1
00:00:00,000 --> 00:00:04,242
Audrow Nash: Hi everyone, Audrow
here. This is collection of live

2
00:00:04,242 --> 00:00:08,289
interviews that I did at a local
robotics event between Austin

3
00:00:08,289 --> 00:00:12,075
and San Antonio Texas. I talk
with Kyle Morgenstein, a PhD

4
00:00:12,075 --> 00:00:15,599
student, working on making
legged robots safe (aka not

5
00:00:15,599 --> 00:00:19,776
destroying everything) in human
environments David Venegas, whos

6
00:00:19,776 --> 00:00:23,823
the Head of Operations at Durin,
on automating tasks in mining

7
00:00:23,823 --> 00:00:27,413
discovery and finally, I talk
with Barrett Ames, whos a

8
00:00:27,413 --> 00:00:31,394
co-founder of Botbuilt, about
their robots in construction. I

9
00:00:31,394 --> 00:00:35,049
think you'll like this interview
for a range of robotics

10
00:00:35,049 --> 00:00:39,291
perspectives from legs to mining
and research to startups. I wish

11
00:00:39,291 --> 00:00:43,207
I could say more about the group
hosting this event. They're

12
00:00:43,207 --> 00:00:47,124
currently in stealth, kind of
like a speakeasy, that is it's

13
00:00:47,124 --> 00:00:51,040
an open secret. Hopefully, I'll
be able to share more in the

14
00:00:51,040 --> 00:00:55,086
near future. What I can say is
that I think what they're doing

15
00:00:55,086 --> 00:00:58,872
is incredibly exciting and
audacious. And I think and hope

16
00:00:58,872 --> 00:01:02,266
that it will be a major
accelerator for the robotics

17
00:01:02,266 --> 00:01:06,312
community. If you're curious,
they're having a big event March

18
00:01:06,312 --> 00:01:10,424
8th in between San Antonio and
Austin Texas. It's free, but you

19
00:01:10,424 --> 00:01:14,014
need to apply to get in. If
you're someone in hardtech,

20
00:01:14,014 --> 00:01:17,865
you'll fit right in. I'll put
the event link in the episode

21
00:01:17,865 --> 00:01:21,716
notes. Alright, with that, I
hope you enjoy the interviews.

22
00:01:21,716 --> 00:01:24,000
Kyle, would you introduce
yourself?

23
00:01:24,000 --> 00:01:27,076
Kyle Morgenstein: Hi everybody.
I think I've met about half of

24
00:01:27,076 --> 00:01:30,153
you. I excited to meet the rest
of you. After this. And also a

25
00:01:30,153 --> 00:01:32,882
huge thank you to Ethan, for
putting this on super cool

26
00:01:32,882 --> 00:01:35,760
event, thrilled to be here. I'm
currently a fifth year PhD

27
00:01:35,760 --> 00:01:38,637
student at UT Austin. I've done
a bunch of stuff in legged

28
00:01:38,637 --> 00:01:41,515
robotics. Humanoids. Quadrupeds,
mostly with reinforcement

29
00:01:41,515 --> 00:01:44,244
learning. Some model based
control as well. I spent the

30
00:01:44,244 --> 00:01:47,321
last year at the Boston Dynamics
AI Institute helping get them

31
00:01:47,321 --> 00:01:50,149
off the ground for a lot of
their RL work. Before that, I

32
00:01:50,149 --> 00:01:53,027
was a NASA JPL. Like I mentioned
before, I was a geologist

33
00:01:53,027 --> 00:01:55,954
originally by training, looking
for queens. Yeah. Know that

34
00:01:55,954 --> 00:01:58,832
that's not a joke. I mostly
undergrad, as in geology. Most

35
00:01:58,832 --> 00:02:01,809
of all, my undergrad research
was, evidence of life on Mars.

36
00:02:01,809 --> 00:02:04,786
So the Mars Perseverance rover
got into JPL doing that stuff

37
00:02:04,786 --> 00:02:07,763
there for three years. And then
the robots we have only have

38
00:02:07,763 --> 00:02:10,740
wheels. Couldn't get to a lot of
the cool rock outcrops down

39
00:02:10,740 --> 00:02:13,866
cave, stuff like that. So that's
like on two legged robots. And

40
00:02:13,866 --> 00:02:16,000
so I've been sort of doing that
ever since.

41
00:02:16,000 --> 00:02:18,711
Audrow Nash: That is awesome. So
tell me what you're doing in

42
00:02:18,711 --> 00:02:21,378
your PhD. Like what's the focus?
What good questions are you

43
00:02:21,378 --> 00:02:22,000
investigating?

44
00:02:22,000 --> 00:02:24,561
Kyle Morgenstein: Yeah. So I
think the common theme for my

45
00:02:24,561 --> 00:02:27,123
PhD thus far has been about
putting these legged robots in

46
00:02:27,123 --> 00:02:29,860
human environments, as sort of
the main promise for a humanoid

47
00:02:29,860 --> 00:02:32,598
robot. It's shaped like us, has
our morphology. So we're going

48
00:02:32,598 --> 00:02:35,071
to put it in things like our
homes and medical space and

49
00:02:35,071 --> 00:02:37,853
stuff like that. And so the big
algorithmic problem that I see,

50
00:02:37,853 --> 00:02:40,503
at least that we aren't paying
off attention to safety. What

51
00:02:40,503 --> 00:02:42,887
does it mean to have these
robots interacting with us,

52
00:02:42,887 --> 00:02:45,625
touching us, bumping into us,
things like that? We know how to

53
00:02:45,625 --> 00:02:48,187
control them generally. There's
still a lot of engineering

54
00:02:48,187 --> 00:02:50,836
challenges to go there, but
there's a lot more work to do to

55
00:02:50,836 --> 00:02:53,000
make them safe and sociable and
things like that.

56
00:02:53,000 --> 00:02:56,084
Audrow Nash: Yeah, so I've seen
like on X, a ton of videos,

57
00:02:56,084 --> 00:02:59,376
quadrupeds and humanoid robots
being very sophisticated. Sounds

58
00:02:59,376 --> 00:03:02,355
like back flipping over stuff
and running on all sorts of

59
00:03:02,355 --> 00:03:05,439
surfaces. Tell me about that.
From a safety perspective and

60
00:03:05,439 --> 00:03:08,000
why that doesn't why mobility
isn't quite solved.

61
00:03:08,000 --> 00:03:10,914
Kyle Morgenstein: Yeah. What's
safety? I think they're really

62
00:03:10,914 --> 00:03:13,541
impressive demos. They're
showing sort of the limits of

63
00:03:13,541 --> 00:03:16,024
the hardware and the
capabilities that the platforms

64
00:03:16,024 --> 00:03:19,034
can obtain. But not one of those
videos. Do they show a person?

65
00:03:19,034 --> 00:03:21,852
There's no people near that
robot. And having worked with a

66
00:03:21,852 --> 00:03:24,909
lot of the unit robots, I would
say that those motors are likely

67
00:03:24,909 --> 00:03:27,727
fried after a lot of those
demos, and the robot drives down

68
00:03:27,727 --> 00:03:30,736
the mountain at a single purpose
demo. And so if you want these

69
00:03:30,736 --> 00:03:33,649
things to be doing work in human
spaces repeatedly, reliably,

70
00:03:33,649 --> 00:03:36,659
safely, it's not just, can I do
a backflip? It's can I hand you

71
00:03:36,659 --> 00:03:39,668
an apple and not fall over and
break my own arm or yours when I

72
00:03:39,668 --> 00:03:42,629
do? These are big metal robots
are not light, at the moment at

73
00:03:42,629 --> 00:03:45,447
least. And so I think there's
yeah, a lot of open questions

74
00:03:45,447 --> 00:03:48,217
there that just again, we're
showing off these very flashy

75
00:03:48,217 --> 00:03:51,083
demos that are remarkable that
that progress we've made. But

76
00:03:51,083 --> 00:03:53,806
there really hasn't been a
concerted effort towards human

77
00:03:53,806 --> 00:03:55,000
environments. Yeah. Okay.

78
00:03:55,000 --> 00:03:58,779
Audrow Nash: So when you want to
make a robot better for safe for

79
00:03:58,779 --> 00:04:02,267
human environments, how do we
start approaching that. How do

80
00:04:02,267 --> 00:04:05,000
we make it safe. What ideas of
safe do we have.

81
00:04:05,000 --> 00:04:07,812
Kyle Morgenstein: Yeah totally.
So there's a couple different

82
00:04:07,812 --> 00:04:10,670
answers to that question. The
first is hardware driven. And so

83
00:04:10,670 --> 00:04:13,574
the when you make contact with
the robot is a kind of reflected

84
00:04:13,574 --> 00:04:16,387
load mass, the reflected load
inertia. And this is the weight

85
00:04:16,387 --> 00:04:19,014
you feel at the end effector.
And so mechanically you can

86
00:04:19,014 --> 00:04:21,780
build robots that can dissipate
the energy in some way, some

87
00:04:21,780 --> 00:04:24,408
compliant mechanisms and soft
robotics, things like that.

88
00:04:24,408 --> 00:04:27,220
There's a lot of work that's
been done in this for many, many

89
00:04:27,220 --> 00:04:30,124
years that there's still lots of
progress being made, even like

90
00:04:30,124 --> 00:04:32,660
SOP, 3D printing. Do these
things very efficiently. But

91
00:04:32,660 --> 00:04:35,518
then on the control side, it's
very, very different. We'd like

92
00:04:35,518 --> 00:04:38,007
to embed these sort of same
ideas of compliance into a

93
00:04:38,007 --> 00:04:40,681
controller, like an impedance
based controller. And so the

94
00:04:40,681 --> 00:04:43,355
question is do you want that
intelligence to come from the

95
00:04:43,355 --> 00:04:46,074
mechanical structure or the
controller or some combination.

96
00:04:46,074 --> 00:04:49,071
So a lot of my work is then more
on the control side. I take this

97
00:04:49,071 --> 00:04:52,021
big metal robot. It doesn't have
safety baked into it. How can I

98
00:04:52,021 --> 00:04:54,787
make it safe enough while we
wait for the soft robotics, the

99
00:04:54,787 --> 00:04:57,000
sort of safer mechanisms to be
embedded as well?

100
00:04:57,000 --> 00:05:00,433
Audrow Nash: And that just seems
like a good approach to me,

101
00:05:00,433 --> 00:05:03,980
where you can make it better by
the hardware, but you can do a

102
00:05:03,980 --> 00:05:07,470
lot of it with software outside,
because you can use existing

103
00:05:07,470 --> 00:05:10,903
robots, but also it's going to
be a lot better to control if

104
00:05:10,903 --> 00:05:14,507
you can do it just with a robot.
That's hard. So how do you how

105
00:05:14,507 --> 00:05:18,169
do you go about this? How do you
go about making it. You say for

106
00:05:18,169 --> 00:05:20,000
like how are you thinking of it.

107
00:05:20,000 --> 00:05:23,000
Kyle Morgenstein: Sure. So the,
the, the way this has been done

108
00:05:23,000 --> 00:05:25,857
thus far is if you have some
let's say I don't want to apply

109
00:05:25,857 --> 00:05:28,762
more than 50 newtons of contact
for. So I have some model for

110
00:05:28,762 --> 00:05:31,714
that contact. Then in a model
based controller I can make that

111
00:05:31,714 --> 00:05:34,619
a constraint, whether a hard
constraint or a soft constraint.

112
00:05:34,619 --> 00:05:37,476
I embed that into my control
approach, that the robot cannot

113
00:05:37,476 --> 00:05:40,333
find a solution that creates
more force than whatever I call

114
00:05:40,333 --> 00:05:43,095
my safety threshold. And this
works fine when you're doing

115
00:05:43,095 --> 00:05:46,095
things like picking up a block
where I can model that as just a

116
00:05:46,095 --> 00:05:49,048
perfect cube. It's a rigid body.
Everything's great when we're

117
00:05:49,048 --> 00:05:51,905
talking about people. I'm not a
rigid body at all. I'm quite

118
00:05:51,905 --> 00:05:54,857
soft, actually. And so all of
our models of contact don't work

119
00:05:54,857 --> 00:05:57,429
here. As soon as that's robust
to any shape, any size,

120
00:05:57,429 --> 00:06:00,381
different material properties,
etc.. And that's where learning

121
00:06:00,381 --> 00:06:03,048
really comes into the picture
that in simulation or with

122
00:06:03,048 --> 00:06:06,048
demonstrations, you can try out
of a wide variety of different,

123
00:06:06,048 --> 00:06:08,524
you know, objects and
interactions and try and teach

124
00:06:08,524 --> 00:06:10,952
the robot that I use
reinforcement. So through this

125
00:06:10,952 --> 00:06:13,714
trial and error to achieve your
goals. But it's not a hard

126
00:06:13,714 --> 00:06:16,238
constraint. The way that model
based controllers have

127
00:06:16,238 --> 00:06:19,143
historically data and things
like the manipulation space of a

128
00:06:19,143 --> 00:06:21,000
lot of the industrial arms that
we see.

129
00:06:21,000 --> 00:06:24,642
Audrow Nash: Tell me. So I'm not
100% sure if everyone's familiar

130
00:06:24,642 --> 00:06:27,500
with like the idea of
reinforcement learning is you

131
00:06:27,500 --> 00:06:31,030
tell me what it is like. What's
the setup , and then how you're

132
00:06:31,030 --> 00:06:34,000
using that in a context to learn
safe movement paths.

133
00:06:34,000 --> 00:06:36,727
Kyle Morgenstein: So at its core
reinforcement learning is about

134
00:06:36,727 --> 00:06:39,326
trial and error. So your robot
tries a thing and then you buy

135
00:06:39,326 --> 00:06:41,967
some reward model that you this
a yes robot. That's very good.

136
00:06:41,967 --> 00:06:44,439
Here is your slice of cake or
that was terrible. Crack the

137
00:06:44,439 --> 00:06:47,080
whip. And so it's it's learning
through this you know positive

138
00:06:47,080 --> 00:06:49,722
and negative reinforcement what
actions it should take. And so

139
00:06:49,722 --> 00:06:52,363
it's a very flexible framework
because as opposed to having to

140
00:06:52,363 --> 00:06:54,835
write down a mathematical
objective function that my robot

141
00:06:54,835 --> 00:06:57,476
must just do this one thing, you
can give it lots of different

142
00:06:57,476 --> 00:07:00,203
factors. I want it to do all of
these different things, and then

143
00:07:00,203 --> 00:07:02,845
it's going to learn the these
behaviors through the reward. So

144
00:07:02,845 --> 00:07:05,529
maybe sometimes on Tuesdays you
should go to this classroom. On

145
00:07:05,529 --> 00:07:08,298
Wednesdays on this classroom, as
long as the robot knows you know

146
00:07:08,298 --> 00:07:10,897
what day it is, you can reward
it each way. If you don't have

147
00:07:10,897 --> 00:07:13,368
to tell it explicitly, it is
Wednesday, so go here. It's a

148
00:07:13,368 --> 00:07:15,925
flow statement as opposed to
with a reward model. I can say,

149
00:07:15,925 --> 00:07:18,439
oh, I want you to go here. I
want you to go here. You know,

150
00:07:18,439 --> 00:07:21,165
depending on all these factors,
the robot's able to observe. And

151
00:07:21,165 --> 00:07:23,722
just to that trial and error,
they're always able to pick up

152
00:07:23,722 --> 00:07:25,000
on it and what it ought to do.

153
00:07:25,000 --> 00:07:28,545
Audrow Nash: Okay. How is this
working like, or what kind of

154
00:07:28,545 --> 00:07:32,209
results have you got so far? And
like, how is the state of the

155
00:07:32,209 --> 00:07:35,755
field in general? Like, are we
going to see, humanoid robots

156
00:07:35,755 --> 00:07:38,000
walking around in our homes next
year?

157
00:07:38,000 --> 00:07:40,769
Kyle Morgenstein: I don't know.
Well, our homes maybe you and I

158
00:07:40,769 --> 00:07:43,407
is home to general people. I
don't think it's quite that far

159
00:07:43,407 --> 00:07:45,956
along yet, but we are getting
there. We've been able to do

160
00:07:45,956 --> 00:07:48,681
locomotion very effectively. The
main way this is leveraged in

161
00:07:48,681 --> 00:07:51,187
robotics research is that as
opposed to having a physical

162
00:07:51,187 --> 00:07:53,780
robot that does these things,
you buy to try it on the real

163
00:07:53,780 --> 00:07:56,593
hardware, that could break or be
inaccurate or whatever we tried

164
00:07:56,593 --> 00:07:59,143
in simulation, we can have, you
know, tens of thousands of

165
00:07:59,143 --> 00:08:01,868
robots at the same time trying
these different things. We take

166
00:08:01,868 --> 00:08:04,549
that control policy we learned
in simulation, and then we put

167
00:08:04,549 --> 00:08:07,231
it on our real robot. And we're
currently starting to see our

168
00:08:07,231 --> 00:08:10,044
robots walk very, very robustly.
They walk like people were able

169
00:08:10,044 --> 00:08:12,813
to do a little bit better than
if you've seen some of the older

170
00:08:12,813 --> 00:08:15,407
robot locomotion videos. They
look very constipated. It's a

171
00:08:15,407 --> 00:08:18,132
very low center of mass. And it
came from an algorithm, called

172
00:08:18,132 --> 00:08:20,945
zero Moment Point from maybe 20
years ago. We started to be able

173
00:08:20,945 --> 00:08:23,538
to overcome that, to do more
natural looking walks that are

174
00:08:23,538 --> 00:08:26,264
very stable to disturbances and
plushies, things like that. So

175
00:08:26,264 --> 00:08:29,033
we're getting there. But I think
getting the robots safe enough

176
00:08:29,033 --> 00:08:31,714
that we can reliably put it in
your home without worry, don't

177
00:08:31,714 --> 00:08:34,396
be hit with a huge lawsuit next
year. We're not quite at that

178
00:08:34,396 --> 00:08:37,209
point yet. We're getting there,
but it's going to be more than a

179
00:08:37,209 --> 00:08:38,000
year, I would say.

180
00:08:38,000 --> 00:08:40,589
Audrow Nash: What would the
timeline like? What would your

181
00:08:40,589 --> 00:08:43,000
optimistic and and pessimistic,
mind? Would you guess?

182
00:08:43,000 --> 00:08:45,441
Kyle Morgenstein: I think
optimistic would be like five

183
00:08:45,441 --> 00:08:48,194
ish years, I think. Will you be
able to buy one to put in your

184
00:08:48,194 --> 00:08:50,813
home if you've got the money for
it? Yeah, within maybe two

185
00:08:50,813 --> 00:08:53,565
years. I, I think even now the
Chinese companies will sell you

186
00:08:53,565 --> 00:08:56,228
one. Yeah. You know, you've got
to do all of it on your own.

187
00:08:56,228 --> 00:08:58,891
Now. You can buy them today. But
in terms of, like, it being

188
00:08:58,891 --> 00:09:01,644
stable enough that you could see
it as a product that is being

189
00:09:01,644 --> 00:09:04,307
marketed for something useful
that isn't just a novelty or a

190
00:09:04,307 --> 00:09:06,793
toy. I think five years is
pretty reasonable. To, to do,

191
00:09:06,793 --> 00:09:09,456
you know, very basic things.
Right? It's not like I buy this

192
00:09:09,456 --> 00:09:12,164
thing. It's my at home, but I
can do everything. We're not at

193
00:09:12,164 --> 00:09:14,739
that point. That's maybe more
like a ten year stretch goal

194
00:09:14,739 --> 00:09:17,402
type thing, but safe enough to
put in your home. I'd like to

195
00:09:17,402 --> 00:09:19,000
see five years. Something like
that.

196
00:09:19,000 --> 00:09:21,783
Audrow Nash: That'd be awesome.
Yeah. So you brought a robot. Do

197
00:09:21,783 --> 00:09:23,000
you want to do a small demo?

198
00:09:23,000 --> 00:09:25,772
Kyle Morgenstein: I'll have it
set up right the second. But

199
00:09:25,772 --> 00:09:28,450
after this talk. Yeah, that
definitely we can, I brought,

200
00:09:28,450 --> 00:09:31,456
one of the Unitree go ones. It's
a little quadruped robot. So if

201
00:09:31,456 --> 00:09:34,040
people want to take a turn
trying it out later, I know.

202
00:09:34,040 --> 00:09:37,000
Please come and see me. I'd love
to, to show that around. Yeah.

203
00:09:37,000 --> 00:09:40,361
Audrow Nash: Oh, yeah. So we'll
open up the Q&A in just a little

204
00:09:40,361 --> 00:09:43,721
bit. But you said one thing that
cracked me up is single purpose

205
00:09:43,721 --> 00:09:46,452
demos. Tell me about that,
because that's hilarious.

206
00:09:46,452 --> 00:09:49,655
Because you assume, like, you
see the video of the robot with

207
00:09:49,655 --> 00:09:52,911
that, like the quadruped with
the wheels skidding down a snowy

208
00:09:52,911 --> 00:09:56,167
hill or something like that. And
it's like, oh my God, they're

209
00:09:56,167 --> 00:09:59,317
amazing. Some, some post on X
was like, how would you outrun

210
00:09:59,317 --> 00:10:00,000
one of these?

211
00:10:00,000 --> 00:10:02,795
Kyle Morgenstein: If you could
run long enough, you could, not

212
00:10:02,795 --> 00:10:05,726
just from a barrier perspective,
but the they're these robots are

213
00:10:05,726 --> 00:10:08,341
very fried. Yeah. They're,
they're they're fairly fragile.

214
00:10:08,341 --> 00:10:10,911
They, they degrade pretty
quickly. So when we do a lot of

215
00:10:10,911 --> 00:10:13,706
our experiments, we do a lot of
multi-robot stuff. So you want

216
00:10:13,706 --> 00:10:16,457
to do some kind of search task?
You want a bunch of different

217
00:10:16,457 --> 00:10:19,072
robots searching through, you
know, let's say you're doing

218
00:10:19,072 --> 00:10:21,957
some fire rescue type stuff. We
only could do our experiments at

219
00:10:21,957 --> 00:10:24,753
night during the summers because
that robots overheats. And so

220
00:10:24,753 --> 00:10:27,368
it's even in 20, 30 minutes, it
doesn't work anymore. Even

221
00:10:27,368 --> 00:10:29,577
though this is like a sold
product using the base

222
00:10:29,577 --> 00:10:32,372
controller that it ships with,
after 20 minutes, it overheats.

223
00:10:32,372 --> 00:10:35,168
And so it's they're not super
reliable yet. They work well for

224
00:10:35,168 --> 00:10:37,963
the first couple months. You own
it. But that performance, you

225
00:10:37,963 --> 00:10:39,000
know, sharply degrades.

226
00:10:39,000 --> 00:10:41,652
Audrow Nash: That's so
interesting. So and then you

227
00:10:41,652 --> 00:10:44,408
mentioned at the beginning so
you've been that Boston

228
00:10:44,408 --> 00:10:47,164
Dynamics. How do you compare
Boston Dynamics which is

229
00:10:47,164 --> 00:10:50,284
probably a good bit pricier
versus Unitree which you can buy

230
00:10:50,284 --> 00:10:52,000
those things and throw them
away.

231
00:10:52,000 --> 00:10:55,108
Kyle Morgenstein: Yeah. So when
I say it is Boston Dynamics AI

232
00:10:55,108 --> 00:10:57,764
Institute I think they've
rebranded as robotics an AI

233
00:10:57,764 --> 00:11:00,972
Institute they can't figure out
their name quite yet. But it's a

234
00:11:00,972 --> 00:11:04,079
separate company. Although both
owned by Hyundai. I'll say the

235
00:11:04,079 --> 00:11:06,936
difference between the Boston
Dynamics spot, for example,

236
00:11:06,936 --> 00:11:09,693
versus the Unitree, so, first
for price comparison, the

237
00:11:09,693 --> 00:11:12,650
cheapest model of this is about
$1,600. The Boston Dynamics

238
00:11:12,650 --> 00:11:15,758
starts at 70,000 that goes up to
about 250,000. The difference

239
00:11:15,758 --> 00:11:18,765
that was made, I think it's the
difference between buying a,

240
00:11:18,765 --> 00:11:22,023
tricycle for your toddler versus
a Porsche for yourself. What the

241
00:11:22,023 --> 00:11:24,980
Boston Dynamics robot, the code
I put on it works the first

242
00:11:24,980 --> 00:11:28,238
time. Every time. Really, really
phenomenal. They've done all the

243
00:11:28,238 --> 00:11:31,095
system identification. I know
that if I ask for 30 Newton

244
00:11:31,095 --> 00:11:34,153
meters of torque, I am getting
30 Newton meters. For that I’m

245
00:11:34,153 --> 00:11:37,311
just praying, and I hope that I
get it. I usually not the first

246
00:11:37,311 --> 00:11:40,418
time I turned it on, it jumped
two feet in the air and hit its

247
00:11:40,418 --> 00:11:43,626
own power switch. I wish I had a
video of it. I it sounds crazy,

248
00:11:43,626 --> 00:11:46,834
but it's short circuited and had
itself turned off. So it's it's

249
00:11:46,834 --> 00:11:49,741
a great, you know, starting
platform. It's it's made it so

250
00:11:49,741 --> 00:11:52,498
that robotics research has
become much more accessible.

251
00:11:52,498 --> 00:11:55,605
Lads all over the world. I've
been able to get into it because

252
00:11:55,605 --> 00:11:58,863
of these platforms. So it's it's
really great in that regard. But

253
00:11:58,863 --> 00:12:02,071
it's really hard to compete with
these sort of really expensive,

254
00:12:02,071 --> 00:12:05,078
battle hard and not battle in a
war sense, but like, they're

255
00:12:05,078 --> 00:12:08,236
very well, engineered to design
the Boston Dynamics robots are,

256
00:12:08,236 --> 00:12:10,692
that's a they're very, very
robust. They've got a

257
00:12:10,692 --> 00:12:13,749
well-documented API in English.
And that helps a lot as well.

258
00:12:13,749 --> 00:12:14,000
Yeah.

259
00:12:14,000 --> 00:12:18,939
Audrow Nash: Okay. Hell yeah.
Okay. With that. I am happy to

260
00:12:18,939 --> 00:12:23,878
open it up to questions. What
questions do you guys have for

261
00:12:23,878 --> 00:12:28,735
Kyle? I guess raise your hand
and then I'll repeat what you

262
00:12:28,735 --> 00:12:33,756
say. So we get it on the video.
But who wants to be the first

263
00:12:33,756 --> 00:12:38,530
person? So reward definition is
hard. How do we get safety

264
00:12:38,530 --> 00:12:41,000
especially. How do we do that.

265
00:12:41,000 --> 00:12:43,493
Kyle Morgenstein: Yeah. So
there's, I'll give you the

266
00:12:43,493 --> 00:12:46,504
answer for what we've done thus
far. Then maybe some of the what

267
00:12:46,504 --> 00:12:49,279
we'll do in the future. Right
now, like you said, there's a

268
00:12:49,279 --> 00:12:52,148
lot of engineering effort that
has to go into designing these

269
00:12:52,148 --> 00:12:54,877
reward models. So the sort of
simplest thing you can do at

270
00:12:54,877 --> 00:12:57,746
least a bit of our software,
there's multi humans interacting

271
00:12:57,746 --> 00:13:00,803
with the robot is that we assume
we have perfect knowledge of all

272
00:13:00,803 --> 00:13:03,485
of the contact forces in our
simulator, whether it's soft

273
00:13:03,485 --> 00:13:06,401
body, rigid body, whatever. And
we, you know, show that to the

274
00:13:06,401 --> 00:13:09,318
robot and we say, great, this is
what we think is a reasonable

275
00:13:09,318 --> 00:13:12,234
pain threshold. We give you more
reward for being below it, no

276
00:13:12,234 --> 00:13:15,103
more penalty for going above it,
and it learns sort of how to

277
00:13:15,103 --> 00:13:18,161
act, whether it should get close
to a person versus not even just

278
00:13:18,161 --> 00:13:20,936
like a little, little hallway
driving robot. It'll learn if

279
00:13:20,936 --> 00:13:23,759
it's able to, see the people's
environment. You give it like

280
00:13:23,759 --> 00:13:26,157
cameras, for example. Then it
will learn to be very

281
00:13:26,157 --> 00:13:29,074
aggressively drive until it sees
a person. It slows down where

282
00:13:29,074 --> 00:13:32,037
the if you don't give it camera
data, then it's going to always

283
00:13:32,037 --> 00:13:35,048
go at whatever the fastest speed
is such that it never would hit

284
00:13:35,048 --> 00:13:37,635
the safety threshold. If it
makes contact with somebody

285
00:13:37,635 --> 00:13:40,269
because it can't prevent it,
it's sort of only reacting.

286
00:13:40,269 --> 00:13:42,809
Those are like the like most
naive approach. You could

287
00:13:42,809 --> 00:13:45,679
possibly take what you think is
going to happen a lot more is

288
00:13:45,679 --> 00:13:48,689
that we're going to embed priors
from the way people acts. We're

289
00:13:48,689 --> 00:13:51,747
going to take a lot of data from
things like video sources, maybe

290
00:13:51,747 --> 00:13:54,099
from YouTube, from
demonstrations, and we're going

291
00:13:54,099 --> 00:13:57,109
to train a reward model based on
that. That takes what the robot

292
00:13:57,109 --> 00:13:59,414
has done and make some
assessment. This outputs a

293
00:13:59,414 --> 00:14:02,189
number. This is good, this is
bad. And we already see this.

294
00:14:02,189 --> 00:14:04,965
And how language models are
trained. But to bring that into

295
00:14:04,965 --> 00:14:07,787
the loop for robotics I think
we'll have some of these more,

296
00:14:07,787 --> 00:14:10,421
you know, less rigid rigidly
defines reward models we're

297
00:14:10,421 --> 00:14:13,244
taking. What do we generally
like about how people interact.

298
00:14:13,244 --> 00:14:16,019
And now can we reward our robot
for doing something that is

299
00:14:16,019 --> 00:14:18,559
similar to that distribution
without telling it to act

300
00:14:18,559 --> 00:14:21,381
exactly the way a person does?
We just wanted to approximate

301
00:14:21,381 --> 00:14:24,298
that distribution in some sense
without having to act. If it's

302
00:14:24,298 --> 00:14:26,885
not learning from demonstration
in the sense that we're

303
00:14:26,885 --> 00:14:29,895
repeating the actions, it's just
taking it as sort of a baseline

304
00:14:29,895 --> 00:14:32,059
to move towards. And a
reinforcement learning,

305
00:14:32,059 --> 00:14:33,000
something like that.

306
00:14:33,000 --> 00:14:36,247
Audrow Nash: Yeah. With this
kind of optimization, do you

307
00:14:36,247 --> 00:14:39,949
have like how hard, how hard has
it been to go from simulation to

308
00:14:39,949 --> 00:14:42,000
physical hardware? How is that
been?

309
00:14:42,000 --> 00:14:45,127
Kyle Morgenstein: Yeah. On
Unitree takes about three months

310
00:14:45,127 --> 00:14:48,465
and spot takes about 30 minutes.
Like out of the box and works.

311
00:14:48,465 --> 00:14:51,751
It's again, the big difference
is about system identification.

312
00:14:51,751 --> 00:14:54,984
Knowing that your motors output
what you expect, knowing that

313
00:14:54,984 --> 00:14:58,429
your, you know, simulation model
matches the real robot. When you

314
00:14:58,429 --> 00:15:01,555
have a model that matches your
real robot, it's the easiest

315
00:15:01,555 --> 00:15:04,470
thing in the world when you
don't have it. Yeah, it's a

316
00:15:04,470 --> 00:15:05,000
nightmare.

317
00:15:05,000 --> 00:15:13,800
Audrow Nash: Yep. Because you're
climbing some wrong

318
00:15:13,800 --> 00:15:16,000
distribution.

319
00:15:16,000 --> 00:15:19,000
Kyle Morgenstein: Yeah. Physical
intelligence.

320
00:15:19,000 --> 00:15:22,098
Audrow Nash: So what do you
think of the foundation model

321
00:15:22,098 --> 00:15:24,000
and how that affects your
timeline?

322
00:15:24,000 --> 00:15:26,691
Kyle Morgenstein: Yeah. So I
think the, the big concern is

323
00:15:26,691 --> 00:15:29,335
about what data this is. They're
doing a learning through

324
00:15:29,335 --> 00:15:32,118
demonstration approach. They
have a very large data set that

325
00:15:32,118 --> 00:15:34,994
they've collected and they, you
know, piece the robots to full

326
00:15:34,994 --> 00:15:37,545
laundry to do the dishes,
whatever may be in their data

327
00:15:37,545 --> 00:15:40,050
set. And the big sort of
breakthrough for them is that

328
00:15:40,050 --> 00:15:42,973
they've got one model that can
view all of these tasks. Most of

329
00:15:42,973 --> 00:15:45,895
the demos that you see, both in
academia and industry, have one

330
00:15:45,895 --> 00:15:48,911
model trained for each task that
they do, and some router to pick

331
00:15:48,911 --> 00:15:51,787
which one they're activating at
a time. The big concern I have

332
00:15:51,787 --> 00:15:54,570
for that, their approach, is
that it's bound by the data for

333
00:15:54,570 --> 00:15:57,307
what they've seen before. So if
my home looks significantly

334
00:15:57,307 --> 00:16:00,090
different than anything that
they've seen during training, I

335
00:16:00,090 --> 00:16:02,781
have no assumption that their
model is going to work in my

336
00:16:02,781 --> 00:16:05,796
home. If it does not generalize,
we can only assume interpolation

337
00:16:05,796 --> 00:16:08,440
within the data that they seen
before. So you can collect

338
00:16:08,440 --> 00:16:11,223
enough data that everybody's
home is represented in the data

339
00:16:11,223 --> 00:16:13,914
set, then yeah, that approach
should work. The question is

340
00:16:13,914 --> 00:16:16,883
scale. I don't have an internet
worth of data to back up here. I

341
00:16:16,883 --> 00:16:19,852
only have whatever they're able
to collect. So it's a phenomenal

342
00:16:19,852 --> 00:16:22,588
approach when the thing that
you're deploying matches close

343
00:16:22,588 --> 00:16:25,372
enough to what you've trained
on. But I have real concern to

344
00:16:25,372 --> 00:16:28,016
go to deploy that into people's
homes that there's a huge

345
00:16:28,016 --> 00:16:30,799
diversity of what people's homes
look like. And I'm not sure

346
00:16:30,799 --> 00:16:33,768
we're at the point where we have
enough data to really make that

347
00:16:33,768 --> 00:16:34,000
jump.

348
00:16:34,000 --> 00:16:38,755
Audrow Nash: Yeah, definitely.
There was related, I think it

349
00:16:38,755 --> 00:16:43,589
was the lab that founded it or a
similar one. They released a

350
00:16:43,589 --> 00:16:47,869
bunch of robots being tele
operated and then doing it,

351
00:16:47,869 --> 00:16:52,307
doing many of the tasks
autonomously. And the thing that

352
00:16:52,307 --> 00:16:57,141
I thought was so funny is in
their videos, they had the robot

353
00:16:57,141 --> 00:17:01,500
push chairs back into the back
against a wall, and they

354
00:17:01,500 --> 00:17:06,017
highlighted specifically that
the third chair or whatever

355
00:17:06,017 --> 00:17:10,852
number was not trained on. So it
shows you how specific is it

356
00:17:10,852 --> 00:17:15,527
that they're calling that out as
a really big thing that we

357
00:17:15,527 --> 00:17:20,282
didn't we didn't do this chair
over here. We just did chairs

358
00:17:20,282 --> 00:17:24,800
over here. So it's kind of
spiky. It doesn't work always.

359
00:17:24,800 --> 00:17:29,000
Probably. Is teleoperation
efficient for these tasks?

360
00:17:29,000 --> 00:17:31,773
Kyle Morgenstein: Yeah. This is
the way the word was sample

361
00:17:31,773 --> 00:17:34,640
efficient. And I want to you
explain that first for people in

362
00:17:34,640 --> 00:17:37,507
the audience that might not be
familiar with the terminology.

363
00:17:37,507 --> 00:17:40,233
So sample efficiency is how
efficient is your algorithm at

364
00:17:40,233 --> 00:17:42,959
replicating a behavior per
sample that it's shown. So if I

365
00:17:42,959 --> 00:17:45,826
show it one demonstration of the
behavior and it can learn it

366
00:17:45,826 --> 00:17:48,176
after just seeing it once,
that's a very efficient

367
00:17:48,176 --> 00:17:50,620
algorithm. So one of the
downsides for reinforcement

368
00:17:50,620 --> 00:17:53,440
learning, the method that I tend
to champion is that they're

369
00:17:53,440 --> 00:17:56,166
very, very simple in a fashion.
You have to see many, many

370
00:17:56,166 --> 00:17:58,751
demonstrations in simulation
before you can use it. The

371
00:17:58,751 --> 00:18:01,618
difference is that in simulation
you can generate essentially

372
00:18:01,618 --> 00:18:04,532
unlimited data, more or less for
free with Teleoperation, when

373
00:18:04,532 --> 00:18:07,353
they even ask about is you have
to physically do it. So it's

374
00:18:07,353 --> 00:18:10,079
very, very sample efficient. You
can show just a couple of

375
00:18:10,079 --> 00:18:13,087
iterations. Maybe you show it 20
to 50 examples, and it has this

376
00:18:13,087 --> 00:18:15,907
sort of standard at the at the
moment in the field, although

377
00:18:15,907 --> 00:18:18,774
people are trying to reduce that
and then it works very, very

378
00:18:18,774 --> 00:18:21,500
well. So it's quite sample
efficient. But it's not time or

379
00:18:21,500 --> 00:18:24,132
energy or money efficient
because that's operator hours.

380
00:18:24,132 --> 00:18:26,435
You've got to build the
infrastructure, build the

381
00:18:26,435 --> 00:18:29,396
hardware, use it to collect that
data, maybe clean that data in

382
00:18:29,396 --> 00:18:32,451
some way. So from an algorithmic
perspective is incredibly sample

383
00:18:32,451 --> 00:18:35,271
efficient from a deployment
perspective. Not sure if it's as

384
00:18:35,271 --> 00:18:37,903
efficient as we'd like it to be
yet. But again, as these

385
00:18:37,903 --> 00:18:40,582
algorithms require fewer and
fewer demonstrations, if you

386
00:18:40,582 --> 00:18:42,932
could show a just one
demonstration that it works,

387
00:18:42,932 --> 00:18:45,000
then maybe those economics work
out for you.

388
00:18:45,000 --> 00:18:49,102
Audrow Nash: Oh yeah. One thing
that I've heard that I found

389
00:18:49,102 --> 00:18:53,272
very funny also is that, a lot
of things that they might want

390
00:18:53,272 --> 00:18:56,827
to do in a learning from
demonstration context, they

391
00:18:56,827 --> 00:19:00,587
don't work multiple times. Like
if you're talking about

392
00:19:00,587 --> 00:19:04,757
assembling furniture, like if
you assemble furniture and then

393
00:19:04,757 --> 00:19:08,928
disassemble it like, let's say
it's Ikea furniture, like it's

394
00:19:08,928 --> 00:19:13,029
eventually all the holes are
going to break and you're going

395
00:19:13,029 --> 00:19:17,131
to you're going to end up going
through a whole bunch of new

396
00:19:17,131 --> 00:19:20,481
furniture. For a lot of these
learn learning from

397
00:19:20,481 --> 00:19:24,720
demonstration tasks. So there
are constraints where your stuff

398
00:19:24,720 --> 00:19:28,548
actually wears out, which is
hilarious as a difficulty I

399
00:19:28,548 --> 00:19:32,787
wouldn't have expected. Okay.
Hell, yeah. How well is working,

400
00:19:32,787 --> 00:19:36,000
training off of YouTube data
work in real life.

401
00:19:36,000 --> 00:19:39,143
Kyle Morgenstein: For a humanoid
robot? Not great. The difference

402
00:19:39,143 --> 00:19:42,092
in the kinematics between, you
know, the dimensions of my arm

403
00:19:42,092 --> 00:19:44,944
or the robot's arm, I end up
making a fairly big difference

404
00:19:44,944 --> 00:19:47,894
there. It can be really helpful
for bootstrapping. So you can

405
00:19:47,894 --> 00:19:50,649
use it to start learning. So
you've got some general idea

406
00:19:50,649 --> 00:19:53,695
for, oh, I've got these things
called arms that I can move kind

407
00:19:53,695 --> 00:19:56,500
of like this. But to get really
good performance on a real

408
00:19:56,500 --> 00:19:59,110
system, you do want to train
with as close to the true

409
00:19:59,110 --> 00:20:01,770
dynamics of your robot as
possible. When I mentioned it

410
00:20:01,770 --> 00:20:04,670
earlier, the context of arms I
like where there's a lot more

411
00:20:04,670 --> 00:20:07,378
promise to that kind of data, is
using it as a reward to

412
00:20:07,378 --> 00:20:10,424
encourage the robot to be more
like the person that can juggle.

413
00:20:10,424 --> 00:20:13,180
So as opposed to having to
figure out how to specify some

414
00:20:13,180 --> 00:20:16,129
reward function explicitly for
what juggling is, that's quite

415
00:20:16,129 --> 00:20:19,030
hard. But it's quite easy. The
videos I'm juggling and and I

416
00:20:19,030 --> 00:20:21,979
show my reward model what my
robot is currently doing, versus

417
00:20:21,979 --> 00:20:24,832
the video of someone juggling
and say, are these two things

418
00:20:24,832 --> 00:20:27,974
the same? Action? If yes, reward
if no penalty. That's where that

419
00:20:27,974 --> 00:20:30,682
kind of data becomes really,
really valuable as a reward

420
00:20:30,682 --> 00:20:33,679
process, but not as valuable for
training the robot's dynamics

421
00:20:33,679 --> 00:20:36,000
itself that you do want to match
pretty closely.

422
00:20:36,000 --> 00:20:41,681
Audrow Nash: Cool idea. Yeah,
yeah. Using it as basically a

423
00:20:41,681 --> 00:20:47,459
way of labeling data, so that's
awesome. Would you introduce

424
00:20:47,459 --> 00:20:49,000
yourself, David?

425
00:20:49,000 --> 00:20:54,268
David Venegas: Hi my name is
David Venegas I'm head of

426
00:20:54,268 --> 00:20:57,000
operations at Durin, Mining.

427
00:20:57,000 --> 00:20:59,947
Audrow Nash: Tell me about what
you guys like., the problem you

428
00:20:59,947 --> 00:21:02,661
guys are trying to solve. And
tell me a bit more about the

429
00:21:02,661 --> 00:21:05,000
company. What stage you're at?
How many employees?

430
00:21:05,000 --> 00:21:08,497
David Venegas: Yeah, we're
small, so pre-seed about five

431
00:21:08,497 --> 00:21:12,306
people right now. And basically
mining. A lot of people don't

432
00:21:12,306 --> 00:21:16,052
really have any idea how it
works. So maybe give me a quick,

433
00:21:16,052 --> 00:21:19,736
like, primer just to understand
the basics. Would be really

434
00:21:19,736 --> 00:21:22,983
good. So it is the most
important industry, the most

435
00:21:22,983 --> 00:21:26,855
fundamental industry. You have
mining, you have energy. That's

436
00:21:26,855 --> 00:21:30,601
basically all that we have to
agriculture. Agriculture would

437
00:21:30,601 --> 00:21:34,473
be a step above because you need
the minerals and you need the

438
00:21:34,473 --> 00:21:38,157
energy to get to that and stuff.
So super fundamental, very

439
00:21:38,157 --> 00:21:41,841
fundamental, same core like
manufacturing construction only

440
00:21:41,841 --> 00:21:45,525
so once thereabout. So yes,
these are the two literally the

441
00:21:45,525 --> 00:21:49,334
most important industries. If
you want to see everything else

442
00:21:49,334 --> 00:21:52,768
grow and yeah, it's been
forgotten. Mining has not been

443
00:21:52,768 --> 00:21:56,077
like it's it's very old
industry, Anyway, to break it

444
00:21:56,077 --> 00:21:59,824
down in a few different steps.
You start with exploration to

445
00:21:59,824 --> 00:22:03,321
really know whether something's
worth doing in an actual

446
00:22:03,321 --> 00:22:06,817
development step. And that's
billions of dollars. You're

447
00:22:06,817 --> 00:22:10,439
building towns beside it.
Running railroads out there like

448
00:22:10,439 --> 00:22:12,000
this is a huge operation.

449
00:22:12,000 --> 00:22:14,620
Audrow Nash: So you're saying,
like, surveying the area,

450
00:22:14,620 --> 00:22:17,333
finding the material, and then
all the logistics of, like,

451
00:22:17,333 --> 00:22:19,673
getting energy, getting
railroads, getting all the

452
00:22:19,673 --> 00:22:20,000
things.

453
00:22:20,000 --> 00:22:23,831
David Venegas: Yes, definitely
going to be the development

454
00:22:23,831 --> 00:22:27,662
side. That's when you know, this
is worth doing and you're

455
00:22:27,662 --> 00:22:31,624
putting literally billions of
dollars. So much time, so much

456
00:22:31,624 --> 00:22:35,257
resources. So that's the
development side to get there.

457
00:22:35,257 --> 00:22:39,088
And you have to go through two
previous steps. Simplifying

458
00:22:39,088 --> 00:22:42,985
this. But it's exploration and
that's finding a needle in a

459
00:22:42,985 --> 00:22:47,146
haystack. You're searching the
entire globe. And that needle is

460
00:22:47,146 --> 00:22:50,976
buried on the ground. So very
difficult. Once you explore,

461
00:22:50,976 --> 00:22:54,873
then it goes to feasibility.
That is the biggest bottleneck

462
00:22:54,873 --> 00:22:58,902
permitting it. It's economics.
Is it like six kilometers away

463
00:22:58,902 --> 00:23:03,063
from a port? is it 600km? Is it
this high from ground? Is it is

464
00:23:03,063 --> 00:23:07,290
whatever. Like all these details
really, really matter. And it's

465
00:23:07,290 --> 00:23:11,385
like, what kind of rock is it?
Is this extractable from like a

466
00:23:11,385 --> 00:23:15,216
chemistry standpoint and a
feasible way. And there's a lot

467
00:23:15,216 --> 00:23:19,179
like, okay, do we have sources
to supply these things? Do we

468
00:23:19,179 --> 00:23:23,273
have power or water, all these
things. And these these last 50

469
00:23:23,273 --> 00:23:24,000
plus years.

470
00:23:24,000 --> 00:23:30,691
Audrow Nash: So it's a big
investment upfront. And then you

471
00:23:30,691 --> 00:23:35,000
might get along at a time of
using it.

472
00:23:35,000 --> 00:23:38,606
David Venegas: Oh, yeah. Yeah,
it's a huge thing. So

473
00:23:38,606 --> 00:23:42,976
exploration that is where Durin
comes in. That is where we are.

474
00:23:42,976 --> 00:23:47,345
I can break down exploration for
you a little bit more as well.

475
00:23:47,345 --> 00:23:51,090
This would be necessarily
explain why we're doing what

476
00:23:51,090 --> 00:23:55,182
we're doing. Basically you can
break it down to prospecting

477
00:23:55,182 --> 00:23:59,273
which is determining okay, like
where it is geology. You're

478
00:23:59,273 --> 00:24:03,296
like, all right, what is this
rift do, where you're taking

479
00:24:03,296 --> 00:24:07,249
soil samples, running planes and
helicopters with gravity

480
00:24:07,249 --> 00:24:10,925
sensors, running electrical
magnetic sensors, gravity

481
00:24:10,925 --> 00:24:15,086
sensors to see like differences.
You're looking for problem.

482
00:24:15,086 --> 00:24:19,455
Reality's like something that's
not common because it's kind of

483
00:24:19,455 --> 00:24:23,686
interesting. It's almost every
single mineral you have a tiny

484
00:24:23,686 --> 00:24:27,569
bit like a five parts per
million or whatever, literally

485
00:24:27,569 --> 00:24:31,869
anywhere on Earth. Like right
now you have uranium, gold, like

486
00:24:31,869 --> 00:24:36,238
basically almost everywhere. But
it's the concentration, right?

487
00:24:36,238 --> 00:24:40,469
Is it economically viable? And
these are extremely high. Like

488
00:24:40,469 --> 00:24:44,699
like sometimes you're moving an
entire ton just to get like a

489
00:24:44,699 --> 00:24:48,930
little tiny, tiny bit. Totally.
Depending on the material and

490
00:24:48,930 --> 00:24:52,883
everything. So yeah, all that
plays into like what is the

491
00:24:52,883 --> 00:24:57,114
grade essentially. So it's a 2%
grade. That means out of, you

492
00:24:57,114 --> 00:25:01,622
know, I'm mining 100 of these to
get to units. In the exploration

493
00:25:01,622 --> 00:25:05,783
process, you do your geophysics,
you do your geology. But to

494
00:25:05,783 --> 00:25:09,805
really know what's down there,
to know if it's worth those

495
00:25:09,805 --> 00:25:14,105
billions of dollars, literally
billions. You have to drill. So

496
00:25:14,105 --> 00:25:18,336
you gotta put a rock, put holes
in the ground to get the rock

497
00:25:18,336 --> 00:25:20,000
out. That's all is dome.

498
00:25:20,000 --> 00:25:22,507
Audrow Nash: So you're taking a
sample basically you're going

499
00:25:22,507 --> 00:25:23,000
and probing?

500
00:25:23,000 --> 00:25:26,548
David Venegas: Yes. Called core
sampling. It's the most

501
00:25:26,548 --> 00:25:30,484
expensive part, upwards of 80%,
85%, oftentimes of the entire

502
00:25:30,484 --> 00:25:34,419
exploration process straight
into drilling all the money, all

503
00:25:34,419 --> 00:25:35,000
the time.

504
00:25:35,000 --> 00:25:38,872
Audrow Nash: All the like
surveying and all of the people

505
00:25:38,872 --> 00:25:43,151
looking. why is it so expensive?
I’m imagining like, drilling a

506
00:25:43,151 --> 00:25:47,294
well. Is it offensive? But it’s
not that expensive. But maybe

507
00:25:47,294 --> 00:25:51,030
this is much harder than that.
Much bigger. You have to

508
00:25:51,030 --> 00:25:53,000
probably probe a lot of area.

509
00:25:53,000 --> 00:25:56,277
David Venegas: You’re going
About 100 holes by the time,

510
00:25:56,277 --> 00:25:59,612
like anything's even considered.
And you're going about a

511
00:25:59,612 --> 00:26:03,357
kilometer, two kilometers deep.
It's it depends. You can do like

512
00:26:03,357 --> 00:26:07,161
300m, but generally you're going
pretty deep. You're paying a lot

513
00:26:07,161 --> 00:26:10,789
of money. About $300 a meter. It
can be as high as a thousand,

514
00:26:10,789 --> 00:26:14,007
sometimes as low as 100. But
generally it's within that

515
00:26:14,007 --> 00:26:17,459
variance of like upwards of 300.
And the deeper you go, the

516
00:26:17,459 --> 00:26:21,146
harder it is, the more materials
you have to use and the longer

517
00:26:21,146 --> 00:26:24,481
it takes. So it just keeps
everything constant. Castling.

518
00:26:24,481 --> 00:26:27,934
So that. Right. And the biggest
issue is emissions that are

519
00:26:27,934 --> 00:26:31,620
doing this. The biggest actual
improvement that we've seen. Has

520
00:26:31,620 --> 00:26:35,014
been back in 1958 someone called
bought one year, invented

521
00:26:35,014 --> 00:26:38,525
something called a wireline
drill which instead of literally

522
00:26:38,525 --> 00:26:42,270
moving your case piping up every
time you wanted to, you're like

523
00:26:42,270 --> 00:26:45,781
three meters of sample and you
have to send a wire down, you

524
00:26:45,781 --> 00:26:49,409
get to send a wire down, which
is the step change improvement.

525
00:26:49,409 --> 00:26:52,978
And now, okay, I just bring that
up and I only trip, which is

526
00:26:52,978 --> 00:26:56,314
called tripping when you’re
moving all your pipe up and I

527
00:26:56,314 --> 00:26:59,415
can take hours. Just when you
need to change a bit or

528
00:26:59,415 --> 00:27:00,000
something.

529
00:27:00,000 --> 00:27:02,819
Audrow Nash: So is it like you
lift up everything by a wire,

530
00:27:02,819 --> 00:27:05,779
for a little bit, and then you
change something down there, and

531
00:27:05,779 --> 00:27:07,000
then you let t down again?

532
00:27:07,000 --> 00:27:10,569
David Venegas: If you knew water
well drilling or oil well

533
00:27:10,569 --> 00:27:13,830
drilling you're basically
casing. You're putting pipe

534
00:27:13,830 --> 00:27:17,768
casing. So it's it's got a hole
in the middle and you're putting

535
00:27:17,768 --> 00:27:21,398
a shot down the middle. I'm
pronouncing it incorrectly, but

536
00:27:21,398 --> 00:27:25,336
that will go down as the drill,
which is it's a tungsten carbide

537
00:27:25,336 --> 00:27:28,782
with industrial diamonds. So
it's sort of a diamond core

538
00:27:28,782 --> 00:27:32,535
drill, basically. That'll cut
through anything. So it'll keep

539
00:27:32,535 --> 00:27:36,411
going, you know, bring rock up
through the middle and you're on

540
00:27:36,411 --> 00:27:40,226
that shot. It's waiting. It's
got a spring load thing. Once it

541
00:27:40,226 --> 00:27:43,918
hits a certain 1.53. You get a
sensor that'll tell you, hey,

542
00:27:43,918 --> 00:27:47,733
the pressure is high and you,
you snap it off and you bring it

543
00:27:47,733 --> 00:27:51,548
up, and then you're laying out
on a trough. You're now hitting

544
00:27:51,548 --> 00:27:55,486
with a hammer to get it out into
a cardboard box, and it's going

545
00:27:55,486 --> 00:27:58,624
off to a final stamp for
exploration, so that's the

546
00:27:58,624 --> 00:28:02,623
chemistry side of things. I want
to know what my grade is at 900m

547
00:28:02,623 --> 00:28:06,315
is it, you know, 2%, is it 1%
whatever. And you're using all

548
00:28:06,315 --> 00:28:09,945
that information to go back into
the model. It's a cyclical

549
00:28:09,945 --> 00:28:13,822
process. It takes years. So how
this actually works is you have

550
00:28:13,822 --> 00:28:17,698
a bunch of exploration companies
that are they're called junior

551
00:28:17,698 --> 00:28:21,451
exploration companies. These
aren't big like conglomerates or

552
00:28:21,451 --> 00:28:25,143
anything. They're aren't like
your Rio Tinto's. Your massive

553
00:28:25,143 --> 00:28:28,712
companies do have arms that'll
do this. But generally they

554
00:28:28,712 --> 00:28:32,404
that's it's really risky. So
they push that all off to these

555
00:28:32,404 --> 00:28:36,157
junior exploration companies to
jump on to $20 million market

556
00:28:36,157 --> 00:28:39,479
caps and their whole like
objective is okay by mineral

557
00:28:39,479 --> 00:28:43,110
rights to a specific piece of
land that we think could have

558
00:28:43,110 --> 00:28:46,309
whatever mineral we're looking
for. And then they’re

559
00:28:46,309 --> 00:28:49,693
contracting out all these
services. So the geology, the

560
00:28:49,693 --> 00:28:53,262
drilling, everything. And the
whole point is to raise more

561
00:28:53,262 --> 00:28:57,139
money to do it again and again.
80% of that money goes straight

562
00:28:57,139 --> 00:28:58,000
into drilling.

563
00:28:58,000 --> 00:29:01,966
Audrow Nash: Wow.. Okay. So it's
hill climbing in a sense,

564
00:29:01,966 --> 00:29:05,795
because they're justifying
themselves through getting an

565
00:29:05,795 --> 00:29:09,624
elevator. Because it’s growing
growing growing. And then

566
00:29:09,624 --> 00:29:13,795
eventually big mine gets put
there. So where do you guys come

567
00:29:13,795 --> 00:29:14,000
in?

568
00:29:14,000 --> 00:29:17,881
David Venegas: We build the
drills. Oh so the drills like I

569
00:29:17,881 --> 00:29:21,762
said before it's biggest step
change has been 1958 when you

570
00:29:21,762 --> 00:29:25,445
get a wire going down. I'm
trying to remember this story

571
00:29:25,445 --> 00:29:29,655
correctly. But basically there's
a there's the company that like

572
00:29:29,655 --> 00:29:33,536
they made a they figure out the
the biggest problem and the

573
00:29:33,536 --> 00:29:36,891
biggest injury for their
drillers was restrained by

574
00:29:36,891 --> 00:29:41,100
moving the pipes. So they found
out by like removing two threads

575
00:29:41,100 --> 00:29:44,784
from the screws that were like
putting into the top. 40%

576
00:29:44,784 --> 00:29:48,533
improvement. I'm saying that's
the biggest changes you're

577
00:29:48,533 --> 00:29:52,677
seeing, like little itty bitty
changes. And again, this isn't a

578
00:29:52,677 --> 00:29:56,427
satellite. This is a really
complicated robot. It's not a

579
00:29:56,427 --> 00:30:00,636
fighter jet which we've built on
the back of napkins in 90 days.

580
00:30:00,636 --> 00:30:04,649
Like we're capable of really
incredible things. Yeah, this is

581
00:30:04,649 --> 00:30:08,530
it goes like this and it goes
down. That's it. He brings up

582
00:30:08,530 --> 00:30:12,411
rock. Okay. Yeah. And it's
really expensive. We're building

583
00:30:12,411 --> 00:30:16,292
more autonomous versions of
these. So not fully autonomous,

584
00:30:16,292 --> 00:30:20,501
not trying to go there. There's,
you know, the 20% that gets you

585
00:30:20,501 --> 00:30:24,448
the 80%. So that's the whole
point is whatever you can do to

586
00:30:24,448 --> 00:30:28,197
make, it just less humans
because, again, safety, that is

587
00:30:28,197 --> 00:30:32,078
the whole thing. Yeah. It's
about 50% of all drilling costs

588
00:30:32,078 --> 00:30:35,959
in north of that is straight
into labor. And this isn't fun

589
00:30:35,959 --> 00:30:39,708
work. It's not a fun job. Yeah,
you're you're losing your

590
00:30:39,708 --> 00:30:43,721
fingers, you're inhaling dust,
your pounding on rods, 12 hour

591
00:30:43,721 --> 00:30:47,865
shifts constantly. Like, yeah,
it's not fun. Very tough. But it

592
00:30:47,865 --> 00:30:51,877
is cool. Rocks are cool. I, I
know all geologists love rocks.

593
00:30:51,877 --> 00:30:55,627
We all love rocks. I like,
fundamentally like, trucks are

594
00:30:55,627 --> 00:30:59,573
sick. When you're a kid, you're
too. You're playing with the

595
00:30:59,573 --> 00:31:03,586
truck. But this industry is in a
really bad place. It's about

596
00:31:03,586 --> 00:31:07,532
been avoiding, change. It's been
avoiding the need to, like,

597
00:31:07,532 --> 00:31:11,611
innovate. This is the way it's
always been done. It's old. The

598
00:31:11,611 --> 00:31:15,492
average age is 55. Half the
workforce is going to retire by

599
00:31:15,492 --> 00:31:19,504
2029. The boomers, 5% of the
workforce is under 30. So nobody

600
00:31:19,504 --> 00:31:23,648
thinks it's cool. Nobody thinks
mining is sexy. You know, we do

601
00:31:23,648 --> 00:31:27,726
like it. It's not appealing to
anyone. And all the like, smart

602
00:31:27,726 --> 00:31:31,607
engineers and talented people
don't want to go up. Don't go

603
00:31:31,607 --> 00:31:35,356
out and do this because it
actually isn't like appealing.

604
00:31:35,356 --> 00:31:39,566
There's no appeal. And that's a
problem. This is one of the most

605
00:31:39,566 --> 00:31:42,000
fundamental, industries in
existence.

606
00:31:42,000 --> 00:31:44,833
Audrow Nash: So what is, I don't
fully understand where your

607
00:31:44,833 --> 00:31:47,618
improvement is to this sort of
automated to a larger extent

608
00:31:47,618 --> 00:31:50,639
makes it fewer people have to do
it. You can have younger people

609
00:31:50,639 --> 00:31:53,000
that don't want to hammer the
fingers off working.

610
00:31:53,000 --> 00:31:55,976
David Venegas: Yeah. You can
have this same person that can

611
00:31:55,976 --> 00:31:59,053
run ten of these sitting in a
little box and monitoring them.

612
00:31:59,053 --> 00:32:02,030
Okay. This one's going to okay.
I'm like, you hear my water

613
00:32:02,030 --> 00:32:05,107
there or whatever. That's a a
bit of an abstraction. But yes,

614
00:32:05,107 --> 00:32:08,234
that is like the idea. Yeah. We
want monitoring. We don't want

615
00:32:08,234 --> 00:32:11,110
24 over seven operation because
what you have is somebody

616
00:32:11,110 --> 00:32:13,834
sitting there waiting for this
thing to be going down.

617
00:32:13,834 --> 00:32:17,012
Somebody's sitting in the cage,
literally like listening to how

618
00:32:17,012 --> 00:32:20,291
the engine and how the generator
goes in and like, okay, how's it

619
00:32:20,291 --> 00:32:23,570
going? Me oh, I know that sound.
And they're adjusting the RPM in

620
00:32:23,570 --> 00:32:26,647
the water flow or the torque.
That's how it gets sucked. That

621
00:32:26,647 --> 00:32:27,000
is why.

622
00:32:27,000 --> 00:32:33,453
Audrow Nash: It's like a person
with a CNC. You do the same

623
00:32:33,453 --> 00:32:39,906
thing there. Because they listen
to it. So how are you guys

624
00:32:39,906 --> 00:32:41,000
improving?

625
00:32:41,000 --> 00:32:45,048
David Venegas: So he goes on in
there. But long you has some

626
00:32:45,048 --> 00:32:48,893
drills that do this in some
ways, but it's not the closed

627
00:32:48,893 --> 00:32:53,143
loop fully. You just close the
loop. The rod handling. Make the

628
00:32:53,143 --> 00:32:57,123
rods do like this. Have the
pipe, the dirt come out. That's

629
00:32:57,123 --> 00:32:58,000
all you need.

630
00:32:58,000 --> 00:33:01,430
Audrow Nash: Wow. Very simple.
Seems very effective. So you

631
00:33:01,430 --> 00:33:03,000
guys are pre-seed you said?

632
00:33:03,000 --> 00:33:09,200
David Venegas: We showed some
numbers. You know TechCrunch and

633
00:33:09,200 --> 00:33:15,300
Wall Street Journal want to do
an article soon. So new stuff.

634
00:33:15,300 --> 00:33:20,000
But 2 million pre-seed. We'll
raise a seed. Q2.

635
00:33:20,000 --> 00:33:21,000
Audrow Nash: Yeah, where are you
guys located?

636
00:33:21,000 --> 00:33:25,683
David Venegas: We’re located in
LA, Ganado, California. So yeah,

637
00:33:25,683 --> 00:33:27,000
it's a good place.

638
00:33:27,000 --> 00:33:29,231
Audrow Nash: Oh, yeah. What
brought you out here? You here

639
00:33:29,231 --> 00:33:30,000
just for this event?

640
00:33:30,000 --> 00:33:33,540
David Venegas: I'm here. Meet
people I connections. We're

641
00:33:33,540 --> 00:33:36,769
looking to hire two people.
Automations and controls

642
00:33:36,769 --> 00:33:39,998
engineer and a robotics
engineer. So people who have

643
00:33:39,998 --> 00:33:43,662
these types of skills, that's
that's very important. We can

644
00:33:43,662 --> 00:33:47,078
hire those, yesterday. So
yesterday, you know. Awesome.

645
00:33:47,078 --> 00:33:50,804
All right, again, if anybody
knows, I also might be a little

646
00:33:50,804 --> 00:33:54,344
easier. There's a site called
learn.durin.com Recommended

647
00:33:54,344 --> 00:33:58,008
Durin is the name. In case I
didn't mention that. Durin and

648
00:33:58,008 --> 00:34:01,858
talking reference. Yeah. That'll
kind of give you a little bit

649
00:34:01,858 --> 00:34:05,336
more to work with. There's
citations and stuff if you're

650
00:34:05,336 --> 00:34:09,000
interested. So yeah, I know
we've probably gone pretty far.

651
00:34:09,000 --> 00:34:12,827
Audrow Nash: No, it's great. I
mean, so this is an industry

652
00:34:12,827 --> 00:34:16,395
like I don't know much about.
And so it's nice to get a

653
00:34:16,395 --> 00:34:20,222
thorough walkthrough of how it
works.. And it sounds like a

654
00:34:20,222 --> 00:34:24,114
valuable thing that you guys are
doing. And it's, it's funny

655
00:34:24,114 --> 00:34:28,200
using like the first thing you
learn in a robotics education is

656
00:34:28,200 --> 00:34:32,416
like a PID controller, you know,
for this kind of thing. And it's

657
00:34:32,416 --> 00:34:36,438
wonderful to apply that kind of
technology to things, work and

658
00:34:36,438 --> 00:34:40,459
adds a lot of value. Well, so
you can have one person watching

659
00:34:40,459 --> 00:34:44,092
ten stations instead of ten
people watching ten station.

660
00:34:44,092 --> 00:34:48,114
Yeah, this kind of thing. Yeah.
Let's see. So I'd like to open

661
00:34:48,114 --> 00:34:51,941
it up for questions. And I have
I want your questions. What

662
00:34:51,941 --> 00:34:55,962
questions do we have? Have you
had trouble like getting on the

663
00:34:55,962 --> 00:34:57,000
excavation site?

664
00:34:57,000 --> 00:35:00,355
David Venegas: No. Everyone
needs this and I know they need

665
00:35:00,355 --> 00:35:03,936
it really bad. Yeah. This is the
linchpin in everything. Again,

666
00:35:03,936 --> 00:35:06,893
all the cost is going this
direction. It's a culture

667
00:35:06,893 --> 00:35:10,304
problem where they can't get the
talent. They don't have the

668
00:35:10,304 --> 00:35:13,659
Silicon Valley money or the
space talent to make it happen.

669
00:35:13,659 --> 00:35:14,000
We do.

670
00:35:14,000 --> 00:35:19,589
Audrow Nash: Let's see other
questions? What do you have so

671
00:35:19,589 --> 00:35:23,000
far and what's the efficiency of
it.

672
00:35:23,000 --> 00:35:27,244
David Venegas: Yeah. We're we're
putting together we're we

673
00:35:27,244 --> 00:35:31,854
literally hired the engineering
team on six weeks ago. So we'll

674
00:35:31,854 --> 00:35:35,805
have holes in the ground
February, mid February. Is it

675
00:35:35,805 --> 00:35:40,195
going very soon. Iterate on
prototypes, manufacture a lot of

676
00:35:40,195 --> 00:35:44,439
these. We're operating them
ourselves. That's where you're

677
00:35:44,439 --> 00:35:47,000
going to see real margins. 3 or
7%.

678
00:35:47,000 --> 00:35:49,188
Audrow Nash: It also sounds like
it's early. You’ll find out a

679
00:35:49,188 --> 00:35:50,000
lot of this eventually.

680
00:35:50,000 --> 00:35:52,938
David Venegas: We know a lot of
this. There's a lot of low

681
00:35:52,938 --> 00:35:56,027
hanging fruit and ridiculous
amount of low hanging fruit when

682
00:35:56,027 --> 00:35:59,168
it comes to this. The biggest
one, again, is, is humans safety

683
00:35:59,168 --> 00:36:01,700
factor, exploration when
companies want to want to

684
00:36:01,700 --> 00:36:04,334
increase their safety factor
because, start drilling

685
00:36:04,334 --> 00:36:07,322
contractors, which we are the
drillers want to increase the

686
00:36:07,322 --> 00:36:10,462
safety factor because they get
more jobs, they get better bids

687
00:36:10,462 --> 00:36:13,603
from the exploration companies,
because if somebody gets hurt,

688
00:36:13,603 --> 00:36:16,642
it's not just the driller who's
paying, it's the exploration

689
00:36:16,642 --> 00:36:19,174
company. And it's really
expensive. So my dropping

690
00:36:19,174 --> 00:36:22,264
safety, you'll actually get a
lot more like not it's actually

691
00:36:22,264 --> 00:36:24,948
more important than like
throughput and speed because

692
00:36:24,948 --> 00:36:27,683
that'll that'll kill an
exploration company if you get

693
00:36:27,683 --> 00:36:29,000
too many safety incidents.

694
00:36:29,000 --> 00:36:33,640
Audrow Nash: And I know also it
sounds like so like grocery

695
00:36:33,640 --> 00:36:38,438
store kiosk kind of thing where
you have one, checkout person

696
00:36:38,438 --> 00:36:42,685
manning ten kiosks that
customers are going to like if

697
00:36:42,685 --> 00:36:46,854
you have one person manning ten
station. That’s a big

698
00:36:46,854 --> 00:36:51,809
multiplier. That sounds like a
very nice thing. Yeah. Let's see

699
00:36:51,809 --> 00:36:56,371
maybe one more question and then
we'll move on to our next

700
00:36:56,371 --> 00:36:57,000
speaker.

701
00:36:57,000 --> 00:37:00,805
David Venegas: Do you simulation
at all? How do you handle tough

702
00:37:00,805 --> 00:37:04,312
things like friction? Yeah. We
will you whatever simulation

703
00:37:04,312 --> 00:37:07,820
test stands we need for for
those sorts of things right now

704
00:37:07,820 --> 00:37:11,446
building the structure, getting
getting it going. It's not as

705
00:37:11,446 --> 00:37:15,251
complicated as you think you're
actually in a really constrained

706
00:37:15,251 --> 00:37:18,817
environment. It's you just need
to make sure that it's okay.

707
00:37:18,817 --> 00:37:22,682
You're hitting this. Let's let's
adjust. Yes. There are like, I'm

708
00:37:22,682 --> 00:37:25,000
not sure if I got the question
correct.

709
00:37:25,000 --> 00:37:28,322
Audrow Nash: I mean, that makes
sense to me. I would imagine for

710
00:37:28,322 --> 00:37:31,696
your specific use case that just
some equations would be what you

711
00:37:31,696 --> 00:37:34,810
want to run everything through
and you can optimize within a

712
00:37:34,810 --> 00:37:38,028
simple model. Not that you need
like a whole visual simulation

713
00:37:38,028 --> 00:37:40,000
and and everything. What do you
think?

714
00:37:40,000 --> 00:37:43,179
David Venegas: Yeah. There's
multiple things to to to attack.

715
00:37:43,179 --> 00:37:46,410
One of them is the rod handling.
So that's robotics. Just move

716
00:37:46,410 --> 00:37:49,642
stuff around. Maybe that's not
what we're going to go for long

717
00:37:49,642 --> 00:37:52,352
term. You want to build
something custom that's very

718
00:37:52,352 --> 00:37:55,739
reliable and the that it can get
it. So we are looking working on

719
00:37:55,739 --> 00:37:56,000
that.

720
00:37:56,000 --> 00:38:02,100
Audrow Nash: Oh yeah. Awesome.
Barrett, you want to introduce

721
00:38:02,100 --> 00:38:03,000
yourself?

722
00:38:03,000 --> 00:38:07,204
Barrett Ames: Yeah, sure. Hey,
everybody. My name is Barrett.

723
00:38:07,204 --> 00:38:11,202
I'm a co-founder at Botbulit.
Been working on robots for a

724
00:38:11,202 --> 00:38:15,130
really long time. When I was a
little kid, I came running

725
00:38:15,130 --> 00:38:19,197
downstairs one morning and I
said, mom, I just want to blow

726
00:38:19,197 --> 00:38:23,332
shit up. And she said, you're
going robot camp. So I've been

727
00:38:23,332 --> 00:38:27,330
building robots since then. I
did my undergrad at Cornell.

728
00:38:27,330 --> 00:38:31,603
Mechanical engineering, computer
science. After that, I got to

729
00:38:31,603 --> 00:38:35,393
work at NASA. Houston, on
Valkyrie, Robonaut, Atlas did

730
00:38:35,393 --> 00:38:39,529
that for a while. Decided I
should go do, PhD in my at Duke.

731
00:38:39,529 --> 00:38:43,733
Focus on the intersection of
artificial intelligence and long

732
00:38:43,733 --> 00:38:47,455
term planning and, like,
geometry of robots. Right? So

733
00:38:47,455 --> 00:38:51,659
just shove all that together. We
started Bob built four years

734
00:38:51,659 --> 00:38:56,001
ago. We build homes with robots.
We do that in a factory. Think

735
00:38:56,001 --> 00:38:58,000
prefab components for a home.

736
00:38:58,000 --> 00:39:01,000
Audrow Nash: Oh, yeah. Okay,
tell me about the need for that.

737
00:39:01,000 --> 00:39:05,560
Barrett Ames: Yeah, yeah, the
need, is a lot like mining,

738
00:39:05,560 --> 00:39:10,440
right? So 40% of contractors are
over the age of 50. Framing,

739
00:39:10,440 --> 00:39:15,000
rough carpentry. Not something
you do well into your 60s.

740
00:39:15,000 --> 00:39:17,857
Audrow Nash: And when we say
framing this to make it totally

741
00:39:17,857 --> 00:39:20,000
clear, that's what. How do you
describe that?

742
00:39:20,000 --> 00:39:23,443
Barrett Ames: Yeah, it's just
the it's all the structural

743
00:39:23,443 --> 00:39:27,067
elements in a residential in a
home. Right. So it's it's all

744
00:39:27,067 --> 00:39:30,510
the two by fours that you stick
together. Right now we're

745
00:39:30,510 --> 00:39:34,255
focused on the walls. We're also
working on on roof trusses as

746
00:39:34,255 --> 00:39:38,000
well. But they're that's kind of
anything that makes it stand.

747
00:39:38,000 --> 00:39:40,930
Audrow Nash: So the idea is you
take all these frames and you

748
00:39:40,930 --> 00:39:43,620
kind of put them all up. And
that's the good start for a

749
00:39:43,620 --> 00:39:46,598
house. They need to do the roof
separately. And you're working

750
00:39:46,598 --> 00:39:49,000
on that. So massive labor
shortages. Tell me more.

751
00:39:49,000 --> 00:39:53,386
Barrett Ames: Yeah. So it's, the
big, big labor win. Is that

752
00:39:53,386 --> 00:39:57,553
right now, when you frame a
home, it takes about 14 to 21

753
00:39:57,553 --> 00:40:01,865
days for a crew on the jobsite
to do it. When they use boat

754
00:40:01,865 --> 00:40:06,251
build panels, it let some stand
up a home in 4 to 8 hours so

755
00:40:06,251 --> 00:40:10,711
that’s massive labor win there,
and we effectively move it to

756
00:40:10,711 --> 00:40:14,000
the robots and we move it off of
the jobsite.

757
00:40:14,000 --> 00:40:16,514
Audrow Nash: So how does that
work? So you say you move it off

758
00:40:16,514 --> 00:40:17,000
the jobsite?

759
00:40:17,000 --> 00:40:18,000
Barrett Ames: Yeah. Yeah.

760
00:40:18,000 --> 00:40:20,362
Audrow Nash: How do you. So I'm
imagining like a big frame for

761
00:40:20,362 --> 00:40:22,000
the side of a house. How do you
ship that.?

762
00:40:22,000 --> 00:40:25,586
Barrett Ames: Yeah, it's
actually pretty easy. Most, most

763
00:40:25,586 --> 00:40:29,361
ceilings are like 9 to 10ft, so
you can you can build a 9 or

764
00:40:29,361 --> 00:40:33,387
10ft wall, stick it on the back
of a truck and ship it. And then

765
00:40:33,387 --> 00:40:37,225
we bad. Yeah. We just flatpack
everything. So the the crew on

766
00:40:37,225 --> 00:40:41,000
site stands it up. We're not
doing, like, big modular stuff.

767
00:40:41,000 --> 00:40:44,860
Audrow Nash: Okay. And so one
thing that I have heard has been

768
00:40:44,860 --> 00:40:48,782
tough for a lot of construction
robotics companies is it's like

769
00:40:48,782 --> 00:40:52,642
there's a balance between cookie
cutter homes that are all the

770
00:40:52,642 --> 00:40:56,191
exact same and some sort of
custom home. How where do you

771
00:40:56,191 --> 00:40:57,000
fall in this?

772
00:40:57,000 --> 00:41:01,254
Barrett Ames: Yeah, we go of
full custom. So you give us a

773
00:41:01,254 --> 00:41:05,655
floor plan and we will build
your house. We've done 25 homes

774
00:41:05,655 --> 00:41:09,910
so far, and 22 of them are
different homes. You just bring

775
00:41:09,910 --> 00:41:14,531
your floor plan, whatever it is.
Like if it's got dimensions on

776
00:41:14,531 --> 00:41:19,005
it and the lines are straight
ish. Will work on it. Okay. And

777
00:41:19,005 --> 00:41:23,406
that's that's like a massive
differentiation, between us and

778
00:41:23,406 --> 00:41:27,000
everyone else who's tried to do
this in the past.

779
00:41:27,000 --> 00:41:30,319
Audrow Nash: Why have other
people not or other companies

780
00:41:30,319 --> 00:41:33,987
not taken this approach? It's it
seems like the ideal like I, I

781
00:41:33,987 --> 00:41:37,306
if I was naively to think a
company should make frames. I

782
00:41:37,306 --> 00:41:40,858
would think you just make it
custom as you're going. And that

783
00:41:40,858 --> 00:41:44,061
would be what customers want.
And you're just basically

784
00:41:44,061 --> 00:41:47,787
replacing the labor that builds
the frame. But I imagine there's

785
00:41:47,787 --> 00:41:50,000
some thorny problems on the way
there.

786
00:41:50,000 --> 00:41:55,046
Barrett Ames: It is a huge pain
in the ass. And it's for a bunch

787
00:41:55,046 --> 00:41:59,934
of different reasons. First and
foremost, we all look at floor

788
00:41:59,934 --> 00:42:04,900
plans, and as engineers, we go.
That's a technical document. It

789
00:42:04,900 --> 00:42:09,946
is not right like it is. One of
the floor plans that we recently

790
00:42:09,946 --> 00:42:14,834
processed, didn't actually form
a closed polygon. So there are

791
00:42:14,834 --> 00:42:19,801
gaps in this house. So we called
up the architect, and we said,

792
00:42:19,801 --> 00:42:24,531
what's what's going on with
this? Like your house is doesn't

793
00:42:24,531 --> 00:42:28,788
house, right? And the response
was, well, I don't like

794
00:42:28,788 --> 00:42:33,203
fractions, so I just rounded
them out. But didn't change

795
00:42:33,203 --> 00:42:37,697
anything. Right? So it's not
really a technical document.

796
00:42:37,697 --> 00:42:42,743
It's mostly like design intent.
Right. And so starting with that

797
00:42:42,743 --> 00:42:47,394
is your framework. You have to
go and build like this whole

798
00:42:47,394 --> 00:42:51,494
generative approach that
understands a deeper level.

799
00:42:51,494 --> 00:42:56,539
Like, yes, we want all of these
constraints to be correct. These

800
00:42:56,539 --> 00:43:01,585
other constraints we don't care
about so much. So if you need to

801
00:43:01,585 --> 00:43:06,000
violate one of those, go ahead
and violate one of those.

802
00:43:06,000 --> 00:43:08,762
Audrow Nash: That is
interesting. So tell me about

803
00:43:08,762 --> 00:43:12,020
the generative approach. It
sounds like you probably need a

804
00:43:12,020 --> 00:43:15,169
way of taking what they give
you, which maybe in a lot of

805
00:43:15,169 --> 00:43:17,930
forms what the builders or
whoever give you or the

806
00:43:17,930 --> 00:43:21,355
architects. And then you turn
that into something that is very

807
00:43:21,355 --> 00:43:24,558
specific so robots can build
it.. Tell me about that whole

808
00:43:24,558 --> 00:43:25,000
process.

809
00:43:25,000 --> 00:43:28,987
Barrett Ames: Yeah. It's,
basically comes in, we have this

810
00:43:28,987 --> 00:43:32,973
PDF, right? We have to extract
all the information that is

811
00:43:32,973 --> 00:43:36,960
basically, door locations,
window locations, dimensions of

812
00:43:36,960 --> 00:43:41,358
walls. And in many cases, those
constraints will be conflicting.

813
00:43:41,358 --> 00:43:45,345
And so we have to say, what's
the, what's the intent here?

814
00:43:45,345 --> 00:43:49,400
Right. And so what we did is we
just took we have one point

815
00:43:49,400 --> 00:43:53,456
7,000,000ft² of plans that we
process and shoved into this.

816
00:43:53,456 --> 00:43:57,786
And that lets us pull out kind
of an intent because that's what

817
00:43:57,786 --> 00:44:01,635
framers do, right? They work
with craptastic plans every

818
00:44:01,635 --> 00:44:05,415
single day. And what they do is
they just lean on their

819
00:44:05,415 --> 00:44:09,058
experience to say, all right,
these two things are in

820
00:44:09,058 --> 00:44:12,907
conflict, but usually people
want it this way. So that's

821
00:44:12,907 --> 00:44:16,000
that's the, the approach that we
really take.

822
00:44:16,000 --> 00:44:19,701
Audrow Nash: That is so cool. So
you are. How did you get this

823
00:44:19,701 --> 00:44:20,000
data?

824
00:44:20,000 --> 00:44:24,571
Barrett Ames: Beg, borrow and
steal. Just like, anything,

825
00:44:24,571 --> 00:44:29,542
anything you can do to get more
floor plan data. Like, I think

826
00:44:29,542 --> 00:44:34,434
anytime I meet a builder, I'm
like, hey, you got floor plans?

827
00:44:34,434 --> 00:44:37,000
Like, can I get some hook me up.

828
00:44:37,000 --> 00:44:40,524
Audrow Nash: That's hilarious.
So when you get all of this data

829
00:44:40,524 --> 00:44:44,161
and then you annotate it in your
system in some way, and that was

830
00:44:44,161 --> 00:44:45,000
time consuming.

831
00:44:45,000 --> 00:44:50,478
Barrett Ames: Yeah. So there's,
a ton of annotation that we had

832
00:44:50,478 --> 00:44:51,000
to do.

833
00:44:51,000 --> 00:44:53,488
Audrow Nash: They're
effectively. Well, I guess

834
00:44:53,488 --> 00:44:56,718
they're like 2D for each one,
but it makes a 3D structure and

835
00:44:56,718 --> 00:45:00,000
there's probably relationships
between the different 2D panel.

836
00:45:00,000 --> 00:45:03,644
Barrett Ames: Right. So the, at
the floor plan level, there's

837
00:45:03,644 --> 00:45:06,810
actually like a bunch of
constraints, a little bit of

838
00:45:06,810 --> 00:45:10,096
structural detail, but it
doesn't actually specify like

839
00:45:10,096 --> 00:45:13,740
this is where this two by four
goes. It doesn't actually tell

840
00:45:13,740 --> 00:45:17,384
you how to build it. Right. The
carpenters do that on the job

841
00:45:17,384 --> 00:45:20,909
site. So we actually have to
generate all of that plan, and

842
00:45:20,909 --> 00:45:23,000
do that according to building
code.

843
00:45:23,000 --> 00:45:26,711
Audrow Nash: So how does that
how do you generate primitives

844
00:45:26,711 --> 00:45:30,237
or like low level things that
you're using to inform your

845
00:45:30,237 --> 00:45:34,072
approach for a new design. Like
it's there's a bunch of ways I

846
00:45:34,072 --> 00:45:37,907
could imagine that you could do
it. But I wonder what you have

847
00:45:37,907 --> 00:45:41,000
done for that. That has been
like the most useful.

848
00:45:41,000 --> 00:45:44,648
Barrett Ames: Yeah. Yeah. I
mean, we we really think about

849
00:45:44,648 --> 00:45:48,547
it as like constraints. And then
we have an objective function

850
00:45:48,547 --> 00:45:52,258
that's just like minimizing
cost, minimizing robot speed or

851
00:45:52,258 --> 00:45:55,780
like robot time, not robot
speed. Yeah. And then we just

852
00:45:55,780 --> 00:45:59,679
it's actually, like, generative
brings in, builds up this nice

853
00:45:59,679 --> 00:46:03,327
set of constraints, and then we
just have, a massive mixed

854
00:46:03,327 --> 00:46:07,164
integer program that we run that
says these are all the right

855
00:46:07,164 --> 00:46:11,000
places to put studs because it
satisfies all the constraints.

856
00:46:11,000 --> 00:46:15,308
Audrow Nash: Cool. That's wild.
So that that's like the crux of

857
00:46:15,308 --> 00:46:19,000
the problem you guys have
solved. Is that fair to say?

858
00:46:19,000 --> 00:46:22,937
Barrett Ames: That's step one.
Step two and step three. Step

859
00:46:22,937 --> 00:46:26,940
two is now that you have this
model of this house, how do you

860
00:46:26,940 --> 00:46:30,877
go and build it. Right. Like
what are the steps of the robot

861
00:46:30,877 --> 00:46:34,879
needs to actually do at a high
level. Right. It's kind of the

862
00:46:34,879 --> 00:46:39,013
task level planning. And if you
guys are familiar with task and

863
00:46:39,013 --> 00:46:42,819
motion planning, like it's
exactly that problem. It's just

864
00:46:42,819 --> 00:46:44,000
construction here.

865
00:46:44,000 --> 00:46:46,000
Audrow Nash: And that's. Then do
that then. Right?

866
00:46:46,000 --> 00:46:49,200
Barrett Ames: Right. But you
have to, you know, you have to,

867
00:46:49,200 --> 00:46:52,400
you have to satisfy all your,
your robot constraints as well

868
00:46:52,400 --> 00:46:54,000
as your task level constraint.

869
00:46:54,000 --> 00:46:56,000
Audrow Nash: Definitely.

870
00:46:56,000 --> 00:47:01,031
Barrett Ames: And then step
three is go and do that. But now

871
00:47:01,031 --> 00:47:05,726
nothing's a perfect poly or
like. So we made a very, big

872
00:47:05,726 --> 00:47:10,840
decision early on, which is to
use the exact same lumber that

873
00:47:10,840 --> 00:47:15,955
everyone is already using. And
so that means it's going to be

874
00:47:15,955 --> 00:47:20,566
borrowed and twisted, warped
wayand like, it's going to

875
00:47:20,566 --> 00:47:25,933
happen. And that lets us operate
at a much lower cost, right? If

876
00:47:25,933 --> 00:47:30,292
you look at a lot of the other
previously mechanized

877
00:47:30,292 --> 00:47:35,323
approaches, they were like, we
need the best wood ten times,

878
00:47:35,323 --> 00:47:37,000
but quite expensive.

879
00:47:37,000 --> 00:47:40,806
Audrow Nash: Yeah. Okay. That
seems like a good approach to

880
00:47:40,806 --> 00:47:41,000
me.

881
00:47:41,000 --> 00:47:43,637
Barrett Ames: Yeah. At that
level, then you have to enable a

882
00:47:43,637 --> 00:47:45,000
whole bunch of computer vision.

883
00:47:45,000 --> 00:47:48,231
Audrow Nash: a lot of real time
response to things, a lot of

884
00:47:48,231 --> 00:47:51,515
like 3D modeling of what you're
getting and then sitting that

885
00:47:51,515 --> 00:47:52,000
together?

886
00:47:52,000 --> 00:47:53,000
Barrett Ames: Right.

887
00:47:52,000 --> 00:47:54,000
Audrow Nash: Wow. Okay.

888
00:47:54,000 --> 00:47:58,312
Barrett Ames: So that's the
those are, you know, first

889
00:47:58,312 --> 00:48:03,262
problem's big. The first one is
actually big enough that we've

890
00:48:03,262 --> 00:48:08,133
found SAS customers for just
that. And then the other two are

891
00:48:08,133 --> 00:48:12,525
like what I thought I was
getting into. Right. Building

892
00:48:12,525 --> 00:48:15,000
and building homes with robots.

893
00:48:15,000 --> 00:48:17,403
Audrow Nash: Yeah. That's wild.
That's. It's so funny how a lot

894
00:48:17,403 --> 00:48:19,729
of times they're like, I want to
do this. And then it's like,

895
00:48:19,729 --> 00:48:22,131
there's so much legwork required
before you can even get to the

896
00:48:22,131 --> 00:48:24,000
part of the problem that they
want to be solving.

897
00:48:24,000 --> 00:48:25,000
Barrett Ames: Yep.

898
00:48:25,000 --> 00:48:27,791
Audrow Nash: Okay. That is so
interesting. And so you said

899
00:48:27,791 --> 00:48:30,920
that you found SAS customers for
this kind of thing. Does it mean

900
00:48:30,920 --> 00:48:33,759
you have like multiple like
that's a revenue stream for you

901
00:48:33,759 --> 00:48:34,000
guys?

902
00:48:34,000 --> 00:48:38,312
Barrett Ames: Yeah. Yeah. So we
have builders who need to

903
00:48:38,312 --> 00:48:42,851
estimate the amount of materials
that are going to go into a

904
00:48:42,851 --> 00:48:47,238
house. We can do that to down to
the stud and to the nail,

905
00:48:47,238 --> 00:48:51,928
because we have to in order to
go build with the robots. Yeah,

906
00:48:51,928 --> 00:48:56,543
right. And so that that level of
detail saves a builder 20 to

907
00:48:56,543 --> 00:49:00,779
40%. Materials. They are
literally using rules of thumb.

908
00:49:00,779 --> 00:49:05,242
Like, you go to the, there's
this crazy incentive alignment

909
00:49:05,242 --> 00:49:09,630
problem where the builder
actually doesn't do his own take

910
00:49:09,630 --> 00:49:14,245
off. He'll take it to a lumber
yard, and the lumber yard will

911
00:49:14,245 --> 00:49:18,935
tell the builder how much lumber
he needs. The great incentive

912
00:49:18,935 --> 00:49:23,549
alignment. If you go into the
lumber yard and say, all right,

913
00:49:23,549 --> 00:49:28,088
how are you guys estimating
that? In many times you're using

914
00:49:28,088 --> 00:49:32,854
their actual thumb, right? It's
not just like a saying. They're

915
00:49:32,854 --> 00:49:37,695
they're actually counting. It's
very literal. Yes. And so that's

916
00:49:37,695 --> 00:49:42,234
how they end up estimating how
much lumber that, that you're

917
00:49:42,234 --> 00:49:47,000
going to get.And so they end up
with just tons of overage. Yes.

918
00:49:47,000 --> 00:49:50,081
Audrow Nash: Now so I want to
move into questions, but I want

919
00:49:50,081 --> 00:49:53,061
to before then just hear about
working with robots for this

920
00:49:53,061 --> 00:49:56,242
task. Like, tell me a little bit
about that and then we'll move

921
00:49:56,242 --> 00:49:57,000
into questions.

922
00:49:57,000 --> 00:50:00,782
Barrett Ames: Yeah. Yeah. I
think, you know, kind of the

923
00:50:00,782 --> 00:50:04,294
interesting bits for using
robots on this particular

924
00:50:04,294 --> 00:50:08,278
problem is that we're using
we're kind of constructing this

925
00:50:08,278 --> 00:50:12,127
like almost worst case motion
planning problem. We've had

926
00:50:12,127 --> 00:50:16,044
these two big by like two big
industrial arms. And they're

927
00:50:16,044 --> 00:50:20,434
usually moving 8 to 10ft lumber.
Right. So you you're always kind

928
00:50:20,434 --> 00:50:23,000
of like in this bug trap type
problem.

929
00:50:23,000 --> 00:50:28,000
Audrow Nash: So they slide
around or anything?

930
00:50:28,000 --> 00:50:32,783
Barrett Ames: Yeah, they're on
tracks on the ground. Purposely

931
00:50:32,783 --> 00:50:37,565
went for seven days off because
I wanted the redundancy wanted

932
00:50:37,565 --> 00:50:42,039
and the ability to do a better
to do linear, movements. So

933
00:50:42,039 --> 00:50:46,668
that's we actually use used
industrial arms. They get kicked

934
00:50:46,668 --> 00:50:51,604
out, in a perfectly usable state
for us most of the time. We had

935
00:50:51,604 --> 00:50:56,156
just found an arm that has five
centimeters of error in it.

936
00:50:56,156 --> 00:51:01,170
Yeah. So that one will be fun to
correct, but they're, like a 90%

937
00:51:01,170 --> 00:51:05,875
price discount, right? So that's
massive for cutting down our

938
00:51:05,875 --> 00:51:10,658
costs. Oh, yeah. Okay. So I'd
like to open it up to questions.

939
00:51:10,658 --> 00:51:14,360
Let me know if you have a
question. What is your

940
00:51:14,360 --> 00:51:18,217
perception stack look like?
Yeah. So right now our

941
00:51:18,217 --> 00:51:22,691
perception stack is is growing.
We seem to always find new

942
00:51:22,691 --> 00:51:27,243
places for more cameras. But
generally right now we've got,

943
00:51:27,243 --> 00:51:31,948
an L5 one five real sense that
looks at lumber. It's a really

944
00:51:31,948 --> 00:51:36,885
interesting the, the particular
wavelength there. Returns really

945
00:51:36,885 --> 00:51:41,050
nicely on lumber and not on our
table. So it makes the

946
00:51:41,050 --> 00:51:45,910
classification problem, a little
bit easier there. But we have,

947
00:51:45,910 --> 00:51:50,461
several convolutional networks
that are that are helping us

948
00:51:50,461 --> 00:51:55,090
classify different, like the
plate versus the stud, pressure

949
00:51:55,090 --> 00:51:59,564
treated versus non pressure
treated. And then actually one

950
00:51:59,564 --> 00:52:04,038
of the most difficult computer
vision problems is, we pull

951
00:52:04,038 --> 00:52:08,975
lumber out of a stack. Right. So
you take a tree that's been cut

952
00:52:08,975 --> 00:52:13,757
up into a whole bunch, two by
fours, and then laid out next to

953
00:52:13,757 --> 00:52:18,000
each other, and simulate each of
those, and pick it up.

954
00:52:18,000 --> 00:52:19,000
Audrow Nash: That sounds hard.

955
00:52:19,000 --> 00:52:20,000
Barrett Ames: It was a tough one
to crack.

956
00:52:20,000 --> 00:52:24,775
Audrow Nash: I bet. On the
robotics, I guess from, like,

957
00:52:24,775 --> 00:52:29,636
the assembly side. Is that,
like, you have to do an extra

958
00:52:29,636 --> 00:52:34,155
process of cutting to the
specific point? Are you pre

959
00:52:34,155 --> 00:52:39,442
sizing everything? So how do you
deal with the lumber being so

960
00:52:39,442 --> 00:52:42,000
difficult. And you process it.

961
00:52:42,000 --> 00:52:46,571
Barrett Ames: So before assembly
we pick it up out of the stack.

962
00:52:46,571 --> 00:52:51,143
We've actually, got, we followed
the input table. It's basically

963
00:52:51,143 --> 00:52:55,786
just, really precise positioning
system. And then we pick up a 16

964
00:52:55,786 --> 00:53:00,429
inch saw blade and we cut it. We
cut it down to a 32nd of an inch

965
00:53:00,429 --> 00:53:04,714
on precision. And we do that
only for top and bottom plates.

966
00:53:04,714 --> 00:53:08,929
Right. So the vertical members
we just call studs. And then

967
00:53:08,929 --> 00:53:13,214
the, the horizontal parts of
your wall, we call those plate.

968
00:53:13,214 --> 00:53:17,500
And so there's a bunch of precut
lumber for stud length. And

969
00:53:17,500 --> 00:53:21,357
plates are have to adjust based
on our design for that

970
00:53:21,357 --> 00:53:25,643
particular house. So the table
like, like really just clamps

971
00:53:25,643 --> 00:53:29,714
down and, and makes it as
straight as it can. Then we cut

972
00:53:29,714 --> 00:53:33,786
it. And when we place it in
there the arms are, you know,

973
00:53:33,786 --> 00:53:38,143
2,600 pounds of cast iron. So
they can just format however it

974
00:53:38,143 --> 00:53:39,000
needs to be.

975
00:53:39,000 --> 00:53:45,051
Audrow Nash: Oh yeah. Other
questions? How margins compared

976
00:53:45,051 --> 00:53:47,000
to pre prefab home?

977
00:53:47,000 --> 00:53:50,975
Barrett Ames: Yeah. Just to give
you an idea, carpenter, a rough

978
00:53:50,975 --> 00:53:54,578
carpenter makes about $31 an
hour nationally in the US. We

979
00:53:54,578 --> 00:53:57,000
can run the robots for a buck 15
hours.

980
00:53:57,000 --> 00:53:58,000
Audrow Nash: Wow.

981
00:53:57,000 --> 00:54:01,000
Barrett Ames: So that's very.
There's, there's a few other

982
00:54:01,000 --> 00:54:05,345
complicating factors in there,
but, one that we're we're better

983
00:54:05,345 --> 00:54:07,000
than, manually doing it.

984
00:54:07,000 --> 00:54:09,000
Audrow Nash: So what's the cost
in transport?

985
00:54:09,000 --> 00:54:12,452
Barrett Ames: Yeah, the the
overall cost. I think a lot of

986
00:54:12,452 --> 00:54:16,262
people would be surprised by the
fact that framing is 20% of the

987
00:54:16,262 --> 00:54:19,595
cost of a home. So it's a
massive portion of it. In that

988
00:54:19,595 --> 00:54:22,929
it's actually in the biggest
portion. It's also the most

989
00:54:22,929 --> 00:54:24,000
dangerous portion.

990
00:54:24,000 --> 00:54:26,492
Audrow Nash: Is that, is that
counting labor costs and

991
00:54:26,492 --> 00:54:27,000
everything?

992
00:54:27,000 --> 00:54:31,716
Barrett Ames: Yeah. So it's,
Yeah. With labor, of that 20%,

993
00:54:31,716 --> 00:54:36,433
10% of it's directly labor. So
they're, they're, it's a big

994
00:54:36,433 --> 00:54:41,149
portion. It's a very dangerous
portion. If you look at most

995
00:54:41,149 --> 00:54:46,025
dangerous jobs in the US. Rough
Carpenter is going to be near

996
00:54:46,025 --> 00:54:51,141
the top. As far as logistics go.
We can ship 500 miles for about

997
00:54:51,141 --> 00:54:55,937
a grand. The typical framing
package is going to be 20 to 40

998
00:54:55,937 --> 00:55:00,174
grand on a house. So in
comparison, relatively small,

999
00:55:00,174 --> 00:55:04,890
for like, a one off house. We'll
we can do that all day. No

1000
00:55:04,890 --> 00:55:09,607
problem. If Dr. Horton wants to
build a whole neighborhood,

1001
00:55:09,607 --> 00:55:14,003
we'll just set up on a
basketball court that's going to

1002
00:55:14,003 --> 00:55:18,000
be in the community anyways, and
just build there.

1003
00:55:18,000 --> 00:55:21,097
Audrow Nash: Oh, yeah. Back
there. You talk about the

1004
00:55:21,097 --> 00:55:24,604
efficiency of time and how time,
money. And then working day

1005
00:55:24,604 --> 00:55:27,000
night. Tell me about the time
efficiency.

1006
00:55:27,000 --> 00:55:31,585
Barrett Ames: Yeah. So from,
from a time efficiency from the

1007
00:55:31,585 --> 00:55:35,406
builder's perspective, time
really is money there.

1008
00:55:35,406 --> 00:55:39,762
Construction loans are very
expensive. Builders will have

1009
00:55:39,762 --> 00:55:44,347
something on the order of $250 a
day. To $750 a day, just in

1010
00:55:44,347 --> 00:55:48,397
interest charges. Right. So
every day matters, to the

1011
00:55:48,397 --> 00:55:52,830
builder. And from a robotic
standpoint. Right. Like we can

1012
00:55:52,830 --> 00:55:57,644
build 24 seven, with these arms.
It's it's, you know, something

1013
00:55:57,644 --> 00:56:01,236
that can be a relatively
constrained within the

1014
00:56:01,236 --> 00:56:02,000
warehouse.

1015
00:56:02,000 --> 00:56:05,578
Audrow Nash: Yeah. So I think
we'll do one more question. You

1016
00:56:05,578 --> 00:56:08,686
mentioned that that company
about four years old. You

1017
00:56:08,686 --> 00:56:12,205
probably got 25 from so far.
Talk about the lifecycle of the

1018
00:56:12,205 --> 00:56:15,666
company where you guys are. It's
still development? Are you

1019
00:56:15,666 --> 00:56:19,243
starting to scale up production
Lifecycle of a company? Where

1020
00:56:19,243 --> 00:56:22,000
you guys are, you've been going
for four years.

1021
00:56:22,000 --> 00:56:26,702
Barrett Ames: Yeah. Yeah. So,
first, kind of like, we actually

1022
00:56:26,702 --> 00:56:31,252
built our first frame three
months after starting. We and we

1023
00:56:31,252 --> 00:56:36,030
were a part of Y Combinator. So
for Demo Day, we actually built

1024
00:56:36,030 --> 00:56:40,960
frames that got stood up as part
of a house in Arkansas. And then

1025
00:56:40,960 --> 00:56:45,814
we, as we learn more and more of
the process, we found all sorts

1026
00:56:45,814 --> 00:56:49,681
of different pieces that we
needed to, to automate.

1027
00:56:49,681 --> 00:56:54,004
Particularly the floor plan
part. Right. So we took took,

1028
00:56:54,004 --> 00:56:58,479
took time to figure that portion
out. And then we've really

1029
00:56:58,479 --> 00:57:02,878
started to hit more of our
production stride this year. We

1030
00:57:02,878 --> 00:57:06,821
just set up a three team
facility in North Carolina,

1031
00:57:06,821 --> 00:57:10,917
that, can pump out, floor a
week. As far so, you know,

1032
00:57:10,917 --> 00:57:15,619
basically 100 panels. And we're
we're going we're scaling that

1033
00:57:15,619 --> 00:57:20,169
up right now. We're just about
to do our series a, so that's

1034
00:57:20,169 --> 00:57:24,719
kind of the, I guess, the, the
rough life cycle from a labor

1035
00:57:24,719 --> 00:57:29,497
standpoint. It's, like roughly
on par with labor right now. But

1036
00:57:29,497 --> 00:57:32,000
we're vastly more cost
effective.

1037
00:57:32,000 --> 00:57:35,813
Audrow Nash: Yeah. That's wild.
The like 20 times improvement

1038
00:57:35,813 --> 00:57:39,563
that you're seeing. Okay. Hell
yeah. So one thing we haven't

1039
00:57:39,563 --> 00:57:43,250
done in the past ones but what
do you hope that people like

1040
00:57:43,250 --> 00:57:46,000
everyone here takes away. This
conversation.

1041
00:57:46,000 --> 00:57:50,385
Barrett Ames: Yeah, I, I hope
everybody realizes that the

1042
00:57:50,385 --> 00:57:54,615
future of construction is in
automation, right? So many

1043
00:57:54,615 --> 00:57:59,615
people don't even realize all of
the pain and suffering that goes

1044
00:57:59,615 --> 00:58:04,231
into their house. And how that
that creates a whole bunch of

1045
00:58:04,231 --> 00:58:08,692
societal ills. The future can be
so much better if we just

1046
00:58:08,692 --> 00:58:10,000
automate housing.

1047
00:58:10,000 --> 00:58:12,600
Audrow Nash: Totally. Hell,
yeah. Alright. Thank you

1048
00:58:12,600 --> 00:58:13,000
Barrett.

1049
00:58:13,000 --> 00:58:17,000
Barrett Ames: Thank you.

1050
00:58:17,000 --> 00:58:20,237
Audrow Nash: That's it. I really
enjoyed this. I just love

1051
00:58:20,237 --> 00:58:23,419
talking as you can tell, and
doing it with an audience is

1052
00:58:23,419 --> 00:58:26,712
super fun because they have
great questions and it was cool

1053
00:58:26,712 --> 00:58:30,116
to feel the excitement in the
room. If you liked this and can

1054
00:58:30,116 --> 00:58:33,577
make it, consider coming to the
event march 8th. I'll probably

1055
00:58:33,577 --> 00:58:37,149
be helping out with panels, and
I'd love to meet you there. I'll

1056
00:58:37,149 --> 00:58:40,721
put the link in the show notes.
That's all for now. See you next

1057
00:58:40,721 --> 00:58:41,000
time.
