1
00:00:00,000 --> 00:00:04,454
Axel Krieger: Ultimately, we're
able to do eight of those back

2
00:00:04,454 --> 00:00:08,693
to back without needing any
human intervention. So, We have

3
00:00:08,693 --> 00:00:12,428
fewer and fewer surgeons
available. We have an aging

4
00:00:12,428 --> 00:00:16,810
society. We have a gap of in the
next 10 years we are missing

5
00:00:16,810 --> 00:00:20,977
20,000 surgeons, My vision was
eventually you might have a

6
00:00:20,977 --> 00:00:25,000
surgeon supervising several
procedures at the same time.

7
00:00:25,000 --> 00:00:26,189
Audrow Nash: Yeah, yeah like a
grocery store checkout model for

8
00:00:26,189 --> 00:00:27,302
surgeons. What a world. would
you be surprised if it wasn't

9
00:00:27,302 --> 00:00:28,472
the case in 20 years from now?
That you could just like go and

10
00:00:28,472 --> 00:00:29,000
have a robot operate on you?

11
00:00:29,000 --> 00:00:39,000
Axel Krieger: I looking at the
trajectory and being in it, I

12
00:00:39,000 --> 00:00:42,000
would be surprised

13
00:00:42,000 --> 00:00:45,000
Audrow Nash: So, hey Axel, would
you introduce yourself?

14
00:00:45,000 --> 00:00:49,552
Axel Krieger: Hello, and thank
you so much for having me. I'm

15
00:00:49,552 --> 00:00:53,582
Axel Krieger. I'm the chief
robotics officer for Inner

16
00:00:53,582 --> 00:00:57,910
Logic. I'm also associate
professor at Johns Hopkins where

17
00:00:57,910 --> 00:01:00,000
I'm currently on sabbatical.

18
00:01:00,000 --> 00:01:04,366
Audrow Nash: Tell me a bit about
Inner Logic and how it relates

19
00:01:04,366 --> 00:01:07,000
to what you're doing at Johns
Hopkins.

20
00:01:07,000 --> 00:01:12,388
Axel Krieger: Fantastic, so my
research is on autonomous

21
00:01:12,388 --> 00:01:18,354
robotic surgery. So currently
robots that I used in the clinic

22
00:01:18,354 --> 00:01:24,319
are operated every step of the
way by an operating surgeon. So

23
00:01:24,319 --> 00:01:29,515
surgeon sits on a console,
controls every step. And my

24
00:01:29,515 --> 00:01:35,576
vision is really that you know
robots can perform some critical

25
00:01:35,576 --> 00:01:41,445
steps autonomously and surgeon
takes on supervisory role. And

26
00:01:41,445 --> 00:01:46,834
so we feel this is coming,
autonomous surgery. And so we

27
00:01:46,834 --> 00:01:52,318
founded the company Inner Logic
with Tito and Matthias, a

28
00:01:52,318 --> 00:01:58,283
neurosurgeon Tito who also has
an MBA. So amazing combination.

29
00:01:58,283 --> 00:02:04,056
And Matthias, a fellow professor
from Hopkins, also a fellow

30
00:02:04,056 --> 00:02:09,733
German American. And he's a
machine learning AI expert. And

31
00:02:09,733 --> 00:02:14,929
so that combination started
Inner Logic and we believe

32
00:02:14,929 --> 00:02:20,798
autonomy is coming. And we want
to provide companies with the

33
00:02:20,798 --> 00:02:26,667
infrastructure to help them on
the way to that. So we support

34
00:02:26,667 --> 00:02:32,055
different medical device
manufacturers with data models,

35
00:02:32,055 --> 00:02:36,000
functions on their ambitious AI
roadmaps.

36
00:02:36,000 --> 00:02:39,519
Audrow Nash: Very very cool.
Tell me how, like tell me more

37
00:02:39,519 --> 00:02:43,218
about Inner Logic, kind of just
the logistics about it. Like I

38
00:02:43,218 --> 00:02:46,916
believe you guys have started
pretty recently, tell me how big

39
00:02:46,916 --> 00:02:50,256
you are, tell me about the
funding you guys have so far,

40
00:02:50,256 --> 00:02:53,000
all the kind of like logistics
of the company.

41
00:02:53,000 --> 00:02:58,518
Axel Krieger: Yeah, we started
fundraising last fall. Kind of

42
00:02:58,518 --> 00:03:03,854
interesting, we received the
first outright from a VC after

43
00:03:03,854 --> 00:03:08,286
publishing a science robotics
study on autonomous

44
00:03:08,286 --> 00:03:13,714
cholecystectomy. And that kind
of really spurred us. We were

45
00:03:13,714 --> 00:03:18,327
already talking in the
background about maybe doing

46
00:03:18,327 --> 00:03:23,844
something together. And that
really got us into gear. We then

47
00:03:23,844 --> 00:03:29,000
did fundraising in the fall,
raised 11 and a half million

48
00:03:29,000 --> 00:03:34,427
dollars a seed by Collet, by
General Catalyst and Bison. And

49
00:03:34,427 --> 00:03:39,854
then really became operational
around January. So moved into

50
00:03:39,854 --> 00:03:45,191
our new space in February and
hired our first five founding

51
00:03:45,191 --> 00:03:50,256
engineers and are now already
close to 20 people down in

52
00:03:50,256 --> 00:03:55,322
Baltimore and hiring also
worldwide and more remote. And

53
00:03:55,322 --> 00:03:59,844
so yeah, and signed our first
contract with device

54
00:03:59,844 --> 00:04:05,000
manufacturers. So exciting,
exciting whirlwind last year.

55
00:04:05,000 --> 00:04:09,060
Audrow Nash: Yeah, really. Okay,
hell yeah. Tell me, so for

56
00:04:09,060 --> 00:04:13,119
context, tell me where the
robotic surgery space is. Like I

57
00:04:13,119 --> 00:04:17,454
like to understand where we're
coming from and then kind of why

58
00:04:17,454 --> 00:04:20,000
you guys are doing what you're
doing.

59
00:04:20,000 --> 00:04:24,700
Axel Krieger: Yeah, so a bit of
context, right? So, you know,

60
00:04:24,700 --> 00:04:29,400
surgical robotics has enabled
you know the shift from open to

61
00:04:29,400 --> 00:04:34,177
minimally invasive surgery. So,
you know, we used to have, you

62
00:04:34,177 --> 00:04:38,646
know, large, you know, scars
opening, having direct access

63
00:04:38,646 --> 00:04:43,577
with your eyes, hands on organs.
And what robots have allowed is

64
00:04:43,577 --> 00:04:47,892
that you can sit on a console
and drive these, you know,

65
00:04:47,892 --> 00:04:52,592
keyhole port based surgeries
with dexterous tools and perform

66
00:04:52,592 --> 00:04:57,138
these, you know, complex, you
know, surgeries now minimally

67
00:04:57,138 --> 00:05:01,838
invasive. And really the main
focus over the last few decades

68
00:05:01,838 --> 00:05:05,767
has been to make the robots
smaller, you know, more

69
00:05:05,767 --> 00:05:10,390
dexterous allowing, you know,
natural orifice or access, for

70
00:05:10,390 --> 00:05:15,167
example, through your mouth and
other, you know, orifices, you

71
00:05:15,167 --> 00:05:19,867
know, to get to the organ of
interest without, you know, much

72
00:05:19,867 --> 00:05:24,259
bleeding, scarring. So, you
know, less invasive. And, you

73
00:05:24,259 --> 00:05:28,882
know, that has been, you know,
the progression in technology

74
00:05:28,882 --> 00:05:33,659
over the last two tech you know
decades. And over the last few

75
00:05:33,659 --> 00:05:38,282
years, and now I think really
projecting into the future has

76
00:05:38,282 --> 00:05:42,289
been the shift with more
software centric, more, you

77
00:05:42,289 --> 00:05:47,220
know, intelligence, smarter, you
know, tools and devices to, you

78
00:05:47,220 --> 00:05:51,766
know, reduce the burden of
assertions and, you know, really

79
00:05:51,766 --> 00:05:56,389
help our patients to, you know,
have fewer complications and

80
00:05:56,389 --> 00:06:01,320
better outcomes. So, it's been a
shift of mechanical innovations

81
00:06:01,320 --> 00:06:06,251
to more software innovations and
intelligence and autonomy. And,

82
00:06:06,251 --> 00:06:11,182
you know, we've been working on
this, you know, since about 2012

83
00:06:11,182 --> 00:06:15,805
on autonomous surgery that was,
you know, considered still a

84
00:06:15,805 --> 00:06:20,351
little crazy at the time. And
now, you know, I think really

85
00:06:20,351 --> 00:06:25,128
that the consensus is that, you
know, AI in surgery is coming,

86
00:06:25,128 --> 00:06:29,905
is helping. And, you know, we
are really helping at, you know,

87
00:06:29,905 --> 00:06:33,989
logic, you know, that
development cycle, you know, to

88
00:06:33,989 --> 00:06:38,072
accelerate and bring, you know,
forward those smarter

89
00:06:38,072 --> 00:06:41,000
intelligence, more autonomous
devices.

90
00:06:41,000 --> 00:06:44,414
Audrow Nash: So just so I can
picture it a bit more clearly,

91
00:06:44,414 --> 00:06:47,884
you said like keyhole and port.
I've seen some of the DaVinci

92
00:06:47,884 --> 00:06:51,298
ones and I've tried them at
different conferences. Not, tell

93
00:06:51,298 --> 00:06:54,199
me if this is a good
representation of more or less

94
00:06:54,199 --> 00:06:57,613
what surgery robots are like. It
was like I was looking in a

95
00:06:57,613 --> 00:07:00,799
screen and I had my thumb and
pointer finger controlling

96
00:07:00,799 --> 00:07:04,155
little robot hands that were
like on the ends of sticks. So

97
00:07:04,155 --> 00:07:07,512
it was like a little claw on the
end of a stick and I would

98
00:07:07,512 --> 00:07:10,698
control that and you could stack
blocks or sort rings or

99
00:07:10,698 --> 00:07:14,168
whatever. That's pretty much
what robot surgery means at this

100
00:07:14,168 --> 00:07:17,810
point, focusing on hardware. Is
it correct and like the problems

101
00:07:17,810 --> 00:07:21,223
have been like making those
hands as small as possible so it

102
00:07:21,223 --> 00:07:24,466
could be minimally invasive and
then maybe providing some

103
00:07:24,466 --> 00:07:27,880
tactile feedback to the user so
that they could feel if they

104
00:07:27,880 --> 00:07:31,293
were gripping something or this
kind of thing. Is that about

105
00:07:31,293 --> 00:07:33,000
right or what do you think? So

106
00:07:33,000 --> 00:07:35,165
Axel Krieger: Yeah, really,
really good summary, right? So,

107
00:07:35,165 --> 00:07:37,404
the operating surgeon sits on a
console, puts their head into

108
00:07:37,404 --> 00:07:39,789
like this, you know, camera, and
then you get like a 3D vision of

109
00:07:39,789 --> 00:07:42,028
the scene because you have 3D
laparoscopes. And then you have

110
00:07:42,028 --> 00:07:44,156
two hand controllers. Robots
typically have four arms that

111
00:07:44,156 --> 00:07:46,174
you can switch to, which one
you're using. And then you

112
00:07:46,174 --> 00:07:48,523
control kind of two arms at the
same time. And it becomes really

113
00:07:48,523 --> 00:07:50,725
an extension of the surgeon,
right? Like, you know, that you

114
00:07:50,725 --> 00:07:52,963
can work inside the body. You
inflate it with CO2 to get like

115
00:07:52,963 --> 00:07:55,128
some distance and some room to
operate in. And then you can

116
00:07:55,128 --> 00:07:57,000
perform this. But every motion
of the robot is, you

117
00:07:57,000 --> 00:08:17,000
Audrow Nash: Cool. Hmm.

118
00:08:17,000 --> 00:08:20,875
Axel Krieger: Know, exactly, you
know, just mimicked what the

119
00:08:20,875 --> 00:08:24,560
surgeon does on the console to
what the, you know, robotic

120
00:08:24,560 --> 00:08:28,372
system does over the patient,
sitting in the same room. And,

121
00:08:28,372 --> 00:08:31,866
you know, currently there's
really no autonomy in these

122
00:08:31,866 --> 00:08:35,678
procedures. That's the standard.
So, for the last, you know,

123
00:08:35,678 --> 00:08:39,553
decades, the robots have gotten
better control, more surgical

124
00:08:39,553 --> 00:08:43,492
tools, you know, more staplers,
more energy devices, you know,

125
00:08:43,492 --> 00:08:44,000
smaller,

126
00:08:44,000 --> 00:08:45,000
Audrow Nash: Cool.

127
00:08:45,000 --> 00:08:47,590
Axel Krieger: You know, more
stable, less, you know,

128
00:08:47,590 --> 00:08:50,579
vibration, you know, and making
them more flexible to enable

129
00:08:50,579 --> 00:08:53,468
some procedures that are, you
know, even further away from

130
00:08:53,468 --> 00:08:56,158
like a laparoscopic entrance,
you know, like more like

131
00:08:56,158 --> 00:08:59,197
snake-like robots. So, a lot of
development on the mechanical

132
00:08:59,197 --> 00:09:02,285
side. And really also like an
explosion of different companies

133
00:09:02,285 --> 00:09:05,324
coming in, you know, Intuitive
is the main, you know, robotic

134
00:09:05,324 --> 00:09:08,163
surgical company that, you know,
a lot of people know the

135
00:09:08,163 --> 00:09:10,953
Intuitive, you know, DaVinci
robot. But now we have, you

136
00:09:10,953 --> 00:09:13,792
know, you know, J&J robot,
Metronic robot, you know, Moon

137
00:09:13,792 --> 00:09:16,880
Surgical, CMR, you know, some
Chinese companies. So, you know,

138
00:09:16,880 --> 00:09:19,919
I think around 60, you know,
surgical robotics companies are,

139
00:09:19,919 --> 00:09:23,007
you know, developing specialty
robots for all different organs

140
00:09:23,007 --> 00:09:25,000
and, you know, bringing those to
market.

141
00:09:25,000 --> 00:09:31,424
Audrow Nash: Wow. Okay. And so
most of these companies have

142
00:09:31,424 --> 00:09:38,393
been innovating on the hardware
side and you guys, you mentioned

143
00:09:38,393 --> 00:09:44,817
infrastructure is where you're
interested in. Tell me a bit

144
00:09:44,817 --> 00:09:51,458
about that and tell me how that
works. It's like you won't be

145
00:09:51,458 --> 00:09:57,665
making the hardware but you'll
be working with all of the

146
00:09:57,665 --> 00:10:03,000
hardware that companies make.
Tell me about that.

147
00:10:03,000 --> 00:10:07,280
Axel Krieger: Yeah, exactly. So,
a lot of the, you know,

148
00:10:07,280 --> 00:10:11,255
companies' expertise, core
expertise, is more on the

149
00:10:11,255 --> 00:10:15,841
mechanical side. And, you know,
and you know we really see a

150
00:10:15,841 --> 00:10:20,580
shift there in the focus and,
you know, like the roadmaps are,

151
00:10:20,580 --> 00:10:25,242
you know, quite ambitious on,
you know, really improving the,

152
00:10:25,242 --> 00:10:29,522
you know, guidance, you know,
decision-making, you know,

153
00:10:29,522 --> 00:10:34,261
control. And we help companies
on their roadmap. So, you know,

154
00:10:34,261 --> 00:10:38,771
how does it look like? So, let's
say you want to, you know,

155
00:10:38,771 --> 00:10:42,669
automate a suction
functionality, right? So then we

156
00:10:42,669 --> 00:10:47,331
help you with, you know, getting
data, with tool tracking. We

157
00:10:47,331 --> 00:10:51,994
can, you know, help with, you
know, foundation models to help

158
00:10:51,994 --> 00:10:56,580
you with, you know, learning
those functions faster. We help

159
00:10:56,580 --> 00:11:01,089
you with testing. So, and all
the way to, you know, really,

160
00:11:01,089 --> 00:11:05,599
you know, co-developing, you
know, functions. So, it starts

161
00:11:05,599 --> 00:11:09,650
often with, you know, data
needs, but then, you know,

162
00:11:09,650 --> 00:11:14,236
really, you know, working on
expanding this to more and more

163
00:11:14,236 --> 00:11:15,000
functions.

164
00:11:15,000 --> 00:11:19,852
Audrow Nash: Mm. That's really
cool. And where, tell me about

165
00:11:19,852 --> 00:11:24,784
the science robotics paper you
all published and how that kind

166
00:11:24,784 --> 00:11:29,000
of kicked things off and where
you think we're going.

167
00:11:29,000 --> 00:11:32,867
Axel Krieger: Yeah, maybe I can,
you know, explain this, you

168
00:11:32,867 --> 00:11:36,992
know, in context to kind of our
research, right? So, we started,

169
00:11:36,992 --> 00:11:40,859
you know, 2012 in autonomous
robotics and started, you know,

170
00:11:40,859 --> 00:11:44,791
with a traditional model-based
approach, right? So, you know,

171
00:11:44,791 --> 00:11:48,143
separate perception tracking,
you know, modeling the

172
00:11:48,143 --> 00:11:51,688
interaction between tool and
tissue and then, you know,

173
00:11:51,688 --> 00:11:55,684
controlling that, you know, and
making adjustments. And so, we

174
00:11:55,684 --> 00:11:59,744
got pretty good in suturing, you
know, with specialized camera,

175
00:11:59,744 --> 00:12:02,000
specialized robotic tool, you
know.

176
00:12:02,000 --> 00:12:05,278
Audrow Nash: Now, I don't really
know medical terms and I'm

177
00:12:05,278 --> 00:12:08,778
supposing people do not either.
Suturing, does that mean sewing

178
00:12:08,778 --> 00:12:10,000
up a wound or a, okay.

179
00:12:10,000 --> 00:12:11,582
Axel Krieger: Exactly, exactly,
exactly. So, you know, suturing

180
00:12:11,582 --> 00:12:13,088
is, you know, where you have a
suture needle and you need to

181
00:12:13,088 --> 00:12:14,544
sew up a wound. And what we've
been working on, yeah, like

182
00:12:14,544 --> 00:12:16,000
they're like kind of half
circles, you know, half circles.

183
00:12:16,000 --> 00:12:25,000
Audrow Nash: Mm-hmm. Like one of
those hook shaped needles.

184
00:12:25,000 --> 00:12:28,690
Axel Krieger: And, you know, the
particular application that we

185
00:12:28,690 --> 00:12:31,736
worked on was on the bowel
anastomosis is called the

186
00:12:31,736 --> 00:12:35,368
closing of a bowel. So, imagine
you have a tumor in the bowel,

187
00:12:35,368 --> 00:12:39,059
you take out a section, right,
and then you need to put it back

188
00:12:39,059 --> 00:12:42,690
together. And putting that back
together requires placing, you

189
00:12:42,690 --> 00:12:46,322
know, 20 little stitches around
the bowel. And if you miss one

190
00:12:46,322 --> 00:12:49,837
little, you know, mark, that's
where you have a leak, right?

191
00:12:49,837 --> 00:12:53,117
And so, you need a lot of
consistency and repeatability.

192
00:12:53,117 --> 00:12:56,925
And so, that's where we started,
you know, technically building a

193
00:12:56,925 --> 00:13:00,615
robot for this anastomosis. And,
you know, we were able to show

194
00:13:00,615 --> 00:13:04,071
that we can perform this in an
animal model, you know, very

195
00:13:04,071 --> 00:13:07,762
consistency in a supervised, you
know, autonomous function. So,

196
00:13:07,762 --> 00:13:11,218
you know, surgeons don't have
to, you know, prescribe every

197
00:13:11,218 --> 00:13:15,025
motion of the robot anymore. The
surgical system, you know, tells

198
00:13:15,025 --> 00:13:18,598
you where it wants to go. You
can make adjustments. And every

199
00:13:18,598 --> 00:13:22,289
now and then it misses a stitch
and you provide a little bit of

200
00:13:22,289 --> 00:13:25,745
help. So, that's, you know, how
we, you know, traditionally

201
00:13:25,745 --> 00:13:29,201
build robots. And, you know, it,
we hit a bit of a ceiling,

202
00:13:29,201 --> 00:13:32,423
right? So, about 70 percent of
stitches, you know, were

203
00:13:32,423 --> 00:13:35,937
autonomous, 30 percent needed
help. And we, you know, really

204
00:13:35,937 --> 00:13:39,510
learned how to do stitching, but
it didn't translate to other

205
00:13:39,510 --> 00:13:43,201
procedures, right? And so, you
know, where we pivoted then, you

206
00:13:43,201 --> 00:13:46,950
know, a few years ago is really
to, you know, a fully, you know,

207
00:13:46,950 --> 00:13:49,000
end-to-end learning-based
approach,

208
00:13:49,000 --> 00:14:04,000
Audrow Nash: Mm-hmm.

209
00:14:04,000 --> 00:14:05,865
Axel Krieger: Where the idea is
that we can use imitation

210
00:14:05,865 --> 00:14:07,828
learning to, you know, teach our
robots how to perform these

211
00:14:07,828 --> 00:14:09,824
procedures. And what is exciting
about imitation learning for

212
00:14:09,824 --> 00:14:11,885
surgery is really we have this
large install base of, you know,

213
00:14:11,885 --> 00:14:13,815
surgical robots. And we have
thousands of incredibly highly

214
00:14:13,815 --> 00:14:15,745
trained surgeons that are, you
know, in their residency and

215
00:14:15,745 --> 00:14:17,643
fellowship training. They are
already using the robot. So,

216
00:14:17,643 --> 00:14:19,606
they grow up with the robot.
They are, you know, deep, deep,

217
00:14:19,606 --> 00:14:21,602
deep specialists on how to use,
you know, robotic technology.

218
00:14:21,602 --> 00:14:23,499
And so, I mean, if you watch
some of the surgical robotics

219
00:14:23,499 --> 00:14:25,331
videos, how they perform
surgery, it's just beautiful. I

220
00:14:25,331 --> 00:14:27,000
mean, it's like an art how every
motion is just so,

221
00:14:27,000 --> 00:14:56,000
Audrow Nash: Mm-hmm. Hmm.

222
00:14:56,000 --> 00:14:59,633
Axel Krieger: You know, you
know, exquisitely, you know,

223
00:14:59,633 --> 00:15:03,655
detailed, great visualization. I
mean, they're really, really,

224
00:15:03,655 --> 00:15:07,808
really good at this, right? And
so, instead of modeling this and

225
00:15:07,808 --> 00:15:11,635
really trying to automate this,
you know, with, you know, a

226
00:15:11,635 --> 00:15:15,723
rule-based, you know, set up the
idea here is where we pivot it

227
00:15:15,723 --> 00:15:19,875
to is that we can learn directly
from those expert surgeons. And

228
00:15:19,875 --> 00:15:23,897
I just feel like for the entire
robotics community, like this,

229
00:15:23,897 --> 00:15:27,271
you know, field is so
interesting because, you know,

230
00:15:27,271 --> 00:15:31,228
what other field has, you know,
so many expert robotic users,

231
00:15:31,228 --> 00:15:34,926
right? Like if we think about,
you know, folding laundry,

232
00:15:34,926 --> 00:15:38,819
right? Like everybody has to
kind of learn it, right? But in

233
00:15:38,819 --> 00:15:42,452
surgical robotics, we have that
pool of, you know, fully

234
00:15:42,452 --> 00:15:46,345
trained, you know, surgical
robotic, you know, surgeons. So,

235
00:15:46,345 --> 00:15:50,237
that's kind of like to, you
know, preface the kind of switch

236
00:15:50,237 --> 00:15:53,740
and the history of our
development in the lab a little

237
00:15:53,740 --> 00:15:54,000
bit.

238
00:15:54,000 --> 00:15:54,593
Audrow Nash: Mm-hmm. I have one
question with all this. With the

239
00:15:54,593 --> 00:15:55,000
imitation learning from the
robot platforms,

240
00:15:55,000 --> 00:16:02,000
Axel Krieger: And, yeah.

241
00:16:02,000 --> 00:16:05,492
Audrow Nash: Do they do very
good logging? So you end up

242
00:16:05,492 --> 00:16:09,297
having a very good ability to
replay everything. You see what

243
00:16:09,297 --> 00:16:13,163
the cameras saw. You see all of
the forces that the surgeon is

244
00:16:13,163 --> 00:16:17,092
putting on things. Maybe you can
even have like a transcript of

245
00:16:17,092 --> 00:16:20,772
it for context or things like
this. Like do you have a very

246
00:16:20,772 --> 00:16:24,202
good data capture of surgeons
from them using the robot

247
00:16:24,202 --> 00:16:27,819
systems and does it vary quite a
bit between the different

248
00:16:27,819 --> 00:16:31,000
companies and their different
robots, or how is it?

249
00:16:31,000 --> 00:16:34,744
Axel Krieger: Yeah. So,
typically, the companies, you

250
00:16:34,744 --> 00:16:39,124
know, really collect and log the
data well, right? And so, you

251
00:16:39,124 --> 00:16:43,645
know, what's very common is that
surgeons capture the video and,

252
00:16:43,645 --> 00:16:47,671
you know, so video data is
available a lot. And companies

253
00:16:47,671 --> 00:16:51,839
collect, you know, also the
kinematics data. So, the robot,

254
00:16:51,839 --> 00:16:55,512
movement data, right? And so,
what is difficult as a

255
00:16:55,512 --> 00:16:59,680
researcher is getting access to
both, right? And so, that's

256
00:16:59,680 --> 00:17:03,000
definitely a hurdle for us.
Yeah. And also like

257
00:17:03,000 --> 00:17:08,000
Audrow Nash: Yeah, yeah. Part of
the reason to start the company

258
00:17:08,000 --> 00:17:12,297
in some sense, because then you
can have collaborations

259
00:17:12,297 --> 00:17:13,000
directly.

260
00:17:13,000 --> 00:17:13,109
Axel Krieger: HIPAA rules,
right? So, this is for patient

261
00:17:13,109 --> 00:17:13,222
privacy, right? So, it's of
course patient data, right? So,

262
00:17:13,222 --> 00:17:13,324
sharing that is not you know,
super easy. So, there's

263
00:17:13,324 --> 00:17:13,437
definitely hurdles around, you
know, the collection of that

264
00:17:13,437 --> 00:17:13,552
data set. But companies have
access to it and there are also

265
00:17:13,552 --> 00:17:13,661
ways of, you know, really
consenting patients and getting

266
00:17:13,661 --> 00:17:13,766
this, you know, ethically
approved to share. Because, I

267
00:17:13,766 --> 00:17:13,881
mean, you know, it's not like
patient identifying. We cannot

268
00:17:13,881 --> 00:17:13,990
really trace that back to a
particular patient, you know.

269
00:17:13,990 --> 00:17:14,000
Yeah.

270
00:17:14,000 --> 00:17:32,975
Audrow Nash: Oh, yeah. Yep.
Yeah. Mm-hmm. Yeah. So cool.

271
00:17:32,975 --> 00:17:53,306
Okay, so that was the context.
And then moving into the next

272
00:17:53,306 --> 00:17:55,000
part.

273
00:17:55,000 --> 00:17:58,081
Axel Krieger: Yeah. So, that's
the context. Why is imitation

274
00:17:58,081 --> 00:18:01,214
learning so exciting? And so, we
started, you know, exploring

275
00:18:01,214 --> 00:18:04,347
imitation learning, you know,
by, you know, just figuring out

276
00:18:04,347 --> 00:18:07,479
how to perform, you know, small
different surgical tasks, you

277
00:18:07,479 --> 00:18:10,407
know, like just grabbing a
needle. And then, you know, in

278
00:18:10,407 --> 00:18:13,539
this paper that we published
last year, you know, really tied

279
00:18:13,539 --> 00:18:16,518
together small little surgical
steps to perform a complete

280
00:18:16,518 --> 00:18:19,342
phase of surgery, you know,
autonomously. We picked the

281
00:18:19,342 --> 00:18:22,270
gallbladder surgery here.
Gallbladder surgery is, was the

282
00:18:22,270 --> 00:18:25,300
first that was converted from
open surgery to laparoscopic.

283
00:18:25,300 --> 00:18:28,073
So, it has a bit of this
lighthouse character of being

284
00:18:28,073 --> 00:18:31,000
the first, you know, and so, it
was a good target for us.

285
00:18:31,000 --> 00:18:38,427
Audrow Nash: Mm-hmm. Cool. Why
is it that way? Is it a simpler

286
00:18:38,427 --> 00:18:46,094
surgery, or does it lend itself
very well to laparoscopic, which

287
00:18:46,094 --> 00:18:53,281
means like the little hands
going in, I guess, right? Little

288
00:18:53,281 --> 00:18:54,000
holes.

289
00:18:54,000 --> 00:18:54,511
Axel Krieger: Exactly. That's
the little keyhole surgery.

290
00:18:54,511 --> 00:18:55,067
Yeah. Yeah. It's, you know, the
it is a, you know, very common

291
00:18:55,067 --> 00:18:55,605
surgery. So, it's done a lot and
it was, you know, it's, you

292
00:18:55,605 --> 00:18:56,000
know, well accessible, right?
And so, it is,

293
00:18:56,000 --> 00:19:11,103
Audrow Nash: Mm-hmm. Is it like
removing gallstones, or what's

294
00:19:11,103 --> 00:19:15,000
the point of it?

295
00:19:15,000 --> 00:19:16,033
Axel Krieger: It's removing the
gallbladder, right? If you have

296
00:19:16,033 --> 00:19:17,000
a gallbladder infection, you
remove the entire gallbladder.

297
00:19:17,000 --> 00:19:21,000
Audrow Nash: Oh, removing it.
Ah, okay.

298
00:19:21,000 --> 00:19:23,440
Axel Krieger: And so, you know,
there's three phases to this.

299
00:19:23,440 --> 00:19:26,000
There's, you have to dissect the
gallbladder from the liver bed.

300
00:19:26,000 --> 00:19:28,280
Then, you have to find and
isolate the little artery that

301
00:19:28,280 --> 00:19:30,760
supplies the gallbladder and the
little duct that connects the

302
00:19:30,760 --> 00:19:33,160
gallbladder to the bowel. And
those two structures, you have

303
00:19:33,160 --> 00:19:35,280
to place like small little
plastic clips in it, three

304
00:19:35,280 --> 00:19:37,000
clips, and then you have to cut
in between.

305
00:19:37,000 --> 00:19:50,000
Audrow Nash: Oh, like clamp it
or something. Mm-hmm.

306
00:19:50,000 --> 00:19:54,882
Axel Krieger: And then you take
it out. So, that's kind of like,

307
00:19:54,882 --> 00:19:59,612
you know, in simple words, how
to perform that surgery. And we

308
00:19:59,612 --> 00:20:04,494
focused in our autonomy work to
make it a bit easier, not on the

309
00:20:04,494 --> 00:20:09,071
entire gallbladder surgery, but
this kind of middle phase of

310
00:20:09,071 --> 00:20:13,571
finding the duct and artery and
then placing clips and then

311
00:20:13,571 --> 00:20:18,225
cutting it. So, that's kind of
the clipping and cutting phase

312
00:20:18,225 --> 00:20:22,573
of the gallbladder surgery. And
so, how did we, you know,

313
00:20:22,573 --> 00:20:27,379
perform this? We found our, you
know, clinical collaborators at

314
00:20:27,379 --> 00:20:32,108
Johns Hopkins to show us how to
perform that. So, they came to

315
00:20:32,108 --> 00:20:36,838
our lab, set up you know a ex
vivo, that means, you know, just

316
00:20:36,838 --> 00:20:41,415
the parts from a butcher, the
gallbladder and the liver. So,

317
00:20:41,415 --> 00:20:45,000
we got those you know, in from
butchers online.

318
00:20:45,000 --> 00:20:50,000
Audrow Nash: Huh. It's like a
cow, or I guess,

319
00:20:50,000 --> 00:20:52,000
Axel Krieger: Yeah, from pigs,
from pigs.

320
00:20:52,000 --> 00:20:53,000
Audrow Nash: For pigs. Oh, yeah,
yeah.

321
00:20:53,000 --> 00:20:53,498
Axel Krieger: They're quite
similar to human organs, right?

322
00:20:53,498 --> 00:20:54,014
So, we could order them and they
showed us how to perform it,

323
00:20:54,014 --> 00:20:54,554
right? How to, you know, isolate
it, you know, visualize it, and

324
00:20:54,554 --> 00:20:55,052
then perform those steps. And
our engineers then, you know,

325
00:20:55,052 --> 00:20:55,542
and students really learned how
to perform it. And then we

326
00:20:55,542 --> 00:20:56,024
collected, you know, data for
the different steps of that

327
00:20:56,024 --> 00:20:56,505
clipping and cutting and
trained, you know, then policies

328
00:20:56,505 --> 00:20:56,944
to perform that and got better
and better. And then,

329
00:20:56,944 --> 00:20:57,417
ultimately, we're able to do
eight of those back to back

330
00:20:57,417 --> 00:20:57,907
without needing any human
intervention. So, we didn't need

331
00:20:57,907 --> 00:20:58,000
any changes

332
00:20:58,000 --> 00:21:34,000
Audrow Nash: Cool.

333
00:21:34,000 --> 00:21:37,297
Axel Krieger: And it was able to
perform all of eight of them

334
00:21:37,297 --> 00:21:38,000
successfully.

335
00:21:38,000 --> 00:21:39,000
Audrow Nash: Hell, yeah.

336
00:21:38,000 --> 00:21:41,059
Axel Krieger: And that's kind of
like the study that we published

337
00:21:41,059 --> 00:21:42,000
in Science Robotics.

338
00:21:42,000 --> 00:21:46,935
Audrow Nash: Okay, and what was
the approach that you guys used?

339
00:21:46,935 --> 00:21:51,100
How did you... Like, so I feel
like I haven't been too

340
00:21:51,100 --> 00:21:55,649
up-to-date with it, but it seems
like imitation learning is

341
00:21:55,649 --> 00:22:00,199
really becoming a good way of
teaching robots to do things.

342
00:22:00,199 --> 00:22:04,595
Like, there's been a big
resurgence in interest in there.

343
00:22:04,595 --> 00:22:09,530
Tell me what's changed, and why
all of a sudden we're seeing all

344
00:22:09,530 --> 00:22:13,000
these really good imitation
learning results.

345
00:22:13,000 --> 00:22:13,840
Axel Krieger: Yes, so the
backbone is an imitation

346
00:22:13,840 --> 00:22:14,849
learning. So, there we used, you
know, in collaboration with

347
00:22:14,849 --> 00:22:15,891
Chelsea Finn in Stanford, you
know, really basing a lot on her

348
00:22:15,891 --> 00:22:16,798
work in the ACT policy. So,
that's the action chunking

349
00:22:16,798 --> 00:22:17,000
transformer,

350
00:22:17,000 --> 00:22:32,000
Audrow Nash: Mm-hmm.

351
00:22:32,000 --> 00:22:36,219
Axel Krieger: You know, that was
really deployed there for more

352
00:22:36,219 --> 00:22:40,371
industrial-type robots. And, you
know, we used that as kind of

353
00:22:40,371 --> 00:22:44,254
like the backbone for our
imitation learning. So, I think,

354
00:22:44,254 --> 00:22:47,000
you know, big, you know, shift
was, yeah.

355
00:22:47,000 --> 00:22:49,636
Audrow Nash: Can you tell me,
like, at a high level how it

356
00:22:49,636 --> 00:22:51,000
works with This ACT, chunking?

357
00:22:51,000 --> 00:22:55,698
Axel Krieger: How do you train a
robot? Yeah, so imagine if you

358
00:22:55,698 --> 00:22:59,799
train a transformer model to
find the next little robot

359
00:22:59,799 --> 00:23:04,124
action step. And you train it
by, you know, pairing videos

360
00:23:04,124 --> 00:23:08,523
with how an expert surgeon moves
the robot to imitate that,

361
00:23:08,523 --> 00:23:13,146
right? So, that's kind of like,
you know, the basic principle.

362
00:23:13,146 --> 00:23:17,919
So, the, you know, system looks
at, you know, the current video,

363
00:23:17,919 --> 00:23:22,169
has, you know, is trained on,
you know, all the different

364
00:23:22,169 --> 00:23:26,643
actions that an expert did and
finds the nice, you know, the

365
00:23:26,643 --> 00:23:31,043
next little robot action to get
closer to the goal. And so,

366
00:23:31,043 --> 00:23:35,218
that's kind of like a little
chunk of action. And so, it

367
00:23:35,218 --> 00:23:39,767
performs that and then find the
next action, the next action,

368
00:23:39,767 --> 00:23:44,017
the next action until you can
like piece that together to

369
00:23:44,017 --> 00:23:47,000
perform, you know, the entire
procedure.

370
00:23:47,000 --> 00:23:51,828
Audrow Nash: Mm-hmm. Achieve it.
How do you segment actions for

371
00:23:51,828 --> 00:23:56,350
this? Is it like a new input,
or... How do we do the action

372
00:23:56,350 --> 00:24:00,182
segmentation? Because I imagine
it could be, like,

373
00:24:00,182 --> 00:24:04,781
infinitesimally small actions,
or large, over big periods of

374
00:24:04,781 --> 00:24:08,000
time, actions. How do you break
them down?

375
00:24:08,000 --> 00:24:09,513
Axel Krieger: Yeah, that's
really, really the trick to make

376
00:24:09,513 --> 00:24:11,077
this work, right? If you try to
automate a long horizon, so a

377
00:24:11,077 --> 00:24:12,590
really, you know, long, you
know, action sequence, then you

378
00:24:12,590 --> 00:24:13,974
have a lot of failures and
problems. And so, you know,

379
00:24:13,974 --> 00:24:15,564
parsing it into like the smaller
fundamental steps is the key,

380
00:24:15,564 --> 00:24:17,000
right? And so, what we deployed
there is a hierarchical,

381
00:24:17,000 --> 00:24:36,000
Audrow Nash: Too much.

382
00:24:36,000 --> 00:24:40,903
Axel Krieger: You know, policy
that, you know, first use, you

383
00:24:40,903 --> 00:24:45,725
know, segment the, you know,
clipping, cutting phase of this

384
00:24:45,725 --> 00:24:50,306
cholocystectomy into small
little fundamental steps using

385
00:24:50,306 --> 00:24:55,048
like a surgical taxonomy, like,
you know, like what are the

386
00:24:55,048 --> 00:24:59,468
fundamental steps of the
surgical procedures, you know,

387
00:24:59,468 --> 00:25:04,210
like, you know, clipping the
first clip into the, you know,

388
00:25:04,210 --> 00:25:08,952
artery, like small little, you
know, sub-steps. And then in

389
00:25:08,952 --> 00:25:13,855
deployment, right, you run this
hierarchical model that first

390
00:25:13,855 --> 00:25:19,079
what's the correct next step and
then executes it under low-level

391
00:25:19,079 --> 00:25:23,901
policy. So, that's how you know,
have a pretty simple policy

392
00:25:23,901 --> 00:25:28,000
achieving, you know, complex,
you know, procedures,

393
00:25:28,000 --> 00:25:30,000
Audrow Nash: Behavior.

394
00:25:30,000 --> 00:25:33,806
Axel Krieger: You know, step by
step in a, you know, like a

395
00:25:33,806 --> 00:25:36,000
surgical, you know, segmentic
way.

396
00:25:36,000 --> 00:25:38,920
Audrow Nash: Mm-hmm. Do you
define the actions, the

397
00:25:38,920 --> 00:25:41,897
breakdown of the actions
manually, using that, like,

398
00:25:41,897 --> 00:25:45,447
surgical tectonomy? So are you
teaching it, "This is the first

399
00:25:45,447 --> 00:25:48,824
part, this is the second part"?
Or is it learned implicitly

400
00:25:48,824 --> 00:25:52,317
through watching? Or how does
that work? And if you are doing

401
00:25:52,317 --> 00:25:55,637
it manually, it sounds like you
probably have to iterate a

402
00:25:55,637 --> 00:25:59,244
little bit and try to figure out
how big or small to break down

403
00:25:59,244 --> 00:26:02,336
the steps to make them
reasonable for the robot. Like,

404
00:26:02,336 --> 00:26:05,828
I'm sure sometimes they don't
perform as well if you're like,

405
00:26:05,828 --> 00:26:09,378
"Well, do the thing." It's like,
it's too big. If you, say, do

406
00:26:09,378 --> 00:26:13,042
this one clip, maybe that works
better. But, or as opposed to do

407
00:26:13,042 --> 00:26:16,649
these clips, this kind of thing.
How do you break that down? Is

408
00:26:16,649 --> 00:26:20,141
it manual? It looks like you're
going to say it's manual. But

409
00:26:20,141 --> 00:26:21,000
yeah, go ahead.

410
00:26:21,000 --> 00:26:23,400
Axel Krieger: Yeah, you can
foresee my answer already. Yeah,

411
00:26:23,400 --> 00:26:25,560
it was very manual and
absolutely trial and error. And

412
00:26:25,560 --> 00:26:28,120
sometimes the step was too large
and then we had to really break

413
00:26:28,120 --> 00:26:29,000
it down further. Yeah.

414
00:26:29,000 --> 00:26:33,142
Audrow Nash: Mm-hmm. Huh. Tell
me about that process. Does that

415
00:26:33,142 --> 00:26:37,152
mean what you do is you try to
take the biggest step and then

416
00:26:37,152 --> 00:26:41,426
you see bad performance and then
you make it smaller and you just

417
00:26:41,426 --> 00:26:45,633
keep doing that until you arrive
at a good working thing? Or how

418
00:26:45,633 --> 00:26:48,000
does, how do we achieve these
steps?

419
00:26:48,000 --> 00:26:50,727
Axel Krieger: So, of course we
started with just teaching small

420
00:26:50,727 --> 00:26:53,281
little steps on its own. And
then we built up an intuition,

421
00:26:53,281 --> 00:26:55,965
what's the right kind of step
size? And so then looking at the

422
00:26:55,965 --> 00:26:58,346
procedure, we were able to
already kind of like, really

423
00:26:58,346 --> 00:27:00,857
have a good idea what could
work. And then if we then have

424
00:27:00,857 --> 00:27:03,411
one step and struggle really
with it, then we often further

425
00:27:03,411 --> 00:27:06,052
divide it into more. So that's
kind of like how it works. You

426
00:27:06,052 --> 00:27:08,606
build an intuition, you do an
initial test run. And then if

427
00:27:08,606 --> 00:27:11,030
you really struggle with a
certain procedure, either you

428
00:27:11,030 --> 00:27:13,325
did something wrong in data
collection, you need some

429
00:27:13,325 --> 00:27:16,009
recovery data. Or you know or
this or it's just too ambitious,

430
00:27:16,009 --> 00:27:18,000
you cannot do it, then you
subdivide it. Yeah.

431
00:27:18,000 --> 00:27:25,714
Audrow Nash: Subdivided. So,
what does it look like to

432
00:27:25,714 --> 00:27:34,429
subdivide the data for this? Is
it like you're segmenting the

433
00:27:34,429 --> 00:27:43,143
video? Would you say, here to
here, this is the step. Here to

434
00:27:43,143 --> 00:27:52,000
here, this is the, that is what
it is? Okay, very interesting.

435
00:27:52,000 --> 00:27:56,836
Axel Krieger: Yes, yeah, so the
subdivision can really happen on

436
00:27:56,836 --> 00:28:01,444
the video. And the video, you
know. and you know the you know

437
00:28:01,444 --> 00:28:06,053
high level policy that selects
what is the current next step,

438
00:28:06,053 --> 00:28:09,000
really runs off of video as
well, yeah.

439
00:28:09,000 --> 00:28:13,537
Audrow Nash: Very interesting.
So, I suppose what you become

440
00:28:13,537 --> 00:28:18,300
good at is you become experts at
to different problems. Is that

441
00:28:18,300 --> 00:28:22,988
kind of like, so first it's the
gallbladder surgery and then I

442
00:28:22,988 --> 00:28:27,600
suspect there are other areas
you're focusing on too. But you

443
00:28:27,600 --> 00:28:32,061
basically become very good at
breaking down the problem for

444
00:28:32,061 --> 00:28:36,824
the robot to learn and then do.
Is that kind of, is that a fair

445
00:28:36,824 --> 00:28:40,000
way to think of this or what do
you think?

446
00:28:40,000 --> 00:28:43,647
Axel Krieger: Yeah, I think
that's definitely a big, you

447
00:28:43,647 --> 00:28:47,879
know a important part of it. The
nice thing is that surgeons have

448
00:28:47,879 --> 00:28:51,526
worked on this idea of the
language of surgery. And also

449
00:28:51,526 --> 00:28:54,000
scientists quite a lot already,
before

450
00:28:54,000 --> 00:28:58,000
Audrow Nash: Quite a lot,
probably. Mm-hmm.

451
00:28:58,000 --> 00:29:01,432
Axel Krieger: Imitation learning
came along, right? So there's a

452
00:29:01,432 --> 00:29:04,702
body of work already in this
taxonomy and like what are small

453
00:29:04,702 --> 00:29:07,866
surgical steps. There has been
some cool work recently from

454
00:29:07,866 --> 00:29:10,976
Andrew Hung, that correlates
these different steps and the

455
00:29:10,976 --> 00:29:14,461
sequence of steps to outcome. to
you know like you know So that's

456
00:29:14,461 --> 00:29:17,571
been really nice where we can
kind of piggyback on body of

457
00:29:17,571 --> 00:29:18,000
research

458
00:29:18,000 --> 00:29:26,000
Audrow Nash: Mm-hmm. All that
existing work. Mm-hmm.

459
00:29:26,000 --> 00:29:28,353
Axel Krieger: Already. And what
we've seen is that the smallest

460
00:29:28,353 --> 00:29:30,743
kind of, it's called a gesture,
that's the surgical term for the

461
00:29:30,743 --> 00:29:32,834
smallest action. That is
actually also a really nice and

462
00:29:32,834 --> 00:29:35,000
small step that we can use for
imitation learning as well.

463
00:29:35,000 --> 00:29:40,033
Audrow Nash: Yeah. Cool. Do you
use, so tell me the role of

464
00:29:40,033 --> 00:29:45,065
language in this. How does, how
does language relate to the

465
00:29:45,065 --> 00:29:49,757
video segments that you're
doing? So it's basically you

466
00:29:49,757 --> 00:29:54,704
labeling what the segment is and
that's used to, I want to

467
00:29:54,704 --> 00:29:59,481
understand how much of the
structure is given by a human

468
00:29:59,481 --> 00:30:04,087
and how much of it is kind of
learned in how to do the

469
00:30:04,087 --> 00:30:09,376
sequence or this kind of thing.
And then also how do the words

470
00:30:09,376 --> 00:30:14,835
relate to the video segments for
this kind of thing? Are we, are

471
00:30:14,835 --> 00:30:20,038
we using LLMs in this? Is it a
visual language model or like,

472
00:30:20,038 --> 00:30:22,000
how does all this work?

473
00:30:22,000 --> 00:30:24,194
Axel Krieger: Yeah, wonderful.
So yeah, language, we use

474
00:30:24,194 --> 00:30:26,545
language to connect the high
level policy from the low level

475
00:30:26,545 --> 00:30:28,936
policy. So how they interact is
through language. And what is

476
00:30:28,936 --> 00:30:31,248
cool about language is that it
also mimics of course, how a

477
00:30:31,248 --> 00:30:33,638
surgeon would you know, oper
operate together with a trainee,

478
00:30:33,638 --> 00:30:36,107
right? If you have a you know n
you know medical student trying

479
00:30:36,107 --> 00:30:38,575
to do a procedure the first time
with experienced surgeons, how

480
00:30:38,575 --> 00:30:40,926
do they interact with language,
right? And so if the trainee

481
00:30:40,926 --> 00:30:43,277
says, I wanna do this next step,
right? Then the supervising

482
00:30:43,277 --> 00:30:45,550
surgeon can say, no, no, do
something else. So it's really

483
00:30:45,550 --> 00:30:47,000
nice from the concept of
supervision.

484
00:30:47,000 --> 00:30:56,648
Audrow Nash: Mm-hmm. Oh cool.
It's a way of doing evaluation

485
00:30:56,648 --> 00:31:06,940
probably too. Is it? Because you
can be like, oh it's suggesting

486
00:31:06,940 --> 00:31:16,749
do this and no, no, no I don't
want to go that path yet. This

487
00:31:16,749 --> 00:31:19,000
kind of thing.

488
00:31:19,000 --> 00:31:22,062
Axel Krieger: Exactly, that
supervision is kind of built

489
00:31:22,062 --> 00:31:25,232
into the structure and then
changing it and then improving

490
00:31:25,232 --> 00:31:28,677
the learning by learning from a
wrong labeling and then getting

491
00:31:28,677 --> 00:31:31,957
it right. So you know, so that's
why this is good and how it

492
00:31:31,957 --> 00:31:34,854
works technically is you
manually label the different

493
00:31:34,854 --> 00:31:38,189
steps and then train a network
to figure out what the correct

494
00:31:38,189 --> 00:31:41,633
step is based on the history of
video that you give. So in this

495
00:31:41,633 --> 00:31:43,000
instance, we gave it four

496
00:31:43,000 --> 00:31:56,000
Audrow Nash: Segments. Mm-hmm.

497
00:31:56,000 --> 00:31:59,000
Axel Krieger: Images, four
seconds apart, and then

498
00:31:59,000 --> 00:32:01,000
Audrow Nash: Aww.

499
00:32:01,000 --> 00:32:04,738
Axel Krieger: Train the network
to figure out what is the next

500
00:32:04,738 --> 00:32:08,476
step that we need to perform.
And that is the basic part, it's

501
00:32:08,476 --> 00:32:11,732
just that language that is
labeled and then we added a

502
00:32:11,732 --> 00:32:15,530
correction part, right? So the
second functionality of the high

503
00:32:15,530 --> 00:32:19,328
level policy is really figuring
out if everything is going well

504
00:32:19,328 --> 00:32:22,584
or if the low level policy is
struggling or needs some

505
00:32:22,584 --> 00:32:26,262
correction. So that's more like
a flag and then a directional

506
00:32:26,262 --> 00:32:29,819
flag, like you need to go a
little bit more left, up, down,

507
00:32:29,819 --> 00:32:30,000
and

508
00:32:30,000 --> 00:32:43,000
Audrow Nash: You have an
evaluator loop. Yeah, yeah.

509
00:32:43,000 --> 00:32:47,902
Axel Krieger: Then the low level
policy takes that step and

510
00:32:47,902 --> 00:32:52,222
takes, do I need to do a
correction? And you need to

511
00:32:52,222 --> 00:32:57,289
change a direction in a certain
way as a language conditioned

512
00:32:57,289 --> 00:33:02,108
low level policy that then takes
this into account by then

513
00:33:02,108 --> 00:33:06,677
producing the current robot
action and that next little

514
00:33:06,677 --> 00:33:10,000
chunk of action that we need to
produce.

515
00:33:10,000 --> 00:33:14,488
Audrow Nash: Cool. Okay, I don't
know that I fully understand all

516
00:33:14,488 --> 00:33:18,769
of it, but the thing that I'm
starting to see is the challenge

517
00:33:18,769 --> 00:33:22,636
with LLMs now is they can't
ingest video. And so part of

518
00:33:22,636 --> 00:33:26,986
what you're doing is you take a
series of frames, you learn how

519
00:33:26,986 --> 00:33:31,336
to talk about it like a surgeon
would, and then you use that to

520
00:33:31,336 --> 00:33:35,617
understand your current state
and to pick what you're doing in

521
00:33:35,617 --> 00:33:39,000
the future for this kind of
thing. Is that right?

522
00:33:39,000 --> 00:33:39,697
Axel Krieger: Yep, yeah. It's
you know, Like very similar to a

523
00:33:39,697 --> 00:33:40,000
vision language model, yes.

524
00:33:40,000 --> 00:33:44,675
Audrow Nash: Super cool. Uh-huh.
What's the, how, so when we're

525
00:33:44,675 --> 00:33:49,276
saying high level and low level,
what exactly do each of those

526
00:33:49,276 --> 00:33:53,580
mean? Is the low level like
literal joint commands for the

527
00:33:53,580 --> 00:33:57,735
robot, or is it like actions
that we're defining for the

528
00:33:57,735 --> 00:34:02,187
robot? And then the high level
is like keeping track of the,

529
00:34:02,187 --> 00:34:06,120
what we're actually doing for
the surgery and how the

530
00:34:06,120 --> 00:34:10,276
progression of steps. And then
how, is there like just a

531
00:34:10,276 --> 00:34:14,951
language model running on this
where it's trying to get through

532
00:34:14,951 --> 00:34:19,700
like a markdown task list of all
the things to do for a surgery,

533
00:34:19,700 --> 00:34:22,000
or how does this actually look?

534
00:34:22,000 --> 00:34:23,120
Axel Krieger: Yeah, so the high
level has this input history of

535
00:34:23,120 --> 00:34:24,062
video and as output, just
language, nothing else. you

536
00:34:24,062 --> 00:34:25,129
know? And the low level has as
input the latest video frame,

537
00:34:25,129 --> 00:34:26,178
the language command of the high
level policy, and then the

538
00:34:26,178 --> 00:34:27,262
output of the low level is the
action of the robot. So that's

539
00:34:27,262 --> 00:34:28,382
kind of like you know, it's kind
of separating the architecture

540
00:34:28,382 --> 00:34:29,467
into one kind of decision making
step and then into an action

541
00:34:29,467 --> 00:34:30,000
step. So it makes it easier to

542
00:34:30,000 --> 00:34:56,000
Audrow Nash: Cool. Mmm.

543
00:34:56,000 --> 00:34:58,000
Axel Krieger: Train each of them
separately,

544
00:34:58,000 --> 00:34:59,000
Audrow Nash: So cool.

545
00:34:59,000 --> 00:35:03,846
Axel Krieger: Makes it easy to
supervise, know what's going on,

546
00:35:03,846 --> 00:35:07,923
and really just was very
successful for us to putting

547
00:35:07,923 --> 00:35:09,000
this together.

548
00:35:09,000 --> 00:35:11,689
Audrow Nash: Mm-hmm. That's such
a cool way to do it. Was there,

549
00:35:11,689 --> 00:35:14,336
so is this very similar to This
ACT framework that Chelsea Finn

550
00:35:14,336 --> 00:35:16,899
and the groups over there have
worked on? Is that pretty much

551
00:35:16,899 --> 00:35:19,000
the same thing, or how is it
similar or different?

552
00:35:19,000 --> 00:35:21,424
Axel Krieger: Yeah, the ACT
framework is the backbone to our

553
00:35:21,424 --> 00:35:23,646
low level policy. So that's
basically taking the latest

554
00:35:23,646 --> 00:35:26,030
villa you know video in and
video framing and producing the

555
00:35:26,030 --> 00:35:28,495
action. That backbone is ACT,
small variation there, and then

556
00:35:28,495 --> 00:35:30,879
also really adapting it to
surgical robots. Surgical robots

557
00:35:30,879 --> 00:35:31,000
are

558
00:35:31,000 --> 00:35:49,000
Audrow Nash: Mm-hmm. Yeah.

559
00:35:49,000 --> 00:35:53,440
Axel Krieger: These long stick
tools. They have a lot of

560
00:35:53,440 --> 00:35:58,118
deflection. They have cable
drives. They are not repeatable

561
00:35:58,118 --> 00:36:03,033
as an industrial robot. So a lot
of our key innovations was to

562
00:36:03,033 --> 00:36:07,315
adapt this, you know, to come up
with the right action

563
00:36:07,315 --> 00:36:11,834
representation to you know
really make this work on these

564
00:36:11,834 --> 00:36:16,432
more cable driven and more
flimsy surgical robots compared

565
00:36:16,432 --> 00:36:20,000
to really well encoded
industrial robot arms.

566
00:36:20,000 --> 00:36:24,506
Audrow Nash: Industrial arms.
Huh. Very cool. Okay, so the

567
00:36:24,506 --> 00:36:29,245
actual tool that you're using is
different, and that makes it

568
00:36:29,245 --> 00:36:33,906
harder because industrial arms
are very stiff and these ones

569
00:36:33,906 --> 00:36:38,334
flex a little bit that you use
in surgery. Were there any

570
00:36:38,334 --> 00:36:43,073
other, I was just going to ask,
differences between like what

571
00:36:43,073 --> 00:36:47,812
you were able to get off the
shelf in some sense and using it

572
00:36:47,812 --> 00:36:52,861
For surgery, like what were some
of those challenging aspects? So

573
00:36:52,861 --> 00:36:57,755
obviously stiff versus kind of
soft actuators, or not soft, but

574
00:36:57,755 --> 00:37:02,805
less rigid, say. Any other major
difficulties in applying kind of

575
00:37:02,805 --> 00:37:07,000
like an industrial solution to a
surgical application?

576
00:37:07,000 --> 00:37:11,519
Axel Krieger: Yeah, surgical
robots have replaceable tools,

577
00:37:11,519 --> 00:37:16,115
right? So our surgical tools are
kind of rated to being used

578
00:37:16,115 --> 00:37:20,634
maybe like 10 times and can be
sterilized 10 times and then

579
00:37:20,634 --> 00:37:25,306
they are thrown away. So they
really wear out quickly, right?

580
00:37:25,306 --> 00:37:29,672
And so, you know, you know
struggle was, for example, you

581
00:37:29,672 --> 00:37:34,191
train on one tool and then you
replace it and have slightly

582
00:37:34,191 --> 00:37:38,787
different kinematics and then
struggle again, right? So that

583
00:37:38,787 --> 00:37:43,077
is certainly a problem. The use
as hardware, the DaVinci

584
00:37:43,077 --> 00:37:47,979
research kit. So these are older
generation DaVinci robots where

585
00:37:47,979 --> 00:37:51,885
the company donates the
hardware. And together with

586
00:37:51,885 --> 00:37:56,481
Hopkins, you know, develop here
at Hopkins and WPI is to you

587
00:37:56,481 --> 00:38:00,923
know ha build a controller
infrastructure that we can then

588
00:38:00,923 --> 00:38:05,136
you know use the hardware
without exposing the software

589
00:38:05,136 --> 00:38:09,809
layer and the control layer of
the original intuitive DaVinci

590
00:38:09,809 --> 00:38:14,021
robot. So it is basically
replacing all the electronics

591
00:38:14,021 --> 00:38:18,540
and software layer. So that is
the robotic test that we are

592
00:38:18,540 --> 00:38:19,000
using.

593
00:38:19,000 --> 00:38:24,000
Audrow Nash: Mm-hmm.

594
00:38:24,000 --> 00:38:27,139
Axel Krieger: There are
different generations of it. And

595
00:38:27,139 --> 00:38:30,614
so another struggle for us was
we you know we recently started

596
00:38:30,614 --> 00:38:33,472
doing you know in vivo, so
living animal tests on a

597
00:38:33,472 --> 00:38:36,667
different generation of this
robot, and there we also saw

598
00:38:36,667 --> 00:38:40,142
large kinematic differences and
our policies did not just work

599
00:38:40,142 --> 00:38:43,000
out of the box. So those are
kind of the struggles.

600
00:38:43,000 --> 00:38:47,337
Audrow Nash: Yeah. Yeah.
Different data distributions.

601
00:38:47,337 --> 00:38:52,558
It's always hard when you change
hardware for everything. How did

602
00:38:52,558 --> 00:38:57,618
you resolve that? Like, did you
just have to retrain on quite a

603
00:38:57,618 --> 00:39:02,277
bit of new data? Or what did you
do when you got different

604
00:39:02,277 --> 00:39:03,000
hardware?

605
00:39:03,000 --> 00:39:03,696
Axel Krieger: Yeah, that's been
kind of like the solution is to,

606
00:39:03,696 --> 00:39:04,359
yeah, exactly. You add more data
with a new configuration and

607
00:39:04,359 --> 00:39:05,000
then can combine it at least to
get best performance. Yeah,

608
00:39:05,000 --> 00:39:10,140
Audrow Nash: Because that's a
common problem. Yeah. Do you do

609
00:39:10,140 --> 00:39:15,449
a lot of like the, I don't know,
maybe something that I've seen

610
00:39:15,449 --> 00:39:20,927
work a little bit is you can try
to leverage the existing data as

611
00:39:20,927 --> 00:39:26,067
much as possible, but then you
fine-tune on some of the newer

612
00:39:26,067 --> 00:39:31,039
hardware to make it work? I
don't know. Is that kind of the

613
00:39:31,039 --> 00:39:35,000
approach that you've taken, or
what did you do?

614
00:39:35,000 --> 00:39:36,470
Axel Krieger: This kind of act
backbone is a little bit small

615
00:39:36,470 --> 00:39:37,819
in parameter size, and so it
doesn't lend itself to fine

616
00:39:37,819 --> 00:39:39,000
tuning. you know, We are doing
fine tuning a lot.

617
00:39:39,000 --> 00:39:47,000
Audrow Nash: Yeah. Oh. You just
restarted, I guess.

618
00:39:47,000 --> 00:39:50,917
Axel Krieger: Yeah, I mean, you
just kind of co-train it with

619
00:39:50,917 --> 00:39:54,000
the new robot data, so that's
been the solution.

620
00:39:54,000 --> 00:39:58,636
Audrow Nash: How much training
is required for it? So you were

621
00:39:58,636 --> 00:40:02,972
working, the science paper was
step two of three, of three

622
00:40:02,972 --> 00:40:07,159
steps in the gallbladder
surgery. How much training data

623
00:40:07,159 --> 00:40:11,720
do you need to be able to do
that reliably? So you said, less

624
00:40:11,720 --> 00:40:16,355
than ten cases successful all in
a row. How much training data

625
00:40:16,355 --> 00:40:18,000
did you need for that?

626
00:40:18,000 --> 00:40:21,775
Axel Krieger: We had about 17
hours of training data, so

627
00:40:21,775 --> 00:40:24,000
that's the order of magnitude.
We

628
00:40:24,000 --> 00:40:26,515
Audrow Nash: How long is one
surgery? Because I mean that

629
00:40:26,515 --> 00:40:28,853
would be, like how many
instances of gallbladder step

630
00:40:28,853 --> 00:40:30,000
two did you end up having?

631
00:40:30,000 --> 00:40:35,056
Axel Krieger: Had 30 different
gallbladders and performed each

632
00:40:35,056 --> 00:40:39,785
of the steps about 15 times. So
it's like 30 variations of

633
00:40:39,785 --> 00:40:44,597
gallbladders, but on each of
them we performed kind of like

634
00:40:44,597 --> 00:40:49,000
each step 15 times to get us to
around this 400 or 500

635
00:40:49,000 --> 00:40:57,000
Audrow Nash: Yep. Instances of
doing it.

636
00:40:57,000 --> 00:40:58,000
Axel Krieger: Instances. Yeah.

637
00:40:58,000 --> 00:41:02,482
Audrow Nash: Okay. And that was,
what's, I don't know if you

638
00:41:02,482 --> 00:41:07,263
probably have an idea for this,
I don't know if this is the only

639
00:41:07,263 --> 00:41:10,998
question. How was the
performance over time as you

640
00:41:10,998 --> 00:41:15,704
were getting those samples? Like
at 100, 200, 300, 400, 500, if

641
00:41:15,704 --> 00:41:19,737
you did 10,000, would it be
unbelievably more? Like it

642
00:41:19,737 --> 00:41:23,846
probably asymptotes in
performance at some point. How's

643
00:41:23,846 --> 00:41:28,253
that, what's your mental model
of how that's happening with

644
00:41:28,253 --> 00:41:29,000
your data?

645
00:41:29,000 --> 00:41:32,165
Axel Krieger: Yeah, we
definitely study that quite a

646
00:41:32,165 --> 00:41:35,696
bit to figure out what's the
right you know amount that we

647
00:41:35,696 --> 00:41:36,000
need.

648
00:41:36,000 --> 00:41:38,000
Audrow Nash: Number to go to.

649
00:41:38,000 --> 00:41:42,895
Axel Krieger: So it really
depends on the complexity of the

650
00:41:42,895 --> 00:41:48,039
procedure and the manipulation
task, but it's typically around

651
00:41:48,039 --> 00:41:53,349
200 to 500 episodes is where we
get some good performance there,

652
00:41:53,349 --> 00:41:58,410
and then it does asymptote. So
we get small improvements, but

653
00:41:58,410 --> 00:42:02,891
not massive improvements
afterwards with these smaller

654
00:42:02,891 --> 00:42:07,537
models. We have, of course,
since the science paper come

655
00:42:07,537 --> 00:42:12,432
out, we're also working a lot
with larger VLA architectures

656
00:42:12,432 --> 00:42:16,000
that are more parameters that
also train on

657
00:42:16,000 --> 00:42:19,000
Audrow Nash: More than the ACT
model, you're saying.

658
00:42:19,000 --> 00:42:21,047
Axel Krieger: A variety of
surgical robotic procedures and

659
00:42:21,047 --> 00:42:23,129
then fine-tune on a particular
case, like you mentioned. So

660
00:42:23,129 --> 00:42:25,000
we've been, of course, working
on these as well. Yep.

661
00:42:25,000 --> 00:42:29,430
Audrow Nash: Yes. Can you tell
me a bit, I have been out of the

662
00:42:29,430 --> 00:42:33,719
loop a little bit. Can you tell
me kind of fundamentally what

663
00:42:33,719 --> 00:42:37,656
visual action language models
are? Like what, how is the

664
00:42:37,656 --> 00:42:41,734
approach, I guess ACT is a much
smaller model for learning

665
00:42:41,734 --> 00:42:45,742
things but these are more
flexible because they're bigger

666
00:42:45,742 --> 00:42:49,961
and I'm curious how you use an
LLM And like what it actually

667
00:42:49,961 --> 00:42:52,000
looks like in implementation.

668
00:42:52,000 --> 00:42:54,859
Axel Krieger: Yeah. So I think,
you know, kind of simply

669
00:42:54,859 --> 00:42:57,463
described is, you know,
relative, you know, similar

670
00:42:57,463 --> 00:43:00,731
between, you know, like act as a
small model and then, you know,

671
00:43:00,731 --> 00:43:04,049
the, you know, larger VLA, a lot
more parameters and, you know, a

672
00:43:04,049 --> 00:43:07,368
kind of like, you know, natural,
you know, you don't have to like

673
00:43:07,368 --> 00:43:10,533
specifically separate in this
high level and low level policy.

674
00:43:10,533 --> 00:43:13,699
That's kind of like integrated
in it, you know, but, you know,

675
00:43:13,699 --> 00:43:16,660
from a functionality, video
input, action output, and, you

676
00:43:16,660 --> 00:43:19,775
know, language to, you know,
give you that context, you know,

677
00:43:19,775 --> 00:43:21,000
what are you performing?

678
00:43:21,000 --> 00:43:26,284
Audrow Nash: All the steps. Oh
yeah, yeah. Is it like a more

679
00:43:26,284 --> 00:43:32,009
general version of what you were
describing with the high and low

680
00:43:32,009 --> 00:43:37,734
level for this? Does it draw its
own abstractions or does it, how

681
00:43:37,734 --> 00:43:43,194
do you, I guess my assumption
with this that you're describing

682
00:43:43,194 --> 00:43:48,919
is you end up needing a lot more
data for it to be performing. Is

683
00:43:48,919 --> 00:43:54,468
that correct? But you can do the
fine tuning and stuff so maybe

684
00:43:54,468 --> 00:43:59,576
that's the direction you're
going for it being more robust

685
00:43:59,576 --> 00:44:04,244
on various hardware because
you're aiming for general

686
00:44:04,244 --> 00:44:09,000
infrastructure for several
companies. Is that correct?

687
00:44:09,000 --> 00:44:12,147
Axel Krieger: Yeah. You're
really correct. So you have, you

688
00:44:12,147 --> 00:44:15,400
know, the more capabilities,
also more training requirements,

689
00:44:15,400 --> 00:44:18,493
but then the ability to, you
know, generalize it and then,

690
00:44:18,493 --> 00:44:21,000
you know, fine-tune it on a
specific, you know,

691
00:44:21,000 --> 00:44:27,000
Audrow Nash: Generalize is much
higher.

692
00:44:27,000 --> 00:44:29,493
Axel Krieger: Work. So, you
know, to give you a good example

693
00:44:29,493 --> 00:44:31,861
of this, we started together
with NVIDIA and 50 different

694
00:44:31,861 --> 00:44:34,230
institutions, this push for
robotics data sets with video

695
00:44:34,230 --> 00:44:36,557
and kinematics. And so we
received, you know, there were

696
00:44:36,557 --> 00:44:38,925
eight companies and involved
giving us also some clinical

697
00:44:38,925 --> 00:44:41,294
data like CMR Surgical, for
example, delivered, you know,

698
00:44:41,294 --> 00:44:42,000
hundreds of hours

699
00:44:42,000 --> 00:44:54,000
Audrow Nash: Cool.

700
00:44:54,000 --> 00:44:54,377
Axel Krieger: And then, you
know, 50 different institutions.

701
00:44:54,377 --> 00:44:54,742
And so we have, you know, this
large data set that is, you

702
00:44:54,742 --> 00:44:55,000
know, very diverse, all
different robots,

703
00:44:55,000 --> 00:45:06,000
Audrow Nash: It's amazing.

704
00:45:06,000 --> 00:45:10,592
Axel Krieger: And there we can,
you know, train, you know, large

705
00:45:10,592 --> 00:45:14,825
reveal architecture on, you
know, general surgical robotics

706
00:45:14,825 --> 00:45:19,058
tasks and then fine-tune them
on, you know, specific robots

707
00:45:19,058 --> 00:45:22,000
and specific application tasks,
you know.

708
00:45:22,000 --> 00:45:26,642
Audrow Nash: So do you think the
core competence of Inner Logic

709
00:45:26,642 --> 00:45:30,989
is going to be effectively
you'll have your visual language

710
00:45:30,989 --> 00:45:35,411
action model that's like the
core thing and then you'll fine

711
00:45:35,411 --> 00:45:39,758
tune it to applications and
procedures. Is that kind of the

712
00:45:39,758 --> 00:45:43,000
play you think or how do you
think about it?

713
00:45:43,000 --> 00:45:44,858
Axel Krieger: So, you know, we,
you know, this imitation

714
00:45:44,858 --> 00:45:47,014
learning, you know, framework is
kind of one, you know, direction

715
00:45:47,014 --> 00:45:48,905
that we are exploring, right?
So, you know, we are really

716
00:45:48,905 --> 00:45:50,000
supporting, you know,
procedural,

717
00:45:50,000 --> 00:45:56,000
Audrow Nash: One aspect.

718
00:45:56,000 --> 00:45:57,285
Axel Krieger: You know, device
manufacturers in a much broader

719
00:45:57,285 --> 00:45:58,528
sense, right? And so it's not
all, you know, action-related.

720
00:45:58,528 --> 00:45:59,731
So, of course, think about,
right, you know, like surgical

721
00:45:59,731 --> 00:46:01,057
robotics, you know, kind of the
last aspects what on the roadmap

722
00:46:01,057 --> 00:46:02,176
is kind of the autonomous
action, right? So, you know,

723
00:46:02,176 --> 00:46:03,523
it's a lot in the analysis side,
it's a lot in the guidance side,

724
00:46:03,523 --> 00:46:04,000
in the decision-making,

725
00:46:04,000 --> 00:46:21,000
Audrow Nash: Yeah. Autonomy.

726
00:46:21,000 --> 00:46:24,932
Axel Krieger: You know, and the
action is kind of the last

727
00:46:24,932 --> 00:46:29,203
piece, you know, a piece. And so
what we've been really finding

728
00:46:29,203 --> 00:46:33,203
is, you know, a lot of needs
already in the kind of earlier

729
00:46:33,203 --> 00:46:37,068
phases. So, you know, we are
certainly, you know, we have

730
00:46:37,068 --> 00:46:41,000
our, you know, research in VLA
and world models, you know,

731
00:46:41,000 --> 00:46:41,931
Audrow Nash: Yep. That's kind of
the end game in some sense but

732
00:46:41,931 --> 00:46:42,877
there's a whole bunch of things
leading up to there that are all

733
00:46:42,877 --> 00:46:43,823
very promising and needed. Yeah
because I mean guidance would be

734
00:46:43,823 --> 00:46:44,000
really cool.

735
00:46:44,000 --> 00:46:52,000
Axel Krieger: We... Exactly,
exactly, exactly, yep.

736
00:46:52,000 --> 00:46:55,546
Audrow Nash: Like there's all
sorts of things if you could

737
00:46:55,546 --> 00:46:58,969
have a more junior surgeon
doing, being advised by an AI

738
00:46:58,969 --> 00:47:02,271
that's like a senior surgeon
that would be super cool.

739
00:47:02,271 --> 00:47:06,000
Especially in like maybe crisis
moments or anything else too.

740
00:47:06,000 --> 00:47:09,078
Axel Krieger: Absolutely. So, an
anatomical, you know,

741
00:47:09,078 --> 00:47:12,383
segmentation, like finding
where's the artery, right? Like

742
00:47:12,383 --> 00:47:15,689
avoid that artery, right? You
know, like that's, you know,

743
00:47:15,689 --> 00:47:17,000
super important, right?

744
00:47:17,000 --> 00:47:19,000
Audrow Nash: Yeah. Don't want to
hit that artery.

745
00:47:19,000 --> 00:47:22,806
Axel Krieger: And yeah, yeah,
and then, you know, just help

746
00:47:22,806 --> 00:47:26,677
what should be the next step,
especially in a bad situation,

747
00:47:26,677 --> 00:47:27,000
yeah.

748
00:47:27,000 --> 00:47:31,046
Audrow Nash: Cool. Okay. So that
all sounds really really cool.

749
00:47:31,046 --> 00:47:34,899
From all the data that you've
been given with all this, how,

750
00:47:34,899 --> 00:47:38,624
tell me some of the challenges
with that because I imagine

751
00:47:38,624 --> 00:47:42,156
you're getting a ton of
different procedures. It's like

752
00:47:42,156 --> 00:47:46,073
I imagine while it's wonderful
to have so much data it's also

753
00:47:46,073 --> 00:47:49,798
probably a challenge to make
sense of that data and use it

754
00:47:49,798 --> 00:47:53,523
productively because there's
probably so much variation in

755
00:47:53,523 --> 00:47:55,000
that. Tell me about it.

756
00:47:55,000 --> 00:47:58,012
Axel Krieger: Yeah, you're so
Right, you know, like, I mean,

757
00:47:58,012 --> 00:48:01,074
that, you know, like, you know,
it sounds awesome to get like

758
00:48:01,074 --> 00:48:04,137
this, you know, totally diverse,
you know, robotic data, some

759
00:48:04,137 --> 00:48:07,349
more endoscopic, some more, you
know, you know, surgical robots,

760
00:48:07,349 --> 00:48:10,110
some more, you know, just the
camera following like all

761
00:48:10,110 --> 00:48:13,072
different, you know, robotic
architectures, but then if you

762
00:48:13,072 --> 00:48:15,984
all try to combine it, you know,
sometimes just, you know,

763
00:48:15,984 --> 00:48:18,594
decrease performance rather
than, you know, increase

764
00:48:18,594 --> 00:48:20,000
adrenalization capabilities.

765
00:48:20,000 --> 00:48:27,000
Audrow Nash: Chaos.

766
00:48:27,000 --> 00:48:30,982
Axel Krieger: So, what we found
often is that we have to like,

767
00:48:30,982 --> 00:48:34,000
you know, use for pre-training
some of the more

768
00:48:34,000 --> 00:48:35,000
Audrow Nash: Specific. Mm hmm.

769
00:48:34,000 --> 00:48:38,820
Axel Krieger: Aligned, you know,
robotic systems. So, you know,

770
00:48:38,820 --> 00:48:43,027
like a subset of the 50
institutions has been like, you

771
00:48:43,027 --> 00:48:47,082
know, the better kind of
training recipe for specific

772
00:48:47,082 --> 00:48:51,672
tasks on some robots, you know.
So, yeah, absolutely. That's

773
00:48:51,672 --> 00:48:56,339
kind of a big, you know, big
task is right, you know, finding

774
00:48:56,339 --> 00:49:01,082
the right, you know, the right
mix of training data, and then,

775
00:49:01,082 --> 00:49:05,290
you know, really finding the
right episodes, right, the

776
00:49:05,290 --> 00:49:09,880
really good episodes to train
from. So, that's a lot of the,

777
00:49:09,880 --> 00:49:14,240
you know, secret sauce going
into the beginning to have a

778
00:49:14,240 --> 00:49:16,000
good, you know, policy.

779
00:49:16,000 --> 00:49:18,500
Audrow Nash: How, can you talk a
bit about how you do it for

780
00:49:18,500 --> 00:49:20,000
that? I know it's a secret sauce
but

781
00:49:20,000 --> 00:49:22,480
Axel Krieger: No, no, I, no, I
mean, you know, secret sauce is

782
00:49:22,480 --> 00:49:24,800
probably the wrong expression,
the challenge, I would say,

783
00:49:24,800 --> 00:49:25,000
yeah.

784
00:49:25,000 --> 00:49:28,000
Audrow Nash: It's the challenge.
Yeah yeah. I see.

785
00:49:28,000 --> 00:49:31,523
Axel Krieger: So, you know, so,
for example, right, if we want

786
00:49:31,523 --> 00:49:34,818
to perform a task with the
DaVinci Research Kit, then, you

787
00:49:34,818 --> 00:49:38,398
know, it really showed that if
we kind of train it on the super

788
00:49:38,398 --> 00:49:41,864
mixed with all different robot
architectures, it did not, you

789
00:49:41,864 --> 00:49:45,443
know, improve the performance,
right. So, but then if you pick,

790
00:49:45,443 --> 00:49:48,000
you know, very similar robotic
architectures,

791
00:49:48,000 --> 00:49:53,000
Audrow Nash: I would expect.
Yeah.

792
00:49:53,000 --> 00:49:56,338
Axel Krieger: Then it's
beneficial to, you know,

793
00:49:56,338 --> 00:50:00,719
pre-train this on the variety of
tasks, and then fine-tuning it

794
00:50:00,719 --> 00:50:04,821
on the specific task versus just
learning, you know, on the

795
00:50:04,821 --> 00:50:09,202
specific task. And that's been
kind of like, you know, it makes

796
00:50:09,202 --> 00:50:13,652
sense, right? I mean, it and it
the data really bore this out as

797
00:50:13,652 --> 00:50:14,000
well.

798
00:50:14,000 --> 00:50:18,246
Audrow Nash: Huh. Do you think
that We're gonna have like it

799
00:50:18,246 --> 00:50:22,705
seems like with language models
like large language models like

800
00:50:22,705 --> 00:50:27,234
similar approaches seem to have
been done for a long time in the

801
00:50:27,234 --> 00:50:31,338
computational linguistics or
whatever it is. But then when

802
00:50:31,338 --> 00:50:35,160
you have internet scale data
then you get pretty crazy

803
00:50:35,160 --> 00:50:39,548
performance seems like you can
reason now it's an awesome tool

804
00:50:39,548 --> 00:50:43,582
for a lot of things. Do you
think that we're missing that

805
00:50:43,582 --> 00:50:47,757
kind of like super large scale
data do you think that would

806
00:50:47,757 --> 00:50:51,932
just kind of solve everything
for you or is it like getting

807
00:50:51,932 --> 00:50:56,037
and I know that the challenge
becomes how to get access to

808
00:50:56,037 --> 00:51:00,000
everything but do you think it's
largely a data problem?

809
00:51:00,000 --> 00:51:00,590
Axel Krieger: I think that's,
you know, a huge part of it, why

810
00:51:00,590 --> 00:51:01,181
it's hard to kind of replicate
those generalis generalization,

811
00:51:01,181 --> 00:51:01,762
you know, results of LAM for
sure. You know, it's a different

812
00:51:01,762 --> 00:51:02,000
magnitude of data, right?

813
00:51:02,000 --> 00:51:17,000
Audrow Nash: The case. Oh yeah
completely different. huh.

814
00:51:17,000 --> 00:51:20,735
Axel Krieger: So, absolutely,
that's, and there I think we

815
00:51:20,735 --> 00:51:24,534
just need to be, you know,
smarter and, you know, smart and

816
00:51:24,534 --> 00:51:28,333
do this, you know, and get
successful at a much, much lower

817
00:51:28,333 --> 00:51:32,068
regime and engineer, you know,
solutions to still get, you

818
00:51:32,068 --> 00:51:34,000
know, really good performance.

819
00:51:34,000 --> 00:51:35,348
Audrow Nash: Yes. Yeah that's
the way because there's not

820
00:51:35,348 --> 00:51:36,768
enough data so you have to like
do feature crafting or break

821
00:51:36,768 --> 00:51:38,258
things down manually or whatever
it might be so you can get the

822
00:51:38,258 --> 00:51:39,725
performance you want with the
limited data that you have. What

823
00:51:39,725 --> 00:51:41,215
other procedures so gallbladder
surgery step two very clear are

824
00:51:41,215 --> 00:51:42,658
you doing what other procedures
are you guys working to solve

825
00:51:42,658 --> 00:51:44,030
for is it the rest of the
gallbladder surgery or what kind

826
00:51:44,030 --> 00:51:45,000
of things are you aiming at as a
company?

827
00:51:45,000 --> 00:51:50,139
Axel Krieger: Exactly. Yeah.
Yeah. So, yeah. Yeah. Over the

828
00:51:50,139 --> 00:51:55,279
last year, really, you know,
we've been pushing to get from

829
00:51:55,279 --> 00:52:00,941
ex vivo, so, you know, just, you
know, piece of the liver and the

830
00:52:00,941 --> 00:52:06,167
god letter to in vivo, right?
So, that's a huge step for us.

831
00:52:06,167 --> 00:52:11,394
And so, we, you know, performed
the first, you know, in vivo

832
00:52:11,394 --> 00:52:16,707
study. Really looks very good,
trained on ex vivo, performing

833
00:52:16,707 --> 00:52:22,021
in vivo, generalizing across
that with breathing. So, that is

834
00:52:22,021 --> 00:52:26,986
super exciting and, you know,
and, you know, amazing. And

835
00:52:26,986 --> 00:52:32,300
then, you know, we are working
to include the other phases of

836
00:52:32,300 --> 00:52:35,000
the god letter surgery, so the,

837
00:52:35,000 --> 00:52:51,000
Audrow Nash: Wow. That is.

838
00:52:51,000 --> 00:52:55,269
Axel Krieger: You know,
dissection, you know, phases, so

839
00:52:55,269 --> 00:52:58,852
that's thick. And then,
expanding to, you know,

840
00:52:58,852 --> 00:53:03,578
different robotic systems and
architecture. Super exciting for

841
00:53:03,578 --> 00:53:07,848
us is kind of endo-luminal,
endo-vascular, so endo means

842
00:53:07,848 --> 00:53:12,345
inside, vascular is, you know,
blood vessels, so, you know,

843
00:53:12,345 --> 00:53:16,614
inside blood vessels. So,
they're really nice new robots

844
00:53:16,614 --> 00:53:21,493
that drive, you know, catheters
into various parts of your body,

845
00:53:21,493 --> 00:53:25,000
especially, you know, in your
brain for stroke

846
00:53:25,000 --> 00:53:27,000
Audrow Nash: Really?

847
00:53:27,000 --> 00:53:29,000
Axel Krieger: Repair and
aneurysm repair. And,

848
00:53:29,000 --> 00:53:31,000
Audrow Nash: That's crazy. I
didn't know that existed.

849
00:53:31,000 --> 00:53:36,053
Axel Krieger: You know, there's,
you know, there's a lot of need

850
00:53:36,053 --> 00:53:40,474
for, you know, guidance for
visualization for, you know,

851
00:53:40,474 --> 00:53:45,211
autonomy to, you know, let's say
you have a stroke somewhere

852
00:53:45,211 --> 00:53:50,026
remotely, how can you know, do
this, you know, without access

853
00:53:50,026 --> 00:53:54,684
to, you know, the high-level
stroke center anywhere nearby.

854
00:53:54,684 --> 00:53:58,000
So, a lot of really cool need
around this.

855
00:53:58,000 --> 00:53:58,817
Audrow Nash: Huh. Man what a
thing. So it's trying to like

856
00:53:58,817 --> 00:53:59,676
democratize healthcare say or
surgery access when it's needed

857
00:53:59,676 --> 00:54:00,000
for this kind of thing.

858
00:54:00,000 --> 00:54:09,000
Axel Krieger: Yeah. Yeah.

859
00:54:09,000 --> 00:54:13,515
Audrow Nash: And that's so cool.
How Well so that sounds like a

860
00:54:13,515 --> 00:54:17,672
very I guess what do you think
the timing for that kind of

861
00:54:17,672 --> 00:54:22,115
thing like it sounds to me like
it may be far out where we are

862
00:54:22,115 --> 00:54:26,200
automating more things and
making like I mean the I don't

863
00:54:26,200 --> 00:54:30,643
know much about the steering and
the scope to block out stroke

864
00:54:30,643 --> 00:54:35,015
and stuff like that's so cool
I've seen some where they do it

865
00:54:35,015 --> 00:54:39,530
in the heart. But in the brain
is like bonkers that can be done

866
00:54:39,530 --> 00:54:43,902
I don't know too much but that
seems crazy to me. What do you

867
00:54:43,902 --> 00:54:48,345
imagine for the timing for these
kinds of things and just like

868
00:54:48,345 --> 00:54:52,000
this area in general where do
you imagine it going?

869
00:54:52,000 --> 00:54:53,630
Axel Krieger: Yeah. So, you
know, I think a good analogy is

870
00:54:53,630 --> 00:54:55,316
there, you know, our journey in
the autonomous driving world,

871
00:54:55,316 --> 00:54:56,974
right? We didn't go from, you
know, zero, you know, autonomy

872
00:54:56,974 --> 00:54:58,604
zero, you know, navigation help
to, you know, self-driving,

873
00:54:58,604 --> 00:55:00,373
fully self-driving car without a
steering wheel, right? Like, we

874
00:55:00,373 --> 00:55:02,003
had like these little small
steps and functions, right? And

875
00:55:02,003 --> 00:55:03,661
so, like a park assist and then
on the highway a bit of, you

876
00:55:03,661 --> 00:55:05,402
know, self-driving, you know, a
brake assist. And so, these are

877
00:55:05,402 --> 00:55:07,116
kind of like also the same kind
of functions that are going to

878
00:55:07,116 --> 00:55:08,000
come into these surgical robots.

879
00:55:08,000 --> 00:55:28,000
Audrow Nash: Yep. Yeah. Mm hmm.

880
00:55:28,000 --> 00:55:32,000
Axel Krieger: So, and, you know,
we already have FDA approved a

881
00:55:32,000 --> 00:55:35,810
autonomous camera following so
that, you know, moon surgical

882
00:55:35,810 --> 00:55:36,000
is,

883
00:55:36,000 --> 00:55:39,000
Audrow Nash: Ooh.

884
00:55:39,000 --> 00:55:41,163
Axel Krieger: You know, the, you
know, surgical device

885
00:55:41,163 --> 00:55:43,447
manufacturer that has FDA
approval for their module that,

886
00:55:43,447 --> 00:55:45,971
you know, autonomously adjusts
the camera position. So, this is

887
00:55:45,971 --> 00:55:48,374
a low-risk procedure, right?
Because you don't interact with

888
00:55:48,374 --> 00:55:50,658
the tool directly and you can
kind of limit the amount of

889
00:55:50,658 --> 00:55:52,982
motions to make it nice and
safe. And so, you know, that's

890
00:55:52,982 --> 00:55:55,345
kind of how the FDA approves
devices is by risk, right? So,

891
00:55:55,345 --> 00:55:57,789
you start with really low-risk,
you know, functions and then,

892
00:55:57,789 --> 00:55:59,872
you know, we get more
confidence, we get better, you

893
00:55:59,872 --> 00:56:02,436
know, and more robust and then,
you know, can climb kind of that

894
00:56:02,436 --> 00:56:05,000
risk profile to, you know, more
and more, longer procedures. So,

895
00:56:05,000 --> 00:56:23,000
Audrow Nash: Yes absolutely.

896
00:56:23,000 --> 00:56:26,094
Axel Krieger: Yeah, adoption of
guidance, decision-making is

897
00:56:26,094 --> 00:56:29,136
going to, you know, tracking
that's, you know, already kind

898
00:56:29,136 --> 00:56:32,385
of coming and, in you know, in
some way, you know, already, you

899
00:56:32,385 --> 00:56:35,273
know, coming and autonomous
action limited in, you know,

900
00:56:35,273 --> 00:56:38,263
functions like, you know, camera
control and now some, you

901
00:56:38,263 --> 00:56:41,460
assistive, you know, traction
or, you know, like some suction,

902
00:56:41,460 --> 00:56:44,348
you know, maybe then some
suturing, maybe some resection

903
00:56:44,348 --> 00:56:47,597
and then eventually, you know,
full procedures, but there needs

904
00:56:47,597 --> 00:56:50,639
to be definitely a lot more, you
know, work done to make it

905
00:56:50,639 --> 00:56:51,000
happen.

906
00:56:51,000 --> 00:56:56,571
Audrow Nash: Some suturing maybe
and. Mm hmm. Regulation. Yeah.

907
00:56:56,571 --> 00:57:01,964
Gotcha. Tell me about so you
mentioned you went from whatever

908
00:57:01,964 --> 00:57:07,269
it is where you have a chunk of
meat that you're like a dead

909
00:57:07,269 --> 00:57:12,663
chunk from the pig to in vivo
alive pig or a live person with

910
00:57:12,663 --> 00:57:17,792
the gallbladder surgery and
breathing. Tell me a bit about

911
00:57:17,792 --> 00:57:23,362
that like I imagine breathing is
interesting because everything

912
00:57:23,362 --> 00:57:28,314
is moving maybe they don't
breathe that fast but because

913
00:57:28,314 --> 00:57:33,000
sedated but tell me a bit about
the challenges there.

914
00:57:33,000 --> 00:57:35,233
Axel Krieger: Yeah, so we work
with, you know, Children's

915
00:57:35,233 --> 00:57:37,780
National, you know, Richard Shaw
and his team there at, you know,

916
00:57:37,780 --> 00:57:39,974
down in DC on, you know,
performing, you know, disc golf

917
00:57:39,974 --> 00:57:42,365
ladder surgery in life peaks and
so that's, you know, ongoing

918
00:57:42,365 --> 00:57:44,637
study, hasn't been published
yet, you know, but, you know,

919
00:57:44,637 --> 00:57:46,988
the initial results look really
promising. So super exciting

920
00:57:46,988 --> 00:57:49,457
that when we had a policy, you
know, trained, you know, just on

921
00:57:49,457 --> 00:57:51,768
ex vivo tissue that it worked
kind of, you know, out of the

922
00:57:51,768 --> 00:57:53,806
box per se, you know, in the
first animal study with

923
00:57:53,806 --> 00:57:56,000
breathing and motion was, you
know, just super exciting.

924
00:57:56,000 --> 00:58:17,000
Audrow Nash: Coming. Yeah. Mm
hmm.

925
00:58:17,000 --> 00:58:19,094
Axel Krieger: You know,
definitely not easy. It's a

926
00:58:19,094 --> 00:58:21,352
different robot. It has
different configuration. It is,

927
00:58:21,352 --> 00:58:23,897
you know, of course, the, you
know, bleeding and breathing is,

928
00:58:23,897 --> 00:58:26,484
you know, different than what we
have trained on. So, you know,

929
00:58:26,484 --> 00:58:29,070
exciting, you know, challenges,
but, you know, really looks, it

930
00:58:29,070 --> 00:58:31,000
looks very promising for the
overall framework.

931
00:58:31,000 --> 00:58:36,844
Audrow Nash: Mm hmm. Super cool.
Where do you guys think you're

932
00:58:36,844 --> 00:58:42,502
going as a company because so
you said funding late last year

933
00:58:42,502 --> 00:58:48,161
I think you said to me when we
talked you started in February

934
00:58:48,161 --> 00:58:53,633
the company. Maybe I'm wrong
with that but around then like

935
00:58:53,633 --> 00:58:59,385
early beginning of this year.
What do you think your next year

936
00:58:59,385 --> 00:59:04,672
looks like. Maybe next three
years too because it's still

937
00:59:04,672 --> 00:59:10,609
pretty early. But it seems like
you're doing very exciting work.

938
00:59:10,609 --> 00:59:12,000
Where we going.

939
00:59:12,000 --> 00:59:15,029
Axel Krieger: Yeah, I mean, it's
been a whirlwind of a year and

940
00:59:15,029 --> 00:59:17,913
with, you know, raising the
funds and then, you know, really

941
00:59:17,913 --> 00:59:20,798
getting started and, you know,
building our first, you know,

942
00:59:20,798 --> 00:59:23,442
product and, you know, and then,
you know, working with

943
00:59:23,442 --> 00:59:25,798
different, you know,
manufacturers. So, you know,

944
00:59:25,798 --> 00:59:27,000
it's, we don't have like,

945
00:59:27,000 --> 00:59:33,000
Audrow Nash: Have you launched
your product. It.

946
00:59:33,000 --> 00:59:34,000
Axel Krieger: You know, we work
with different, you

947
00:59:34,000 --> 00:59:35,818
Audrow Nash: The beta testing it
or alpha testing it kind of

948
00:59:35,818 --> 00:59:36,000
thing.

949
00:59:36,000 --> 00:59:36,302
Axel Krieger: Know, we sign
deals with, you know,

950
00:59:36,302 --> 00:59:36,671
manufacturers and we, you know,
are deploying, but it's, you

951
00:59:36,671 --> 00:59:37,034
know, not like a, you know,
strict, you know, product line.

952
00:59:37,034 --> 00:59:37,366
It's really customized for, you
know, different device

953
00:59:37,366 --> 00:59:37,717
manufacturers, you know. So,
it's, you know, a, you know,

954
00:59:37,717 --> 00:59:38,000
customizable, you know, a
product line around,

955
00:59:38,000 --> 00:59:58,000
Audrow Nash: Yep. Nice.

956
00:59:58,000 --> 01:00:02,570
Axel Krieger: You know, a lot of
data. So, you know, synthetic

957
01:00:02,570 --> 01:00:07,066
data and then, you know, also
the other, you know, is what we

958
01:00:07,066 --> 01:00:11,194
talked about the struggle with
how do you operationally,

959
01:00:11,194 --> 01:00:15,838
operationalize, you know, real
data. How do you know, structure

960
01:00:15,838 --> 01:00:20,408
it? How do you know, make it so
you can, you know, train with?

961
01:00:20,408 --> 01:00:24,978
And so, we have, you know, these
two, you know, you know, data

962
01:00:24,978 --> 01:00:28,000
platforms as, you know, initial
products.

963
01:00:28,000 --> 01:00:31,634
Audrow Nash: So that makes
sense. That's exciting that it

964
01:00:31,634 --> 01:00:35,460
sounds to me a little bit like
consulting work but it's like

965
01:00:35,460 --> 01:00:39,030
while you fit specific customers
what they need and then

966
01:00:39,030 --> 01:00:42,664
eventually you generalize and
that becomes a more general

967
01:00:42,664 --> 01:00:46,617
product for this kind of thing
is that kind of the leaning you

968
01:00:46,617 --> 01:00:47,000
think.

969
01:00:47,000 --> 01:00:51,130
Axel Krieger: Yeah, let me just
explain this a little bit

970
01:00:51,130 --> 01:00:55,188
better. On the, you know,
synthetic, you know, data set,

971
01:00:55,188 --> 01:00:59,246
what we have is a large cohort
of, you know, CT scans of

972
01:00:59,246 --> 01:01:03,594
patients, so tens of thousands
of patient data sets with all

973
01:01:03,594 --> 01:01:08,014
various anatomies. And so, what
we can do, for example, is to

974
01:01:08,014 --> 01:01:12,362
generate, you know, a large
cohort of, you know, data with a

975
01:01:12,362 --> 01:01:16,710
certain medical device, you
know, from a company, right? So,

976
01:01:16,710 --> 01:01:21,130
of course, there's customization
by, you know, using our, you

977
01:01:21,130 --> 01:01:22,000
know, cohort

978
01:01:22,000 --> 01:01:23,000
Audrow Nash: Oh I see.

979
01:01:23,000 --> 01:01:24,261
Axel Krieger: With a specific
device data. But, of course, all

980
01:01:24,261 --> 01:01:25,460
the, you know, staging and
platform is reusable, right? And

981
01:01:25,460 --> 01:01:26,701
so, we can, you know, have the
platform set up, you know, and

982
01:01:26,701 --> 01:01:27,961
then we can, you know, you know
customize it around the device

983
01:01:27,961 --> 01:01:29,222
needs for, you know, a specific,
you know, company. So, if you

984
01:01:29,222 --> 01:01:30,503
want to work with this on a, you
know, bone fracture, you know,

985
01:01:30,503 --> 01:01:31,519
versus on a, you know,
neurovascular guidance, you

986
01:01:31,519 --> 01:01:32,739
know, device, of course, it's,
you know, slightly different,

987
01:01:32,739 --> 01:01:34,000
but it's still the backbone of
it is this large data set, and,

988
01:01:34,000 --> 01:01:35,240
you know, then we can visualize
it on your specific needs and

989
01:01:35,240 --> 01:01:36,481
with your, you know, in mind.
And similarly, on the real, you

990
01:01:36,481 --> 01:01:37,762
know, data intake, right? If you
work on an endovascular robot,

991
01:01:37,762 --> 01:01:39,002
it looks slightly different than
on a laparoscopic, you know,

992
01:01:39,002 --> 01:01:40,283
robot, but it's still, you know,
like, you know, annotating the

993
01:01:40,283 --> 01:01:41,543
anatomy, annotating, you know,
the tools, you know, annotating

994
01:01:41,543 --> 01:01:42,662
the different steps of surgery
are also, you know, very

995
01:01:42,662 --> 01:01:43,821
related, right? So, we support a
wide range of procedural

996
01:01:43,821 --> 01:01:45,000
devices. It becomes a little bit
customized per, you know,

997
01:01:45,000 --> 01:02:41,000
Audrow Nash: Mm hmm. Cool. Yeah.
Application

998
01:02:41,000 --> 01:02:41,462
Axel Krieger: Application and
customer, but, you know, on the

999
01:02:41,462 --> 01:02:41,917
back end, on our side, it's
definitely, you know, general as

1000
01:02:41,917 --> 01:02:42,000
a platform.

1001
01:02:41,000 --> 01:02:45,440
Audrow Nash: Customer and
everything. It's fairly general.

1002
01:02:45,440 --> 01:02:50,185
Cool. Tell me about the role of
synthetic data for this like I

1003
01:02:50,185 --> 01:02:54,701
guess what you're saying with
the CT scans is that like you

1004
01:02:54,701 --> 01:02:59,141
have real patient data where
they did a CT scan that gives

1005
01:02:59,141 --> 01:03:03,963
you like a CAT scan. Is that one
of those things where they sit

1006
01:03:03,963 --> 01:03:08,785
in like a donut and they go scan
through it and that gets there

1007
01:03:08,785 --> 01:03:12,000
like 3D body decomposition or
composition.

1008
01:03:12,000 --> 01:03:14,337
Axel Krieger: Yes, exactly. So,
it's basically an X-ray machine,

1009
01:03:14,337 --> 01:03:16,199
but it rotates around you and
has a very, very high

1010
01:03:16,199 --> 01:03:18,463
resolution. So, you get, like,
submillimeter resolution of the

1011
01:03:18,463 --> 01:03:20,691
full body, the full anatomy of
tens of thousands of patients.

1012
01:03:20,691 --> 01:03:22,846
So, imagine, like, you know,
like knowing all the different

1013
01:03:22,846 --> 01:03:25,000
anatomical variations that your
device could encounter. So,

1014
01:03:25,000 --> 01:03:34,000
Audrow Nash: Cool.

1015
01:03:34,000 --> 01:03:35,151
Axel Krieger: You know, like, if
you do test your device in a

1016
01:03:35,151 --> 01:03:36,340
cadaver lab, you can do it maybe
dozens of times, you know, but

1017
01:03:36,340 --> 01:03:37,491
you cannot get the scales of,
you know, of tens of thousands,

1018
01:03:37,491 --> 01:03:38,623
right? And so, that's, and our
simulation is, of course, not

1019
01:03:38,623 --> 01:03:39,717
just, you know, a statistical
atlas of something, but it's

1020
01:03:39,717 --> 01:03:40,736
real patient anatomy and
capturing the real variations

1021
01:03:40,736 --> 01:03:41,868
that we would see. So, that's
the cool thing that, you know,

1022
01:03:41,868 --> 01:03:42,981
we can use this for, you know,
you know training models and

1023
01:03:42,981 --> 01:03:44,094
then, you know, a lot of the
testing. So, that's, you know,

1024
01:03:44,094 --> 01:03:45,000
the exciting part to see this
across variations.

1025
01:03:45,000 --> 01:03:52,840
Audrow Nash: Simulation. Mm hmm.
How is it used in all of this?

1026
01:03:52,840 --> 01:03:59,187
Because okay so you have these
really high fidelity

1027
01:03:59,187 --> 01:04:06,529
three-dimensional models of
people. How do you use those or

1028
01:04:06,529 --> 01:04:14,493
where do you use them? Is it in
simulating surgeries? Is it like

1029
01:04:14,493 --> 01:04:21,711
I just the thing that I maybe
naively think is the body is

1030
01:04:21,711 --> 01:04:28,556
very deformable. And I think
that would make simulation

1031
01:04:28,556 --> 01:04:35,773
really hard but maybe you can
just use it as a you kind of

1032
01:04:35,773 --> 01:04:43,116
juggle or shift the organs
around to where they are in this

1033
01:04:43,116 --> 01:04:50,707
particular person and then you
use that for variation and see

1034
01:04:50,707 --> 01:04:58,422
if it will fail. How do you use
simulation here? Like where is

1035
01:04:58,422 --> 01:05:06,262
it really valuable in learning
things for I don't know robotics

1036
01:05:06,262 --> 01:05:09,000
or other applications?

1037
01:05:09,000 --> 01:05:11,699
Axel Krieger: You're absolutely
right. the you know depending on

1038
01:05:11,699 --> 01:05:14,399
the application, it's either you
know easy to use it directly or

1039
01:05:14,399 --> 01:05:16,972
you need to do additional steps
to make it usable, right? And

1040
01:05:16,972 --> 01:05:19,545
So, you know, like, having a
rigid, you know, body, you know,

1041
01:05:19,545 --> 01:05:22,118
helps you already for kind of
orthopedic devices, right? Your

1042
01:05:22,118 --> 01:05:24,775
bones are not going to move and
deform, right? So, you know, so

1043
01:05:24,775 --> 01:05:27,137
we work, you know, with
companies where, you know, like,

1044
01:05:27,137 --> 01:05:29,626
you know, you have different,
you know, orthopedic implants

1045
01:05:29,626 --> 01:05:32,283
that you want to test on a large
cohort of patients, right? So,

1046
01:05:32,283 --> 01:05:33,000
that's, you know,

1047
01:05:33,000 --> 01:05:45,418
Audrow Nash: Oh yeah you
mentioned bones. Mm hmm. Oh

1048
01:05:45,418 --> 01:05:49,000
that's so cool.

1049
01:05:49,000 --> 01:05:53,149
Axel Krieger: So that's, you
know, like, and then they get

1050
01:05:53,149 --> 01:05:57,442
implanted on X, on the X-ray
guidance. So, we can, you know,

1051
01:05:57,442 --> 01:06:00,661
then, you know, do a
physics-based, you know,

1052
01:06:00,661 --> 01:06:05,096
simulation of a specific, you
know, X-ray machine and can show

1053
01:06:05,096 --> 01:06:09,317
you at these different angles
how this would look like, you

1054
01:06:09,317 --> 01:06:13,824
know, on a large cohort. We can
vary the imaging parameters and

1055
01:06:13,824 --> 01:06:17,902
then, you know, you know really
help you in designing and

1056
01:06:17,902 --> 01:06:22,051
accelerating your design and
then, you know, building, you

1057
01:06:22,051 --> 01:06:26,415
know, then models that help you
track, you know, devices and,

1058
01:06:26,415 --> 01:06:30,493
you know, so, and work towards,
you know, more, you know,

1059
01:06:30,493 --> 01:06:35,000
functions. So, yeah, that's easy
on kind of bony anatomy. Then,

1060
01:06:35,000 --> 01:06:37,000
Audrow Nash: Yeah harder on
squishy anatomy.

1061
01:06:37,000 --> 01:06:39,222
Axel Krieger: Yes, and some, you
know, are also a little bit

1062
01:06:39,222 --> 01:06:41,481
easier. We call this kind of
like just deformation, right? If

1063
01:06:41,481 --> 01:06:43,556
you don't cut, if you don't,
like, you know, make larger

1064
01:06:43,556 --> 01:06:45,630
changes. So, for example,
endovascular, right, where you

1065
01:06:45,630 --> 01:06:47,963
put like a catheter in, it like
forms, but it's just, you know,

1066
01:06:47,963 --> 01:06:48,000
a

1067
01:06:48,000 --> 01:06:57,000
Audrow Nash: Yeah. Following
something.

1068
01:06:57,000 --> 01:06:59,273
Axel Krieger: Elastic
deformation goes back to the

1069
01:06:59,273 --> 01:07:02,091
original state, right? And so,
it's also not that difficult to

1070
01:07:02,091 --> 01:07:04,545
simulate. So, we can now
simulate, for example, like a

1071
01:07:04,545 --> 01:07:05,000
stand in a

1072
01:07:05,000 --> 01:07:10,000
Audrow Nash: Huh.

1073
01:07:10,000 --> 01:07:10,317
Axel Krieger: Vessel, right? We
can have, you know, deploy, you

1074
01:07:10,317 --> 01:07:10,618
know, different stands, you
know, in these tens of thousands

1075
01:07:10,618 --> 01:07:10,920
of, you know, models and it
looks, you know, incredibly, you

1076
01:07:10,920 --> 01:07:11,000
know, realistic,

1077
01:07:11,000 --> 01:07:23,000
Audrow Nash: So cool.

1078
01:07:23,000 --> 01:07:26,399
Axel Krieger: You know, in that
model. So, we can, you know,

1079
01:07:26,399 --> 01:07:29,516
functionalize it and then, you
know, for the particular

1080
01:07:29,516 --> 01:07:33,085
application, you know, increase
the complexity of that model to

1081
01:07:33,085 --> 01:07:36,258
really mimic the procedure. Of
course, what we discussed

1082
01:07:36,258 --> 01:07:39,657
earlier, the cholecystectomy,
like cutting and stuff, that's

1083
01:07:39,657 --> 01:07:43,000
kind of like the most difficult
part, right? And so, there,

1084
01:07:43,000 --> 01:07:45,000
Audrow Nash: Yeah I imagine.

1085
01:07:45,000 --> 01:07:46,047
Axel Krieger: You know, we also
work on really cool simulations

1086
01:07:46,047 --> 01:07:47,061
starting with these, you know,
large cohorts of data sets and

1087
01:07:47,061 --> 01:07:48,075
new simulation techniques to
model, you know, how to cut, how

1088
01:07:48,075 --> 01:07:49,122
to suture, how to do this, and
then how to render it, you know,

1089
01:07:49,122 --> 01:07:50,102
that it's visually, you know,
really similar to what we see

1090
01:07:50,102 --> 01:07:51,000
under camera guidance. So, yeah,
that's exciting work.

1091
01:07:51,000 --> 01:07:56,328
Audrow Nash: Yep. Close. Is
that? So what I'm imagining and

1092
01:07:56,328 --> 01:08:01,566
I don't know if it's correct is
this seems like kind of an

1093
01:08:01,566 --> 01:08:06,894
adjacent value add to kind of
the main quest of the company

1094
01:08:06,894 --> 01:08:11,861
but it seems to me like
eventually having access to all

1095
01:08:11,861 --> 01:08:17,279
this data and having a lot of
videos of surgery you can kind

1096
01:08:17,279 --> 01:08:22,517
of put the two together. So I
mean all of these body scans

1097
01:08:22,517 --> 01:08:28,116
with a lot of surgery videos and
you can use that maybe to get

1098
01:08:28,116 --> 01:08:33,263
very accurate simulations or
something like you could and

1099
01:08:33,263 --> 01:08:37,959
then you would be less
constrained by messy variable

1100
01:08:37,959 --> 01:08:43,016
real world data for this kind of
thing because you could

1101
01:08:43,016 --> 01:08:48,344
procedurally try to generate a
bunch of it or something. Is

1102
01:08:48,344 --> 01:08:53,311
that kind of does it feel
somewhat correct that there's

1103
01:08:53,311 --> 01:08:58,729
kind of separate things that
will eventually hopefully merge

1104
01:08:58,729 --> 01:08:59,000
or?

1105
01:08:59,000 --> 01:09:02,635
Axel Krieger: Yeah, absolutely.
So, you know, we still need the

1106
01:09:02,635 --> 01:09:06,154
real data to really capture the
surgical intent and technique

1107
01:09:06,154 --> 01:09:08,000
and skills, right? So, that's...

1108
01:09:08,000 --> 01:09:11,685
Audrow Nash: Absolutely. And
also how the body reacts and the

1109
01:09:11,685 --> 01:09:14,946
deformable materials and when
you cut what happens and

1110
01:09:14,946 --> 01:09:17,000
breathing or whatever it might
be.

1111
01:09:17,000 --> 01:09:19,339
Axel Krieger: Yeah, yeah, but we
are limited on how much we can,

1112
01:09:19,339 --> 01:09:21,312
you know, collect there and we
are limited on how many

1113
01:09:21,312 --> 01:09:23,505
variations we're going to see.
always going to be some crazy

1114
01:09:23,505 --> 01:09:25,661
edge cases that we have to deal
with, right? And so, that's

1115
01:09:25,661 --> 01:09:27,890
where, you know, simulation, I
think, comes really into place

1116
01:09:27,890 --> 01:09:28,000
to,

1117
01:09:28,000 --> 01:09:34,000
Audrow Nash: Oh yeah.

1118
01:09:34,000 --> 01:09:38,545
Axel Krieger: You know, have,
you know, a much larger cohort

1119
01:09:38,545 --> 01:09:43,242
and include all the different
variations that you have to, you

1120
01:09:43,242 --> 01:09:47,939
know, deal with and encounter.
So, that's where, you know, the

1121
01:09:47,939 --> 01:09:52,485
two things do come together and
really complement each other

1122
01:09:52,485 --> 01:09:54,000
really, really well.

1123
01:09:54,000 --> 01:09:57,111
Audrow Nash: How far out do you
think that is the coming

1124
01:09:57,111 --> 01:10:00,444
together? Like is it are there a
ton of really hard unsolved

1125
01:10:00,444 --> 01:10:03,833
problems there? Is it a data
limitation? Is it all the things

1126
01:10:03,833 --> 01:10:07,333
like how do you how do you think
about the two coming together?

1127
01:10:07,333 --> 01:10:10,667
Because it seems like that will
be an amazing place to be in

1128
01:10:10,667 --> 01:10:13,833
when that does occur. You'll
just you'll be able to solve

1129
01:10:13,833 --> 01:10:17,056
surgery at least for a lot of
cases and then you can solve

1130
01:10:17,056 --> 01:10:20,444
more from what you can solve
because you're getting more data

1131
01:10:20,444 --> 01:10:23,389
and they're being more widely
used and it'll be a big

1132
01:10:23,389 --> 01:10:26,889
flywheel. But when do you think
that kind of thing occurs or is

1133
01:10:26,889 --> 01:10:30,000
it very researchy and what are
the problems before that?

1134
01:10:30,000 --> 01:10:32,930
Axel Krieger: Yeah, and it
really is going to depend on the

1135
01:10:32,930 --> 01:10:35,662
complexity of the surgery,
right? So, what we discussed

1136
01:10:35,662 --> 01:10:38,791
earlier, the kind of, you know,
bony surgery, we already showed

1137
01:10:38,791 --> 01:10:41,374
that we can train something
purely and it works on a

1138
01:10:41,374 --> 01:10:44,354
cadaver, you know, like that,
you know, segmentation of, you

1139
01:10:44,354 --> 01:10:45,000
know, like...

1140
01:10:45,000 --> 01:10:48,637
Audrow Nash: Do bone surgery.
What are bone surgeries? Like

1141
01:10:48,637 --> 01:10:52,151
what do you mean? Is it putting
in an implant? Is it like

1142
01:10:52,151 --> 01:10:54,000
cutting something off or yeah?

1143
01:10:54,000 --> 01:10:57,625
Axel Krieger: Yeah, exactly,
exactly. Or in evertebrae, if

1144
01:10:57,625 --> 01:11:01,000
you need, like, you know,
injections or, you know, you

1145
01:11:01,000 --> 01:11:02,000
know that, yeah.

1146
01:11:02,000 --> 01:11:05,500
Audrow Nash: Ee. You have to not
be squeamish with all the

1147
01:11:05,500 --> 01:11:09,000
things. With all the years in
surgery robotics. Yeah yeah.

1148
01:11:09,000 --> 01:11:12,714
Axel Krieger: Yeah, so, you
know, the simulating the

1149
01:11:12,714 --> 01:11:17,214
guidance, you know, for these,
you know, procedures under X-ray

1150
01:11:17,214 --> 01:11:21,143
work out of the box seem to
real. So, that's, you know,

1151
01:11:21,143 --> 01:11:23,000
already super, super cool.

1152
01:11:23,000 --> 01:11:26,000
Audrow Nash: That's amazing.

1153
01:11:26,000 --> 01:11:28,882
Axel Krieger: You know, then
this, like, if you add some

1154
01:11:28,882 --> 01:11:31,765
elastic deformation, what I
described as, like, the stem

1155
01:11:31,765 --> 01:11:33,000
deployment or a catheter

1156
01:11:33,000 --> 01:11:36,000
Audrow Nash: Yep. Things where
you're just viewing it.

1157
01:11:36,000 --> 01:11:37,367
Axel Krieger: Going up, you
know, into a vessel, so that is,

1158
01:11:37,367 --> 01:11:38,689
you know, small, you know, a
simulation. That is also, you

1159
01:11:38,689 --> 01:11:40,124
know, feasible to simulate the
physics a little bit harder. And

1160
01:11:40,124 --> 01:11:41,377
we like to combine the real
physics of real experiments

1161
01:11:41,377 --> 01:11:42,744
with, you know, some that have
more autonomous variations in

1162
01:11:42,744 --> 01:11:44,203
synthetic world, but not as, you
know, high fidelity on the, you

1163
01:11:44,203 --> 01:11:45,000
know, physics interaction side.
So,

1164
01:11:45,000 --> 01:12:05,000
Audrow Nash: Simulate.

1165
01:12:05,000 --> 01:12:07,454
Axel Krieger: Combining them
together is really effective

1166
01:12:07,454 --> 01:12:10,123
there. So, that's going to be
the next one is these procedures

1167
01:12:10,123 --> 01:12:12,706
like catheter guidance, like
stem deployment with an elastic

1168
01:12:12,706 --> 01:12:15,247
deformation. That is where the
two, you know, going to come

1169
01:12:15,247 --> 01:12:17,916
together next and where we still
have to show, you know, this,

1170
01:12:17,916 --> 01:12:20,413
but that's going to be the
easier case than the, you know,

1171
01:12:20,413 --> 01:12:23,039
real surgery cutting, you know,
bleeding where the, you know,

1172
01:12:23,039 --> 01:12:25,708
our current way of simulating it
is still not quite there yet,

1173
01:12:25,708 --> 01:12:27,000
you know, like, you know, it's

1174
01:12:27,000 --> 01:12:41,000
Audrow Nash: Mm-hmm. Cutting and
whatnot. Mm-hmm.

1175
01:12:41,000 --> 01:12:44,695
Axel Krieger: Progressing, we're
working it and the company is as

1176
01:12:44,695 --> 01:12:48,220
well. So, you know, it's really,
you know, I think a matter of

1177
01:12:48,220 --> 01:12:51,517
time before it becomes, you
know, strong enough to really,

1178
01:12:51,517 --> 01:12:54,928
you know, mimic it. A lot of
cool work with the world and on

1179
01:12:54,928 --> 01:12:58,623
the world model side, especially
I find and on the material point

1180
01:12:58,623 --> 01:13:01,806
methods also the kind of two
building blocks that you're

1181
01:13:01,806 --> 01:13:03,000
really bullish about.

1182
01:13:03,000 --> 01:13:07,532
Audrow Nash: Hell yeah. Really
cool. Let's see. Going a bit

1183
01:13:07,532 --> 01:13:11,450
bigger picture. How do you
imagine surgery robotics

1184
01:13:11,450 --> 01:13:16,289
proceeding? Like what you guys
are doing? Any other interesting

1185
01:13:16,289 --> 01:13:20,437
companies or research that's
going on? Like how do you

1186
01:13:20,437 --> 01:13:24,739
imagine it progressing like
three years, five years, ten

1187
01:13:24,739 --> 01:13:29,578
years from now? I don't know if
we can even guess really but it

1188
01:13:29,578 --> 01:13:34,187
seems to me like everything
seems to be speeding up in terms

1189
01:13:34,187 --> 01:13:38,489
of progress. I think these
language models and just like

1190
01:13:38,489 --> 01:13:43,251
intelligence being kind of on
tap is making it so advancements

1191
01:13:43,251 --> 01:13:48,014
seem faster. So I might make it
extra hard to predict but what

1192
01:13:48,014 --> 01:13:52,777
do you think? Like or even like
chronologically how the things

1193
01:13:52,777 --> 01:13:56,925
will fall. Like how do you
imagine looking out in this

1194
01:13:56,925 --> 01:13:58,000
general field?

1195
01:13:58,000 --> 01:14:02,367
Axel Krieger: Y yeah, I think
it's gonna really transform. I

1196
01:14:02,367 --> 01:14:06,734
think it's gonna accelerate, I
think it's going to come into

1197
01:14:06,734 --> 01:14:11,320
the clinic from really, really
low risk to high risk. I think a

1198
01:14:11,320 --> 01:14:15,905
lot of data analytics and then
guidance, then decision support,

1199
01:14:15,905 --> 01:14:20,199
and then action. I think you
know that it's already pouring

1200
01:14:20,199 --> 01:14:24,712
in. you know. So you know, so
yeah, If I see where the big you

1201
01:14:24,712 --> 01:14:29,225
know innovations over the next
10-20 years, I think it's going

1202
01:14:29,225 --> 01:14:33,592
to be a bit less on the new
mechanical improvements. I think

1203
01:14:33,592 --> 01:14:37,377
we are a little bit saturating
there and more on the

1204
01:14:37,377 --> 01:14:41,962
intelligence autonomy side. So,
you know, I What we're going to

1205
01:14:41,962 --> 01:14:44,000
see in three, five years are

1206
01:14:44,000 --> 01:14:57,000
Audrow Nash: Probably.

1207
01:14:57,000 --> 01:15:01,166
Axel Krieger: Some you know you
know proof that these autonomous

1208
01:15:01,166 --> 01:15:05,071
subfunctions are helpful in
surgery. Imagine if you have you

1209
01:15:05,071 --> 01:15:08,846
know, a one of the robotic arms
you know s doing the blood

1210
01:15:08,846 --> 01:15:12,556
suction to clear the field
autonomously. So that th Those

1211
01:15:12,556 --> 01:15:16,071
are functions that are
technically feasible in a short

1212
01:15:16,071 --> 01:15:19,000
amount of time and super helpful
to surgeons.

1213
01:15:19,000 --> 01:15:23,375
Audrow Nash: Mm-hmm. So cool.
Mm-hmm. Yeah because if you can

1214
01:15:23,375 --> 01:15:27,749
make it so they don't have to
pay attention to all the things

1215
01:15:27,749 --> 01:15:32,339
and it's like they have a trusty
assistant that just does it and

1216
01:15:32,339 --> 01:15:36,713
doesn't get distracted. Do you
think autonomous cars are kind

1217
01:15:36,713 --> 01:15:41,159
of the good metaphor for this?
You brought it up. It does seem

1218
01:15:41,159 --> 01:15:45,749
like it'll be a slow creep of or
not maybe not slow but it'll be

1219
01:15:45,749 --> 01:15:49,478
a creep of features until the
whole thing is largely

1220
01:15:49,478 --> 01:15:53,637
automated. And then maybe
surgeons will be supervisors for

1221
01:15:53,637 --> 01:15:55,000
this kind of thing.

1222
01:15:55,000 --> 01:15:56,583
Axel Krieger: I think that
analogy is really good. We are

1223
01:15:56,583 --> 01:15:58,250
not going to be you know, fully
manual and then suddenly you

1224
01:15:58,250 --> 01:15:59,972
know, the you know, three years
later it's like you don't need

1225
01:15:59,972 --> 01:16:01,000
to search anymore. That's just
not...

1226
01:16:01,000 --> 01:16:09,027
Audrow Nash: Won't be pow
autonomously. Probably still

1227
01:16:09,027 --> 01:16:12,000
want the surgeon but

1228
01:16:12,000 --> 01:16:13,000
Axel Krieger: Absolutely.

1229
01:16:13,000 --> 01:16:15,032
Audrow Nash: It's like a power
failure or the thing needs to be

1230
01:16:15,032 --> 01:16:16,000
restarted or something stupid.

1231
01:16:16,000 --> 01:16:20,110
Axel Krieger: Yes, no, I mean,
we really, there's no desire to

1232
01:16:20,110 --> 01:16:24,088
replace surgeons that's you know
really making surgeons more

1233
01:16:24,088 --> 01:16:27,602
efficient, effective, and
performing this. We're just

1234
01:16:27,602 --> 01:16:28,000
really

1235
01:16:28,000 --> 01:16:33,000
Audrow Nash: Mm-hmm. I mean
maybe.

1236
01:16:33,000 --> 01:16:36,792
Axel Krieger: Helping with this
crisis. We have fewer and fewer

1237
01:16:36,792 --> 01:16:40,523
surgeons available. We have an
aging society. We have a gap of

1238
01:16:40,523 --> 01:16:44,014
in the next 10 years we are
missing 20,000 surgeons, so we

1239
01:16:44,014 --> 01:16:46,000
need to make them more
efficient.

1240
01:16:46,000 --> 01:16:49,000
Audrow Nash: How many surgeons
are there?

1241
01:16:49,000 --> 01:16:52,000
Axel Krieger: I would have to
look it up, but I

1242
01:16:52,000 --> 01:16:52,656
Audrow Nash: I'm just wondering
if it's like if it's 10% more

1243
01:16:52,656 --> 01:16:53,000
than we have or 20% or 30%. It's

1244
01:16:53,000 --> 01:16:57,069
Axel Krieger: Don't know it.
It's a big amount, but I don't

1245
01:16:57,069 --> 01:16:59,000
know off the top of my hand.

1246
01:16:59,000 --> 01:17:03,276
Audrow Nash: A large amount. I
mean you see the labor shortages

1247
01:17:03,276 --> 01:17:07,551
everywhere. Yeah, yeah. Uh-huh.
What a time. The I wonder. It's

1248
01:17:07,551 --> 01:17:11,420
good. I think it maybe will
parallel programming a little

1249
01:17:11,420 --> 01:17:15,424
bit too in terms of like I mean
like I'm a programmer. I am

1250
01:17:15,424 --> 01:17:19,360
writing a lot of code with
agents now which is really cool

1251
01:17:19,360 --> 01:17:23,296
when I supervise and I don't
like that and I adjust things

1252
01:17:23,296 --> 01:17:27,504
and whatnot. And it's like maybe
eventually. Do you think that

1253
01:17:27,504 --> 01:17:31,780
surgeons will kind of just like
they'll watch the operation and

1254
01:17:31,780 --> 01:17:35,377
they'll get results. Maybe
they'll even do like a few

1255
01:17:35,377 --> 01:17:39,517
surgeries at once because the
robots and AIs are doing a good

1256
01:17:39,517 --> 01:17:43,724
bit of it and they check in and
make sure things look good. Do

1257
01:17:43,724 --> 01:17:48,000
you imagine that would be kind
of similar or what do you think?

1258
01:17:48,000 --> 01:17:52,774
Axel Krieger: Yeah, I had at our
surgical robotics conference

1259
01:17:52,774 --> 01:17:57,391
this summer SRS a nice debate
with Intuitive, with you know

1260
01:17:57,391 --> 01:18:01,852
John Walsh there. It was
interesting where we debated how

1261
01:18:01,852 --> 01:18:06,861
a supervision could look like in
the future, right? And you know

1262
01:18:06,861 --> 01:18:11,243
and so My vision was eventually
you might have a surgeon

1263
01:18:11,243 --> 01:18:15,000
supervising several procedures
at the same time.

1264
01:18:15,000 --> 01:18:17,198
Audrow Nash: Yeah, yeah like a
grocery store checkout model for

1265
01:18:17,198 --> 01:18:18,000
surgeons. What a world.

1266
01:18:18,000 --> 01:18:19,928
Axel Krieger: I think that's
kind of like if you're

1267
01:18:19,928 --> 01:18:22,271
incredibly successful and this
it really, you know, is as good

1268
01:18:22,271 --> 01:18:24,577
as it could be, that's kind of
like the ultimate limit. Where

1269
01:18:24,577 --> 01:18:26,959
you have a surgeon kind of like
you know sitting on a couch and

1270
01:18:26,959 --> 01:18:29,000
watching several procedures
going on at the same time.

1271
01:18:29,000 --> 01:18:34,244
Audrow Nash: Yeah. And like
talking to the AI be like oh on

1272
01:18:34,244 --> 01:18:39,489
screen two you need to be doing
this or that looks a little

1273
01:18:39,489 --> 01:18:45,000
worrying to me in this one. Make
sure we address it. Whatever.

1274
01:18:45,000 --> 01:18:49,678
Axel Krieger: But it's still
super far away and right now I

1275
01:18:49,678 --> 01:18:54,674
think you know like really the
small little functions where you

1276
01:18:54,674 --> 01:18:59,273
have direct supervision and
surgeons can take over. That's

1277
01:18:59,273 --> 01:19:03,000
kind of like the vision for the
next few years.

1278
01:19:03,000 --> 01:19:07,178
Audrow Nash: Mm-hmm. I have a
friend who's a surgeon and what

1279
01:19:07,178 --> 01:19:11,288
he was telling me is that a lot
of the older surgeons do not

1280
01:19:11,288 --> 01:19:15,466
like the robot. Like I think
they use them but they're not as

1281
01:19:15,466 --> 01:19:19,301
proficient with the robots
whereas he is like super good

1282
01:19:19,301 --> 01:19:23,411
with the robots for actually
doing surgery. Really likes it.

1283
01:19:23,411 --> 01:19:27,726
This kind of thing. How do you
imagine this changing over time?

1284
01:19:27,726 --> 01:19:31,699
Like I'm thinking with a lot of
AI it's like technology is

1285
01:19:31,699 --> 01:19:35,192
actually getting more
accessible. Do you think that

1286
01:19:35,192 --> 01:19:38,959
this will help to narrow this
gap where it's like newer

1287
01:19:38,959 --> 01:19:43,068
surgeons embrace it real easily
and kind of older ones may I

1288
01:19:43,068 --> 01:19:47,110
don't know be hesitant to go
down this robotic path or what

1289
01:19:47,110 --> 01:19:48,000
do you think?

1290
01:19:48,000 --> 01:19:49,263
Axel Krieger: Yeah, y yeah, I
mean, you know, surgical robots

1291
01:19:49,263 --> 01:19:50,506
are already in a lot of the you
know disciplines are already

1292
01:19:50,506 --> 01:19:51,728
kind of part of the training
program, right? If you look at

1293
01:19:51,728 --> 01:19:52,846
the urology trainings, the
residents kind of like, you

1294
01:19:52,846 --> 01:19:54,027
know, use it very early on in
the training. So I think of

1295
01:19:54,027 --> 01:19:55,000
course the you know that that's
changing how we

1296
01:19:55,000 --> 01:20:14,000
Audrow Nash: They're everywhere
yeah.

1297
01:20:14,000 --> 01:20:18,120
Axel Krieger: Train you know
surgeons. So yeah, I think,

1298
01:20:18,120 --> 01:20:22,240
yeah, there's kind of like
co-role in helping, you know,

1299
01:20:22,240 --> 01:20:26,875
and learning from the beginning
how it can improve and what are

1300
01:20:26,875 --> 01:20:31,583
the pitfalls. I think it's going
to be also important to include

1301
01:20:31,583 --> 01:20:35,997
that in the training program.
You know, of course, you know,

1302
01:20:35,997 --> 01:20:40,338
becoming a surgeon is super
difficult, you know, like years

1303
01:20:40,338 --> 01:20:41,000
and also,

1304
01:20:41,000 --> 01:20:45,000
Audrow Nash: Takes years. Yeah.
Ten years or something more.

1305
01:20:45,000 --> 01:20:48,689
Axel Krieger: You know, it also
is a combination, right,

1306
01:20:48,689 --> 01:20:52,509
between, you know, all the
cognitive skills, but then also

1307
01:20:52,509 --> 01:20:55,934
all the kind of fine
manipulation skills, right? And

1308
01:20:55,934 --> 01:20:59,359
so maybe if we, you know,
capture some of those fine

1309
01:20:59,359 --> 01:21:02,850
manipulation skills with, you
know, with more robotic

1310
01:21:02,850 --> 01:21:06,671
technology, maybe it opens up
the profession for also, you

1311
01:21:06,671 --> 01:21:07,000
know,

1312
01:21:07,000 --> 01:21:08,000
Audrow Nash: More folks.

1313
01:21:07,000 --> 01:21:07,797
Axel Krieger: Other talents, you
know. So yeah, maybe there's a

1314
01:21:07,797 --> 01:21:08,544
way of, you know, long term
shift, but I, you know, haven't

1315
01:21:08,544 --> 01:21:09,000
really deeply thought about it
or...

1316
01:21:09,000 --> 01:21:13,737
Audrow Nash: Hell yeah. Seems
like it could be. I mean

1317
01:21:13,737 --> 01:21:19,263
especially like I don't know I
wonder what do you think will be

1318
01:21:19,263 --> 01:21:24,439
the way that features will fall
in the space? Like and then

1319
01:21:24,439 --> 01:21:30,140
because I imagine all these will
be very valuable and I mean like

1320
01:21:30,140 --> 01:21:34,877
when will we get autonomous
suturing? When will we get

1321
01:21:34,877 --> 01:21:39,614
autonomous like simple
procedures like an appendectomy

1322
01:21:39,614 --> 01:21:45,316
like cutting off the appendix or
something like that or when will

1323
01:21:45,316 --> 01:21:50,754
like these kinds of things? What
are a few milestones and like

1324
01:21:50,754 --> 01:21:55,930
I'm sure it's very hard to
predict the years and everything

1325
01:21:55,930 --> 01:22:01,368
but like I don't know ballpark
of how you see them in terms of

1326
01:22:01,368 --> 01:22:04,000
difficulty and maybe timeline.

1327
01:22:04,000 --> 01:22:06,611
Axel Krieger: Yeah, I think a
lot of that, you know, depends,

1328
01:22:06,611 --> 01:22:09,265
of course, you know, on kind of
the funding and who's going to

1329
01:22:09,265 --> 01:22:11,833
like, you know, really, you
know, take this on and with, you

1330
01:22:11,833 --> 01:22:14,486
know, how much of an effort a
specific task. So it's like hard

1331
01:22:14,486 --> 01:22:15,000
for me to...

1332
01:22:15,000 --> 01:22:19,251
Audrow Nash: Wait. Mm-hmm. How
is the funding climate for this

1333
01:22:19,251 --> 01:22:23,160
would you think like you
mentioned investors reaching out

1334
01:22:23,160 --> 01:22:27,549
to you they're looking for other
things which is really awesome.

1335
01:22:27,549 --> 01:22:31,663
How and for you all it was like
a pre-seed funding or like I

1336
01:22:31,663 --> 01:22:35,571
know it's a medical related one
so probably a much larger

1337
01:22:35,571 --> 01:22:39,000
pre-seed than most. Was it
pre-seed or seed round?

1338
01:22:39,000 --> 01:22:41,000
Axel Krieger: A seed round, seed
round. Yep.

1339
01:22:41,000 --> 01:22:45,560
Audrow Nash: Okay but for like
Series B appetite and like

1340
01:22:45,560 --> 01:22:50,040
things where they get the
companies become much and much

1341
01:22:50,040 --> 01:22:54,760
more established. How is the
general appetite for investing

1342
01:22:54,760 --> 01:22:59,720
in this like medical technology?
It's probably growing I would

1343
01:22:59,720 --> 01:23:04,040
imagine but what do you think?
What's your read on the

1344
01:23:04,040 --> 01:23:05,000
environment?

1345
01:23:05,000 --> 01:23:06,177
Axel Krieger: Oh, I mean, I
think it's, you know, a

1346
01:23:06,177 --> 01:23:07,608
fantastic environment growing. I
mentioned that there are more

1347
01:23:07,608 --> 01:23:08,877
surgical robotics companies
coming in. So it's a highly

1348
01:23:08,877 --> 01:23:10,331
dynamic, you know, environment.
You know, this kind of surgical

1349
01:23:10,331 --> 01:23:11,577
robotics conference, the SRS
conference, has been like

1350
01:23:11,577 --> 01:23:12,985
seemingly doubling in size every
year. It's kind of crazy how

1351
01:23:12,985 --> 01:23:14,000
it's expanding. I mean, it's
really massive,

1352
01:23:14,000 --> 01:23:32,000
Audrow Nash: Yeah. So cool.

1353
01:23:32,000 --> 01:23:35,115
Axel Krieger: You know. So
there's a lot of, you know,

1354
01:23:35,115 --> 01:23:38,058
appetite for an excitement
around, you know, AI and

1355
01:23:38,058 --> 01:23:41,000
surgery. So, you know, I think
that's definitely...

1356
01:23:41,000 --> 01:23:45,236
Audrow Nash: Mm-hmm. And big
aging population need a lot of

1357
01:23:45,236 --> 01:23:49,615
medical treatments are at the
end of people's lives which big

1358
01:23:49,615 --> 01:23:53,851
aging population so there's
gonna be a lot of diet or a lot

1359
01:23:53,851 --> 01:23:58,159
of appetite for lots and lots of
surgeries probably. And big

1360
01:23:58,159 --> 01:24:02,538
aging population means probably
lots of investment in medical

1361
01:24:02,538 --> 01:24:06,487
stuff. It's probably it's
probably a very good time for

1362
01:24:06,487 --> 01:24:09,000
surgical robotics. Would you
think?

1363
01:24:09,000 --> 01:24:12,575
Axel Krieger: Yeah, wonderful
time. Yeah. And I think, you

1364
01:24:12,575 --> 01:24:16,151
know, our approach is not to
become, you know, the 56, you

1365
01:24:16,151 --> 01:24:19,726
know, vertically integrated
robotic system, you know, but,

1366
01:24:19,726 --> 01:24:23,486
you know, horizontally support,
you know, the AI roadmaps of,

1367
01:24:23,486 --> 01:24:26,877
you know, you know, s all
different, you know, robotics

1368
01:24:26,877 --> 01:24:30,144
surgical companies. So for us,
because we are in this

1369
01:24:30,144 --> 01:24:33,966
supporting role, I cannot give a
prediction how long a certain

1370
01:24:33,966 --> 01:24:37,849
function will take because, you
know, that really is driven by,

1371
01:24:37,849 --> 01:24:41,486
you know, the, you know, you
know a certain robotic company

1372
01:24:41,486 --> 01:24:45,000
like I'm gonna, you know, really
focus on this, you know.

1373
01:24:45,000 --> 01:24:47,000
Audrow Nash: Do this one but
yeah.

1374
01:24:47,000 --> 01:24:48,806
Axel Krieger: That's why it's
hard for me to predict with a

1375
01:24:48,806 --> 01:24:50,000
really exact time scale, you
know. Yep.

1376
01:24:50,000 --> 01:24:51,174
Audrow Nash: With any certainty.
Yeah cuz you don't have control

1377
01:24:51,174 --> 01:24:52,312
of it. Do you have any do you
have any guests though like what

1378
01:24:52,312 --> 01:24:53,431
all the milestones be over the
next few years? Doesn't need a

1379
01:24:53,431 --> 01:24:54,000
timeline but what do you think?

1380
01:24:54,000 --> 01:24:54,763
Axel Krieger: Yep. Yeah. Yeah.
What I mentioned is that, you

1381
01:24:54,763 --> 01:24:55,564
know, autonomous camera control
is already FDA approved, right?

1382
01:24:55,564 --> 01:24:56,352
And so I really feel like in the
next year or two, I would be,

1383
01:24:56,352 --> 01:24:57,025
you know, really shocked if
they're not more of these

1384
01:24:57,025 --> 01:24:57,775
smaller low risk assistive
functions coming in. So what are

1385
01:24:57,775 --> 01:24:58,513
some Targets are, you know, like
some, yeah, yeah, exactly

1386
01:24:58,513 --> 01:24:59,326
soaking up blood. And then, you
know, like the kind of grasping,

1387
01:24:59,326 --> 01:25:00,127
you know, like small changes of
the grasper can also be kind of

1388
01:25:00,127 --> 01:25:00,915
limited in the amount. So like
the active grasping. So it kind

1389
01:25:00,915 --> 01:25:01,614
of like exposes, you know,
different anatomy for you to

1390
01:25:01,614 --> 01:25:02,377
exactly, you know, like holding
the liver up, something like

1391
01:25:02,377 --> 01:25:03,203
that in a, in a gentle way, like
for a certain procedure. So that

1392
01:25:03,203 --> 01:25:03,979
is, I think, you know, coming. I
I'd be shocked if that's not

1393
01:25:03,979 --> 01:25:04,742
coming in the next couple of
years, you know, but, you know,

1394
01:25:04,742 --> 01:25:05,492
after that, you know, full, you
know, suturing anastomosis,

1395
01:25:05,492 --> 01:25:06,000
that's a little bit, you know,
exciting.

1396
01:25:06,000 --> 01:25:25,852
Audrow Nash: Are you guys cool?
Soaking up blood or whatever.

1397
01:25:25,852 --> 01:25:43,751
Yep. Got like peek under
something? Yeah. That would be

1398
01:25:43,751 --> 01:25:59,373
exciting. What's the
esteromosis? What does that

1399
01:25:59,373 --> 01:26:01,000
mean?

1400
01:26:01,000 --> 01:26:01,761
Axel Krieger: And as the most is
the closing of tubular

1401
01:26:01,761 --> 01:26:02,522
structures so that, you know,
like, yeah, yeah, exactly

1402
01:26:02,522 --> 01:26:03,381
suturing off, you know, tubular
structures together, where you

1403
01:26:03,381 --> 01:26:04,266
have to do a like suture all the
way around. If you know, have a

1404
01:26:04,266 --> 01:26:05,000
little gap, that's where you
have a leak. Yeah. Yeah.

1405
01:26:05,000 --> 01:26:10,727
Audrow Nash: Oh like intestines
or something. No good. Gotcha.

1406
01:26:10,727 --> 01:26:16,177
Yeah suturing. I feel like I
don't know anything but I feel

1407
01:26:16,177 --> 01:26:21,904
like suturing will be a big win
when we get there because it's

1408
01:26:21,904 --> 01:26:27,815
so repetitive and I do feel like
quite a lot of errors happen in

1409
01:26:27,815 --> 01:26:28,000
it

1410
01:26:28,000 --> 01:26:32,000
Axel Krieger: Hmm. Absolutely.

1411
01:26:32,000 --> 01:26:35,444
Audrow Nash: Overall. Just it's
so tedious and you can have so

1412
01:26:35,444 --> 01:26:39,000
many required. So I don't know.
That would be great to automate.

1413
01:26:39,000 --> 01:26:41,327
Axel Krieger: Yeah. Absolutely.
I think that's a, you know, a

1414
01:26:41,327 --> 01:26:43,654
very good goal and there, you
know, there are some procedures

1415
01:26:43,654 --> 01:26:46,095
that are a little bit lower risk
like closing off the port after

1416
01:26:46,095 --> 01:26:48,270
you take out that little port,
you have to close it up so

1417
01:26:48,270 --> 01:26:50,520
that's, you know, a low risk,
you know, suturing procedure,

1418
01:26:50,520 --> 01:26:52,809
where would be a really good
starting point, you know, yeah.

1419
01:26:52,809 --> 01:26:53,000
Yeah.

1420
01:26:53,000 --> 01:27:00,500
Audrow Nash: Yes. Yes. Hell
yeah. Awesome. What do you think

1421
01:27:00,500 --> 01:27:08,000
is after suturing? Like what
gets a little harder than that?

1422
01:27:08,000 --> 01:27:10,307
Axel Krieger: I, you know, and
Johns Hopkins we work a lot in

1423
01:27:10,307 --> 01:27:12,727
resections, so, you know, tumor
resections, you know, like, that

1424
01:27:12,727 --> 01:27:14,958
kind of, you know, precision
tumor resection you know where

1425
01:27:14,958 --> 01:27:17,000
you pay attention to not, you
know, hit any, you know,

1426
01:27:17,000 --> 01:27:25,676
Audrow Nash: Oh so removing
something. Arteries or veins or

1427
01:27:25,676 --> 01:27:27,000
whatever.

1428
01:27:27,000 --> 01:27:30,072
Axel Krieger: Vessel or artery
or, you know, any, you know,

1429
01:27:30,072 --> 01:27:33,039
nerve, and, you know, need to
fully remove the tumor. You

1430
01:27:33,039 --> 01:27:35,798
know, it's so hard for surgeons
right you have like a

1431
01:27:35,798 --> 01:27:39,182
preoperative scan like an MRI or
CT, and you mentally kind of map

1432
01:27:39,182 --> 01:27:42,046
this together in your head
together what you see during

1433
01:27:42,046 --> 01:27:45,377
surgery. And so I really think a
robot that, you know, does this

1434
01:27:45,377 --> 01:27:48,553
registration really well learns
how to do these cuts can, you

1435
01:27:48,553 --> 01:27:51,000
know, potentially, you know, do
this, you know,

1436
01:27:51,000 --> 01:27:58,971
Audrow Nash: Feels very well
suited for it. Way better,

1437
01:27:58,971 --> 01:28:01,000
faster. Mmhmm.

1438
01:28:01,000 --> 01:28:03,250
Axel Krieger: You know, a bit, a
bit better. So I think there's

1439
01:28:03,250 --> 01:28:04,000
definitely potential.

1440
01:28:04,000 --> 01:28:07,943
Audrow Nash: Because you can
have it map directly the scan

1441
01:28:07,943 --> 01:28:12,021
over what it sees which that'll
be like a perfect mapping. I

1442
01:28:12,021 --> 01:28:16,032
don't know. I just you hear
things about like the wrong leg

1443
01:28:16,032 --> 01:28:20,042
being amputated or the wrong
like things like this. So it's

1444
01:28:20,042 --> 01:28:24,053
like I know surgeon like those
are freak cases but they are

1445
01:28:24,053 --> 01:28:28,131
busy. They work long shifts.
Surgeries can take a long time.

1446
01:28:28,131 --> 01:28:32,142
I don't know a robot that just
transposes it and is like oh

1447
01:28:32,142 --> 01:28:36,017
we're gonna cut this out like
that could be really really

1448
01:28:36,017 --> 01:28:40,231
good. Hell yeah. And then what's
harder than that? Maybe we'll

1449
01:28:40,231 --> 01:28:44,378
stop at this one but what's the
what do you think is the next

1450
01:28:44,378 --> 01:28:48,592
thing after removing chunks? I
don't know how to phrase things

1451
01:28:48,592 --> 01:28:49,000
right.

1452
01:28:49,000 --> 01:28:50,145
Axel Krieger: I think, you know,
you know, removing the tumor is

1453
01:28:50,145 --> 01:28:51,219
still like one step out of a
full procedure right because if

1454
01:28:51,219 --> 01:28:52,311
you do the procedure right you
have to like plan where you go

1455
01:28:52,311 --> 01:28:53,438
in, then you have to like kind
of dissect to get to the anatomy

1456
01:28:53,438 --> 01:28:54,530
of the tumor. You have to like,
you know, remove all the fat,

1457
01:28:54,530 --> 01:28:55,676
you know, create that exposure,
you know, then, you know, maybe,

1458
01:28:55,676 --> 01:28:56,767
you know, you know, stabilize it
so like it just becomes this

1459
01:28:56,767 --> 01:28:57,000
kind of long,

1460
01:28:57,000 --> 01:29:18,000
Audrow Nash: Yep. Mmhmm.

1461
01:29:18,000 --> 01:29:18,799
Axel Krieger: You know, thing
that all you know needs to kind

1462
01:29:18,799 --> 01:29:19,559
of fit together, you know,
beautifully. So that's, I think

1463
01:29:19,559 --> 01:29:20,319
that, you know, that the most
difficult endpoint, the peak

1464
01:29:20,319 --> 01:29:21,000
point would be to do like full
procedures, you know.

1465
01:29:21,000 --> 01:29:24,908
Audrow Nash: Mmhmm. The peak
thing. That would be so cool.

1466
01:29:24,908 --> 01:29:28,210
Imagine like you just go
somewhere and you get an

1467
01:29:28,210 --> 01:29:32,455
operation you needed this kind
of thing. Like buy an autonomous

1468
01:29:32,455 --> 01:29:36,498
system that'd be so cool. I
wonder if that'll happen. Do you

1469
01:29:36,498 --> 01:29:40,743
think that'll happen in like...
Would you be surprised, I guess

1470
01:29:40,743 --> 01:29:44,247
it's kind of a ridiculous
question, but would you be

1471
01:29:44,247 --> 01:29:48,424
surprised if it wasn't the case
in 20 years from now? That you

1472
01:29:48,424 --> 01:29:52,602
could just like go and have a
robot operate on you? I mean the

1473
01:29:52,602 --> 01:29:56,240
like market side of it to be
determined but like fully

1474
01:29:56,240 --> 01:30:00,351
automated surgeries in 20 years.
Would you be surprised if it

1475
01:30:00,351 --> 01:30:04,326
wasn't the case? It's hard to
predict without technology is

1476
01:30:04,326 --> 01:30:05,000
now but...

1477
01:30:05,000 --> 01:30:07,351
Axel Krieger: Yeah, yeah, I
mean, right now with where, you

1478
01:30:07,351 --> 01:30:09,701
know, looking at the trajectory
and being in it, I would be

1479
01:30:09,701 --> 01:30:11,972
surprised if there are not at
least some procedures where

1480
01:30:11,972 --> 01:30:14,522
that's the case, you know, like
what we discussed before there's

1481
01:30:14,522 --> 01:30:15,000
like a nice,

1482
01:30:15,000 --> 01:30:19,000
Audrow Nash: Better.

1483
01:30:19,000 --> 01:30:20,000
Axel Krieger: You know, like

1484
01:30:20,000 --> 01:30:21,000
Audrow Nash: There's some easier
ones. Mmhmm.

1485
01:30:21,000 --> 01:30:21,791
Axel Krieger: Scale of
complexity, you know. So some of

1486
01:30:21,791 --> 01:30:22,683
the endovascular procedures, you
know, might be the first ones

1487
01:30:22,683 --> 01:30:23,388
and, you know, even with the
kind of surgery like

1488
01:30:23,388 --> 01:30:24,266
cholecystectomy, there's also
different scales of complexity.

1489
01:30:24,266 --> 01:30:25,086
Some are elective where you
don't have much inflammation.

1490
01:30:25,086 --> 01:30:25,921
And so those are easier to
perform than the ones that are,

1491
01:30:25,921 --> 01:30:26,799
you know, where your gallbladder
is all angry and all red and

1492
01:30:26,799 --> 01:30:27,619
swollen up. Those are super
hard, you know, so, you know,

1493
01:30:27,619 --> 01:30:28,482
there's, I think these, you
know, simpler, you know, ones, I

1494
01:30:28,482 --> 01:30:29,000
can certainly see in that
timeframe.

1495
01:30:29,000 --> 01:30:35,960
Audrow Nash: Mmhmm. Yeah.
There's levels. Mmhmm. How would

1496
01:30:35,960 --> 01:30:43,640
you... So you're obviously in an
amazing position in this really

1497
01:30:43,640 --> 01:30:50,600
exciting time for robotic
surgery. How would you... If you

1498
01:30:50,600 --> 01:30:57,680
were someone say you're like
done with college or finishing

1499
01:30:57,680 --> 01:31:05,120
college or something like this
or early in your career and you

1500
01:31:05,120 --> 01:31:12,200
wanted to go in the direction of
surgery robots or surgical

1501
01:31:12,200 --> 01:31:19,520
robots, how... What advice do
you have for them I guess? Like

1502
01:31:19,520 --> 01:31:26,960
how would you start to get on
this? Because it seems like such

1503
01:31:26,960 --> 01:31:29,000
an exciting time.

1504
01:31:29,000 --> 01:31:32,103
Axel Krieger: Yeah, I mean, I
can just speak from experience

1505
01:31:32,103 --> 01:31:35,103
that this is really my passion,
right? I feel in robotics,

1506
01:31:35,103 --> 01:31:38,155
right? You know, working in
healthcare and helping patients

1507
01:31:38,155 --> 01:31:41,414
is just such a, you know, really
cool field to be in, right? So

1508
01:31:41,414 --> 01:31:44,569
if you're kind of like thinking,
I love robots and, you know,

1509
01:31:44,569 --> 01:31:47,672
which field is for You know, I'm
so incredibly happy to have

1510
01:31:47,672 --> 01:31:50,310
landed, you know, in the
medical, surgical, robotic

1511
01:31:50,310 --> 01:31:53,517
space. It's just, you know, so
rewarding. You know, through my

1512
01:31:53,517 --> 01:31:56,776
PhD, I built a robot, you know,
and we, you know, was used on a

1513
01:31:56,776 --> 01:31:59,828
patient and I talked to that
patient afterwards, right? And

1514
01:31:59,828 --> 01:32:03,086
that was kind of a transforming
moment for me, right? Like kind

1515
01:32:03,086 --> 01:32:06,345
of like, you know, so that high
of that, you know, feeling that

1516
01:32:06,345 --> 01:32:09,293
is something where, you know,
it's just not, you know, in

1517
01:32:09,293 --> 01:32:11,000
other fields, you know, the
case.

1518
01:32:11,000 --> 01:32:18,356
Audrow Nash: Whoa. Oh I imagine.
Ha! I mean I like when people

1519
01:32:18,356 --> 01:32:24,288
use my software. That's
completely different. Like

1520
01:32:24,288 --> 01:32:31,051
you've built a robot and
operated on someone. Like that's

1521
01:32:31,051 --> 01:32:32,000
bonkers.

1522
01:32:32,000 --> 01:32:39,403
Axel Krieger: Yes. Yeah, yeah,
that was, you know, incredible.

1523
01:32:39,403 --> 01:32:40,000
So...

1524
01:32:40,000 --> 01:32:42,655
Audrow Nash: I bet you're one of
the few people in the whole

1525
01:32:42,655 --> 01:32:45,000
world that can say something
like that. That's crazy.

1526
01:32:45,000 --> 01:32:48,823
Axel Krieger: Yeah, really,
really lucky. So I, yeah, don't

1527
01:32:48,823 --> 01:32:52,776
take it for granted. And I just
absolutely, you know, love to

1528
01:32:52,776 --> 01:32:56,535
share the passion and why I
think this is so exciting. So,

1529
01:32:56,535 --> 01:33:00,163
yeah, I think it's super
important to find your passion,

1530
01:33:00,163 --> 01:33:03,987
right? And then, you know, if
you want to get into surgical

1531
01:33:03,987 --> 01:33:07,810
robotics, right? Yeah, it's
great because there are several

1532
01:33:07,810 --> 01:33:11,115
entry ways, right? More
clinically, more, you know,

1533
01:33:11,115 --> 01:33:15,068
hardware oriented, more from a,
you know, ML, AI side, right?

1534
01:33:15,068 --> 01:33:18,891
There are so different entry
points into it, right? And so,

1535
01:33:18,891 --> 01:33:23,103
you know, that's, I think, super
exciting. And you have to have a

1536
01:33:23,103 --> 01:33:26,797
bit of this kind of system
mindset, you know, for all the

1537
01:33:26,797 --> 01:33:29,000
different things to work
together.

1538
01:33:29,000 --> 01:33:32,015
Audrow Nash: Hmm. Any advice on
growing that system mindset?

1539
01:33:32,015 --> 01:33:34,879
Because that seems like a hard
thing. Like just mentoring

1540
01:33:34,879 --> 01:33:37,844
students and stuff. It's like
that's a really hard thing to

1541
01:33:37,844 --> 01:33:39,000
teach in my experience.

1542
01:33:39,000 --> 01:33:42,700
Axel Krieger: Yeah, you know, I
think, you know, we, our

1543
01:33:42,700 --> 01:33:46,928
company's right next to Hopkins
Hospital and, you know, with, of

1544
01:33:46,928 --> 01:33:50,429
course, you know, this great
relationship, we can see

1545
01:33:50,429 --> 01:33:54,591
procedures, right? So, you know,
I think that's, you know, just

1546
01:33:54,591 --> 01:33:58,489
so amazing. If you know, can,
you know, really see that how

1547
01:33:58,489 --> 01:34:02,519
this is being used, right? And,
you know, now with, you know,

1548
01:34:02,519 --> 01:34:06,548
YouTube videos, right? You can
also get a really, really good

1549
01:34:06,548 --> 01:34:10,182
feeling for the different
procedures. You can get a lot

1550
01:34:10,182 --> 01:34:14,146
of, you know, access to, you
know, how things are performed.

1551
01:34:14,146 --> 01:34:18,176
So, yeah, it's, you know, dive
deep into it, you know, on the

1552
01:34:18,176 --> 01:34:22,139
use case, you know, understand,
you know, that the language,

1553
01:34:22,139 --> 01:34:25,971
right? So you can talk with
researchers, understand video,

1554
01:34:25,971 --> 01:34:30,067
right? So that's, I think, you
know, a really important aspect

1555
01:34:30,067 --> 01:34:34,163
of it. But you can find your way
into this, you know, more on,

1556
01:34:34,163 --> 01:34:37,796
you know, from a clinical
perspective, more from a, you

1557
01:34:37,796 --> 01:34:41,694
know, device, you know, robot
architecture perspective, or,

1558
01:34:41,694 --> 01:34:45,724
you know, more from a AI, ML,
you know, training perspective.

1559
01:34:45,724 --> 01:34:49,820
But it needs a bit of the three
in combined, in combination to

1560
01:34:49,820 --> 01:34:52,000
be, you know, really e
effective.

1561
01:34:52,000 --> 01:34:59,960
Audrow Nash: Mmhmm. Yes. Yeah
you can't forget one. It just

1562
01:34:59,960 --> 01:35:06,976
doesn't work. A total
non-starter. It's like a stool

1563
01:35:06,976 --> 01:35:13,992
fallover or something. huh.
Let's see. So if someone

1564
01:35:13,992 --> 01:35:20,603
watching this takes away one
thing from our whole

1565
01:35:20,603 --> 01:35:26,000
conversation, what do you want
it to be?

1566
01:35:26,000 --> 01:35:27,260
Axel Krieger: So, I it's
something think less on the, you

1567
01:35:27,260 --> 01:35:28,586
know, s robotic architecture
side. It's really the next, you

1568
01:35:28,586 --> 01:35:29,934
know, decade will be on, you
know, intelligence autonomy. And

1569
01:35:29,934 --> 01:35:30,000
so,

1570
01:35:30,000 --> 01:35:39,000
Audrow Nash: Hardware side?
Mmhmm.

1571
01:35:39,000 --> 01:35:43,538
Axel Krieger: Yeah, that's the,
you know, direction and the

1572
01:35:43,538 --> 01:35:47,692
change, the transformative
change that is happening. I

1573
01:35:47,692 --> 01:35:52,077
think it's, you know, fantastic
for, you know, access for

1574
01:35:52,077 --> 01:35:56,846
patients. And, you know, super
excited to be, you know, a part

1575
01:35:56,846 --> 01:35:58,000
of this change.

1576
01:35:58,000 --> 01:36:00,810
Audrow Nash: Oh yeah. Alright
thank you. It's been a lot of

1577
01:36:00,810 --> 01:36:01,000
fun.

1578
01:36:01,000 --> 01:36:03,438
Axel Krieger: Thank you so much.
Thank you so much for having me.

1579
01:36:03,438 --> 01:36:04,000
It was awesome.

1580
01:36:04,000 --> 01:36:07,000
Audrow Nash: Oh yeah. Alright
bye everyone.
