> 00:27:54,341
with the, aloha work. But now,
you know, you see companies like

502
00:27:54,341 --> 00:27:57,846
physical intelligence, and Dana
really showing off that even

503
00:27:57,846 --> 00:28:00,885
with simple two fingered
grippers, they can do a lot

504
00:28:00,885 --> 00:28:04,156
really complicated folding,
manipulation and things like

505
00:28:04,156 --> 00:28:07,720
that. Really, I think there's a
lot of, a lot of, low hanging

506
00:28:07,720 --> 00:28:11,342
fruit available in being able to
unlock dexterous manipulation

507
00:28:11,342 --> 00:28:14,673
with what we can do with
relatively simple grippers. Now,

508
00:28:14,673 --> 00:28:18,120
I think that the same techniques
might also unlock a lot of

509
00:28:18,120 --> 00:28:21,099
utility out of really
complicated hands. But really

510
00:28:21,099 --> 00:28:24,605
complicated, multi-step hands
means a lot of failure points.

511
00:28:24,605 --> 00:28:27,000
It means that when you move to
high, low.

512
00:28:27,000 --> 00:28:28,000
Audrow Nash: Very expensive,
those degrees of.

513
00:28:28,000 --> 00:28:30,000
Pras Velagapudi: Freedom are now
expensive because they.

514
00:28:30,000 --> 00:28:31,000
Audrow Nash: Have to.

515
00:28:30,000 --> 00:28:33,197
Pras Velagapudi: Deal with
dealing with the large forces of

516
00:28:33,197 --> 00:28:36,286
a big payload or something like
that. Totally. So I think

517
00:28:36,286 --> 00:28:39,808
there's a there's this notion of
dexterity, and we all agree that

518
00:28:39,808 --> 00:28:42,409
more dexterity is more
capability, but does more

519
00:28:42,409 --> 00:28:45,498
degrees of freedom equal? More
dexterity is maybe an open

520
00:28:45,498 --> 00:28:48,749
question versus maybe you just
need like a gummy hand pad or

521
00:28:48,749 --> 00:28:51,675
like a big clamp thing, you
know, that's sort of open,

522
00:28:51,675 --> 00:28:52,000
right?

523
00:28:52,000 --> 00:28:55,784
Audrow Nash: Yeah. Yeah. I think
the thing that I always think of

524
00:28:55,784 --> 00:28:59,044
when I see those, like, really
ornate hands with all the

525
00:28:59,044 --> 00:29:02,711
degrees of freedom, it's like, I
don't know, I, I heard it some

526
00:29:02,711 --> 00:29:06,378
friend or something was like, it
just keeps nerd sniping people

527
00:29:06,378 --> 00:29:09,871
where they, like, want to go
build the perfect hand replica.

528
00:29:09,871 --> 00:29:13,306
And it's like the utility of
it's not that good, but it's a

529
00:29:13,306 --> 00:29:16,682
very interesting problem for
this kind of thing. Is fun to

530
00:29:16,682 --> 00:29:20,000
work on. Yeah, but, it's a
beautiful engineering problem.

531
00:29:20,000 --> 00:29:22,747
Pras Velagapudi: Right? But then
when I go take their hand and

532
00:29:22,747 --> 00:29:25,582
try and pick up a 50 pound tote
and break all of the fingers off

533
00:29:25,582 --> 00:29:27,000
of it, and then it's like, yeah.

534
00:29:27,000 --> 00:29:29,239
Audrow Nash: And then it's so
expensive and time consuming to

535
00:29:29,239 --> 00:29:31,000
repair and repairs take a while.
So yeah, we'll.

536
00:29:31,000 --> 00:29:34,375
Pras Velagapudi: See where that
is. But there's a second piece

537
00:29:34,375 --> 00:29:37,641
to it, right? Is that there's
plenty of things that we do in

538
00:29:37,641 --> 00:29:40,798
our own environment where we
really could do it a lot more

539
00:29:40,798 --> 00:29:43,847
effectively if we had tool
changing, like if I'm using a

540
00:29:43,847 --> 00:29:47,113
drill right, the entire handle
assembly and trigger assembly

541
00:29:47,113 --> 00:29:50,597
and speed setting and all of the
features of that drill that are

542
00:29:50,597 --> 00:29:54,026
just there because a human can't
plug into it and have a direct

543
00:29:54,026 --> 00:29:57,347
IO connection that says drill,
please turn it this speed with

544
00:29:57,347 --> 00:29:58,000
this torque.

545
00:29:58,000 --> 00:29:59,000
Audrow Nash: So freaking cool.

546
00:29:59,000 --> 00:30:03,223
Pras Velagapudi: Yeah, it's just
that's a sort of local problem

547
00:30:03,223 --> 00:30:06,777
to us, right? Like a robot
doesn't have to have those

548
00:30:06,777 --> 00:30:10,598
problems and so we can just
engineer our way around them.

549
00:30:10,598 --> 00:30:11,000
Right.

550
00:30:11,000 --> 00:30:14,930
Audrow Nash: Super, super cool.
Yeah, it's really interesting

551
00:30:14,930 --> 00:30:18,925
because it, I guess it's cool to
me to see robotics going from

552
00:30:18,925 --> 00:30:22,856
very specialized towards the
application to more general form

553
00:30:22,856 --> 00:30:26,979
factors now, like humanoids. And
then it's super cool to see the

554
00:30:26,979 --> 00:30:30,588
ability for humanoids or
whatever form factor to do tool

555
00:30:30,588 --> 00:30:34,454
changes, to better suit itself
to whatever the task is. It's

556
00:30:34,454 --> 00:30:36,000
doing super, super cool.

557
00:30:36,000 --> 00:30:39,013
Pras Velagapudi: Yeah. I feel
like we're seeing the

558
00:30:39,013 --> 00:30:42,381
convergence of all of these
different seeds of technology

559
00:30:42,381 --> 00:30:46,104
that have been advancing in the
background for years. So things

560
00:30:46,104 --> 00:30:49,472
like energy storage getting
better, things like actuators

561
00:30:49,472 --> 00:30:52,839
getting more power dense and
more, mass efficient. And of

562
00:30:52,839 --> 00:30:56,326
course, on the compute side,
just more powerful compute and

563
00:30:56,326 --> 00:30:59,575
better AI modeling and
techniques to be able to control

564
00:30:59,575 --> 00:31:03,179
complex systems and interpret
complex perception data. All of

565
00:31:03,179 --> 00:31:06,902
these things have been sort of
raising their water once and for

566
00:31:06,902 --> 00:31:10,506
all, crossing over to create
this new generation of much more

567
00:31:10,506 --> 00:31:14,110
complicated physical systems.
And so it's really cool because

568
00:31:14,110 --> 00:31:17,774
essentially the water line is
just rising, and we know that we

569
00:31:17,774 --> 00:31:21,378
can do these things now. And
it's just created this new level

570
00:31:21,378 --> 00:31:25,159
of capability that that I think
is really going to change how we

571
00:31:25,159 --> 00:31:29,000
approach automating stuff in the
future. We just can do this now.

572
00:31:29,000 --> 00:31:32,991
Audrow Nash: Yeah. I think the
thing that's interesting to me

573
00:31:32,991 --> 00:31:36,785
is it's like before it felt
like, oh, that's not feasible.

574
00:31:36,785 --> 00:31:40,579
This won't work now. It's like,
it's not efficient kind of

575
00:31:40,579 --> 00:31:43,000
thing, which is an interesting
thing.

576
00:31:43,000 --> 00:31:45,717
Pras Velagapudi: Yeah, it's
really funny because having gone

577
00:31:45,717 --> 00:31:48,525
through that, you know, Agility
sort of went through that path

578
00:31:48,525 --> 00:31:50,925
where when Agility first
started, you were convincing

579
00:31:50,925 --> 00:31:53,507
people that a humanoid could
even work and you'd be like,

580
00:31:53,507 --> 00:31:56,360
come to our facility like, look,
we'll show you one that works.

581
00:31:56,360 --> 00:31:59,122
And now it's a lot of people
complaining that they work fine.

582
00:31:59,122 --> 00:32:01,568
They totally believe that.
They're like, coming to our

583
00:32:01,568 --> 00:32:04,285
facility. We 100% believe you
can solve this problem. And of

584
00:32:04,285 --> 00:32:06,867
course, they'll give us some
incredibly complicated, very

585
00:32:06,867 --> 00:32:09,448
difficult problems. And then
they're like, why aren't you

586
00:32:09,448 --> 00:32:12,392
doing it faster? Like go faster,
be cheaper. And then you're sort

587
00:32:12,392 --> 00:32:14,928
of like guys like five years
ago, you wouldn't have even

588
00:32:14,928 --> 00:32:17,826
believed that you could do this.
And now you're complaining that

589
00:32:17,826 --> 00:32:20,000
it's not good enough and you
should be doing it.

590
00:32:20,000 --> 00:32:23,759
Audrow Nash: That you know.
Yeah, it's an interesting thing.

591
00:32:23,759 --> 00:32:27,580
Let's see. So, well, tell me,
you alluded to it a little bit.

592
00:32:27,580 --> 00:32:31,213
So you have your version four
robot. And so that's the one

593
00:32:31,213 --> 00:32:34,659
that you're doing now and that's
that one that has that

594
00:32:34,659 --> 00:32:38,230
interesting gripper that you
mentioned. And you mentioned

595
00:32:38,230 --> 00:32:41,988
that you have a version five
coming, didn't you, before, and

596
00:32:41,988 --> 00:32:45,998
you mentioned a little bit, tell
me a bit about that and tell me

597
00:32:45,998 --> 00:32:47,000
what's changing.

598
00:32:47,000 --> 00:32:50,745
Pras Velagapudi: Yeah. So with
the version five, the big

599
00:32:50,745 --> 00:32:54,558
changes, version four is
designed to work out in industry

600
00:32:54,558 --> 00:32:58,504
and be compliant with regulatory
standards for working in a

601
00:32:58,504 --> 00:33:02,784
facility with an external safety
system with some form of a work

602
00:33:02,784 --> 00:33:06,864
cell around it. So we need to
separate the robot from humans.

603
00:33:06,864 --> 00:33:10,810
Now, the robot has an onboard
safety system, but it doesn't

604
00:33:10,810 --> 00:33:14,689
necessarily have all of the
onboard capabilities to detect

605
00:33:14,689 --> 00:33:18,234
humans and to be able to
dynamically change what it's

606
00:33:18,234 --> 00:33:22,180
doing in response to humans
getting close to it. So we need

607
00:33:22,180 --> 00:33:26,326
to either put up walls around
it, or somehow segment where the

608
00:33:26,326 --> 00:33:30,339
robot works from where humans
are. And if a human enters the

609
00:33:30,339 --> 00:33:34,686
space, we basically kind of shut
down the robot before we let the

610
00:33:34,686 --> 00:33:38,432
human enter the space. Our big
design change with our V5

611
00:33:38,432 --> 00:33:42,712
platform, which we're working on
right now, to come out in about

612
00:33:42,712 --> 00:33:46,992
18 months or less, is to be able
to integrate all of that safety

613
00:33:46,992 --> 00:33:50,938
onboard the robot and make that
robot safe and compliant to

614
00:33:50,938 --> 00:33:55,018
operate cooperatively around
humans, which is to say that the

615
00:33:55,018 --> 00:33:58,897
robot and humans can move
through the same spaces together

616
00:33:58,897 --> 00:34:02,843
without specific guarding or
protection around them. All of

617
00:34:02,843 --> 00:34:06,856
that's included onboard the
robot, so they can't necessarily

618
00:34:06,856 --> 00:34:10,401
work on the same thing at the
same time. That type of

619
00:34:10,401 --> 00:34:14,280
collaborative safety night,
because that requires a little

620
00:34:14,280 --> 00:34:18,427
bit more than what exists right
now in the safety industry. It

621
00:34:18,427 --> 00:34:22,373
requires some fairly advanced
technology there that doesn't

622
00:34:22,373 --> 00:34:26,720
quite exist. But pushing what we
can do now to its boundaries, we

623
00:34:26,720 --> 00:34:30,800
can do cooperative safety, which
is that the robot can detect

624
00:34:30,800 --> 00:34:34,879
humans in its vicinity and limit
its motion with respect that

625
00:34:34,879 --> 00:34:38,959
using onboard safety systems.
And so that's really pushing us

626
00:34:38,959 --> 00:34:42,705
to do this, reinvention of the
Digit technology stack to

627
00:34:42,705 --> 00:34:46,584
incorporate safety at every
level, from the physics of the

628
00:34:46,584 --> 00:34:50,663
robot itself all the way through
the drive systems, the drive

629
00:34:50,663 --> 00:34:54,476
electronics, and the low and
high level software to build

630
00:34:54,476 --> 00:34:58,355
this new platform. So it's
pretty exciting. And we've been

631
00:34:58,355 --> 00:35:01,766
doing a lot of work, not just
from the robot design

632
00:35:01,766 --> 00:35:05,578
perspective, where we've been
developing out a lot of new

633
00:35:05,578 --> 00:35:09,858
technology for, AI based safety
and on board dynamics stability,

634
00:35:09,858 --> 00:35:13,336
but also on the safety side
itself, where we're also

635
00:35:13,336 --> 00:35:17,082
developing out new standards
that can use these types of

636
00:35:17,082 --> 00:35:20,961
models as part of a safety
strategy that we can take to, a

637
00:35:20,961 --> 00:35:24,572
nationally recognized testing
laboratory to basically,

638
00:35:24,572 --> 00:35:28,585
guarantee and certify that our
robots are in compliance with

639
00:35:28,585 --> 00:35:32,665
industry standards. So that's
been a huge effort as well. And

640
00:35:32,665 --> 00:35:37,012
we're pretty excited about that.
And we just recently, basically,

641
00:35:37,012 --> 00:35:40,958
created a new, ISO standard
that's in development. So we're

642
00:35:40,958 --> 00:35:44,770
now working on a specific ISO
standard towards safety for

643
00:35:44,770 --> 00:35:48,382
industrial robots that use
dynamic stability, which is

644
00:35:48,382 --> 00:35:52,462
robots like Digit, where the
dynamic stability has to be part

645
00:35:52,462 --> 00:35:54,000
of the safety strategy.

646
00:35:54,000 --> 00:35:56,000
Audrow Nash: Okay. Tell me about
that.

647
00:35:56,000 --> 00:35:59,780
Pras Velagapudi: So in previous
standards, things like NCR 1508

648
00:35:59,780 --> 00:36:03,080
which deals with autonomous
mobile robots, a lot of the

649
00:36:03,080 --> 00:36:06,620
standards, basically have gaps
in them when it comes to how

650
00:36:06,620 --> 00:36:09,980
they're applied to a robot
that's dynamically balancing.

651
00:36:09,980 --> 00:36:13,760
They kind of make the assumption
that you more or less probably

652
00:36:13,760 --> 00:36:17,540
have a statically stable robot.
It's either bolted to something

653
00:36:17,540 --> 00:36:20,000
or it's on wheels or something
like that.

654
00:36:20,000 --> 00:36:21,000
Audrow Nash: Because that was
true in the.

655
00:36:21,000 --> 00:36:24,140
Pras Velagapudi: Past, right? It
was true for industrial arms. It

656
00:36:24,140 --> 00:36:27,039
was true for most collaborative
arms and for amrs autonomous

657
00:36:27,039 --> 00:36:29,599
mobile robots, which are
primarily wheeled bases that

658
00:36:29,599 --> 00:36:32,449
have, you know, shelves and
things mounted on them. So that

659
00:36:32,449 --> 00:36:35,058
means that a lot of the
standards essentially say, oh,

660
00:36:35,058 --> 00:36:37,908
well, in order to guarantee
safety, like, here's how you're

661
00:36:37,908 --> 00:36:40,565
going to power the entire robot
in this in this type of

662
00:36:40,565 --> 00:36:41,000
scenario.

663
00:36:41,000 --> 00:36:42,000
Audrow Nash: Safely, right? Yes.

664
00:36:42,000 --> 00:36:45,512
Pras Velagapudi: Which works if
locking up your wheels is a

665
00:36:45,512 --> 00:36:49,262
thing that makes you more safe.
It doesn't work. If turning off

666
00:36:49,262 --> 00:36:52,833
all of your motors is a thing
that makes you so that you can

667
00:36:52,833 --> 00:36:56,226
no longer keep yourself upright,
which is the case in any

668
00:36:56,226 --> 00:36:59,500
dynamically stable robot. So
that's the case in Digits,

669
00:36:59,500 --> 00:37:02,893
certainly with legs. But it
would have been the case, for

670
00:37:02,893 --> 00:37:06,464
example, in the Boston Dynamics
handle robot. Or if you were

671
00:37:06,464 --> 00:37:09,917
trying to make a Segway that
could work, you know, with an

672
00:37:09,917 --> 00:37:13,250
industrial arm on it or
something like that. That's also

673
00:37:13,250 --> 00:37:16,881
a dynamically stable robot and
would have some of the similar

674
00:37:16,881 --> 00:37:20,333
gaps in specification in the
safety standards. So this new

675
00:37:20,333 --> 00:37:23,845
effort, which we're doing in
collaboration with other folks

676
00:37:23,845 --> 00:37:27,238
in Boston Dynamics, is also
participating, with, with our

677
00:37:27,238 --> 00:37:30,155
tool, it's a collaborative
effort with a bunch of

678
00:37:30,155 --> 00:37:33,429
different, providers that are
working on these types of

679
00:37:33,429 --> 00:37:36,524
platforms to basically to find
out how do we do this

680
00:37:36,524 --> 00:37:40,274
dynamically stable safety. What
do we need to, specify? What do

681
00:37:40,274 --> 00:37:43,905
we need to, describe so that we
can ensure that a dynamically

682
00:37:43,905 --> 00:37:47,357
stable robot that's doing
certain things at certain levels

683
00:37:47,357 --> 00:37:50,274
of reliability could be
considered safe enough to

684
00:37:50,274 --> 00:37:53,488
operate in an industrial
environment. And it really is

685
00:37:53,488 --> 00:37:57,060
kind of like clarifying some of
those details to get this to

686
00:37:57,060 --> 00:38:00,690
match with broader, you know,
industrial standards for safety

687
00:38:00,690 --> 00:38:04,381
of machinery as a whole. So all
of these standards are kind of

688
00:38:04,381 --> 00:38:07,595
building up. So there's,
different levels of standards

689
00:38:07,595 --> 00:38:10,988
like type A through type C, and
the idea is that you have

690
00:38:10,988 --> 00:38:14,143
broader standards that
essentially say like make this

691
00:38:14,143 --> 00:38:17,476
robot safe and then more
specific ones which are around.

692
00:38:17,476 --> 00:38:20,810
And here's how you make a
wheeled robot safe, and here's

693
00:38:20,810 --> 00:38:24,321
how you make a collaborative
robot arm safe. And now here's

694
00:38:24,321 --> 00:38:27,000
how you make a dynamically
stable robot safe.

695
00:38:27,000 --> 00:38:31,000
Audrow Nash: So how do you make
a dynamically stable robot safe.

696
00:38:31,000 --> 00:38:34,567
Pras Velagapudi: Yeah. It comes
down to a few things. And

697
00:38:34,567 --> 00:38:38,447
there's a little bit of secret
sauce in there that we're, that

698
00:38:38,447 --> 00:38:42,202
we're building out. So I can't
talk about all of the detail.

699
00:38:42,202 --> 00:38:45,957
But what we think is kind of
primarily important is that you

700
00:38:45,957 --> 00:38:49,774
need to understand where people
are in your environments, and

701
00:38:49,774 --> 00:38:53,716
then you need to be able to be
maintaining, essentially, a safe

702
00:38:53,716 --> 00:38:57,659
operating envelope for your own
robot that you can guarantee in

703
00:38:57,659 --> 00:39:00,851
the events of faults or
failures. That's within the

704
00:39:00,851 --> 00:39:04,731
space of where people are. So
you have to essentially say, oh,

705
00:39:04,731 --> 00:39:08,673
how do I keep my robot such that
it's bounded in how it's going

706
00:39:08,673 --> 00:39:12,553
to move or exert forces, or even
trip and fall and things like

707
00:39:12,553 --> 00:39:15,995
that? So it'll always end up
within this, this bounding

708
00:39:15,995 --> 00:39:19,687
radius or this bounding sphere
or whatever it is. And I can

709
00:39:19,687 --> 00:39:20,000
make.

710
00:39:20,000 --> 00:39:21,000
Audrow Nash: Sure that there's
no person.

711
00:39:21,000 --> 00:39:24,606
Pras Velagapudi: That people, as
they're approaching, I can

712
00:39:24,606 --> 00:39:28,150
shrink my radius down by doing
things like slowing down my

713
00:39:28,150 --> 00:39:31,939
motion or getting closer to the
ground. And so it's this whole

714
00:39:31,939 --> 00:39:35,422
dynamic process of how do I
enforce that as a person gets

715
00:39:35,422 --> 00:39:39,089
closer to the robot, the robot
is reducing its hazard by the

716
00:39:39,089 --> 00:39:43,000
things that it can control about
it's balancing and its posture.

717
00:39:43,000 --> 00:39:46,613
Audrow Nash: That sounds really
cool because so I'm imagining

718
00:39:46,613 --> 00:39:50,403
that a robot is moving through a
space and it's estimating where

719
00:39:50,403 --> 00:39:53,957
there is risk for it falling and
doing damage. And what it's

720
00:39:53,957 --> 00:39:57,569
doing is it's making sure people
aren't in that zone for this

721
00:39:57,569 --> 00:40:00,945
kind of thing. And there's
probably some risk negotiation

722
00:40:00,945 --> 00:40:04,558
there where it's a tight space
and I have to. So I'll go real

723
00:40:04,558 --> 00:40:08,289
slow or something like that. Is
it it's similar to that kind of

724
00:40:08,289 --> 00:40:09,000
thing. Yeah.

725
00:40:09,000 --> 00:40:12,588
Pras Velagapudi: It's this idea
of basically trying to figure

726
00:40:12,588 --> 00:40:16,235
out where your various levels of
hazard can be and then, yeah,

727
00:40:16,235 --> 00:40:19,588
what can you do to, you know, if
you got to go faster and

728
00:40:19,588 --> 00:40:23,118
increase that radius or you can,
you know, do things like go

729
00:40:23,118 --> 00:40:26,706
slower or use lower forces and
shrink that radius. And now if

730
00:40:26,706 --> 00:40:29,941
you twist that lever, right, if
I can combine that with

731
00:40:29,941 --> 00:40:33,588
detecting where people are, and
I think detecting where people

732
00:40:33,588 --> 00:40:37,118
are is kind of an important
piece here. Yeah. Because you're

733
00:40:37,118 --> 00:40:40,706
on a platform whose job. Right.
The thing that you're getting

734
00:40:40,706 --> 00:40:44,353
paid for to do is exert large
forces on objects. Right. I need

735
00:40:44,353 --> 00:40:47,765
to pick up a 50 pound tote.
That's a huge force. That same

736
00:40:47,765 --> 00:40:49,000
force, if I apply it.

737
00:40:49,000 --> 00:40:50,000
Audrow Nash: Did you pick up a
50 pound tote?

738
00:40:50,000 --> 00:40:53,434
Pras Velagapudi: So the newer
versions of Digit can pick up,

739
00:40:53,434 --> 00:40:56,754
up to because that's heavy.
Yeah. That's impressive. 23-20

740
00:40:56,754 --> 00:40:59,845
five kilograms. That's what
we're targeting for our V5

741
00:40:59,845 --> 00:41:03,337
platform. The existing robot can
within a higher workspace do

742
00:41:03,337 --> 00:41:06,657
16kg. And it's higher as you go
a little bit closer to the

743
00:41:06,657 --> 00:41:07,000
robot.

744
00:41:07,000 --> 00:41:08,000
Audrow Nash: 30 pounds.

745
00:41:08,000 --> 00:41:10,503
Pras Velagapudi: Yeah. Like 33
pounds or 35. Yeah. And it's

746
00:41:10,503 --> 00:41:13,006
doing that. It's doing that, you
know, all day every day at

747
00:41:13,006 --> 00:41:15,000
facilities. It's actually
handling those loads.

748
00:41:15,000 --> 00:41:18,000
Audrow Nash: It's awesome. So
yeah we 50 pounds is all we.

749
00:41:18,000 --> 00:41:20,035
Pras Velagapudi: Have Digit you
know. Oh yeah doing, doing

750
00:41:20,035 --> 00:41:22,000
little dumbbell curls with
sandbags and things like the.

751
00:41:22,000 --> 00:41:24,880
Audrow Nash: Barbell and
whatever. Hilarious. I would

752
00:41:24,880 --> 00:41:27,000
love to see those videos. That's
great.

753
00:41:27,000 --> 00:41:30,733
Pras Velagapudi: Yeah. Okay. So
basically once you're doing that

754
00:41:30,733 --> 00:41:34,525
right, you can't you can't limit
yourself just by saying, oh, I'm

755
00:41:34,525 --> 00:41:38,142
really close to something that
could be a person or a shelf or

756
00:41:38,142 --> 00:41:41,000
something. I need to turn down
all of my, forces.

757
00:41:41,000 --> 00:41:42,000
Audrow Nash: That'd be useless.

758
00:41:41,000 --> 00:41:45,107
Pras Velagapudi: And you'd never
be able to carry anything exact

759
00:41:45,107 --> 00:41:48,572
for sure. So that's why
understanding where people are

760
00:41:48,572 --> 00:41:52,487
is such a big part of having a
functional robot that can also

761
00:41:52,487 --> 00:41:53,000
be safe.

762
00:41:53,000 --> 00:41:57,134
Audrow Nash: So that sounds like
a lot of perception is required.

763
00:41:57,134 --> 00:42:01,076
I'm imagining almost like I'm
thinking of Tesla and their cars

764
00:42:01,076 --> 00:42:04,638
and how they see everything
around. And so I'm imagining

765
00:42:04,638 --> 00:42:08,771
that you need something similar.
So the robot has no blind spots,

766
00:42:08,771 --> 00:42:12,459
so that you can go and look at
everything and assess where

767
00:42:12,459 --> 00:42:16,466
things are, or you need a good
map of the environment, but then

768
00:42:16,466 --> 00:42:20,218
always a person could sneak up
behind you if, say, you only

769
00:42:20,218 --> 00:42:23,906
have forward facing eyes and
that would make your estimate

770
00:42:23,906 --> 00:42:27,595
bad. So do you basically have
like super aware perception,

771
00:42:27,595 --> 00:42:31,092
like, is that part of the
standard or tell me about the

772
00:42:31,092 --> 00:42:33,000
perception side of this. Yeah.

773
00:42:33,000 --> 00:42:36,157
Pras Velagapudi: So this there's
two pieces here to making a safe

774
00:42:36,157 --> 00:42:39,266
robot. The first is the standard
where we're essentially writing

775
00:42:39,266 --> 00:42:42,277
the rules. So in that part of
things we're not really defining

776
00:42:42,277 --> 00:42:45,191
like should you use cameras or
should you use later lidar or

777
00:42:45,191 --> 00:42:48,106
things like that. You're really
defining like what would you

778
00:42:48,106 --> 00:42:50,000
need to prove in or and what
would you.

779
00:42:50,000 --> 00:42:52,745
Audrow Nash: Standard as
modular? In a sense, because of

780
00:42:52,745 --> 00:42:55,735
that, you're making there's no
assumption on how you get this

781
00:42:55,735 --> 00:42:58,873
data about the environment. You
just said once you have it, this

782
00:42:58,873 --> 00:43:00,000
is the case for safety.

783
00:43:00,000 --> 00:43:03,204
Pras Velagapudi: Yeah, it's
defining. How do you go about,

784
00:43:03,204 --> 00:43:06,408
how do you go about guaranteeing
that that are dynamically

785
00:43:06,408 --> 00:43:09,611
stable? Robot is safe like
what's what are the rules. What

786
00:43:09,611 --> 00:43:13,202
are the what are the checks that
you need to kind of meet. Right.

787
00:43:13,202 --> 00:43:16,737
But not the technology. And then
on the other side, you know, at

788
00:43:16,737 --> 00:43:19,830
Agility, we're obviously
developing our platform here to

789
00:43:19,830 --> 00:43:23,145
try and be able to meet those
rules. Right. And that's where

790
00:43:23,145 --> 00:43:26,404
we're deciding things like,
okay, how many cameras or other

791
00:43:26,404 --> 00:43:29,829
types of sensors might we need?
Where would we place them. You

792
00:43:29,829 --> 00:43:32,977
know, you can you can, for
example, choose to, have fewer

793
00:43:32,977 --> 00:43:36,015
sensors but then need to
maintain more state about your

794
00:43:36,015 --> 00:43:39,385
environment. So essentially you
could say, well, as long as I

795
00:43:39,385 --> 00:43:42,810
remember that I passed by this
area and a person wasn't there,

796
00:43:42,810 --> 00:43:46,069
I could assume that someone
couldn't have snuck in, because

797
00:43:46,069 --> 00:43:49,272
I can see around all of the
places that they could come in

798
00:43:49,272 --> 00:43:52,642
from, even if I can't see that
specific spot. Right. So there

799
00:43:52,642 --> 00:43:56,122
is some amount of that. But then
you, you then if you lose that

800
00:43:56,122 --> 00:43:58,000
line of sight or you get that,
or.

801
00:43:58,000 --> 00:44:00,000
Audrow Nash: They were hiding in
a shelf or something.

802
00:44:00,000 --> 00:44:02,507
Pras Velagapudi: Yeah. So then
there's edge cases you have to

803
00:44:02,507 --> 00:44:04,849
deal with where you have to
maintain that state. And that

804
00:44:04,849 --> 00:44:07,233
might be better served by just
adding an additional sensor

805
00:44:07,233 --> 00:44:09,000
there or something. Probably in
some cases.

806
00:44:09,000 --> 00:44:10,000
Audrow Nash: Yeah.

807
00:44:09,000 --> 00:44:12,301
Pras Velagapudi: Clears the
space for you. But you have to

808
00:44:12,301 --> 00:44:15,943
make those trade offs right. And
so we do plan on having a bunch

809
00:44:15,943 --> 00:44:19,187
of sensors. That's I think an
important part of this. But

810
00:44:19,187 --> 00:44:22,488
there is also some amount of
understanding exactly how the

811
00:44:22,488 --> 00:44:25,789
environment evolves from a
safety perspective and clearing

812
00:44:25,789 --> 00:44:28,976
certain areas or reasoning about
where could a person be

813
00:44:28,976 --> 00:44:32,561
realistically, given that this
is what I've been doing? Like if

814
00:44:32,561 --> 00:44:35,805
I was just in that space and I
could see all of the areas

815
00:44:35,805 --> 00:44:39,106
around it that a person could
have moved from, so a person

816
00:44:39,106 --> 00:44:42,350
couldn't have, you know,
teleported in there, right? They

817
00:44:42,350 --> 00:44:45,935
would have to have already been
in there, in which case I would

818
00:44:45,935 --> 00:44:49,463
have seen them before or after
something like that. Right. You

819
00:44:49,463 --> 00:44:53,049
can have that type of reasoning
now. Yeah. For sure. It'll be a

820
00:44:53,049 --> 00:44:56,520
combination of, of some amount
of understanding that from the

821
00:44:56,520 --> 00:44:59,935
safety side. Right. And also
just having really good sensing

822
00:44:59,935 --> 00:45:03,463
and a variety of sensing that
can reliably detect humans under

823
00:45:03,463 --> 00:45:05,000
a variety of circumstances.

824
00:45:05,000 --> 00:45:08,600
Audrow Nash: Yeah. And one thing
that's really interesting to me

825
00:45:08,600 --> 00:45:11,638
with this, from how you
described it, I think you said

826
00:45:11,638 --> 00:45:15,013
it like on agent or onboard
safety for this kind of thing as

827
00:45:15,013 --> 00:45:18,050
opposed to with cameras
everywhere. And that gives the

828
00:45:18,050 --> 00:45:21,369
robot the perfect, observer that
it gives the world perfect

829
00:45:21,369 --> 00:45:24,969
observability for the robot. And
maybe that would work. But it's

830
00:45:24,969 --> 00:45:28,400
less flexible in its ability to
deploy because you need power

831
00:45:28,400 --> 00:45:32,000
and data, ability to stream data
if you have cameras everywhere.

832
00:45:32,000 --> 00:45:35,027
Pras Velagapudi: So you can
actually pair even the V4

833
00:45:35,027 --> 00:45:38,511
version of Digit, right? So the
v4 version of Digits designed

834
00:45:38,511 --> 00:45:42,053
for an external safety system,
you can pair it if you have one

835
00:45:42,053 --> 00:45:45,080
with a safety system that's
using external cameras or

836
00:45:45,080 --> 00:45:48,564
external, laser or in various
forms, whether it's 3D lidar or

837
00:45:48,564 --> 00:45:52,163
things like that. And people do
make systems for doing external

838
00:45:52,163 --> 00:45:55,818
safety where they are detecting
people from a third party, not a

839
00:45:55,818 --> 00:45:58,960
third party, but from an
external, perspective. You put

840
00:45:58,960 --> 00:46:02,044
cameras all through your
facility or some area, right?

841
00:46:02,044 --> 00:46:05,300
And then you, you essentially
detect where people are and

842
00:46:05,300 --> 00:46:08,670
where robots are. And you can
use that to then write a rule

843
00:46:08,670 --> 00:46:12,040
that says, hey, if something's
getting close to a V4 robot,

844
00:46:12,040 --> 00:46:15,295
shut it down, right. So you
could do that. Now, it's just

845
00:46:15,295 --> 00:46:18,037
that those systems require you
to build out that

846
00:46:18,037 --> 00:46:21,521
infrastructure, which a lot of
our customers are coming to us

847
00:46:21,521 --> 00:46:24,834
because they don't want to have
that infrastructure in any

848
00:46:24,834 --> 00:46:28,375
specific place because they want
Digit to kind of come into an

849
00:46:28,375 --> 00:46:31,745
area and be ready to go and
maybe move it to someplace else

850
00:46:31,745 --> 00:46:35,229
later. And they don't want to
necessarily put a camera system

851
00:46:35,229 --> 00:46:37,000
here and a camera system there.

852
00:46:37,000 --> 00:46:39,609
Audrow Nash: And you see you
have a 30,000 square foot

853
00:46:39,609 --> 00:46:42,556
warehouse or something or three,
like some massive, 1,000,000

854
00:46:42,556 --> 00:46:45,406
square foot warehouse. It's
like, that's going to be pretty

855
00:46:45,406 --> 00:46:47,000
costly to put cameras
everywhere.

856
00:46:47,000 --> 00:46:50,281
Pras Velagapudi: Yeah, you might
be able to do it in a certain

857
00:46:50,281 --> 00:46:53,510
work cell area so that for that
use case it might make sense.

858
00:46:53,510 --> 00:46:56,791
And when you can do it, if you
can afford the cost because the

859
00:46:56,791 --> 00:47:00,125
systems are not cheap, you can
certainly do it. And it might be

860
00:47:00,125 --> 00:47:03,407
even better than having onboard
safety. But it's really about,

861
00:47:03,407 --> 00:47:06,000
again, the infrastructure versus
the flexibility.

862
00:47:06,000 --> 00:47:09,231
Audrow Nash: Yes, definitely.
Yeah. And you guys are heavily

863
00:47:09,231 --> 00:47:12,354
betting on the flexibility,
which I think is a good bet in

864
00:47:12,354 --> 00:47:13,000
all of this.

865
00:47:13,000 --> 00:47:16,433
Pras Velagapudi: It's a part of
the space. There's always going

866
00:47:16,433 --> 00:47:19,212
to be some part of these
industries that needs that

867
00:47:19,212 --> 00:47:22,699
flexibility, whether it's a type
of industry where the amount of

868
00:47:22,699 --> 00:47:25,969
flow changes really seasonally.
You know, one of the reasons

869
00:47:25,969 --> 00:47:29,238
that grows so interested in us
is as a third party logistics

870
00:47:29,238 --> 00:47:32,617
provider, they get new customers
all the time. And so a single

871
00:47:32,617 --> 00:47:36,104
facility might be changing what
fulfillment it's doing, or might

872
00:47:36,104 --> 00:47:39,210
add or remove a customer
throughout the course of a year.

873
00:47:39,210 --> 00:47:42,698
And so you really don't want to
be in a situation where you paid

874
00:47:42,698 --> 00:47:45,967
for a whole bunch of automation
that's at a single rate or a

875
00:47:45,967 --> 00:47:49,237
single size, and now I add a
customer and it's now the wrong

876
00:47:49,237 --> 00:47:52,615
size, and I have to rip it up or
add a really big expansion or

877
00:47:52,615 --> 00:47:56,048
something. So that flexibility
for them is really valuable. And

878
00:47:56,048 --> 00:47:59,590
there's other cases in which you
have the use of that flexibility

879
00:47:59,590 --> 00:48:02,859
is because you might have a
really big facility. We're using

880
00:48:02,859 --> 00:48:05,693
different parts of it at
different times. Right. You

881
00:48:05,693 --> 00:48:08,690
might do something in the
morning and something else in

882
00:48:08,690 --> 00:48:11,905
the afternoon. Or there's things
like cruise shifts where a

883
00:48:11,905 --> 00:48:15,120
single person rotates between
different stations. So that's

884
00:48:15,120 --> 00:48:18,280
really common in certain types
of manufacturing. So I have

885
00:48:18,280 --> 00:48:21,604
another customer, Scheffler, who
does automotive parts, where

886
00:48:21,604 --> 00:48:24,874
that's a really big part of
where we can bring value to them

887
00:48:24,874 --> 00:48:28,089
is that they have, you know,
parts, machining equipment. So

888
00:48:28,089 --> 00:48:31,249
things like washers and dryers
and heaters and things like

889
00:48:31,249 --> 00:48:34,573
that, heat treatment systems
that basically batch up work. So

890
00:48:34,573 --> 00:48:37,843
you load them up and then you
leave, right. And they go, you

891
00:48:37,843 --> 00:48:41,384
know, clean parts for 90 minutes
or something, and you don't need

892
00:48:41,384 --> 00:48:44,000
to be there for those 90
minutes, and it'd be a.

893
00:48:44,000 --> 00:48:45,000
Audrow Nash: Waste to.

894
00:48:45,000 --> 00:48:48,495
Pras Velagapudi: Something else.
Robot arm to be sitting there

895
00:48:48,495 --> 00:48:51,539
for 90 minutes, after 90
minutes, then a bunch of bins

896
00:48:51,539 --> 00:48:55,034
come out and you need to take
them and load them into the next

897
00:48:55,034 --> 00:48:58,248
thing. Right. And so they have
people that often do these

898
00:48:58,248 --> 00:49:01,517
functions right now. So it's
really a very useful platform

899
00:49:01,517 --> 00:49:04,787
for being able to go between
these different stations that

900
00:49:04,787 --> 00:49:08,000
need that type of service, but
not be there all the time.

901
00:49:08,000 --> 00:49:11,000
Audrow Nash: Yeah. Where you can
go have them do something else.

902
00:49:11,000 --> 00:49:14,947
Pras Velagapudi: Right or man
multiple stations or something

903
00:49:14,947 --> 00:49:16,000
like that. Yeah.

904
00:49:16,000 --> 00:49:20,189
Audrow Nash: So one thing that
I'm thinking with your so it is

905
00:49:20,189 --> 00:49:23,500
a new ISO standard, a new
international standards

906
00:49:23,500 --> 00:49:27,486
organization standard. Why make
a standard I so the way I'm

907
00:49:27,486 --> 00:49:31,270
thinking, which I'm happy you
guys are making a standard

908
00:49:31,270 --> 00:49:35,257
because I think it'll improve
everything for everyone, but,

909
00:49:35,257 --> 00:49:39,446
why? Like I'm imagining it could
be a competitive advantage. I

910
00:49:39,446 --> 00:49:43,297
don't know what the other
humanoid companies are doing or

911
00:49:43,297 --> 00:49:47,351
dynamic robot companies. But.
And maybe having a third party

912
00:49:47,351 --> 00:49:51,203
validate things. I. Can you tell
me some of the strategic

913
00:49:51,203 --> 00:49:55,392
thinking behind it? Because I'm
wondering if you didn't make a

914
00:49:55,392 --> 00:49:59,176
standard and just had an
internal secret sauce for these

915
00:49:59,176 --> 00:50:03,095
safety things. Maybe it would be
a significant competitive

916
00:50:03,095 --> 00:50:06,000
advantage. But how do you think
about this?

917
00:50:06,000 --> 00:50:08,879
Pras Velagapudi: Yeah. So this
is standard. I mean, to be

918
00:50:08,879 --> 00:50:12,061
clear. Right. We still need the
secret sauce about how us to do

919
00:50:12,061 --> 00:50:14,788
it. And anyone else that's
trying to do it is going to

920
00:50:14,788 --> 00:50:17,566
achieve the standard. The
standard itself is laying out

921
00:50:17,566 --> 00:50:20,596
the rules, right? The rules for
what you need to do. So what

922
00:50:20,596 --> 00:50:21,000
they're.

923
00:50:21,000 --> 00:50:23,653
Audrow Nash: Probably well
thought out. And that saves a

924
00:50:23,653 --> 00:50:26,589
lot of early work and back and
forth with customers and things

925
00:50:26,589 --> 00:50:29,574
like this, which I'm probably,
I'm sure you did in creating the

926
00:50:29,574 --> 00:50:30,000
standard.

927
00:50:30,000 --> 00:50:33,620
Pras Velagapudi: Right. But one
of the reasons that we're

928
00:50:33,620 --> 00:50:37,494
pursuing it is because in order
to deploy robots outside of a

929
00:50:37,494 --> 00:50:41,432
work cell, we already need to
meet standards, right? So within

930
00:50:41,432 --> 00:50:45,497
the US, you know, OSHA, while it
doesn't specifically, you know,

931
00:50:45,497 --> 00:50:48,863
name a certain standard, is
basically putting out the

932
00:50:48,863 --> 00:50:52,865
expectation that any automation
deployed within the US is going

933
00:50:52,865 --> 00:50:56,485
to meet a set of prevailing
standards. And so you already

934
00:50:56,485 --> 00:51:00,232
have ISO standards for things
like the safety of machinery,

935
00:51:00,232 --> 00:51:04,106
the placement of stocks. And
you're already required to be in

936
00:51:04,106 --> 00:51:08,044
compliance with them. They just
don't have good specifications

937
00:51:08,044 --> 00:51:11,982
for what it takes for our robot
to be in compliance with them.

938
00:51:11,982 --> 00:51:15,920
Yeah. Outside of a work. So it's
really that in order to be in

939
00:51:15,920 --> 00:51:19,540
production in the US and
especially in Europe, where it's

940
00:51:19,540 --> 00:51:23,224
a lot more important, it's
basically a requirement for you

941
00:51:23,224 --> 00:51:27,288
to be able to certify your robot
at least as partially completed

942
00:51:27,288 --> 00:51:30,655
machinery, to go into a
production environment. Well,

943
00:51:30,655 --> 00:51:34,465
you really need to have the
document that specifies what you

944
00:51:34,465 --> 00:51:38,213
need to do to meet that, right?
Because otherwise you don't

945
00:51:38,213 --> 00:51:41,769
really have a path forward to
say, hey, I'm meeting this

946
00:51:41,769 --> 00:51:45,580
requirement. I'm, I'm complying
with, industry standards for

947
00:51:45,580 --> 00:51:49,010
best practices here because
there, there aren't really

948
00:51:49,010 --> 00:51:52,439
completely defined industry
standards. So for us, it's

949
00:51:52,439 --> 00:51:56,123
really important that we get
these standards out there and

950
00:51:56,123 --> 00:51:59,997
written and specified so that
with our customers together, we

951
00:51:59,997 --> 00:52:04,062
can go out and ensure that we're
in compliance to meet the legal

952
00:52:04,062 --> 00:52:08,000
obligations that we have, when
operating in the U.S or abroad.

953
00:52:08,000 --> 00:52:11,584
Audrow Nash: How do you think it
works with other humanoid

954
00:52:11,584 --> 00:52:15,169
companies? Well you say like
figure with their I think BMW

955
00:52:15,169 --> 00:52:19,000
partnership for Tesla with their
one or the Tesla or one X or.

956
00:52:19,000 --> 00:52:22,230
Pras Velagapudi: Yeah. So it
right now the way that pretty

957
00:52:22,230 --> 00:52:25,460
much all humanoids can be
deployed is kind of in two ways.

958
00:52:25,460 --> 00:52:28,579
One is within the US you can
deploy them you know not it

959
00:52:28,579 --> 00:52:31,530
directly into production
applications. You can deploy

960
00:52:31,530 --> 00:52:34,816
them into basically research
environments or other types of

961
00:52:34,816 --> 00:52:38,269
new innovation spaces. And those
don't have the same standards

962
00:52:38,269 --> 00:52:41,499
requirements. Obviously, you
still want to make the robots

963
00:52:41,499 --> 00:52:44,785
safe, but you're not being
specifically directed to like an

964
00:52:44,785 --> 00:52:47,847
ISO standard for, for an
industrial standard, if you're

965
00:52:47,847 --> 00:52:51,412
deploying it in essentially like
a research lab or an innovation

966
00:52:51,412 --> 00:52:54,976
part of a of a facility. And so
in a lot of cases, and certainly

967
00:52:54,976 --> 00:52:58,094
even in our earlier POCs, we'd
often be deployed into an

968
00:52:58,094 --> 00:53:01,324
innovation center or into a
special facility that has some

969
00:53:01,324 --> 00:53:03,719
portion of the facility
essentially running

970
00:53:03,719 --> 00:53:06,949
non-production workflows. And
the difference there is that

971
00:53:06,949 --> 00:53:09,845
when you're deploying in a
research environment, the

972
00:53:09,845 --> 00:53:13,242
requirements are a little bit
different because your exposure

973
00:53:13,242 --> 00:53:16,416
is a little bit different in
terms of the people that you

974
00:53:16,416 --> 00:53:19,535
expect to be around the robot,
and how the robot's being

975
00:53:19,535 --> 00:53:22,821
operated. When we're talking
about why it's so important to

976
00:53:22,821 --> 00:53:25,884
meet industrial safety
standards, part of it is because

977
00:53:25,884 --> 00:53:29,002
once a robot's deployed in a
facility working full shift

978
00:53:29,002 --> 00:53:32,511
work, the types of people that
are around it and their level of

979
00:53:32,511 --> 00:53:35,630
training with respect to the
robot is different, and the

980
00:53:35,630 --> 00:53:38,692
amount of attention that they're
paying to the robot is

981
00:53:38,692 --> 00:53:42,145
different. These are people who
are now they're not there just

982
00:53:42,145 --> 00:53:45,709
to work with the robot. It's not
like an innovation center where

983
00:53:45,709 --> 00:53:48,828
the whole point is to look at
the robot and see how it's

984
00:53:48,828 --> 00:53:52,169
doing. They're trying to do
their own work, right? Their job

985
00:53:52,169 --> 00:53:55,455
is not to follow the robot
around. It might be to walk past

986
00:53:55,455 --> 00:53:58,685
the robot and do something else.
Right. And so they're not

987
00:53:58,685 --> 00:54:01,860
necessarily paying the same
level of attention, certainly

988
00:54:01,860 --> 00:54:05,201
over 8 or 16 hours. And so for
sure, that's why it's so much

989
00:54:05,201 --> 00:54:07,985
more important that you're
ensuring that the robot

990
00:54:07,985 --> 00:54:11,271
performed a certain level of
reliability and how it's doing

991
00:54:11,271 --> 00:54:14,780
the safety. So there's this one
aspect which is that, you know,

992
00:54:14,780 --> 00:54:18,065
you can deploy into different
environments that are not the

993
00:54:18,065 --> 00:54:20,738
industrial production
environments. And a lot of

994
00:54:20,738 --> 00:54:23,857
people start out there. The
second piece is that you can

995
00:54:23,857 --> 00:54:27,031
deploy humanoid robots, and
we're doing this right now in

996
00:54:27,031 --> 00:54:30,540
conjunction with the work. So if
you have another system that's

997
00:54:30,540 --> 00:54:34,048
around the robot, that's really
what's keeping humans safe. And

998
00:54:34,048 --> 00:54:37,000
maybe that is plexi glass panels
or, a light curtain.

999
00:54:37,000 --> 00:54:38,000
Audrow Nash: Or something that
bounds it.

1000
00:54:38,000 --> 00:54:41,371
Pras Velagapudi: Then you can do
it right. You can put that

1001
00:54:41,371 --> 00:54:44,571
around the robot and then
achieve, some, compliance with

1002
00:54:44,571 --> 00:54:47,886
safety standards, at least in
North America right now. And

1003
00:54:47,886 --> 00:54:51,429
that's why you see that that's
what a lot of people are doing.

1004
00:54:51,429 --> 00:54:54,914
They'll deploy into parts of
facilities, or they'll deploy in

1005
00:54:54,914 --> 00:54:57,943
use cases where the robots
really, not moving between

1006
00:54:57,943 --> 00:55:01,543
stations or it's moving around
within some restricted zone. And

1007
00:55:01,543 --> 00:55:05,143
that's how we operate right now
at GCC. So we have a restricted

1008
00:55:05,143 --> 00:55:08,571
area that the robot operates
within. And it has a link to an

1009
00:55:08,571 --> 00:55:12,000
external safety system that
helps guarantee that we can shut

1010
00:55:12,000 --> 00:55:15,543
down the robot from outside of
that work cell if we need to go

1011
00:55:15,543 --> 00:55:16,000
into it.

1012
00:55:16,000 --> 00:55:19,587
Audrow Nash: Cool. Do you do you
know if any of the other

1013
00:55:19,587 --> 00:55:23,552
companies making humanoids have
their robots in production? And

1014
00:55:23,552 --> 00:55:27,517
if they do, they're probably in
work cells, I would assume. But

1015
00:55:27,517 --> 00:55:31,231
yeah. What do you what do you
think is the current state of

1016
00:55:31,231 --> 00:55:34,000
humanoid robot companies other
than Agility?

1017
00:55:34,000 --> 00:55:37,601
Pras Velagapudi: I think that it
basically because of the way

1018
00:55:37,601 --> 00:55:41,260
that the standards are defined,
everybody has to at best be in

1019
00:55:41,260 --> 00:55:44,683
production in a work cell, or if
they're not in the works,

1020
00:55:44,683 --> 00:55:48,166
they'll be in a non-production
environment. And in general,

1021
00:55:48,166 --> 00:55:51,707
that's just everybody has to
follow the same rules there, at

1022
00:55:51,707 --> 00:55:55,426
least within North America. And
Europe. Yeah, right. I think we

1023
00:55:55,426 --> 00:55:58,908
often see, you know, one off,
demonstrations of technology.

1024
00:55:58,908 --> 00:56:02,627
And certainly we can teach them
to. Right. Like we did a recent

1025
00:56:02,627 --> 00:56:06,169
demo for, for Google, where we
work together using their new

1026
00:56:06,169 --> 00:56:09,828
Gemini models to go pick items
off a grocery shelf. Right. And

1027
00:56:09,828 --> 00:56:13,251
that was in our innovation
space. And the robot, you know,

1028
00:56:13,251 --> 00:56:16,911
grabbed a little basket. And I
think I think we told it to get

1029
00:56:16,911 --> 00:56:20,630
the ingredients for spaghetti or
something. And so it's picking

1030
00:56:20,630 --> 00:56:24,348
tomato sauce and pasta off the
shelf. Right. The robot could do

1031
00:56:24,348 --> 00:56:27,595
that. Right. But it can't do
that in production. Right.

1032
00:56:27,595 --> 00:56:30,782
That's not a production
capability. And that's in part

1033
00:56:30,782 --> 00:56:34,382
because that's not something
that it could do, without a work

1034
00:56:34,382 --> 00:56:37,452
cell around it. Right. And
that's the same for us as

1035
00:56:37,452 --> 00:56:40,816
everybody else. Anyone who's
doing this legally has to be

1036
00:56:40,816 --> 00:56:43,000
following the same sort of
practices.

1037
00:56:43,000 --> 00:56:46,939
Audrow Nash: Make sense. So you
needed to do this to enable a

1038
00:56:46,939 --> 00:56:51,007
lot of high potential things for
you guys to do this, to create

1039
00:56:51,007 --> 00:56:55,140
the standard, because there was
a gap in the existing standards,

1040
00:56:55,140 --> 00:56:59,079
because these robots are so new.
And with this kind of thing,

1041
00:56:59,079 --> 00:57:02,889
with this additional standard,
once you prove that once you

1042
00:57:02,889 --> 00:57:06,828
write the standard, then prove
that you need it, then you are

1043
00:57:06,828 --> 00:57:10,703
able to have your robot expand
it to a much wider, number of

1044
00:57:10,703 --> 00:57:14,448
applications, probably, or a
much larger application where

1045
00:57:14,448 --> 00:57:18,000
you are working outside of a
work cell for all of this.

1046
00:57:18,000 --> 00:57:21,377
Pras Velagapudi: Yeah. And
that's exactly why we went after

1047
00:57:21,377 --> 00:57:24,696
it, right now. Because what
we're really seeing is that as

1048
00:57:24,696 --> 00:57:28,188
we're moving into the space,
there's a lot of capability that

1049
00:57:28,188 --> 00:57:31,736
that we can show that others are
showing of the platforms. You

1050
00:57:31,736 --> 00:57:35,285
know, we could pick groceries
off of a shelf. You know, we can

1051
00:57:35,285 --> 00:57:38,661
we can walk around outside, you
know, Digit can walk around

1052
00:57:38,661 --> 00:57:42,324
outside. Yeah, totally. We don't
have the safety rules about how

1053
00:57:42,324 --> 00:57:45,873
one does that for a lot of these
environments. So we're really

1054
00:57:45,873 --> 00:57:48,792
seeing that safety is a key
gating item for getting

1055
00:57:48,792 --> 00:57:52,169
dynamically stable robots out
into the world. It's actually

1056
00:57:52,169 --> 00:57:55,717
probably going to be, lagging
behind these capability of these

1057
00:57:55,717 --> 00:57:59,323
to do things in space. So we'll
be able to pick groceries off a

1058
00:57:59,323 --> 00:58:02,814
shelf. We can pick groceries off
a shelf, but we have to come

1059
00:58:02,814 --> 00:58:06,248
back with, well, what are the
safety, safety cases for doing

1060
00:58:06,248 --> 00:58:09,625
that in a retail environment?
You know, what are the safety

1061
00:58:09,625 --> 00:58:12,830
cases for doing that in a
construction environment? In a

1062
00:58:12,830 --> 00:58:16,378
home eventually. Right. Those
have to be worked out. And homes

1063
00:58:16,378 --> 00:58:19,583
are particular challenging.
They're sort of hardest case

1064
00:58:19,583 --> 00:58:23,189
because not only do you have all
of the lack of structure of an

1065
00:58:23,189 --> 00:58:26,452
alarm about your home, but
mine's not the neatest. Right.

1066
00:58:26,452 --> 00:58:30,057
Chaos. Right, exactly. But, you
know, two to exactly your point

1067
00:58:30,057 --> 00:58:33,320
you were mentioning earlier,
you've got a small child and

1068
00:58:33,320 --> 00:58:36,525
you've got an animal in your
house, and yeah, in both of

1069
00:58:36,525 --> 00:58:40,130
those, as a small children and
animals. Pets, right, are things

1070
00:58:40,130 --> 00:58:43,564
that you have to be safe around.
If you were a robot and you

1071
00:58:43,564 --> 00:58:46,827
can't expect them to either be
trained or avoid a hazard,

1072
00:58:46,827 --> 00:58:50,375
right? You can't be like, no, we
trained all of the puppies to

1073
00:58:50,375 --> 00:58:53,809
never come close to the robot
like, you can't, I don't know,

1074
00:58:53,809 --> 00:58:57,415
I, I have to make my robot never
do anything that could cause a

1075
00:58:57,415 --> 00:59:00,906
hazard to your puppy, right?
That's the way. Oh, yeah. Right.

1076
00:59:00,906 --> 00:59:04,340
And that's a pretty high bar.
That's way harder than in a in

1077
00:59:04,340 --> 00:59:06,000
an industrial facility where.

1078
00:59:06,000 --> 00:59:07,000
Audrow Nash: Very much pretty.

1079
00:59:07,000 --> 00:59:10,631
Pras Velagapudi: Much adults,
they are probably receiving some

1080
00:59:10,631 --> 00:59:14,146
sort of training. There might be
PPE restrictions on certain

1081
00:59:14,146 --> 00:59:17,250
areas anyway. And those are
reasonable. What's called

1082
00:59:17,250 --> 00:59:20,999
administrative controls that you
can put in your robot still has

1083
00:59:20,999 --> 00:59:24,337
to have a certain level of
safety, but you can do that in

1084
00:59:24,337 --> 00:59:27,793
conjunction with some of these
administrative controls. And

1085
00:59:27,793 --> 00:59:31,307
that really simplifies things
quite a bit. Right? Yeah. In a

1086
00:59:31,307 --> 00:59:34,939
home like you have to be taking
care of you, everything has to

1087
00:59:34,939 --> 00:59:38,394
be on your side. You can't
induce a hazard that could, that

1088
00:59:38,394 --> 00:59:41,791
could harm any, any, any of
these individuals. An adult, a

1089
00:59:41,791 --> 00:59:44,954
child, a pet. And you have to
assume that they are not

1090
00:59:44,954 --> 00:59:48,000
necessarily doing anything to
make your life easier.

1091
00:59:48,000 --> 00:59:51,144
Audrow Nash: Yeah. Very much. We
also have chickens and those

1092
00:59:51,144 --> 00:59:54,289
would be chaos near the robot.
And it's funny because they go

1093
00:59:54,289 --> 00:59:57,536
to you when you go outside. So
they would go right to the robot

1094
00:59:57,536 --> 00:59:58,000
probably.

1095
00:59:58,000 --> 00:59:59,855
Pras Velagapudi: Yeah. Exactly.
Just running around its feet you

1096
00:59:59,855 --> 01:00:00,000
know.

1097
01:00:00,000 --> 01:00:05,531
Audrow Nash: Hey hilarious.
Okay. That's a very interesting

1098
01:00:05,531 --> 01:00:11,063
thing. So do you think that how
do you imagine these safety

1099
01:00:11,063 --> 01:00:16,406
standards? So is this ISO
standard, are you working on it

1100
01:00:16,406 --> 01:00:18,000
or is it done or.

1101
01:00:18,000 --> 01:00:21,923
Pras Velagapudi: So we just got
what's the status of it. Yeah.

1102
01:00:21,923 --> 01:00:25,909
We just got approved to create
the working group for it. So now

1103
01:00:25,909 --> 01:00:29,578
there is a working group, a
team, an international team of

1104
01:00:29,578 --> 01:00:33,185
ourselves and other, robot
companies and other industrial

1105
01:00:33,185 --> 01:00:36,981
automation companies that are
participating to define it and

1106
01:00:36,981 --> 01:00:40,714
build it and ratify it. So
that's where we are right now is

1107
01:00:40,714 --> 01:00:44,510
we've got the group together.
We've got approval to go after

1108
01:00:44,510 --> 01:00:48,622
this. Everyone's recognized that
this is an area that needs a new

1109
01:00:48,622 --> 01:00:52,228
standard. And so now we're
building it out together with,

1110
01:00:52,228 --> 01:00:54,000
other folks in the industry.

1111
01:00:54,000 --> 01:00:57,410
Audrow Nash: What does the
timeline look like for this? I

1112
01:00:57,410 --> 01:01:00,821
guess the whole thing would be
interesting to getting the

1113
01:01:00,821 --> 01:01:04,590
working. I guess it's the that's
logistics and you just have to

1114
01:01:04,590 --> 01:01:08,000
be approved. What what's the
timeline look like for this?

1115
01:01:08,000 --> 01:01:11,398
Pras Velagapudi: Yeah. So it
does take a it does take a

1116
01:01:11,398 --> 01:01:15,166
while. We do have the benefit
of, one of the key folks on our

1117
01:01:15,166 --> 01:01:18,749
team is Melanie Wise, who
largely did this work before for

1118
01:01:18,749 --> 01:01:22,517
the AR 1508 standard, which is
the one that is used for Amrs.

1119
01:01:22,517 --> 01:01:25,977
So luckily, having hers is
hugely valuable because she's

1120
01:01:25,977 --> 01:01:29,869
been through this before with a
different type of robot. And so

1121
01:01:29,869 --> 01:01:33,637
she's been working really hard,
with, with the other folks on

1122
01:01:33,637 --> 01:01:37,220
our safety team, like Kevin
Reese, who's also phenomenally

1123
01:01:37,220 --> 01:01:41,236
knowledgeable about the space to
basically build out the standard

1124
01:01:41,236 --> 01:01:45,127
and sort of get all the pieces
in place. But that is definitely

1125
01:01:45,127 --> 01:01:48,649
accelerating the process,
because it can take a number of

1126
01:01:48,649 --> 01:01:52,293
years for a safety standard to
really go through all of its

1127
01:01:52,293 --> 01:01:56,000
paces. But luckily, with some
experienced folks on the team.

1128
01:01:56,000 --> 01:01:57,000
Audrow Nash: It will be a bit.
It should be.

1129
01:01:57,000 --> 01:02:00,974
Pras Velagapudi: It should be
easier. And we really hope to be

1130
01:02:00,974 --> 01:02:04,949
able to get our vibes out in the
world and get them, CE marked

1131
01:02:04,949 --> 01:02:07,000
within the next couple of years.

1132
01:02:07,000 --> 01:02:09,000
Audrow Nash: See, is like, what
would a c?

1133
01:02:09,000 --> 01:02:13,289
Pras Velagapudi: So C that's a
marking that is used in, in the

1134
01:02:13,289 --> 01:02:17,164
European side of things, to
basically mark that you have

1135
01:02:17,164 --> 01:02:20,000
approval under certain sets of
standards.

1136
01:02:20,000 --> 01:02:23,368
Audrow Nash: So for the
standards so you'll make the

1137
01:02:23,368 --> 01:02:26,737
robots, the robots will be
capable of fulfilling the

1138
01:02:26,737 --> 01:02:30,688
standard. The standard will be
ratified at some point. Do you

1139
01:02:30,688 --> 01:02:34,316
have like do you have like an
optimistic and pessimistic

1140
01:02:34,316 --> 01:02:36,000
estimate for the standard?

1141
01:02:36,000 --> 01:02:40,295
Pras Velagapudi: I think it's
probably, you know, I think I

1142
01:02:40,295 --> 01:02:44,881
don't necessarily know exactly.
Melanie is the person who has a

1143
01:02:44,881 --> 01:02:49,395
much better intuition of it. But
I mean, we're really talking,

1144
01:02:49,395 --> 01:02:53,617
you know, single Digit year,
couple of years, few years at

1145
01:02:53,617 --> 01:02:55,000
most. We're really.

1146
01:02:55,000 --> 01:02:57,000
Audrow Nash: It's like less than
five, hopefully.

1147
01:02:57,000 --> 01:02:59,803
Pras Velagapudi: Yeah,
definitely. I mean, I'd say, you

1148
01:02:59,803 --> 01:03:02,809
know, 1 to 3 is what I would
hope. Right? Like, it might be

1149
01:03:02,809 --> 01:03:05,713
quite quick, to get to an
initial version of the standard

1150
01:03:05,713 --> 01:03:08,618
and then we might. Yeah, you
know, see where it goes from

1151
01:03:08,618 --> 01:03:11,777
there or maybe, you know, takes
a little bit longer, but we'll

1152
01:03:11,777 --> 01:03:14,834
have enough of it sort of worked
out that we'll have maybe a

1153
01:03:14,834 --> 01:03:17,637
working draft or something that
we can start getting in

1154
01:03:17,637 --> 01:03:20,490
compliance with. But that's
really more of her specialty

1155
01:03:20,490 --> 01:03:21,000
than mine.

1156
01:03:21,000 --> 01:03:25,210
Audrow Nash: Yeah. Okay. Yeah.
I've had Melanie on the podcast

1157
01:03:25,210 --> 01:03:29,420
a few times. Melanie. Awesome.
Hearing her perspective on that

1158
01:03:29,420 --> 01:03:33,562
would be very interesting, but
super cool. I really like that

1159
01:03:33,562 --> 01:03:37,500
you guys are working on the
standard, do you? So with your

1160
01:03:37,500 --> 01:03:41,574
knowledge of standards, I know
that she's kind of like she's

1161
01:03:41,574 --> 01:03:45,716
been leading, right? Like she
has the experience with the Amr

1162
01:03:45,716 --> 01:03:49,790
one, but for robots to enter the
home, or say like a grocery

1163
01:03:49,790 --> 01:03:53,660
store or places where there are
more people that are less

1164
01:03:53,660 --> 01:03:58,006
trained, how do you imagine that
will occur? Do you imagine that

1165
01:03:58,006 --> 01:04:02,216
we're going to wait on standards
like this? Or will it be that

1166
01:04:02,216 --> 01:04:05,883
companies just kind of go with
it and suffer the legal

1167
01:04:05,883 --> 01:04:09,957
consequences when they, like,
step on a kid or something? Or

1168
01:04:09,957 --> 01:04:14,302
how do you how do you think this
space, like, how do you imagine

1169
01:04:14,302 --> 01:04:16,000
the space moving forward?

1170
01:04:16,000 --> 01:04:19,568
Pras Velagapudi: Yeah, that's a
good question. I think that we

1171
01:04:19,568 --> 01:04:23,194
should be building out standards
as we go. I do think that once

1172
01:04:23,194 --> 01:04:26,475
we have one good reference
point, that does make it a lot

1173
01:04:26,475 --> 01:04:30,043
easier to move to the next one
in the next one, because we can

1174
01:04:30,043 --> 01:04:33,266
largely build off of that or
reference it. Right? It can

1175
01:04:33,266 --> 01:04:36,777
leverage it. I would hope that
we're not going down the route

1176
01:04:36,777 --> 01:04:40,173
of people just winging it and
trying to get insured anyway,

1177
01:04:40,173 --> 01:04:43,281
and then having something
horrible show up in the news

1178
01:04:43,281 --> 01:04:46,561
about, a robot stepping on
someone's toes or something. I

1179
01:04:46,561 --> 01:04:48,000
don't want us to. I mean.

1180
01:04:48,000 --> 01:04:50,000
Audrow Nash: That already
happened in some places, but.

1181
01:04:50,000 --> 01:04:53,588
Pras Velagapudi: It's definitely
happened in some, at some

1182
01:04:53,588 --> 01:04:57,175
conferences and trade shows
where they've been showing off

1183
01:04:57,175 --> 01:05:01,072
smaller humanoids doing stuff,
because those humanoids largely,

1184
01:05:01,072 --> 01:05:04,660
run without too much safety
system. Yes. Oversight at all.

1185
01:05:04,660 --> 01:05:08,433
For sure. But I don't want that
to happen. I mean, I think it

1186
01:05:08,433 --> 01:05:12,144
could be a big, black mark on
the industry. If that's how we

1187
01:05:12,144 --> 01:05:16,165
approach the problem. I'd really
love to, get there in a safe way

1188
01:05:16,165 --> 01:05:19,258
and with a pretty good
understanding of what we're

1189
01:05:19,258 --> 01:05:22,907
doing now. I think there's two
pieces that might accelerate

1190
01:05:22,907 --> 01:05:26,742
that. One is that obviously with
an example reference of like,

1191
01:05:26,742 --> 01:05:30,268
here's how we're doing it in
industry X going to industry

1192
01:05:30,268 --> 01:05:34,103
wide, it's a lot simpler like we
said before. And then I think

1193
01:05:34,103 --> 01:05:37,938
the other thing is if we really
show a lot of capability, then

1194
01:05:37,938 --> 01:05:41,155
there's a lot of industry
pressure to figure out the

1195
01:05:41,155 --> 01:05:44,742
standards and that sort of like
naturally, accelerates the

1196
01:05:44,742 --> 01:05:48,454
development and the, the working
groups and really the whole

1197
01:05:48,454 --> 01:05:52,165
process, really if a lot of
people want a thing, right, then

1198
01:05:52,165 --> 01:05:55,629
it tends to move faster. And so
I think we will see that

1199
01:05:55,629 --> 01:05:59,464
hopefully that the capabilities
that we're able to demonstrate

1200
01:05:59,464 --> 01:06:02,928
in some sense, builds a lot of,
goodwill and interest in

1201
01:06:02,928 --> 01:06:06,701
building out safety frameworks
that can support doing it, and

1202
01:06:06,701 --> 01:06:10,660
maybe there'll be interim stuff
that comes out of it, similar to

1203
01:06:10,660 --> 01:06:14,186
how, drones initially had some
experimental airworthiness

1204
01:06:14,186 --> 01:06:17,711
certificates and things like
that that let people try out

1205
01:06:17,711 --> 01:06:20,742
things under special
qualifications, or that, you

1206
01:06:20,742 --> 01:06:24,701
know, we'll see, changes in the
regulatory space to kind of like

1207
01:06:24,701 --> 01:06:28,351
provide, ways to phase in the
technology in earlier stages.

1208
01:06:28,351 --> 01:06:32,062
Right? In controlled settings
and things like that. We might

1209
01:06:32,062 --> 01:06:35,649
see some of that and we might
also just see that, hey, you

1210
01:06:35,649 --> 01:06:39,361
know, because it's such a high
demand that the standards for

1211
01:06:39,361 --> 01:06:43,196
things like, retail industry,
just move really quickly because

1212
01:06:43,196 --> 01:06:47,031
there's so much interest in it
that a lot of people will spend

1213
01:06:47,031 --> 01:06:50,000
a lot of energy, working on it.
Yeah. Moving it.

1214
01:06:50,000 --> 01:06:54,132
Audrow Nash: Interesting. How do
you feel like, you feel like

1215
01:06:54,132 --> 01:06:58,129
there's significant energy and,
desire for robots to be out

1216
01:06:58,129 --> 01:07:01,923
there by these different
industries? I know you guys are

1217
01:07:01,923 --> 01:07:05,445
most focused on material
handling. But do you have a

1218
01:07:05,445 --> 01:07:09,239
feeling for these other
industries, like, is there a lot

1219
01:07:09,239 --> 01:07:11,000
of pull from other places?

1220
01:07:11,000 --> 01:07:14,434
Pras Velagapudi: I think there's
a lot of interest. Yes. I mean,

1221
01:07:14,434 --> 01:07:17,439
we certainly see a lot of
interest. We have folks coming

1222
01:07:17,439 --> 01:07:20,820
to us all the time asking, can
you use Digit over here? Can you

1223
01:07:20,820 --> 01:07:22,000
use it in my facility?

1224
01:07:22,000 --> 01:07:23,000
Audrow Nash: That's a spot.

1225
01:07:23,000 --> 01:07:26,565
Pras Velagapudi: Oh yeah, it's
great. It's great to be there.

1226
01:07:26,565 --> 01:07:30,189
But unfortunately right now we
have to focus on a lot of those

1227
01:07:30,189 --> 01:07:33,872
folks, which is unfortunate. And
we're hoping making a big list

1228
01:07:33,872 --> 01:07:37,554
of folks that we'd like to come
back to, you know, in a year or

1229
01:07:37,554 --> 01:07:40,944
two where we could say, hey,
yeah, actually, now we can go

1230
01:07:40,944 --> 01:07:44,334
look at that for you. And I
think that's, you know, that's

1231
01:07:44,334 --> 01:07:48,017
part of it. There's no shortage
of interest in the platform. It

1232
01:07:48,017 --> 01:07:51,640
can it can do a lot and it can
do a lot very easily. In easily

1233
01:07:51,640 --> 01:07:54,855
in that sense, not, obviously
there's work to get. It's

1234
01:07:54,855 --> 01:07:58,537
capable robots, but like, it's
yeah, it's not the same level of

1235
01:07:58,537 --> 01:08:01,928
work as, as deploying another
type of automation where you

1236
01:08:01,928 --> 01:08:05,376
might really have to. Yeah,
design what you're doing around

1237
01:08:05,376 --> 01:08:09,000
it. You can kind of slot Digit
into the process that you have.

1238
01:08:09,000 --> 01:08:12,757
Audrow Nash: Interesting. What
do you think? So kind of a

1239
01:08:12,757 --> 01:08:16,976
tangent. But it seems like a lot
of my interviews, it eventually

1240
01:08:16,976 --> 01:08:20,733
ends up that companies are
planning on doing this kind of

1241
01:08:20,733 --> 01:08:24,885
thing. So if it sounds like you
guys are having a ton of demand

1242
01:08:24,885 --> 01:08:28,576
for things that you can't
service yourself, any interest

1243
01:08:28,576 --> 01:08:32,465
in, like a robot app store or
any sort of things where it's

1244
01:08:32,465 --> 01:08:36,354
like, you let developers make
applications for your app for

1245
01:08:36,354 --> 01:08:40,177
your platform to solve some of
the problems that customers

1246
01:08:40,177 --> 01:08:44,000
might have. How would you think
around that kind of thing?

1247
01:08:44,000 --> 01:08:47,000
Pras Velagapudi: So we are
definitely going to be.

1248
01:08:47,000 --> 01:08:48,000
Audrow Nash: Moving in that.

1249
01:08:48,000 --> 01:08:51,297
Pras Velagapudi: Oh, we're
already basically moving in that

1250
01:08:51,297 --> 01:08:54,818
direction. So one of the things
we launched, last year was Arc,

1251
01:08:54,818 --> 01:08:58,060
which is our cloud management
platform for the robots. And

1252
01:08:58,060 --> 01:09:00,966
within arc, you can basically
build out what we call

1253
01:09:00,966 --> 01:09:04,263
workflows, which is essentially
telling the robot how to do

1254
01:09:04,263 --> 01:09:07,337
different types of work by
graphically putting together

1255
01:09:07,337 --> 01:09:10,578
building blocks. And those
building blocks are essentially

1256
01:09:10,578 --> 01:09:14,043
skills they might be like pick
up a tote or look for something

1257
01:09:14,043 --> 01:09:17,564
in the environment or move to a
location. But they're basically

1258
01:09:17,564 --> 01:09:21,197
drag and drop high level skills,
and sometimes they're powered by

1259
01:09:21,197 --> 01:09:24,606
AI models that are trained on
things. Sometimes they're their

1260
01:09:24,606 --> 01:09:28,238
own, sort of specialized motions
like docking sequences or things

1261
01:09:28,238 --> 01:09:31,759
like that. But we're designing
that platform with the intent of

1262
01:09:31,759 --> 01:09:35,168
being able to hand it off to
integrators and customers in the

1263
01:09:35,168 --> 01:09:38,354
future and let them edit the
workflows that Digit does on

1264
01:09:38,354 --> 01:09:41,763
their own. They're very high
level. They're at the level that

1265
01:09:41,763 --> 01:09:44,725
you might write a work
instruction for a person doing

1266
01:09:44,725 --> 01:09:48,358
the same task, which is like, go
to this work cell and, you know,

1267
01:09:48,358 --> 01:09:51,711
look for a tote over here, you
know, read the barcode on it,

1268
01:09:51,711 --> 01:09:55,232
figure out which of these three
places it goes and move it over

1269
01:09:55,232 --> 01:09:58,473
there. Right. You can drag and
drop that flow together. Or

1270
01:09:58,473 --> 01:10:01,715
maybe in the future you'll be
able to use a generic AI and

1271
01:10:01,715 --> 01:10:03,000
just type that request.

1272
01:10:03,000 --> 01:10:04,000
Audrow Nash: In and it'll what
you want to drag.

1273
01:10:04,000 --> 01:10:07,256
Pras Velagapudi: The box the
blocks together for you. Right.

1274
01:10:07,256 --> 01:10:10,458
Super cool. But that's our
that's 100% our intent. We don't

1275
01:10:10,458 --> 01:10:13,769
want Agility employees to have
to touch every deployed robot.

1276
01:10:13,769 --> 01:10:17,296
We want to be able to ship you a
kit that, you know, an installer

1277
01:10:17,296 --> 01:10:20,498
can put in for you, and that you
can configure on your own.

1278
01:10:20,498 --> 01:10:23,483
That's our that's our goal.
Right. And so, you know, we

1279
01:10:23,483 --> 01:10:27,011
actually used to have, you know,
this, this internal tagline like

1280
01:10:27,011 --> 01:10:30,213
we're going to create the App
Store for labor. But then now

1281
01:10:30,213 --> 01:10:33,469
with AI, it's like, well, you
don't really have to write the

1282
01:10:33,469 --> 01:10:36,726
apps anymore. So now we're just
a store for labor, you know,

1283
01:10:36,726 --> 01:10:40,199
maybe, that's the intent, right?
We don't we want you to be able

1284
01:10:40,199 --> 01:10:43,618
to we want to democratize, you
know, democratize the automation

1285
01:10:43,618 --> 01:10:46,983
of labor activities. Right? We
don't want people to have to do

1286
01:10:46,983 --> 01:10:50,293
work, and we want people to have
to spend time working on the

1287
01:10:50,293 --> 01:10:53,387
automation either. Right. We
want you to just get a get a

1288
01:10:53,387 --> 01:10:56,589
piece of equipment, get a kit
that you can set up that just

1289
01:10:56,589 --> 01:10:58,000
does these things for you.

1290
01:10:58,000 --> 01:11:02,043
Audrow Nash: Yeah, yeah. One of
the things I've been thinking

1291
01:11:02,043 --> 01:11:06,086
that I would imagine robotics
will arrive at some point, will

1292
01:11:06,086 --> 01:11:09,997
be that you can have developers
that are building robots or

1293
01:11:09,997 --> 01:11:13,974
using robots for applications,
but they're not building them

1294
01:11:13,974 --> 01:11:17,885
themselves. Kind of like what
you're describing. But one of

1295
01:11:17,885 --> 01:11:21,994
the things that I imagine is in
specific domains, you're going

1296
01:11:21,994 --> 01:11:25,573
to find that Digit needs to
interact with something or

1297
01:11:25,573 --> 01:11:29,683
there's something that's really
hard perception wise, that you

1298
01:11:29,683 --> 01:11:33,660
guys haven't solved, because
it's another thing that you run

1299
01:11:33,660 --> 01:11:37,769
into frequently or some thorny
algorithm or something. Are you

1300
01:11:37,769 --> 01:11:42,012
adding developer support so you
can have like a module that, has

1301
01:11:42,012 --> 01:11:44,000
custom code written in it, or.

1302
01:11:44,000 --> 01:11:46,333
Pras Velagapudi: How does all
that work? Yeah. So that's

1303
01:11:46,333 --> 01:11:48,667
something that we've been
playing around. It's something

1304
01:11:48,667 --> 01:11:49,000
that in.

1305
01:11:49,000 --> 01:11:50,000
Audrow Nash: Turn that's hard to
do.

1306
01:11:50,000 --> 01:11:53,487
Pras Velagapudi: Yeah. Well it's
I mean, I'm sure you're very

1307
01:11:53,487 --> 01:11:57,088
familiar with all of the various
attempts at like, low code. No

1308
01:11:57,088 --> 01:12:00,232
code. Robot. Oh, yeah. That's
right. And yeah, there's,

1309
01:12:00,232 --> 01:12:03,777
there's always the thorny edge
cases of like, here's the thing

1310
01:12:03,777 --> 01:12:07,149
that isn't one of the building
blocks that you offer, but I

1311
01:12:07,149 --> 01:12:10,122
need you. Right. There's two
things that are kind of

1312
01:12:10,122 --> 01:12:13,666
interesting about it for us from
our perspective. The first is

1313
01:12:13,666 --> 01:12:17,096
that we certainly can do the
route of, hey, we'll give you a

1314
01:12:17,096 --> 01:12:20,469
little sandbox block. That is
something that you write your

1315
01:12:20,469 --> 01:12:24,013
own, low level API calls into at
some point. Right? Right now,

1316
01:12:24,013 --> 01:12:27,271
it's not super easy for people
to do that because Digits,

1317
01:12:27,271 --> 01:12:30,873
control system and so forth are
relatively complex. But this is

1318
01:12:30,873 --> 01:12:34,074
where the second piece comes in,
which is now that we're

1319
01:12:34,074 --> 01:12:37,732
leveraging AI models a lot more
for Digits, low level skills and

1320
01:12:37,732 --> 01:12:40,819
controls, that's actually
changing the game of how you

1321
01:12:40,819 --> 01:12:44,249
specify things in this category
of new skills in particular.

1322
01:12:44,249 --> 01:12:47,393
Now, a lot of what we're
starting to explore are things

1323
01:12:47,393 --> 01:12:50,766
like physically demonstrating
how to do a task to Digit and

1324
01:12:50,766 --> 01:12:53,910
then turning into a block.
Right. And that's where it's

1325
01:12:53,910 --> 01:12:57,397
actually quite different from
anything we've ever seen in the

1326
01:12:57,397 --> 01:13:00,941
past where, hey, if you don't
have a block you like, you might

1327
01:13:00,941 --> 01:13:04,314
not author it by using code
anymore, right? Which is really

1328
01:13:04,314 --> 01:13:05,000
interesting.

1329
01:13:05,000 --> 01:13:07,921
Audrow Nash: What a trip. Yeah,
just a learning from

1330
01:13:07,921 --> 01:13:11,067
demonstration based approach,
but it's hard to make them

1331
01:13:11,067 --> 01:13:14,551
general. But there's a lot of
good work going on in that space

1332
01:13:14,551 --> 01:13:15,000
I think.

1333
01:13:15,000 --> 01:13:17,222
Pras Velagapudi: Yeah. And then
there's also the question of

1334
01:13:17,222 --> 01:13:19,593
exactly how general you need to
make it, because it can still be

1335
01:13:19,593 --> 01:13:21,000
a building block in a workflow.
Right?

1336
01:13:21,000 --> 01:13:24,167
Audrow Nash: So yeah, if you
take a bunch of actions that

1337
01:13:24,167 --> 01:13:27,444
already exist and just compose
them with AI that watched it

1338
01:13:27,444 --> 01:13:28,000
say, yeah.

1339
01:13:28,000 --> 01:13:31,700
Pras Velagapudi: Or you just use
AI for one piece of the actions,

1340
01:13:31,700 --> 01:13:35,230
right? I don't need to use AI to
like, figure out a new way to

1341
01:13:35,230 --> 01:13:38,589
move from point A to point B,
like that's already a skill I

1342
01:13:38,589 --> 01:13:42,290
have, but maybe I'll use it for,
you know, stacking a new type of

1343
01:13:42,290 --> 01:13:45,705
toad or interacting with a new
piece of machinery. I haven't

1344
01:13:45,705 --> 01:13:49,064
seen that before, but I can
demonstrate that and make a new

1345
01:13:49,064 --> 01:13:52,082
skill. So that's all stuff that
we're, exploring very

1346
01:13:52,082 --> 01:13:55,156
aggressively. It's not in
production. Right. But it is

1347
01:13:55,156 --> 01:13:58,515
stuff that we're exploring very
aggressively internally and

1348
01:13:58,515 --> 01:14:01,988
working with folks like Nvidia
and Google, using their latest

1349
01:14:01,988 --> 01:14:05,574
models on seeing how we can do
on this type of new pipeline for

1350
01:14:05,574 --> 01:14:08,819
creating new skills. And if
generalist models do start to

1351
01:14:08,819 --> 01:14:12,121
work well, we're selling a robot
solution. Right. So we're

1352
01:14:12,121 --> 01:14:14,000
perfectly happy to be
leveraging.

1353
01:14:14,000 --> 01:14:15,000
Audrow Nash: You're very well
positioned right.

1354
01:14:15,000 --> 01:14:18,152
Pras Velagapudi: We can leverage
any generalist models that come

1355
01:14:18,152 --> 01:14:21,254
out because as long as they help
us do something, we're able to

1356
01:14:21,254 --> 01:14:24,159
use it as part of our skill set.
But at the end of the day,

1357
01:14:24,159 --> 01:14:27,261
someone has to ship you a robot
in a box and have a way for you

1358
01:14:27,261 --> 01:14:30,117
to get that up and running and
connected to your warehouse

1359
01:14:30,117 --> 01:14:33,023
management system or whatever
orchestration platform you've

1360
01:14:33,023 --> 01:14:35,928
got. And we're doing all of
that. We're the person that you

1361
01:14:35,928 --> 01:14:38,932
can call on the phone that'll
ship you in your arm when yours

1362
01:14:38,932 --> 01:14:41,000
breaks, or whatever you might
need. Right?

1363
01:14:41,000 --> 01:14:44,553
Audrow Nash: Yeah, yeah, yeah.
It's a good moat to have the

1364
01:14:44,553 --> 01:14:48,407
physical hardware for this kind
of thing. Because there's one AI

1365
01:14:48,407 --> 01:14:51,960
generalist solution, and then
another one comes up and it's

1366
01:14:51,960 --> 01:14:55,693
like they don't have any moat
that you just switch from one to

1367
01:14:55,693 --> 01:14:58,945
the next. But if you are
actually making the hardware,

1368
01:14:58,945 --> 01:15:02,558
that's a much stronger position
to be in. And any AI general

1369
01:15:02,558 --> 01:15:06,231
solution that comes along is
like, oh, great, throw it on our

1370
01:15:06,231 --> 01:15:09,964
thing. It's real easy. That's
great. So hell yeah. How are you

1371
01:15:09,964 --> 01:15:13,457
viewing all this AI activity?
And I guess breaking it down

1372
01:15:13,457 --> 01:15:16,889
into different areas. And then
how do you think about it?

1373
01:15:16,889 --> 01:15:20,623
Melanie when I talked to her, it
was one of the early podcasts

1374
01:15:20,623 --> 01:15:24,236
when I started this podcast, was
thinking of it kind of like

1375
01:15:24,236 --> 01:15:27,909
really good search. How are you
thinking? And that's specific

1376
01:15:27,909 --> 01:15:31,643
for the language models in this
kind of thing. And I know that

1377
01:15:31,643 --> 01:15:34,955
Transformers are really
interesting and they give you a

1378
01:15:34,955 --> 01:15:38,447
lot of, like, interesting
abilities across, say, vision or

1379
01:15:38,447 --> 01:15:42,000
all sorts of other domains. But
how are you thinking of it?

1380
01:15:42,000 --> 01:15:45,661
Pras Velagapudi: Yeah. So we've
definitely, been growing and

1381
01:15:45,661 --> 01:15:49,139
evolving our AI strategy
internally, and we use now a lot

1382
01:15:49,139 --> 01:15:52,739
more AI inside of the robot
than, than we did even when you

1383
01:15:52,739 --> 01:15:56,339
last talked to Melanie. Because
the field is it's moving so

1384
01:15:56,339 --> 01:15:59,634
rapidly and, and we're
benefiting from that greatly. I

1385
01:15:59,634 --> 01:16:03,173
think there's a couple of
different layers to it. At least

1386
01:16:03,173 --> 01:16:06,956
that's how we think about it, is
there's a couple of different

1387
01:16:06,956 --> 01:16:10,800
layers to the AI models that we
employ. And that comes not from

1388
01:16:10,800 --> 01:16:14,583
necessarily what the models can
do, but in terms of the places

1389
01:16:14,583 --> 01:16:18,183
that you use them. And what
types of requirements there are

1390
01:16:18,183 --> 01:16:21,844
around how you use them. So at
the very lowest levels of the

1391
01:16:21,844 --> 01:16:25,810
robot, there's the local idea of
what am I doing with my arms and

1392
01:16:25,810 --> 01:16:29,532
legs, right? Like, what do I
need to do to not fall over? And

1393
01:16:29,532 --> 01:16:33,315
what do I need to do to interact
with the environment to exert

1394
01:16:33,315 --> 01:16:37,037
forces on that? And that's a
controller that needs to be very

1395
01:16:37,037 --> 01:16:40,881
close to the robot and work at a
very high rate, because that's

1396
01:16:40,881 --> 01:16:44,420
the rate at which disturbances
to that model or definitely

1397
01:16:44,420 --> 01:16:48,020
happen rates that might be
hundreds of hertz, right? That's

1398
01:16:48,020 --> 01:16:51,742
running right on the robot at
hundreds of Hertz, right inside

1399
01:16:51,742 --> 01:16:54,000
of our control loops. Right?
Yeah, a.

1400
01:16:54,000 --> 01:16:55,000
Audrow Nash: Lot of times a.

1401
01:16:55,000 --> 01:16:57,985
Pras Velagapudi: Second. Right.
And so that is taking as inputs,

1402
01:16:57,985 --> 01:17:00,458
you know, some idea of what am I
trying to do with my

1403
01:17:00,458 --> 01:17:03,443
environment. And then it's going
to figure out what do I do with

1404
01:17:03,443 --> 01:17:06,382
my arms and legs. So the layer
above it, that's like a controls

1405
01:17:06,382 --> 01:17:09,181
AI model. Then a layer above
that is like a skills AI model,

1406
01:17:09,181 --> 01:17:11,000
which is really thinking, what
are you.

1407
01:17:11,000 --> 01:17:14,932
Audrow Nash: Doing at the, at
the so at that lowest level,

1408
01:17:14,932 --> 01:17:18,797
what kind of AI thing are you
doing? Like a lot of like I

1409
01:17:18,797 --> 01:17:23,068
talked to Jonathan Hurst at some
point and a lot of it was like

1410
01:17:23,068 --> 01:17:26,864
really good application of
existing control theory, if I

1411
01:17:26,864 --> 01:17:30,932
remember correctly. But maybe
that's not true any longer for

1412
01:17:30,932 --> 01:17:35,000
these kinds of low level command
or low level control loops.

1413
01:17:35,000 --> 01:17:36,000
Pras Velagapudi: Yeah.

1414
01:17:35,000 --> 01:17:37,432
Audrow Nash: So are you using
either and what kind of things

1415
01:17:37,432 --> 01:17:38,000
are you doing.

1416
01:17:38,000 --> 01:17:41,727
Pras Velagapudi: So we are using
AI there. And we do have, we do

1417
01:17:41,727 --> 01:17:44,871
have a good model based
controller. So like if you see

1418
01:17:44,871 --> 01:17:48,597
the robot working in production,
it's right now flipping between

1419
01:17:48,597 --> 01:17:51,800
the two. It actually has a
really good model predictive

1420
01:17:51,800 --> 01:17:55,003
controller. But then if things
get out of scope of that

1421
01:17:55,003 --> 01:17:58,438
controller, it actually switches
to a reinforcement learned

1422
01:17:58,438 --> 01:18:01,000
controller that's more robust,
that handles.

1423
01:18:01,000 --> 01:18:02,877
Audrow Nash: One that you just
perturb a lot and simulate the

1424
01:18:02,877 --> 01:18:03,000
one.

1425
01:18:03,000 --> 01:18:06,433
Pras Velagapudi: That's been
trained for a month. Yeah. It's

1426
01:18:06,433 --> 01:18:09,809
been trained against a huge
variety of disturbances. And so

1427
01:18:09,809 --> 01:18:13,357
it'll actually switch to that.
But what we're seeing over time

1428
01:18:13,357 --> 01:18:16,561
is that that RL model is
covering this pretty good, more

1429
01:18:16,561 --> 01:18:19,766
of what the model based
controller was originally tasked

1430
01:18:19,766 --> 01:18:23,027
with. And so we think that
that's so cool. Likely to sort

1431
01:18:23,027 --> 01:18:26,289
of subsume a lot of that
functionality in the future. Now

1432
01:18:26,289 --> 01:18:29,493
that RL model is it's a
relatively small model, but it's

1433
01:18:29,493 --> 01:18:32,812
one that's trained across just a
huge variety of different

1434
01:18:32,812 --> 01:18:35,959
environmental conditions,
especially around things like

1435
01:18:35,959 --> 01:18:39,507
ground contact forces and slip
and so forth. And it also takes

1436
01:18:39,507 --> 01:18:43,112
advantage of still knowing a lot
about the robot. So we do have

1437
01:18:43,112 --> 01:18:46,545
the advantage of because we
build the robot, we can actually

1438
01:18:46,545 --> 01:18:50,093
co-design the robot and the RL
together. Like sometimes we can

1439
01:18:50,093 --> 01:18:53,183
actually make some design
decisions to say, hey, we're

1440
01:18:53,183 --> 01:18:56,787
going to change this actuation
in our next version of the robot

1441
01:18:56,787 --> 01:19:00,106
because we want to make it so
that it's easier to model in

1442
01:19:00,106 --> 01:19:03,768
reinforcement learning. Or hey,
like, we can go get the actuator

1443
01:19:03,768 --> 01:19:06,000
parameters or change the control
layer.

1444
01:19:06,000 --> 01:19:07,000
Audrow Nash: Coupling is so
cool.

1445
01:19:07,000 --> 01:19:08,000
Pras Velagapudi: Yeah, it.

1446
01:19:08,000 --> 01:19:09,000
Audrow Nash: Is very cool.

1447
01:19:09,000 --> 01:19:12,114
Pras Velagapudi: And it leads to
very good. All right. That sim

1448
01:19:12,114 --> 01:19:15,030
to real type transfer. So we can
do what's called zero shot

1449
01:19:15,030 --> 01:19:18,194
transfer, which is we can train
a controller and simulation, run

1450
01:19:18,194 --> 01:19:21,110
it the first and then run it
first time on the robot. Yeah,

1451
01:19:21,110 --> 01:19:22,000
it's on the robot.

1452
01:19:22,000 --> 01:19:23,000
Audrow Nash: And so that's super
cool.

1453
01:19:23,000 --> 01:19:26,540
Pras Velagapudi: That loop of
being able to just say, yeah,

1454
01:19:26,540 --> 01:19:30,140
robot RL, robot RL is really
powerful and we think is one of

1455
01:19:30,140 --> 01:19:33,980
the things that, you know, we're
really excited about doing well

1456
01:19:33,980 --> 01:19:37,640
here at Agility scale. Yeah. So
that all is that first layer.

1457
01:19:37,640 --> 01:19:38,000
Right.

1458
01:19:38,000 --> 01:19:41,931
Audrow Nash: And I have another
question there too. And that's

1459
01:19:41,931 --> 01:19:45,925
so, so cool. So I was I was in a
like a local locomotion lab at

1460
01:19:45,925 --> 01:19:49,159
University of Michigan for a
while, and we did some

1461
01:19:49,159 --> 01:19:53,026
reinforcement learning around
gaits for a bipedal robot. And,

1462
01:19:53,026 --> 01:19:56,133
it was very hard to use
reinforcement learning to

1463
01:19:56,133 --> 01:20:00,190
converge on stable gaits that we
could transfer to the robot. So

1464
01:20:00,190 --> 01:20:04,121
we'd learn it in simulation. We
could get it doing really cool

1465
01:20:04,121 --> 01:20:08,115
stuff. And then it was very hard
to take that set of parameters

1466
01:20:08,115 --> 01:20:11,793
that we gained, and transfer
that to the robot to actually

1467
01:20:11,793 --> 01:20:15,533
make it work like it did in
simulation. Why have we seen so

1468
01:20:15,533 --> 01:20:19,210
we've and then when there was
this sim to real gap that we

1469
01:20:19,210 --> 01:20:23,141
talked about a lot in robotics,
how have we advanced so far so

1470
01:20:23,141 --> 01:20:26,818
fast that we can do really good
training in simulation and

1471
01:20:26,818 --> 01:20:30,622
transfer it to the robot for
zero shot learning in this kind

1472
01:20:30,622 --> 01:20:34,744
of thing? Is it just, we know to
vary absolutely everything about

1473
01:20:34,744 --> 01:20:38,738
the model and the contact forces
and things like that, and that

1474
01:20:38,738 --> 01:20:42,478
creates a robust policy or what
have we done to make such a

1475
01:20:42,478 --> 01:20:44,000
significant advancement.

1476
01:20:44,000 --> 01:20:46,965
Pras Velagapudi: Yeah. So I
think it's been basically

1477
01:20:46,965 --> 01:20:50,042
recently two factors. One has
been that there have been

1478
01:20:50,042 --> 01:20:52,000
improvements and, and more
compute.

1479
01:20:52,000 --> 01:20:53,000
Audrow Nash: Well, that's that
was the other one.

1480
01:20:53,000 --> 01:20:56,388
Pras Velagapudi: Definitely part
of it. So one of one of them is

1481
01:20:56,388 --> 01:20:59,194
that we have made some
advancements in the techniques

1482
01:20:59,194 --> 01:21:02,529
like we're still using like PPO
optimization like it. That part

1483
01:21:02,529 --> 01:21:05,812
has actually stayed relatively
constant for these types of low

1484
01:21:05,812 --> 01:21:09,253
level controllers. But there are
some new techniques. And for PPO

1485
01:21:09,253 --> 01:21:11,000
proximal policy transition.
Yeah.

1486
01:21:11,000 --> 01:21:12,836
Audrow Nash: I'm surprised I
have that in my head still.

1487
01:21:12,836 --> 01:21:13,000
Okay.

1488
01:21:13,000 --> 01:21:16,818
Pras Velagapudi: You go I mean,
it's a yeah, you probably know

1489
01:21:16,818 --> 01:21:20,513
it better than I do if you're a
guy. It's been a while. Lab.

1490
01:21:20,513 --> 01:21:23,901
Right. But, the key thing there
is that, like that core

1491
01:21:23,901 --> 01:21:27,596
fundamental piece actually does
work pretty well still, with

1492
01:21:27,596 --> 01:21:30,983
other techniques like teacher
student systems, and, and

1493
01:21:30,983 --> 01:21:34,678
basically other things that
we've learned about how to build

1494
01:21:34,678 --> 01:21:38,373
a curriculum over time to learn
about stuff like locomotion,

1495
01:21:38,373 --> 01:21:41,699
because a lot of people have
been exploring it in this

1496
01:21:41,699 --> 01:21:45,394
intervening period and found
what works and what doesn't. So

1497
01:21:45,394 --> 01:21:49,150
that's definitely helped. And I
think that, you know, a rich,

1498
01:21:49,150 --> 01:21:52,537
academic, history there has
really helped us figure out

1499
01:21:52,537 --> 01:21:55,986
these are the techniques that
work really well for doing

1500
01:21:55,986 --> 01:21:59,189
dynamically stable legged
locomotion. And we've been

1501
01:21:59,189 --> 01:22:03,130
learning that this whole time on
the other side, which is really

1502
01:22:03,130 --> 01:22:06,702
where just the benefit of
everybody pouring so much energy

1503
01:22:06,702 --> 01:22:10,397
into AI has been the compute
that we have available. And GPU

1504
01:22:10,397 --> 01:22:14,339
accelerated simulation is really
a game changer, because now you

1505
01:22:14,339 --> 01:22:17,972
can also just say, well, let's
also just run it like 100, a

1506
01:22:17,972 --> 01:22:21,975
thousand more. Yeah, trials than
we ever could before in like one

1507
01:22:21,975 --> 01:22:25,362
at a time. Right? You can
simulate really fast. You can

1508
01:22:25,362 --> 01:22:29,057
simulate on really powerful
hardware. And that means to your

1509
01:22:29,057 --> 01:22:32,691
earlier point, when you're
talking about how do I make sure

1510
01:22:32,691 --> 01:22:36,694
that I've captured the parameter
space of the real robot, you can

1511
01:22:36,694 --> 01:22:40,266
just domain randomize over that
space and make sure you've

1512
01:22:40,266 --> 01:22:43,714
captured it somewhere in that
cluster of parameters that

1513
01:22:43,714 --> 01:22:46,794
you're, that you're covering.
Right. And so domain

1514
01:22:46,794 --> 01:22:50,304
randomization in conjunction
with really powerful compute

1515
01:22:50,304 --> 01:22:54,307
help solve for some of this. And
there is some amount of just hey

1516
01:22:54,307 --> 01:22:58,125
if you train it for longer you
can get better results. And now

1517
01:22:58,125 --> 01:23:01,697
training it for longer doesn't
take as long. And so that's

1518
01:23:01,697 --> 01:23:05,639
really kind of all come together
to make it such that, it really

1519
01:23:05,639 --> 01:23:09,087
is highly effective to build
out, legged locomotion in a

1520
01:23:09,087 --> 01:23:12,351
reinforcement learning
environment now and just have,

1521
01:23:12,351 --> 01:23:16,108
an RL system basically trained
up from whole cloth, you know,

1522
01:23:16,108 --> 01:23:19,803
how to move in an environment.
And of course, you can always

1523
01:23:19,803 --> 01:23:23,560
add in additional stuff on top
of that. Like you can, do some

1524
01:23:23,560 --> 01:23:27,378
amount of imitation learning on
existing gates and things like

1525
01:23:27,378 --> 01:23:31,196
that to seed the optimization.
But once you get into that base

1526
01:23:31,196 --> 01:23:34,830
of attraction, pure RL can
really go a long way these days.

1527
01:23:34,830 --> 01:23:36,000
And so that's, you.

1528
01:23:36,000 --> 01:23:38,000
Audrow Nash: Know, climbing keep
getting better.

1529
01:23:38,000 --> 01:23:40,857
Pras Velagapudi: Yeah. And it's
been super impressive. Some of

1530
01:23:40,857 --> 01:23:43,806
the results that that, research
labs and, and different parts of

1531
01:23:43,806 --> 01:23:46,433
the industry have been able to
show off with just really,

1532
01:23:46,433 --> 01:23:48,000
really impressive dynamic
motions.

1533
01:23:48,000 --> 01:23:51,383
Audrow Nash: Yeah, I guess,
yeah. It just seems like our

1534
01:23:51,383 --> 01:23:55,128
whole industry has learned more
compute better. And it like is

1535
01:23:55,128 --> 01:23:57,000
very true, which is just crazy.

1536
01:23:57,000 --> 01:24:00,933
Pras Velagapudi: I think we're,
we're seeing the same trend

1537
01:24:00,933 --> 01:24:04,933
that's hit many other, areas
before us. Right. Which is just

1538
01:24:04,933 --> 01:24:08,800
that, there's a lot of things
that, that we were trying to

1539
01:24:08,800 --> 01:24:13,000
solve, you know, by changing the
problem. And it turns out that

1540
01:24:13,000 --> 01:24:17,267
we could just change our sort of
that our cluster size and solve

1541
01:24:17,267 --> 01:24:18,000
it instead.

1542
01:24:18,000 --> 01:24:20,691
Audrow Nash: Yeah. Very much.
It's so exciting and also a

1543
01:24:20,691 --> 01:24:23,524
little disheartening to me for
some reason, which is kind of

1544
01:24:23,524 --> 01:24:26,309
funny because it's like you
learn all these advanced things

1545
01:24:26,309 --> 01:24:29,000
and like I just simulate it more
and it works fine. Well.

1546
01:24:29,000 --> 01:24:31,723
Pras Velagapudi: So there's a
balance there. There's

1547
01:24:31,723 --> 01:24:35,023
definitely learning structurally
about like the techniques that

1548
01:24:35,023 --> 01:24:38,322
work and those that work at any
scale and are really important.

1549
01:24:38,322 --> 01:24:41,412
But then there's also the stuff
where, you know, like I did

1550
01:24:41,412 --> 01:24:44,502
this, you know, in my PhD at
Carnegie Mellon, I was doing a

1551
01:24:44,502 --> 01:24:47,487
bunch of simulation with
multi-agent, I think it was like

1552
01:24:47,487 --> 01:24:50,472
Dec Pomdps, which is partially
observable Markov decision

1553
01:24:50,472 --> 01:24:53,510
processes. And then, yeah,
decentralized versions of those

1554
01:24:53,510 --> 01:24:56,600
where they're spread across
machines. But now, I don't know

1555
01:24:56,600 --> 01:24:59,637
that I would have to do any of
that because I could easily

1556
01:24:59,637 --> 01:25:02,046
probably do that same
optimization on a single

1557
01:25:02,046 --> 01:25:04,927
machine, and therefore I
wouldn't have had to do nearly

1558
01:25:04,927 --> 01:25:07,703
as work, as I was doing to
solve, these decentralized

1559
01:25:07,703 --> 01:25:10,740
policies because I wouldn't need
them to be decentralized.

1560
01:25:10,740 --> 01:25:13,882
Really. So there's definitely
this, this evolution of, like,

1561
01:25:13,882 --> 01:25:16,972
the techniques that that we
needed to do back then, a large

1562
01:25:16,972 --> 01:25:20,272
portion of what we were solving
for was actually just trying to

1563
01:25:20,272 --> 01:25:22,000
work around limitations,
compute.

1564
01:25:22,000 --> 01:25:23,000
Audrow Nash: Limitations.

1565
01:25:23,000 --> 01:25:25,880
Pras Velagapudi: And
distribution of information. And

1566
01:25:25,880 --> 01:25:29,196
nowadays, some of those problems
have been solved and some of

1567
01:25:29,196 --> 01:25:32,402
them are still fundamental
problems. And, and yeah, benefit

1568
01:25:32,402 --> 01:25:33,000
from these.

1569
01:25:33,000 --> 01:25:37,067
Audrow Nash: Very true. Yes. For
sure. Okay. So higher blocks

1570
01:25:37,067 --> 01:25:38,000
you're saying.

1571
01:25:38,000 --> 01:25:41,312
Pras Velagapudi: Yeah. All
right. Let's call block number

1572
01:25:41,312 --> 01:25:44,915
one. So block number two above
that is the skill layer. So the

1573
01:25:44,915 --> 01:25:48,228
controls block again is I am
getting some input of like I

1574
01:25:48,228 --> 01:25:51,773
need to modify my environment
somehow. I need to exert forces

1575
01:25:51,773 --> 01:25:55,376
or do something. Let me figure
out how to do that at the joint

1576
01:25:55,376 --> 01:25:58,688
level. A layer above that is
what am I trying to do, like

1577
01:25:58,688 --> 01:26:01,943
some task specification? And
then what does that mean? I

1578
01:26:01,943 --> 01:26:04,906
should be doing a my
environment. So I need to pick

1579
01:26:04,906 --> 01:26:08,568
up a tote. The totes over here
maybe I like draw a mask over it

1580
01:26:08,568 --> 01:26:11,938
in an image or something like
that. Or I give it a pose or

1581
01:26:11,938 --> 01:26:15,715
something. Now I need some layer
that's again pretty close to the

1582
01:26:15,715 --> 01:26:19,377
robot because it's got to take
into account changing perception

1583
01:26:19,377 --> 01:26:22,805
information, but maybe not as
fast, right? So maybe 10Hz or

1584
01:26:22,805 --> 01:26:26,292
30Hz and it's going to take that
and generate out a new what

1585
01:26:26,292 --> 01:26:29,546
should I be doing in my
environment? I should be putting

1586
01:26:29,546 --> 01:26:32,917
my hands here. I should be
grabbing it like this. I should

1587
01:26:32,917 --> 01:26:36,346
be exerting these types of
forces. I don't know all the way

1588
01:26:36,346 --> 01:26:40,123
how that's going to end up being
joint level controls. But that's

1589
01:26:40,123 --> 01:26:43,610
what I'm trying to do with the
world. So we find that that's

1590
01:26:43,610 --> 01:26:47,213
like a useful layer to sort of
separate out and be training is

1591
01:26:47,213 --> 01:26:50,816
idea of a skill, which is some
sort of task specification down

1592
01:26:50,816 --> 01:26:54,303
to some sort of specification of
what needs to happen in the

1593
01:26:54,303 --> 01:26:55,000
environment.

1594
01:26:55,000 --> 01:26:59,122
Audrow Nash: Yeah. Okay. And
then how are you benefiting from

1595
01:26:59,122 --> 01:27:00,000
AI with that.

1596
01:27:00,000 --> 01:27:03,876
Pras Velagapudi: Well, that's an
area where historically you

1597
01:27:03,876 --> 01:27:07,751
would get someone to do. And we
still have quite a few, very

1598
01:27:07,751 --> 01:27:11,304
talented folks here who write
out that specification in

1599
01:27:11,304 --> 01:27:15,115
something like Python. It's
maybe, hey, if you see the toad

1600
01:27:15,115 --> 01:27:19,055
here, move to this location,
push the hands against it, close

1601
01:27:19,055 --> 01:27:22,608
the fingers, pick it up, and
then that's, you know, the

1602
01:27:22,608 --> 01:27:26,419
pickup totes primitive, right?
And we had to hand off those

1603
01:27:26,419 --> 01:27:30,230
blocks and you might make a
block for picking up a tote and

1604
01:27:30,230 --> 01:27:33,782
placing a tote and moving
between areas. And, and human

1605
01:27:33,782 --> 01:27:37,916
skilled humans would write those
right and still do for a lot of

1606
01:27:37,916 --> 01:27:41,792
the more complex tasks. The
interesting piece with the newer

1607
01:27:41,792 --> 01:27:45,409
AI technologies, especially
around things like, behavior

1608
01:27:45,409 --> 01:27:49,026
cloning, is that you can do
learning from demonstrate or

1609
01:27:49,026 --> 01:27:52,708
imitation learning or even,
doing things like RL to solve

1610
01:27:52,708 --> 01:27:56,648
for how to do that layer of the
stack. How do I get this task

1611
01:27:56,648 --> 01:28:00,589
done? By modifying something in
my environment. And so that's

1612
01:28:00,589 --> 01:28:04,722
new. And because that has always
been, a large chunk of building

1613
01:28:04,722 --> 01:28:08,663
out the capability of a robot,
we're now in this space where,

1614
01:28:08,663 --> 01:28:12,667
hey, the thing that was the most
expensive sort of incremental

1615
01:28:12,667 --> 01:28:16,091
cost of going into a new
environment is now working a

1616
01:28:16,091 --> 01:28:19,967
totally different way. Yeah,
it's, operating under a totally

1617
01:28:19,967 --> 01:28:21,000
different model.

1618
01:28:21,000 --> 01:28:23,000
Audrow Nash: With the 10% effort
instead of 90.

1619
01:28:23,000 --> 01:28:26,360
Pras Velagapudi: Or something,
like just the type of data that

1620
01:28:26,360 --> 01:28:29,396
you collect, the type of, the
type of person that can be

1621
01:28:29,396 --> 01:28:32,322
collecting data. So if you're
talking about collecting

1622
01:28:32,322 --> 01:28:35,575
something like tally up data,
right. You might be looking at

1623
01:28:35,575 --> 01:28:38,772
someone who's a technician
collecting that data rather than

1624
01:28:38,772 --> 01:28:41,699
someone who is a software
engineer. Right. And so that

1625
01:28:41,699 --> 01:28:44,897
was, you know, definitely move
more quickly in building out

1626
01:28:44,897 --> 01:28:47,932
these skill sets. And then
you're software engineers are

1627
01:28:47,932 --> 01:28:50,805
more dealing with the
infrastructure behind it rather

1628
01:28:50,805 --> 01:28:54,328
than tuning each specific skill,
that you have. Yeah. So that's a

1629
01:28:54,328 --> 01:28:57,634
huge game changer. And I think
that everyone is scrambling to

1630
01:28:57,634 --> 01:29:01,157
figure out how well and what the
right techniques are for this or

1631
01:29:01,157 --> 01:29:04,247
that. The space. This is the
really exciting, oh my gosh,

1632
01:29:04,247 --> 01:29:07,499
this could really change. Open
up a lot of things. Yeah. How

1633
01:29:07,499 --> 01:29:10,209
robots are scaling into
environments. And so we're

1634
01:29:10,209 --> 01:29:13,461
really interested in this third
parties that, you know, like

1635
01:29:13,461 --> 01:29:16,604
physical intelligence and
skilled are really interested in

1636
01:29:16,604 --> 01:29:19,585
solving for just the model part
of it. But everybody is

1637
01:29:19,585 --> 01:29:22,566
scrambling to figure out how
this is going to be a game

1638
01:29:22,566 --> 01:29:23,000
changer.

1639
01:29:23,000 --> 01:29:26,631
Audrow Nash: Do you think that
so, I mean, one of the things

1640
01:29:26,631 --> 01:29:30,443
that we've been hearing for the
last, I don't know, 20 years or

1641
01:29:30,443 --> 01:29:34,256
something is like data is gold
for all of this. And it seems to

1642
01:29:34,256 --> 01:29:37,947
be true with all the processes
that require a lot of compute.

1643
01:29:37,947 --> 01:29:41,699
Do you think that we could enter
a data is cheap point of view

1644
01:29:41,699 --> 01:29:45,269
when it's like, oh, you just
need to show it once it learns

1645
01:29:45,269 --> 01:29:49,082
how to do it. And then for a lot
of applications, it's like the

1646
01:29:49,082 --> 01:29:52,470
companies that have been
collecting an endless amount of

1647
01:29:52,470 --> 01:29:56,343
data, they're not really sitting
on a mountain of gold like they

1648
01:29:56,343 --> 01:29:59,914
thought for this kind of thing.
Do you think we could enter

1649
01:29:59,914 --> 01:30:03,000
that? Or is it a silly idea?
What are you thinking?

1650
01:30:03,000 --> 01:30:06,738
Pras Velagapudi: I think it's
somewhere in the middle. Maybe.

1651
01:30:06,738 --> 01:30:10,108
So in order to get to models
that do really well with a

1652
01:30:10,108 --> 01:30:14,029
little bit of data, they tend to
have to be pre-trained on a lot

1653
01:30:14,029 --> 01:30:17,828
of data. And so you end up with
this situation where depending

1654
01:30:17,828 --> 01:30:21,811
on where those models are coming
from and how those are commodity

1655
01:30:21,811 --> 01:30:25,672
sized, you may end up in a space
where a little data is all you

1656
01:30:25,672 --> 01:30:29,593
need, or you may not end up in a
space where you need to collect

1657
01:30:29,593 --> 01:30:33,392
a lot of data to build out that
model. That lets you then move

1658
01:30:33,392 --> 01:30:37,130
from point to point with a much
smaller amount of data. And I

1659
01:30:37,130 --> 01:30:40,623
think one of the big differences
here is that in physical

1660
01:30:40,623 --> 01:30:44,176
intelligence or certain of
physical in physical AI, as the

1661
01:30:44,176 --> 01:30:47,853
terminology has, has become
known. Yeah. Physical AI doesn't

1662
01:30:47,853 --> 01:30:51,652
have this long standing corpus
of really well curated data for

1663
01:30:51,652 --> 01:30:55,206
the types of robot platforms
that we're creating today. If

1664
01:30:55,206 --> 01:30:59,127
anything, the largest data sets
are probably single arm overhead

1665
01:30:59,127 --> 01:31:03,110
tabletop problems coming out of,
academia or some of historically

1666
01:31:03,110 --> 01:31:06,848
the largest data sets. And so
now we're seeing a lot of folks

1667
01:31:06,848 --> 01:31:10,525
scramble to build up these data
sets and build models off of

1668
01:31:10,525 --> 01:31:14,201
them. But it's pretty different
because we can't just go and

1669
01:31:14,201 --> 01:31:17,632
basically download the entirety
of everything that's, on

1670
01:31:17,632 --> 01:31:21,431
Wikipedia. Some folks are trying
to figure out if maybe we can

1671
01:31:21,431 --> 01:31:24,985
use some portion of the videos
on YouTube. You know, maybe

1672
01:31:24,985 --> 01:31:28,172
that's a large corpus we could
unlock with the right

1673
01:31:28,172 --> 01:31:31,725
transformation. But that's a
right now, a huge, difference

1674
01:31:31,725 --> 01:31:35,402
between where we are with LMS
and Vlaams, that are, that are

1675
01:31:35,402 --> 01:31:38,956
working in kind of just abstract
human reasoning and human

1676
01:31:38,956 --> 01:31:42,816
language to, systems and models
that are dealing with, physical

1677
01:31:42,816 --> 01:31:43,000
or.

1678
01:31:43,000 --> 01:31:44,000
Audrow Nash: Physical.

1679
01:31:43,000 --> 01:31:46,567
Pras Velagapudi: Space. Right.
And there there's a lot of

1680
01:31:46,567 --> 01:31:50,260
differences in the data sets
that people have available and

1681
01:31:50,260 --> 01:31:54,203
how well curated they are, and
the types of, robots and sensors

1682
01:31:54,203 --> 01:31:57,958
that have been used to collect
the data and how that affects

1683
01:31:57,958 --> 01:32:01,088
how well it works on other
systems and how well it

1684
01:32:01,088 --> 01:32:04,530
transfers. So I think we're
going to see some amount of

1685
01:32:04,530 --> 01:32:08,473
perhaps commoditization of parts
of models. There. So it may be

1686
01:32:08,473 --> 01:32:11,727
the case that there become
established players, that

1687
01:32:11,727 --> 01:32:15,608
produce, good models that other
people can basically fine tune

1688
01:32:15,608 --> 01:32:19,551
to their applications with very
small amounts of data. It could

1689
01:32:19,551 --> 01:32:23,431
be the case that, a really big
entity like an Nvidia, puts out

1690
01:32:23,431 --> 01:32:27,374
an open source model that sort
of levels the playing fields, in

1691
01:32:27,374 --> 01:32:28,000
the space.

1692
01:32:28,000 --> 01:32:30,565
Audrow Nash: That would be like,
amazing. Yeah. And they're

1693
01:32:30,565 --> 01:32:31,000
certainly.

1694
01:32:31,000 --> 01:32:33,746
Pras Velagapudi: Working on it
with, with, with, efforts like

1695
01:32:33,746 --> 01:32:36,492
Groot. Right. That's, that is
the intent of efforts like that

1696
01:32:36,492 --> 01:32:39,283
is to basically try and produce,
models that sort of level the

1697
01:32:39,283 --> 01:32:41,714
playing field, because of
course, for Nvidia, the more

1698
01:32:41,714 --> 01:32:44,280
people that are playing, the
better they're doing. Right?

1699
01:32:44,280 --> 01:32:45,000
Because they're.

1700
01:32:45,000 --> 01:32:47,000
Audrow Nash: They're still in
trouble.

1701
01:32:47,000 --> 01:32:51,093
Pras Velagapudi: Right? Exactly.
So we could see that happen to

1702
01:32:51,093 --> 01:32:54,860
the space. I don't think we
really know exactly where it's

1703
01:32:54,860 --> 01:32:58,953
going to land. But I think what
we're what we're seeing is that

1704
01:32:58,953 --> 01:33:03,110
at the very least, we have the,
we have the amount of data where

1705
01:33:03,110 --> 01:33:06,943
we can get, single skills or
like small, generalized skills

1706
01:33:06,943 --> 01:33:11,035
in certain local areas to work
really well and really reliably.

1707
01:33:11,035 --> 01:33:15,063
And so we know that that at the
very least seems tractable and

1708
01:33:15,063 --> 01:33:18,831
that seems like it would work.
And that alone is already a

1709
01:33:18,831 --> 01:33:20,000
pretty big unlock.

1710
01:33:20,000 --> 01:33:22,269
Audrow Nash: Oh yeah. Okay. So
then above the skill level I

1711
01:33:22,269 --> 01:33:24,615
suppose you have something like
a planning level or something

1712
01:33:24,615 --> 01:33:25,000
like that.

1713
01:33:25,000 --> 01:33:28,382
Pras Velagapudi: Pretty much at
that point you start getting,

1714
01:33:28,382 --> 01:33:31,763
it's a, you know what you might
call a cognition or reasoning

1715
01:33:31,763 --> 01:33:35,256
level. And that's really where
Vlaams and La Amazing are really

1716
01:33:35,256 --> 01:33:38,305
just, they're, they're right
there, this at this level,

1717
01:33:38,305 --> 01:33:41,520
you're basically starting to
think agent like I where your

1718
01:33:41,520 --> 01:33:44,736
skill is the API that you're
interacting with. And so here

1719
01:33:44,736 --> 01:33:47,896
you're really just leveraging
web scale models. You don't

1720
01:33:47,896 --> 01:33:51,000
necessarily really need to build
anything that's to, no.

1721
01:33:51,000 --> 01:33:52,556
Audrow Nash: Specific really of
questions on this one. This one

1722
01:33:52,556 --> 01:33:53,000
was kind of clear.

1723
01:33:53,000 --> 01:33:57,506
Pras Velagapudi: Yeah. Yeah. The
really interesting part, though,

1724
01:33:57,506 --> 01:34:01,179
that I find particularly
exciting is that because the

1725
01:34:01,179 --> 01:34:05,408
space is so exciting to so many
people, you really see really

1726
01:34:05,408 --> 01:34:09,567
large players, spending time
making models that are specific

1727
01:34:09,567 --> 01:34:13,448
to, physical systems. So,
Gemini, for example, you know,

1728
01:34:13,448 --> 01:34:17,677
Google has been training on,
robot data sets, which makes VMs

1729
01:34:17,677 --> 01:34:21,974
that, you know, their versions
of, of Gemini that are actually

1730
01:34:21,974 --> 01:34:26,272
able to answer, or better answer
queries that, that pertain to

1731
01:34:26,272 --> 01:34:30,431
things like, objects in the
world. Right. So historically is

1732
01:34:30,431 --> 01:34:34,382
really hard to ask a GLM for
things like, hey, what's the

1733
01:34:34,382 --> 01:34:38,679
object that's, you know, to the
left of my hand? Or like, what

1734
01:34:38,679 --> 01:34:43,046
would I need to do to go pick up
this object in the back of the

1735
01:34:43,046 --> 01:34:47,205
scene or behind the table and
move it to close to me like it

1736
01:34:47,205 --> 01:34:50,948
did really poorly around
egocentric reasoning and, you

1737
01:34:50,948 --> 01:34:55,107
know, relative reasoning about
things in the scene. But now,

1738
01:34:55,107 --> 01:34:59,197
because so many people are
applying these films to physical

1739
01:34:59,197 --> 01:35:03,287
applications, that's now a
specific area of focus that, the

1740
01:35:03,287 --> 01:35:07,445
providers are tuning towards and
trying to collect data and.

1741
01:35:07,445 --> 01:35:08,000
Correct.

1742
01:35:08,000 --> 01:35:11,287
Audrow Nash: Oh, cool. Wow.
Okay. Yeah. What a thing. It's

1743
01:35:11,287 --> 01:35:14,801
amazing. All of this feels like
it's moving so fast for all of

1744
01:35:14,801 --> 01:35:18,428
this, where it's like, I have no
idea where we'll be, like, even

1745
01:35:18,428 --> 01:35:22,055
like you think back to 2020 and
where the state of robotics was,

1746
01:35:22,055 --> 01:35:25,569
because that to me feels like a
big indicator in time, because

1747
01:35:25,569 --> 01:35:28,912
of Covid and everything. It's
like where we were then. It's

1748
01:35:28,912 --> 01:35:32,256
like everything has like now we
have this whole AI boom and

1749
01:35:32,256 --> 01:35:35,826
everything, and now there's all
this interest in humanoids, and

1750
01:35:35,826 --> 01:35:39,000
it's just it's so interesting
where everything is going.

1751
01:35:39,000 --> 01:35:42,688
Pras Velagapudi: Yeah, I think
the that's the most exciting

1752
01:35:42,688 --> 01:35:46,438
part about all of it is that
regardless of what the specific

1753
01:35:46,438 --> 01:35:50,438
outcomes are or exactly what the
robots of the future look like,

1754
01:35:50,438 --> 01:35:54,000
what's increasingly clear is
that we can have them. Yeah.

1755
01:35:54,000 --> 01:35:57,389
Audrow Nash: Super cool. What
are you so going back to

1756
01:35:57,389 --> 01:36:00,653
Agility, where are you guys
headed? I know it's hard

1757
01:36:00,653 --> 01:36:04,418
because, like, okay, we think
back five years and everything

1758
01:36:04,418 --> 01:36:08,310
has changed. Where are you guys
heading into the next, like, 2

1759
01:36:08,310 --> 01:36:09,000
to 5 years.

1760
01:36:09,000 --> 01:36:10,000
Pras Velagapudi: So yeah.

1761
01:36:10,000 --> 01:36:13,159
Audrow Nash: So for us, the
standard and unstructured

1762
01:36:13,159 --> 01:36:16,616
environment or less, I don't
know, less work environments.

1763
01:36:16,616 --> 01:36:19,000
But yeah. Where are you going
otherwise.

1764
01:36:19,000 --> 01:36:22,354
Pras Velagapudi: Well I think
the biggest thing for us is, is

1765
01:36:22,354 --> 01:36:25,817
building out our V5 platform to,
to really be the platform that

1766
01:36:25,817 --> 01:36:29,171
we can scale out of the work
cell with. That's really what we

1767
01:36:29,171 --> 01:36:32,524
think is a gating function to
there being a lot of robots out

1768
01:36:32,524 --> 01:36:35,988
there doing a lot of different
tasks. We can already, you know,

1769
01:36:35,988 --> 01:36:38,682
we already have a fleet
management system. We can

1770
01:36:38,682 --> 01:36:42,035
already deploy new over-the-air
updates to the robot and give

1771
01:36:42,035 --> 01:36:45,609
them new capabilities out in the
field. They don't require a huge

1772
01:36:45,609 --> 01:36:48,522
amount of customization for
different specialty tasks

1773
01:36:48,522 --> 01:36:51,931
currently. Right? We don't we
don't do that. We only build one

1774
01:36:51,931 --> 01:36:55,449
platform and it's called Digit,
right. So from that perspective,

1775
01:36:55,449 --> 01:36:58,803
you know, that's all ready to
go. And so if we could just get

1776
01:36:58,803 --> 01:37:02,101
these last few unlocks around
things like the safety, around

1777
01:37:02,101 --> 01:37:05,565
things like, you know, opening
up our manipulation capabilities

1778
01:37:05,565 --> 01:37:08,754
with the tool changer, we're
really trying to unlock these

1779
01:37:08,754 --> 01:37:12,272
last few pieces to kind of open
the floodgates for all the stuff

1780
01:37:12,272 --> 01:37:15,461
that we've been working on.
That's almost. You ready to be

1781
01:37:15,461 --> 01:37:17,000
unlocked, right. All of the.

1782
01:37:17,000 --> 01:37:18,000
Audrow Nash: Building.

1783
01:37:17,000 --> 01:37:21,512
Pras Velagapudi: So for us, that
2 to 5 year timeline is really

1784
01:37:21,512 --> 01:37:25,951
is a scaling timeline. It's get
these last few pieces unlocked

1785
01:37:25,951 --> 01:37:29,460
and then just open the
floodgates. And so we hope

1786
01:37:29,460 --> 01:37:33,757
that's what our next 2 to 5
years look like. We hope that we

1787
01:37:33,757 --> 01:37:38,412
are going to be out there in the
world that Digits will be doing,

1788
01:37:38,412 --> 01:37:42,780
lots of different tasks, that
they'll be easy to deploy, that

1789
01:37:42,780 --> 01:37:46,862
they'll be out working full
shifts, in the places that we

1790
01:37:46,862 --> 01:37:50,872
can deploy them now from
regulatory standpoint. And then

1791
01:37:50,872 --> 01:37:55,384
that will basically be the, the,
the driving function for other

1792
01:37:55,384 --> 01:37:59,609
industries to want it. Right.
They'll see them. And that'll

1793
01:37:59,609 --> 01:38:03,905
become kind of a fire that that
gets lit to say, hey, I want

1794
01:38:03,905 --> 01:38:08,345
these in, in my stores, in my
facilities, in my use cases. And

1795
01:38:08,345 --> 01:38:12,427
that'll help really drive the,
the ecosystem to, to scale

1796
01:38:12,427 --> 01:38:13,000
further.

1797
01:38:13,000 --> 01:38:16,892
Audrow Nash: Yeah, it'll be a
flywheel. It'll be spun up at

1798
01:38:16,892 --> 01:38:20,917
that point. That'll be nuts. So
wrapping up, what do you hope

1799
01:38:20,917 --> 01:38:25,007
that our listeners and watchers
take away from this? Or do you

1800
01:38:25,007 --> 01:38:29,097
have any calls to action, like
you guys are hiring or anything

1801
01:38:29,097 --> 01:38:32,000
else that you'd like to, wrap up
with? Yeah.

1802
01:38:32,000 --> 01:38:36,214
Pras Velagapudi: So certainly,
yeah. If you, if you see open

1803
01:38:36,214 --> 01:38:39,866
roles, you know, we're always
looking for talent. We

1804
01:38:39,866 --> 01:38:43,798
definitely, you know, if any
listeners are interested, I

1805
01:38:43,798 --> 01:38:48,293
encourage you to go take a look
at our job postings. I would say

1806
01:38:48,293 --> 01:38:52,296
if you are a humanoid robotics
provider yourself, I would

1807
01:38:52,296 --> 01:38:56,159
encourage you to understand
safety and start, you know,

1808
01:38:56,159 --> 01:39:00,302
working with the standards
bodies that are that are working

1809
01:39:00,302 --> 01:39:04,446
on that. You know, if you want
to join our, our ISO working

1810
01:39:04,446 --> 01:39:08,449
group, I believe it's 25785-1,
you can get involved, get,

1811
01:39:08,449 --> 01:39:12,592
informed at least. Right? Don't
get stuck where you build a

1812
01:39:12,592 --> 01:39:16,455
perfectly working robot and you
can't legally deploy it

1813
01:39:16,455 --> 01:39:18,000
anywhere. Like that's.

1814
01:39:18,000 --> 01:39:19,000
Audrow Nash: That'd be crazy.

1815
01:39:19,000 --> 01:39:22,433
Pras Velagapudi: Exactly. Like,
don't end up there. Everybody

1816
01:39:22,433 --> 01:39:25,754
should be up to date on this
stuff. Everybody should really

1817
01:39:25,754 --> 01:39:29,187
understand more. I think about
what it takes to safely deploy

1818
01:39:29,187 --> 01:39:32,508
robots in production and at
scale, because I feel like it's

1819
01:39:32,508 --> 01:39:35,603
something that, you know, I
certainly didn't even think

1820
01:39:35,603 --> 01:39:39,093
about when I was in academia. I
didn't even know most of these

1821
01:39:39,093 --> 01:39:42,639
standards existed. And it's been
such a core theme of what I've

1822
01:39:42,639 --> 01:39:45,734
had to do to take robots to
production since then in my

1823
01:39:45,734 --> 01:39:48,999
professional career. So I think
I think that's if you're a

1824
01:39:48,999 --> 01:39:52,319
roboticist and you don't know,
what safety standards are or

1825
01:39:52,319 --> 01:39:55,640
like which ones are applicable,
like go read up on them, go

1826
01:39:55,640 --> 01:39:59,073
learn about them more, because
they could play a big role. If

1827
01:39:59,073 --> 01:40:02,000
you ever try to build a robot as
a product. Oh yeah.

1828
01:40:02,000 --> 01:40:05,291
Audrow Nash: How do you learn up
on those? Well, like, what would

1829
01:40:05,291 --> 01:40:06,000
you recommend?

1830
01:40:06,000 --> 01:40:09,589
Pras Velagapudi: Yeah, I mean, I
would recommend, doing some

1831
01:40:09,589 --> 01:40:12,818
reading on, a field called
functional safety and also,

1832
01:40:12,818 --> 01:40:16,167
maybe starting with. I mean,
it's, there's a lot of good

1833
01:40:16,167 --> 01:40:20,055
primers on safety for industrial
robots is a field that's existed

1834
01:40:20,055 --> 01:40:23,763
for a long time. And even just
seeing what that looks like and

1835
01:40:23,763 --> 01:40:27,591
what the standards are and what
they mean for different types of

1836
01:40:27,591 --> 01:40:31,239
robots is, is, I think, a useful
thing just knowing that they

1837
01:40:31,239 --> 01:40:34,888
exist and maybe just ask the
question of, hey, I was thinking

1838
01:40:34,888 --> 01:40:38,416
about using a robot in ECS
environment. Like what standards

1839
01:40:38,416 --> 01:40:42,124
apply in that environment? You
know what? What would I have to

1840
01:40:42,124 --> 01:40:45,833
certify my robot to, you know,
even just starting that thought

1841
01:40:45,833 --> 01:40:49,122
process? And then once you
really get into it, there's,

1842
01:40:49,122 --> 01:40:52,830
there's third parties that can
help you with that. And there's

1843
01:40:52,830 --> 01:40:56,478
certainly a lot more, knowledge
that's available, but I think

1844
01:40:56,478 --> 01:41:00,246
it's just something that's not
necessarily on everyone's radar.

1845
01:41:00,246 --> 01:41:04,134
Like, every robot should have an
app that you can get to stop the

1846
01:41:04,134 --> 01:41:07,663
robot. Like if you don't have
one, put one in and make sure

1847
01:41:07,663 --> 01:41:11,371
that if the robot starts going
haywire, you could actually get

1848
01:41:11,371 --> 01:41:14,660
to it, right? Stuff like that.
Right. So I think that's

1849
01:41:14,660 --> 01:41:18,249
important and not always on
everyone's radar until something

1850
01:41:18,249 --> 01:41:21,000
happens and then and then it's
on their radar.

1851
01:41:21,000 --> 01:41:23,000
Audrow Nash: Yes. Big, big on
their radar.

1852
01:41:23,000 --> 01:41:26,785
Pras Velagapudi: Yeah. But, I
think, you know, this is an

1853
01:41:26,785 --> 01:41:30,569
exciting time. I, you know,
safety is a huge component of

1854
01:41:30,569 --> 01:41:34,421
this. I also think that, you
know, if there's still any AI

1855
01:41:34,421 --> 01:41:38,471
naysayers out there, I all three
of you, like, come on, guys,

1856
01:41:38,471 --> 01:41:42,256
like, get on the train. Right.
This is this is happening,

1857
01:41:42,256 --> 01:41:46,306
right? It's not a question of
if. It's just a question of how

1858
01:41:46,306 --> 01:41:50,024
much and how fast. And really,
you know, I came from the

1859
01:41:50,024 --> 01:41:53,942
position of, of doing a lot of
classical AI, some more like

1860
01:41:53,942 --> 01:41:57,926
optimization and search type
strategies in the past and over

1861
01:41:57,926 --> 01:42:01,511
the past 3 or 4 years, I've
really, had an eye opening

1862
01:42:01,511 --> 01:42:04,632
experience of just
understanding, you know, how

1863
01:42:04,632 --> 01:42:08,748
much is different and how much,
capability, and flexibility is

1864
01:42:08,748 --> 01:42:12,467
available in the latest
generation of AI, model building

1865
01:42:12,467 --> 01:42:16,716
systems and, and simulation and
training platforms. So if you're

1866
01:42:16,716 --> 01:42:20,966
not thinking about that, that's
also definitely something that's

1867
01:42:20,966 --> 01:42:24,883
worth, trying to although that
field move so fast that like

1868
01:42:24,883 --> 01:42:28,867
every three months, like you
have to relearn half of it, but

1869
01:42:28,867 --> 01:42:29,000
I.

1870
01:42:29,000 --> 01:42:30,000
Audrow Nash: Know.

1871
01:42:30,000 --> 01:42:32,969
Pras Velagapudi: At least start
to keep your finger on the pulse

1872
01:42:32,969 --> 01:42:35,799
and don't assume that something
won't get solved because that

1873
01:42:35,799 --> 01:42:38,582
almost categorically end up
being solved very soon after you

1874
01:42:38,582 --> 01:42:39,000
say that.

1875
01:42:39,000 --> 01:42:42,085
Audrow Nash: I know. And how
would you keep your finger on

1876
01:42:42,085 --> 01:42:45,543
the pulse of AI? Because that to
me it's like, I don't know, like

1877
01:42:45,543 --> 01:42:48,628
I don't really have a good way
other than talking to great

1878
01:42:48,628 --> 01:42:49,000
people.

1879
01:42:49,000 --> 01:42:53,011
Pras Velagapudi: Yeah, well, I
do think that's a core part of

1880
01:42:53,011 --> 01:42:57,219
it is, is, you know, pick a few
sources of folks that, you know,

1881
01:42:57,219 --> 01:43:01,427
are good, knowledgeable, sources
that can parse the information.

1882
01:43:01,427 --> 01:43:05,570
So I actually one of the things
that I do personally is, is one

1883
01:43:05,570 --> 01:43:09,186
of, our, lead AI engineers,
Chris Paxton, runs a really

1884
01:43:09,186 --> 01:43:13,000
active, social media presence.
And so I'm often following.

1885
01:43:13,000 --> 01:43:14,000
Audrow Nash: He's excellent.

1886
01:43:13,000 --> 01:43:16,557
Pras Velagapudi: Yeah. And I
think, you know, following him,

1887
01:43:16,557 --> 01:43:20,054
following, you know, Jim Fan at
Nvidia and other folks like

1888
01:43:20,054 --> 01:43:23,611
that. So that's how I get a lot
of my news is by seeing what

1889
01:43:23,611 --> 01:43:27,286
they're responding to and their,
assessments of different work

1890
01:43:27,286 --> 01:43:30,843
that's coming out. I rely on a
lot on that to get kind of, a

1891
01:43:30,843 --> 01:43:34,222
filtered feed, because you're
right, it is a firehose. If

1892
01:43:34,222 --> 01:43:36,000
you're just trying to look at.

1893
01:43:36,000 --> 01:43:39,829
Audrow Nash: Every and there's
so much hype and there's so much

1894
01:43:39,829 --> 01:43:43,414
because I mean, like, there's a
lot of promise, but there's

1895
01:43:43,414 --> 01:43:47,182
definitely a lot of hype. Also.
And so having great folks like

1896
01:43:47,182 --> 01:43:50,890
Chris filter it out, seems like
a good thing. All right. Hell

1897
01:43:50,890 --> 01:43:54,779
yeah. Well, this has been a lot
of fun. Really cool to hear what

1898
01:43:54,779 --> 01:43:58,000
you guys are doing. And that
sounds really promising.

1899
01:43:58,000 --> 01:44:01,464
Pras Velagapudi: Thanks. I mean,
this is just a really exciting

1900
01:44:01,464 --> 01:44:04,818
time to be doing this type of
robot, to be doing this type of

1901
01:44:04,818 --> 01:44:08,227
work. So, you know, it is really
awesome to just come into the

1902
01:44:08,227 --> 01:44:11,636
lab and just Digits off picking
groceries or putting Halloween

1903
01:44:11,636 --> 01:44:14,000
candy in a basket, and we're
just like, oh.

1904
01:44:14,000 --> 01:44:15,000
Audrow Nash: That's so cool.

1905
01:44:14,000 --> 01:44:18,531
Pras Velagapudi: That's just
what we're doing today. Like,

1906
01:44:18,531 --> 01:44:19,000
great.

1907
01:44:19,000 --> 01:44:23,000
Audrow Nash: Super cool. Hell
yeah. Alright. Thank you Pras.

1908
01:44:23,000 --> 01:44:25,000
Pras Velagapudi: Yeah. Thanks so
much for having me. Bye.

1909
01:44:25,000 --> 01:44:26,000
Audrow Nash: Bye everyone.
