Latest / Automated with Brian Heater / Rodney Brooks on Physical AI, Robot Safety and Real-World Deployment

Transcript

Rodney Brooks on Physical AI, Robot Safety and Real-World Deployment · Automated with Brian Heater

0:00
SPEAKER_00: If we had dexterist manipulation, it would change everything.
0:03
SPEAKER_00: Are we going to get there anytime soon?
0:06
SPEAKER_00: A lot of work and a lot of wasted capital is going into collecting data sets, which I don't think are going to be useful in the long run.
0:13
SPEAKER_01: You have carved out a very interesting spot for yourself in the industry as the pragmatist, as the voice of reason.
0:20
SPEAKER_01: People are often reaching out to you for sort of a counterpoint.
0:23
SPEAKER_00: The humanoid conferences don't want the counterpoint at the moment.
0:26
SPEAKER_00: That would spoil the fun.
0:28
SPEAKER_00: We're likely to see co-evolutions rather than take the world exactly as is, put this humanoid robot which can do anything because it looks like a human, then everything's gonna be the same.
0:37
SPEAKER_00: It's gonna be much more complex.
0:38
SPEAKER_00: People have fallen in love with end-to-end learning.
0:41
SPEAKER_00: Everyone thinks we'll just get more data and the robots will get better, but they're not as good as the robots that we know how to program the general purpose methods into to do tasks like manipulation.
0:51
SPEAKER_01: It seems like a given that Bidexter's manipulation is going to play a role in the industrial setting.
0:57
SPEAKER_00: Oh, yeah.
0:58
SPEAKER_00: Yeah.
0:58
SPEAKER_01: So I assume that some of the work that you're doing, probably that you can't talk about right now, is sort of continuing to evolve along those lines.
1:05
SPEAKER_00: If I was doing it, I couldn't talk about it.
1:08
SPEAKER_00: Okay.
1:09
SPEAKER_00: Blink twice if you're doing it.
1:12
SPEAKER_00: The wheel devices over the next 50 years will be much more like my current robots across all sorts of industries.
1:19
SPEAKER_00: By the way, I do want to say I'm completely positive about robots.
1:22
SPEAKER_00: I just think people have gotten beyond themselves with hype and promising stuff they're not going to deliver anytime soon.
1:48
SPEAKER_01: Joining us this week is returning champion Rodney Brooks, the co-founder and CTO of Robust AI.
1:56
SPEAKER_01: Rodney was a day one guest at Automated and is now the first returning guest of the show.
2:02
SPEAKER_01: This conversation was recorded in front of a live audience at Circuit Launch, a very cool accelerator space in Mountain View, California.
2:12
SPEAKER_01: Thanks to them.
2:12
SPEAKER_01: Thanks to Silicon Valley Robotics for facilitating the event.
2:16
SPEAKER_01: Thanks to everyone who showed up.
2:18
SPEAKER_01: The waiting list was more than 100 people.
2:20
SPEAKER_01: So apologies to those who weren't able to make it inside.
2:24
SPEAKER_01: The good news is that the following interview is probably more convenient than standing in a packed house.
2:30
SPEAKER_01: The other good news is that the event was uh such an overwhelming success that we are going to be announcing another live episode very soon.
2:37
SPEAKER_01: It's a great conversation, full of the sort of pragmatism that we've come to appreciate from Rodney's many decades in the industry.
2:46
SPEAKER_01: Uh if you are enjoying the episode, if you're enjoying the show, please like and subscribe.
2:50
SPEAKER_01: Both help a lot.
2:51
SPEAKER_01: Tell a friend.
2:52
SPEAKER_01: And don't forget to subscribe to the newsletter over at automated.fm.
2:57
SPEAKER_01: And with that, please enjoy this conversation with Rodney Brooks of Robust AI.
3:03
SPEAKER_01: Hey folks, I wanted to tell you about something that we've been working on here at A3 EditorialHQ, which is actually a series of uh work from home settings decentralized on the Web 3.0 blockchain.
3:16
SPEAKER_01: I'm excited to report that in addition to our recently launched automated.fm news page, we now have a news center for all of A3's editorial content that includes automated, Aaron's new business of automation, all of the podcasts, the market intelligence reports.
3:33
SPEAKER_01: We've grown this team really fast, and frankly, it's useful for us to have all of this information in one place.
3:40
SPEAKER_01: Check it out over at automationnewscenter.com.
3:44
SPEAKER_01: Bookmark the site.
3:45
SPEAKER_01: Uh, we're continuing to roll out new features throughout the end of the year.
3:49
SPEAKER_01: Thank you for reading.
3:51
SPEAKER_01: Do you find that you're getting invited to a lot of these humanoid events these days?
3:56
SPEAKER_00: Um Not so many.
3:58
SPEAKER_00: Not so many?
3:58
SPEAKER_00: Yeah.
3:59
SPEAKER_01: Um I I know, obviously, I know that like right now that at Robust you're not making a humanoid, but it you you have carved out a very interesting spot for yourself in the industry as the pragmatist, as the voice of reason.
4:11
SPEAKER_01: People are are often reaching out to you for sort of a counterpoint.
4:15
SPEAKER_00: Aaron Powell They are l yeah, but the humanoid conferences don't want the counterpoint at the moment.
4:19
SPEAKER_00: That would spoil the fun.
4:21
SPEAKER_00: So um no, I haven't I haven't been to uh any of them, uh nor being invited.
4:26
SPEAKER_00: But I do uh cherish the invitations I do get to to other places.
4:31
SPEAKER_01: Aaron Powell I mean one thing we should say is I mean you are generally speaking or or have been historically, you know, bullish on humanoids or the eventuality of humanoids.
4:39
SPEAKER_00: Trevor Burrus Yeah, I started building uh uh first humanoid that I started building was in 1992 uh at MIT, and we built COG and a bunch of other humanoids.
4:48
SPEAKER_00: And for instance, I know uh you recently talked to Aaron Edsinger, he built a humanoid for his PhD with me.
4:55
SPEAKER_00: So we had a whole fleet of uh humanoids, although we did not do walking.
4:58
SPEAKER_00: Walking was down in the basement, and that was um That was uh Mark Mark Rabert and then Gil Pratt.
5:05
SPEAKER_00: Um and um Jerry Pratt was a student of Gil Pratt, unrelated, but Jerry Pratt is now in the humanoid space.
5:12
SPEAKER_00: And so they were building walking robots.
5:14
SPEAKER_00: Trevor Burrus, Jr.
5:14
SPEAKER_01: Yeah, that was what that was something that uh yeah, so as as you mentioned, we uh visited Aaron uh Edzinger's offices.
5:20
SPEAKER_01: He's in Martinez with with Hello Robot, and yeah, he said, well, you know, we know Boston Dynamics is working on the legs, we figured they'll figure that part out, and we can figure out the upper half of the robot.
5:31
SPEAKER_00: Yeah, and so what we've seen, I think, in the last few years is uh a lot of people coming into humanoids, and they've been struggling with um uh Newton's universal law of gravity and dealing with that, uh things falling down.
5:47
SPEAKER_00: They've gotten better at that, but they've been sort of ignoring Newton's third law of uh of of motion, and that there's an equal and opposite reaction to every force that's applied.
5:58
SPEAKER_00: And so in dexterity, I think they've sort of not they've concentrated on the trajectory of someone's hand or the trajectory of someone's, you know, doing a task and collecting that data.
6:10
SPEAKER_00: But that's not what we do as people.
6:14
SPEAKER_00: The trajectory that our system takes is a result of the forces that are applied backwards when we apply forces.
6:21
SPEAKER_00: And so they've been ignoring force to a large degree, trying to collect lots of data on what it uh what the motion appears to be.
6:29
SPEAKER_00: And I think that's that's uh why the things are a lot slower than than a lot of people thought for getting humanoids to do useful work.
6:38
SPEAKER_01: Is this a problem that to a certain extent can be addressed with uh onboard sensors, force sensors on the other?
6:43
SPEAKER_00: Well, i i uh uh force sensors, touch sensors, so there's a bunch of companies working on those, and there's some some little startups which are really concentrating on force-based manipulation and a lot concentrating on haptic-based uh manipulation.
6:57
SPEAKER_00: I'm advisors to a bunch of small companies doing that.
7:00
SPEAKER_00: So people are working on it, but it we don't it's not just let's collect a lot of video of people doing stuff and then we'll know how to do stuff.
7:07
SPEAKER_00: It's it's the there's actually some physics involved and and some math involved, and it's gonna take a while, a lot longer than I think a lot of people have thought and are thinking.
7:19
SPEAKER_00: You might remember, yeah, Optimus, Optimus started five years ago this August, and it was gonna take over the world in 10 years.
7:28
SPEAKER_01: Are there not 50,000 units being deployed every year?
7:32
SPEAKER_00: No, no, it doesn't seem to be.
7:34
SPEAKER_00: Yeah.
7:35
SPEAKER_01: Yeah.
7:36
SPEAKER_01: Um one one of the things uh that that we were talking about earlier uh that that is uh I would say a solved problem is is navigation, certainly navigation within the warehouse.
7:46
SPEAKER_00: Aaron Powell Oh, within warehouses, yes.
7:48
SPEAKER_00: Visual Slam.
7:49
SPEAKER_00: We've had SLAM for a long time in warehouses using 1D LIDARs, but then you don't see a lot of stuff.
7:56
SPEAKER_00: So vision-based SLAM is now completely practical.
7:59
SPEAKER_00: And uh there are open source libraries you can get.
8:03
SPEAKER_00: Um then there are things that, for instance, my company we do on top of that, which make it even more practical.
8:10
SPEAKER_01: Aaron Ross Powell So I I mean again that that's effectively sort of the primary thing that your robot has to do as far as interacting with the space is is autonomously navigating around.
8:21
SPEAKER_00: Trevor Burrus It's got to know where it is.
8:22
SPEAKER_00: Yeah.
8:22
SPEAKER_00: It's got to deal it's got to deal with not just planning how to get somewhere, but also dealing with dynamic obstacles.
8:28
SPEAKER_00: And there are two primary dynamic obstacles in warehouses.
8:32
SPEAKER_00: There are forklift trucks and there are people.
8:34
SPEAKER_00: Trevor Burrus, Jr.
8:35
SPEAKER_00: Sometimes other robots or sometimes other robots, yeah.
8:38
SPEAKER_00: Um and and especially when you have a fleet, even own robots are around.
8:43
SPEAKER_00: Um but you need to treat those two classes very differently.
8:48
SPEAKER_00: Um uh forklift trucks are dangerous to the robots, and robots should not be dangerous to humans.
8:55
SPEAKER_00: So that it's a an asymmetric um um situation in in each case.
9:01
SPEAKER_01: Aaron Ross Powell We we have all of these companies, and many of them, if not most of them, are are based in Silicon Valley or you know, we there there are, I guess there are several in in uh Pittsburgh out of CMU as well, these physical AI companies that are building these massive models, um largely for general-purpose robots, humanoids.
9:20
SPEAKER_01: Do these models that they're currently compiling, I mean, is is something like that of any use to the system that you currently have deployed?
9:27
SPEAKER_00: We use certain models.
9:29
SPEAKER_00: We use the YOLO models of humans, so we get the full body frame and we know which way they're looking, which way they're what they're doing.
9:37
SPEAKER_00: We can recognize the tasks they're doing, and that informs our robot of what's likely to happen in the next 20 seconds or so, if uh and and and know how to how to interact with those people.
9:50
SPEAKER_00: Aaron Powell But but these but these like giant general purpose models.
9:53
SPEAKER_00: Aaron Powell Giant general purpose models are not really deployed or deployable at this stage.
10:00
SPEAKER_00: The the they're um st but I I just this morning I got uh I got an unsolicited thing saying we can send you lots and lots of data of of uh egocentric data of people doing tasks.
10:12
SPEAKER_00: And uh you need that for your robots.
10:14
SPEAKER_00: Aaron Powell Yeah.
10:15
SPEAKER_01: Well you you you've got you recently uh hit a 100 deployed.
10:19
SPEAKER_00: Yeah, about a month ago we we uh got to a hundred deployed, and it's more than that now.
10:24
SPEAKER_00: It's really picked up.
10:25
SPEAKER_00: So it took us two years to get to a hundred from our first deployment.
10:29
SPEAKER_01: But but you know, obviously you you've had these systems out in the world for some time.
10:32
SPEAKER_01: You've got quite a few of them out there prior to that.
10:35
SPEAKER_01: Um you know, you've got you've got a warehouse space in your offices.
10:39
SPEAKER_01: So you do have robots driving around collecting a a a ton of data.
10:44
SPEAKER_01: But in spite of the fact that that that um, as we said, it's a solved problem.
10:47
SPEAKER_01: Like are they still constantly refining navigation?
10:51
SPEAKER_00: Yeah.
10:51
SPEAKER_00: It turns out that the each warehouse is different in some unique ways.
10:56
SPEAKER_00: There's different stuff on the floor that you have to worry about.
11:00
SPEAKER_00: So we do tune up our models that we have, but they're not they're not these large models.
11:04
SPEAKER_00: They're very specific sorts of models for improving vision, finding the floor, et cetera, or finding people.
11:12
SPEAKER_00: Or then we've got other stuff which is useful data that we collect, which is stuff that we can tell the operators about that they don't necessarily know.
11:21
SPEAKER_00: So we can notice when there's this it's called a slot.
11:23
SPEAKER_00: In a warehouse, there's um aisles, there's bays, there's levels, A, B, C, D, and then the slots where you know a particular product is in a slot.
11:35
SPEAKER_00: And we can detect when those slots are empty, and then tell the warehouse management system we notice it's empty.
11:41
SPEAKER_00: And that may be news to the warehouse management system, so we can help them know about exactly what's happening in the vast spaces.
11:49
SPEAKER_01: The the the other piece of news that came along with the the 100 robot deployment was uh a new CEO at the company, and it seems like the timing between those two things aren't coincidental.
12:01
SPEAKER_01: Obviously, 100, you know, it's a big round number.
12:03
SPEAKER_01: That's the kind of thing that you put in i in a press release.
12:05
SPEAKER_01: But is is there significance to hitting a number like that beyond just you know being able to get some press around it?
12:14
SPEAKER_00: The number per se not.
12:16
SPEAKER_00: Sure.
12:16
SPEAKER_00: Well, you know, what I think is more important is you know, we've got um we're just about to go live with our third deployment that has over 20 robots in it.
12:25
SPEAKER_00: So that's 20 robots in particular places.
12:28
SPEAKER_01: Um is so it's is I mean, of the 100 that are out there, are they kind of clustered in warehouses in a lot of cases?
12:35
SPEAKER_01: So like twelve twenty, that'd be like a a fifth of the Yeah.
12:39
SPEAKER_00: Yeah.
12:39
SPEAKER_00: Yeah.
12:40
SPEAKER_00: And then there's the then the smaller deployments, uh down to four or five robots in a in a a warehouse or uh or a manufacturing site.
12:48
SPEAKER_00: We also operate in manufacturing sites.
12:50
SPEAKER_01: Aaron Powell Yeah.
12:51
SPEAKER_01: How how does the I guess how does the nature of the company change when you scale to that kind of number?
12:58
SPEAKER_00: Aaron Powell Oh, now we have to we have to do things differently.
13:01
SPEAKER_00: Up until now, we've you know flown our engineers out when there's a problem or for deployment.
13:07
SPEAKER_00: We have to get out of the deployment per se, our own people.
13:10
SPEAKER_00: So we're starting to work with other partners around the country who can provide w whose technicians we can train and are local in cities, because if it takes a day for someone to get there when there's a problem, that's not real good.
13:23
SPEAKER_00: And we're not a big company, so we have to have relationships all over the country with people who can go and and uh fix problems when they uh come about.
13:35
SPEAKER_00: Anecdotally, uh um, you know, the biggest problems we see are actually usually on the other side of the warehouse management system getting confused and uh Wi-Fi issues.
13:48
SPEAKER_00: Wi-Fi is just a terrible thing in warehouses.
13:51
SPEAKER_00: And the reason Wi-Fi is so terrible, well, there's two reasons.
13:54
SPEAKER_00: One is um there's a lot of uh metal in the in the racking, and that means there's uh dropout spots all over.
14:03
SPEAKER_00: And the other reason is the local IT guy at each warehouse who doesn't want anything else on his network.
14:11
SPEAKER_00: And so that's a that's a problem that that we have to work through.
14:14
SPEAKER_01: Aaron Powell Well, this is interesting.
14:15
SPEAKER_01: I mean, obviously, again, if we're looking at the potential of like more general purpose robots, like they're they're obviously going to require a to a certain extent a constant connection.
14:25
SPEAKER_00: Um, yeah, well, no, well well we just don't can't get that in a warehouse.
14:29
SPEAKER_00: So our robots may operate for minutes without talking to anything else.
14:32
SPEAKER_00: So they've got to do everything on board.
14:34
SPEAKER_00: And then when they do get some good Wi-Fi, then they lots of messages go back and forth to the warehouse management system to update what's the robots have done, what the people have put on the robot, what the new orders are for the person who's with that robot.
14:47
SPEAKER_00: Trevor Burrus, Jr.
14:48
SPEAKER_01: Yeah, I mean, because the you know, best case scenario I'm hearing from it from a few companies right now is maybe it's like some form of like very small model on-edge processing for these systems if they lose connectivity for a little bit.
14:58
SPEAKER_01: But the thing that um when you and I spoke a number of years ago and you first you know, you're hyping up your hardware before you announce it, you said this is different.
15:07
SPEAKER_01: And and I think the thing that really was distinct about it was, you know, that there was uh passive functionality as well.
15:15
SPEAKER_01: Trevor Burrus, Jr.
15:15
SPEAKER_01: In in our robots.
15:16
SPEAKER_00: Yeah.
15:16
SPEAKER_00: Yeah.
15:17
SPEAKER_00: They're they're passive in that a person can take over at any time and grab the robot and move it and do with it what they want.
15:24
SPEAKER_00: Um which has been interesting to see how different sites have different sort of patterns of behavior.
15:33
SPEAKER_00: Aaron Ross Powell, Jr.
15:33
SPEAKER_00: You called them cultures, which I thought was interesting.
15:35
SPEAKER_00: Yeah, there's sort of a cult, you know, there's a local culture in in different warehouses of how they end up using our robots.
15:40
SPEAKER_00: Where also for the places that we are some some places that we've got our robots deployed, they're air gapped.
15:47
SPEAKER_00: There's no data is allowed in and out of that facility.
15:50
SPEAKER_00: That's not in warehouses, that's in manufacturing.
15:52
SPEAKER_00: Um but where we are allowed to get particular sorts of data, and that varies from place to place what they're willing for us to get.
16:00
SPEAKER_00: We see how individuals who we don't have we don't get any video of them, but we, you know, associate a a badge number, we see how over time they get better at interacting with the robot and how long it takes them.
16:18
SPEAKER_00: And and then we see wide variations in productivities of the human workers uh uh uh uh just from looking at the data.
16:25
SPEAKER_00: So the data is very, very rich.
16:28
SPEAKER_00: Another thing, we have an emergency stop on the robot, and I always say that if anyone ever presses an emergency stop, it means that we have failed.
16:37
SPEAKER_00: Because we should never get to a place where uh a human has to emergency stop the robot.
16:42
SPEAKER_00: But at one place, the emergency stops are being pressed 20, 30, 50 times a day.
16:49
SPEAKER_00: What's going on?
16:50
SPEAKER_00: Well, yeah, it's like it's like force quitting out of your apps instead of doing it the right way.
16:54
SPEAKER_00: Yeah, well, what it what it is is it's when the worker wants to go and do something else for a bit, but doesn't want someone else to steal their robot or doesn't want the robot to get new orders to go somewhere else.
17:05
SPEAKER_00: So they just hit the emergency stop and wander off and do something for 15 minutes and come back and then they know that they can power on the robot.
17:14
SPEAKER_00: Which is sort of annoying to us because we didn't mean it.
17:17
SPEAKER_00: But it does it does tell us that there's something about not wanting to lose their robot that we have to accommodate better, and maybe we have to accommodate it without them having to, you know, use the emergency stop.
17:31
SPEAKER_00: Because that's the tool they saw.
17:32
SPEAKER_00: Yeah, that's what I can use.
17:34
SPEAKER_00: And then once one person starts doing it in a place, they all learn to do that.
17:38
SPEAKER_01: Yeah, I mean, but you know, you I I think as far as I'm aware, any industrial robot needs an emergency stop that's just a lot of things.
17:46
SPEAKER_01: No, no, no.
17:46
SPEAKER_01: I'm saying and and and the uh the the point that I'm getting at is that uh when your emergency stop is uh uh is hit, the robot doesn't collapse to the ground.
17:56
SPEAKER_01: Some other robots that we're talking about now do.
17:58
SPEAKER_00: Oh yeah, yeah.
18:00
SPEAKER_00: It turns out wheels are really good.
18:02
SPEAKER_00: Wheels don't fall down when the when the motors stop turning.
18:07
SPEAKER_01: Again, for from you know, the first time that I visited the offices uh from day one, you had the these uh robots working in this uh sort of like fake warehouse setting in and around your people.
18:21
SPEAKER_01: And I mean everybody I'm sure in in the room knows this, and probably most of the people watching the show know this too, but none of the humanoids that we're seeing right now are working like anywhere near people.
18:32
SPEAKER_01: They're all in cages and behind fences.
18:34
SPEAKER_01: They're not safe to work around in in a large extent.
18:38
SPEAKER_01: Yeah, and uh uh is but safety has been a big piece for the company.
18:44
SPEAKER_00: Aaron Powell Well safety's always been important to me.
18:46
SPEAKER_00: I mean, we we we we de at iRobot, we deployed 55 million robots.
18:51
SPEAKER_00: And um we were rare you know, when we started deploying them, there were no standards for robots in the home.
18:58
SPEAKER_00: So we had to think about the safety case ourselves.
19:01
SPEAKER_00: And one of the things we that we were worried about, it only weighs, you know, uh three or four kilos, but how could it be dangerous?
19:08
SPEAKER_00: Well, if it fell down the stairs and hit a kid downstairs.
19:12
SPEAKER_00: So we were really uh concentrating on making that sure that couldn't happen, and we never have had a reported case of a rumor falling downstairs.
19:20
SPEAKER_00: Because we had triple redundancy, and one of them was a purely passive mechanical system that doesn't require computation.
19:28
SPEAKER_00: So they just didn't fall downstairs.
19:30
SPEAKER_00: So safety's always been really important to me.
19:33
SPEAKER_00: I think you remember Baxter and Sawyer, the upper torso humanoids that we sold to factories around the world.
19:40
SPEAKER_00: My testing for those was to go and put my head in front of the robots and have it hit me with its arm at full velocity.
19:47
SPEAKER_00: And maybe that was that explains a lot.
19:49
SPEAKER_00: Yeah, maybe I was foolish.
19:51
SPEAKER_00: Uh but, you know, the engineers knew that I was gonna do that, and they, you know, it wasn't gonna be good if if uh if that turned out to be uh even painful, let alone head injury.
20:03
SPEAKER_01: So one of the things that I was telling you before as as we were visiting um Aaron Edsinger's office uh offices is uh in the prep to that conversation, we were watching videos of of Domo.
20:15
SPEAKER_01: Domo, that was his robot prescription.
20:18
SPEAKER_00: Trevor Burrus, Jr.
20:19
SPEAKER_00: Yeah, Domo Arigato.
20:21
SPEAKER_00: It means thank you, Mr.
20:22
SPEAKER_00: Robot.
20:22
SPEAKER_00: Yeah.
20:23
SPEAKER_01: What was interesting to me is you know the the first thing you think when you're watching this is uh uh all of the like incredible breakthroughs, everything that's changed in the last 20 years.
20:31
SPEAKER_01: But then you start to watch it and you realize, oh, in a lot of ways, a lot of the fundamentals actually haven't really changed all that much in 20 years.
20:40
SPEAKER_00: Aaron Ross Powell Yeah, although uh you know I I don't want to harp on this too much, but I think people have fallen in love with end-to-end learning, which is something of a misnomer.
20:52
SPEAKER_00: Uh and everyone thinks we'll just give more data and the robots will get better, but they're not as good as the robots that we know how to program the the general purpose methods into, which is what Domo had, to do tasks like manipulation.
21:06
SPEAKER_00: Trevor Burrus, Jr.
21:07
SPEAKER_01: You know, the the kind of the breakthrough that we're talking about for physical AI right now is the fact that you don't have to like hard code.
21:14
unknown: Okay.
21:15
SPEAKER_00: That's the that's the plan.
21:17
SPEAKER_00: That's the plan.
21:20
SPEAKER_00: I'll keep quiet.
21:20
SPEAKER_00: That has not been that has not been shown to be true.
21:24
SPEAKER_00: You know, we we've had three fantastic successes in so-called end-to-end learning.
21:31
SPEAKER_00: We had it with um in 2012 for deep learning, image labeling.
21:35
SPEAKER_00: What's what's in an image?
21:36
SPEAKER_00: And that was for the big training set.
21:38
SPEAKER_00: You know, it was set up uh by Faye Fee at Stanford, and that Image Net was what drove that.
21:44
SPEAKER_00: Second one is uh speech to text.
21:47
SPEAKER_00: That was, you know, 25 years ago I had an office which was all speech controlled, and it was just hell for me to do anything in my office.
21:54
SPEAKER_01: Anyone who's ever owned an Alexa knows that pain.
21:57
SPEAKER_00: Yeah.
21:58
SPEAKER_00: And um so that that That also worked.
22:01
SPEAKER_00: And then LLMs, language.
22:03
SPEAKER_00: Now, in each of those cases, it wasn't pure end-to-end learning.
22:07
SPEAKER_00: We had years and years of technology which had been built to extract the important part of the signal.
22:13
SPEAKER_00: So for speech, you know, ATT, Claude Shannon, after whom today's Claude is named, worried about information theory and how to put lots of voices on a single undersea cable.
22:26
SPEAKER_00: And so we figured out what we needed to extract from the voice signal so that it was still understandable at the other end, but use the minimal number of bits.
22:36
SPEAKER_00: And that's what was the starting front end to speech to text learning systems, where we had a recording of people, and we had a transcription from places like the Canadian Parliament, which also was cross-language, English and French, which was great.
22:56
SPEAKER_00: And then we then we had added a whole lot of other things to it, pre-processing.
23:02
SPEAKER_00: Likewise with vision, we had cameras, and cameras are really tuned to our visual system because when we see an image, uh, even a photograph, our stereo system in our eyes tells us it's flat, but we've figured out how to have enough richness in the image so that uh the rest of our uh brain can reconstruct the depth.
23:24
SPEAKER_00: And and likewise with language, we had you know had written language for all languages and and got minimal stuff.
23:32
SPEAKER_00: And all three of those cases were about providing for humans a remote experience of perceiving things as though it was local and in real time, but it was remote remote in time and place.
23:45
SPEAKER_00: We had those technologies in place.
23:47
SPEAKER_00: For manipulation, we have no technology for remote experience of feeling, touch, or forces.
23:54
SPEAKER_00: We just don't have that technology.
23:56
SPEAKER_00: But people think, oh well, learning worked in those cases, but I'm not I I don't think we've demonstrated that it works in this case with touch and forces yet.
24:05
SPEAKER_00: Watch watch any humanoid video.
24:07
SPEAKER_00: Uh it's really terrible manipulation.
24:11
SPEAKER_01: A little bit of that is is is force touch sensors, uh tactility on the vice itself, um, and is that translating to like teleoperation on the other side as far as Yeah, and people are people are amazing at teleoperation w even without that sensing.
24:26
SPEAKER_00: You can buy you can buy a remote gripper, you know, if you see people picking up stuff off the ground, you know, wrappers and things.
24:33
SPEAKER_00: People can still do that because they feel the forces in their hand remotely, and we're good at using tools.
24:41
SPEAKER_00: Um so people are still using a lot of their sensing, but uh extending it remotely.
24:47
SPEAKER_00: But we have not had a way to record that in an interesting way to replay.
24:51
SPEAKER_01: Trevor Burrus, Jr.
24:51
SPEAKER_01: A lot of the physic the uh the the data that they're talking about right now, or a lot of a lot of sort of the pre-training sets that uh are in discussion right now are uh uh so there are companies trying to like comb the internet for video.
25:03
SPEAKER_00: Trevor Burrus Yeah.
25:04
SPEAKER_00: So so so so one you know, one of the things that's going into uh you know large robot models is and and I've heard about companies being funded to do this, is collecting data from first-person shooter games.
25:18
SPEAKER_00: Okay?
25:19
SPEAKER_00: Oh god.
25:19
SPEAKER_00: So now you're training your robot on how to move around the world, like the person playing the video game who doesn't uh uh care about uh what are they non-r non-role-playing characters?
25:31
SPEAKER_00: NPCs.
25:31
SPEAKER_00: NPCs.
25:32
SPEAKER_00: They don't care about most of the humans here because they're NPCs.
25:35
SPEAKER_01: We were I I just as an aside, earlier today we were we were at a company and they're showing us a robot uh chopping up celery, and he was uh telling us that um there's a reason why they don't use horror movies for training data.
25:48
SPEAKER_00: Yeah.
25:48
SPEAKER_00: Yeah.
25:49
SPEAKER_00: Anyway, so a lot of it is first-person data from from cameras.
25:54
SPEAKER_00: Egocentric.
25:54
SPEAKER_00: Egocentric.
25:55
SPEAKER_00: And John Aliamonis at at University of Maryland got a fantastic increase in capability by doing this amazing thing, changing to object-centered coordinate instead of human-centered coordinates, uh, which is sort of the first obvious thing that one would do if one were programming a robot to do something, it would be about the object coordinates.
26:15
SPEAKER_00: So a lot of work and a lot of wasted capital is going into collecting data sets, which I don't think are going to be useful in the long run.
26:23
SPEAKER_00: Interesting.
26:24
SPEAKER_01: Yeah.
26:24
SPEAKER_01: I one one thing I'm curious about is um, you know, you and I have talked it a number of times over the years, and obviously we um any anyone who's in and around the space was was expecting Gil Pratt called it, I think, the Cambrian explosion in robotics, right?
26:39
SPEAKER_01: The moment when when all of this comes together.
26:41
SPEAKER_01: And and I think I don't know, is it safe to say that we're at least at sort of maybe a beginnings of that, or at least acceleration?
26:46
SPEAKER_00: Yeah, except we've also gone the other way.
26:48
SPEAKER_00: There's been this thought, everything will be a humanoid.
26:51
SPEAKER_00: If we make things that look like a human, they'll be able to do everything a human can do.
26:55
SPEAKER_00: So why would we need any other robot form?
26:58
SPEAKER_00: Um I recently uh at a talk I gave a few weeks ago, I had a picture of me commuting to work on a humanoid robot running down 101.
27:08
SPEAKER_00: It's not the ideal form.
27:09
SPEAKER_00: I think the Waymo is a much better form for for as a robot for carrying people around.
27:15
SPEAKER_01: Yeah, I'm I'm I'm I'm thinking about it from the standpoint of like, you know, something that that you've been thinking about happening for a long time, um, you know, and and the way that it might play out hypothetically and how different it's actually play is playing out right now.
27:31
SPEAKER_01: Is is is it playing out in a like fundamentally different way than you anticipated?
27:35
SPEAKER_00: Aaron Powell Well, uh, it's not clear that it's actually going to play out yet.
27:39
SPEAKER_00: Aaron Powell How do you define playing out?
27:41
SPEAKER_00: Um where we have lots and lots of robots deployed.
27:44
SPEAKER_00: Sure.
27:45
SPEAKER_01: Um because we're not seeing that the deployments are I mean that that's it it that's an inevitability.
27:51
SPEAKER_00: But I think it's more likely to happen with special forms of robots.
27:54
SPEAKER_00: Trevor Burrus, Jr.
27:54
SPEAKER_00: Okay.
27:55
SPEAKER_00: Yes.
27:55
SPEAKER_00: Uh single, multi-purpose robots.
27:57
SPEAKER_00: Trevor Burrus, Jr.: Yeah, maybe maybe some multiple purpose, but not all purpose.
28:01
SPEAKER_00: So for instance, um we have a tremendous pull from aging populations.
28:08
SPEAKER_00: We don't, you know, the the demographic inversion, we can't have the same ratio of caregivers to old people that we've had in the past.
28:15
SPEAKER_00: Trevor Burrus, Jr.
28:15
SPEAKER_00: And this is why Japan was so far ahead of everybody on robotics.
28:19
SPEAKER_00: Exactly.
28:19
SPEAKER_00: One-third of their population is over 65 already.
28:22
SPEAKER_00: But what do caregivers do for humans, old humans?
28:27
SPEAKER_00: They help them stand up, sit down, they help them get into bed, out of bed, they help them to sit on the on the toilet.
28:34
SPEAKER_00: Sometimes they wipe their bottoms.
28:37
SPEAKER_00: We don't have humanoid robots that you want anywhere near an old person at the moment.
28:42
SPEAKER_00: So the the task there is to provide physical support, and using a balancing machine to provide physical support is probably not the best idea.
28:51
SPEAKER_00: In fact, in elder care homes, they have wield things with with winches to help, and so that people help and touch the humans, put wrap things around them, and then use this other winch to get them up because even balancing real humans are not very safe for helping people do these tasks.
29:13
SPEAKER_01: Aaron Powell Is there a path in a sense?
29:16
SPEAKER_01: Is there a path from a warehouse to uh uh you know a uh elder care facility into the home?
29:23
SPEAKER_01: Does that make sense on a fundamental level?
29:26
SPEAKER_00: Aaron Powell Well, uh for the people who don't know about my current robots, the big feature of them is that they're safe to be around.
29:34
SPEAKER_00: You can backdrive them, you can grab the magic handlebar, and when you gr grab the magic handlebar, you have infinite strength over that robot.
29:42
SPEAKER_00: Except, and I always do this demo, if you come to the to the company now, I'll say, grab one of my robots, push it, and you can push it with hardly any any force at all, and hit me with it.
29:54
SPEAKER_00: And no one can hit me with it.
29:56
SPEAKER_00: No one no one can hit me with it because we have vision seeing that I'm a person and it will not hit me.
30:02
SPEAKER_01: Aaron Powell I'm just I don't want to play therapists, but I just feel like I'm starting to notice a trend across your companies.
30:09
SPEAKER_01: Yeah.
30:10
SPEAKER_01: But but but no, but I mean obviously that that's you know, this is this this is so safe that I'm willing to put my body on the line.
30:15
SPEAKER_00: So so so where I was going with that is once you have a robot that that is safe, really safe to be around humans, and it can be around crowd, crowds of humans, lots of people.
30:26
SPEAKER_00: We we now in our new facility, the robots not uh are not only in our warehouse, they just come and and go through around all the cubicles all the time because we want to test them and make sure that they're safe, and multiple robots moving around at all times.
30:41
SPEAKER_00: So once you have that, you can imagine all sorts of places they can be deployed.
30:44
SPEAKER_00: They can be deployed in retail.
30:46
SPEAKER_00: If you if you're in a a large store, you see people pushing carts around to restock the shelves while people are there.
30:53
SPEAKER_00: Well, that could be done by robots in the future.
30:55
SPEAKER_00: Um then um um on the on the uh at at airports on the uh tarmac, you know, there's things moving luggage things around.
31:05
SPEAKER_00: Well, if and you notice when the when they have those those trailer after trailer and the the the person has to loop around to try and get them in a straight line.
31:15
SPEAKER_00: Well, if you just go and grab one of those and move it any way you want, that's uh a much better way of operating.
31:22
SPEAKER_00: So I think that wheeled devices over the next 50 years will be much more like, all of them will be much more like my current robots uh across all sorts of industries.
31:33
SPEAKER_00: We're in Waymo land here.
31:35
SPEAKER_00: The the first Waymo I took was in 2012, just on the freeway right here from X, Google X.
31:43
SPEAKER_00: Um but one of the big frustrations with with Waymo's is when they get stuck, they get stuck and no one can get them to do anything.
31:51
SPEAKER_00: It would be pretty good if you could just go up and grab them and move them out of the way.
31:55
SPEAKER_01: I I don't know if this is still the case, but I I was reading a story not long ago that apparently they have to um hire people on Task Rabbit to physically close the door because a lot of people just get out of the car.
32:06
SPEAKER_00: Yeah, it's it's it's it it's called honk is the app.
32:09
SPEAKER_00: If you want if you want to work for Waymo, get the honk app, and you get paid to go and shut a door when someone has left it open and other things like that.
32:18
SPEAKER_00: And there's cleaning up vomit and all sorts of other.
32:20
SPEAKER_00: So that's a that's a problem that people d don't don't think about.
32:24
SPEAKER_00: They they see a robot doing or they think about a it's you know it's a particular sort of task, but there's all the ancillary tasks that a human does that the robot may not be able to do.
32:35
SPEAKER_01: Trevor Burrus, Jr.
32:35
SPEAKER_01: It seems it strikes me that like in the grand scheme of things, you know, versus uh uh uh building a you know level five autonomous highway capable car that opening and closing an automatic door is probably a relatively easy.
32:48
SPEAKER_00: Well I think as as they build new ones, they're putting that in there.
32:50
SPEAKER_00: But the you know right.
32:52
SPEAKER_01: But these are things that they just didn't even have to do.
32:54
SPEAKER_00: But I back in 2017, I started interviewing um boober drivers.
32:59
SPEAKER_00: I asked them, have you ever had to clean up vomit in the back of your car?
33:02
SPEAKER_00: And they say, oh, every Saturday.
33:04
SPEAKER_00: Uh well, uh that's a problem.
33:07
SPEAKER_01: Well you you were why were you interviewing them?
33:09
SPEAKER_01: What was that for a project?
33:11
SPEAKER_01: No, just when I was riding around.
33:12
SPEAKER_00: So you're just riding around collecting data.
33:14
SPEAKER_01: Yeah.
33:14
SPEAKER_01: Well, no, I specifically because you know, I and and I think that this is really important.
33:18
SPEAKER_01: Um a valuable thing as you know, you're you're a CTO, but you know, if you're the CEO or any like C-suite management level of any of these companies, um it's important to not only talk with customers, but if you can, uh go out to to the warehouse and actually talk with the human beings who are using the robot.
33:39
SPEAKER_00: Yeah, yeah, absolutely.
33:41
SPEAKER_00: Um I've spent a lot of time in warehouses and manufacturing sites.
33:46
SPEAKER_00: Manufacturing sites back from the days with Rethink, but still today and now warehouses.
33:51
SPEAKER_00: They've got the valuable feedback versus like management, right?
33:54
SPEAKER_00: Yeah.
33:54
SPEAKER_00: Yeah.
33:55
SPEAKER_00: Well, you're really when when you're putting a robot in a warehouse or a factory, you're really you've got three customers.
34:00
SPEAKER_00: You've got corporate, you've got the manager of the facility, and you've got the workers.
34:04
SPEAKER_00: And all th if any one of those three doesn't like the robots, they're not going to get deployed.
34:10
SPEAKER_00: So you have to figure out how it's valuable to all of them.
34:14
SPEAKER_01: One of the things I I was thinking about as as somebody who who has a perspective that you have that must be very frustrating, somebody who's been working on um was working on humanoids uh so long ago, um, is watching people uh start with the form factor, abstract it down, pull out the essentials, rebuild that that abstract from scratch, and uh just keep moving in, it seems like in the same cycles over and over again.
34:46
SPEAKER_00: Yeah.
34:46
SPEAKER_00: So one of the things that we discovered at at MIT was that um gaze direction was a cue that people would pick up directly.
34:56
SPEAKER_00: And you see, most of the the humanoid companies now do not have any indication of where they're looking.
35:02
SPEAKER_00: They're looking everywhere.
35:03
SPEAKER_00: There's no uh uh what you you can't understand what they're paying attention to.
35:08
SPEAKER_00: And so that's why at Rethink we had a screen with fake eyes just to sh show what the robot was, you know, because if if I'm gonna reach for this glass, I glance that way first and then I reach.
35:21
SPEAKER_00: Yeah.
35:22
SPEAKER_00: And once we started doing that, we went to factories and with a deployed robot, we'd ask someone two or three days after the robot had got there, what's with the eyes?
35:30
SPEAKER_00: And they'd tell us, oh, it it looks where it's about to reach.
35:33
SPEAKER_00: They they pick that up immediately.
35:35
SPEAKER_00: And the current generation of humanoids is not giving it any indication.
35:39
SPEAKER_00: The Waymo's are.
35:40
SPEAKER_00: The Waymo's have on their top, they they show that there's a person walking ahead.
35:45
SPEAKER_00: So they're giving some feedback.
35:46
SPEAKER_00: I think they're a lot more advanced in giving feedback than any of the humanoid companies.
35:51
SPEAKER_01: Aaron Powell Yeah, well, so so the original, if people remember, the original digit didn't have a head, right?
35:57
SPEAKER_01: You know, it was it was from the Cassie Robot, and the first generation, maybe generation or two, didn't have a head.
36:02
SPEAKER_01: They added a head, and one of the things that they added on to this is agility.
36:05
SPEAKER_01: One of the things that they added onto the head were the eyes, and the reason why they added the eyes on is because the eyes look in the direction it moves in.
36:11
SPEAKER_00: Oh, I hadn't noticed that.
36:12
SPEAKER_00: And that but that but that's a useful, that's a useful.
36:15
SPEAKER_00: Yeah, another another another positive about agility, which I th I think uh s says something about their seriousness, is they don't have five-fingered hands.
36:23
SPEAKER_00: They have a uh a gripper like this, because they're gonna grab totes and things or push their hands together.
36:29
SPEAKER_00: So they they're not claiming they're gonna get they're not m showing a hand that says you'll be able to do anything a human can do.
36:35
SPEAKER_00: Because they they won't.
36:36
SPEAKER_00: They're very concentrated on reality of what is plausible at the moment.
36:40
SPEAKER_01: Aaron Powell What's the distinction between humanoid theater and humanoid deployment theater?
36:45
SPEAKER_00: Aaron Powell Okay, so humanoid humanoid theater is the videos you see of black and white robots in a black and white kitchen, uh opening a dishwasher uh which has four dishes in it, picking up one at a time, and then dropping it at the last.
37:01
SPEAKER_01: So these are like PR2 or like like MIT, CMU, the things that we've been seeing for like 20 years now, like those sorts of no I'm talking about I'm not gonna name the companies.
37:09
SPEAKER_00: I'm talking about companies for humanoid theater.
37:13
SPEAKER_00: Showing that's doing this task, but it's not really doing it.
37:15
SPEAKER_00: Humanoid deployment theater is saying that one robot in a million square foot facility, uh, you know, in a cage with no one around it, is uh is a deployment when it may just be a a a uh uh what it what's the word for it?
37:30
SPEAKER_00: Pilot.
37:31
SPEAKER_00: Pilot.
37:32
SPEAKER_00: And uh and a pilot isn't as demanding as real deployment.
37:36
SPEAKER_00: Yeah.
37:36
SPEAKER_00: Uh we did a first pilot for six weeks.
37:39
SPEAKER_00: We knew it was going to be six weeks, but we did actual picks, we pick tours.
37:43
SPEAKER_00: Um but many of these deployment, so-called deployments are not real deployment.
37:48
SPEAKER_00: I uh let me tell you what I think deployment is about and what I use as my definition.
37:51
SPEAKER_01: Trevor Burrus, Jr.: Which by the way, and and and the the this is, I think, part of the problem is that that you have to have a specific definition for it that, you know, everybody I talk to has a different definition for pilot, and that that's kind of an issue.
38:04
SPEAKER_00: Let me let me say for deployment.
38:05
SPEAKER_01: Trevor Burrus, Jr.: Yeah.
38:05
SPEAKER_01: But but but I'm uh the fact that people have different definitions is not a problem.
38:09
SPEAKER_00: Trevor Burrus, Jr.: They're trying to they're trying to say they've got deployments where it's maybe not a repeatable model yet.
38:14
SPEAKER_00: Uh deployment to me is something where you have to have an organization which can take a phone call at 3 a.m.
38:21
SPEAKER_00: on a Sunday morning if there's a problem, and tell the customer what you're gonna do about it and when.
38:27
SPEAKER_00: So you have to have a that that means that it's important enough to the customer that they need satisfaction at 3 a.m.
38:35
SPEAKER_00: on a Sunday morning.
38:36
SPEAKER_00: Whereas a lesser deployment, a trial run of some sort, that doesn't matter.
38:40
SPEAKER_00: We'll fix it next week.
38:42
SPEAKER_00: Um the second thing is they have to have measurable ROI.
38:46
SPEAKER_00: Yeah.
38:46
SPEAKER_00: And and and then they they like the measurable ROI, and then they decide to buy more, and so you're getting repeat business from the same place, getting more robots.
38:55
SPEAKER_00: And the third thing is you as a company can't be shelling out cash for everyone you ship.
39:01
SPEAKER_00: You you may not be making a profit over all yet, because it takes a while to become a profitable company, but you can't the marginal cost of shipping a robot should not be negative.
39:11
SPEAKER_00: It should be positive.
39:13
SPEAKER_00: And if you've got those three things, then you're on the path to growing a real business.
39:18
SPEAKER_01: Yeah, well, I I think uh one of the distinctions that I would draw, and I don't I don't know if you necessarily agree with this, but I've noticed that um obviously this is not something that a lot of people talk about, but uh pilots are often like over to the side somewhere.
39:34
SPEAKER_01: Exactly.
39:34
SPEAKER_01: And um one of the things that I learned at at A3 is um it's uh it's obvious when I say it, but um, if I'm running a factory, uh the last thing in the world I want is something slowing down the line.
39:47
SPEAKER_01: And and and if that if that's in you know, however long it takes to integrate a system or certainly to test a system, and that that takes a lot of confidence for a company for them to say, okay, let's start actually putting these robots in the real world into our workflow.
40:04
SPEAKER_01: Aaron Powell Yeah.
40:04
SPEAKER_00: And and most of the ones you'll see uh for humanoids today uh not like that at all.
40:10
SPEAKER_00: They're not in the workflow.
40:11
SPEAKER_01: Aaron Powell Would you say that you do demos that robust as demos um as far as um I I don't I don't know this isn't this is another sort of thing that I've been grappling with a little bit lately, is is like how how we would sort of like define it.
40:26
SPEAKER_00: Well when I when we said we had a hundred robots deployed, that meant that we've got a hundred robots that are paying, we're getting paid every month.
40:33
SPEAKER_01: I mean from the standpoint of sort of like a like like a video demo, right?
40:36
SPEAKER_01: Like this is the thing right now with humanoids.
40:39
SPEAKER_01: We're seeing really flashy, um like I call them like the car commercialization of uh of of robots and and uh offering these promises, unrealistic expectations, um you know, these shots that took a hundred takes to get right.
40:56
SPEAKER_01: And um it it's hard not to feel like without more transparency that they're not doing a complete disservice to the industry.
41:06
SPEAKER_00: Aaron Powell I think so.
41:07
SPEAKER_00: And I think uh another place that uh people are getting will get caught up on is some of the companies are saying we're doing deployments into home.
41:15
SPEAKER_00: Oh, and by the way, they're being teleoped remotely.
41:18
SPEAKER_00: So there's some as someone in my company says, some dude somewhere looking around your home operating the robot.
41:26
SPEAKER_01: Yeah, that's it's a teleop conver Yeah, it's a whole nother but that's an interesting conversation as well, is because I was, you know, not long ago I I kind of was of the uh feeling that that that teleop was in a way almost like a failure because it wasn't full autonomy.
41:42
SPEAKER_01: But Teleop does, especially for like post-training and other things.
41:46
SPEAKER_01: Like Teleop will have a role that's important, but there's just a privacy perspective of having a privacy perspective and whether it's whether it's gonna really happen.
41:55
SPEAKER_00: I talked to a CEO, a former CEO of one of the food delivery robot companies uh earlier this year, and he told me that um 7% of the revenue came from delivering the meals, and 93% of the revenue came from the customer who wanted to look techno-futuristic.
42:16
SPEAKER_00: And that's what they're selling.
42:17
SPEAKER_00: They're selling we look at us, we've got future technology.
42:21
SPEAKER_00: It's not actually delivery.
42:23
SPEAKER_00: And and you can go online, and I've I've I've I've found the websites, you can go online to where you can learn how to do training to operate these robots.
42:32
SPEAKER_00: And some companies are very upfront.
42:34
SPEAKER_00: Oh, our robots are all operated by people in Mexico.
42:37
SPEAKER_00: Um as if Mexico doesn't want their own automation and you know it's low-cost labor doing that.
42:45
SPEAKER_00: It's not, but it's presented as a robot doing the task.
42:49
SPEAKER_00: So then someone says, Well, why can't your robot do the stuff that I've seen on humanoid robot theater or in these teleopt systems?
42:59
SPEAKER_00: It's because they're being presented as something different than what they are.
43:03
SPEAKER_01: Trevor Burrus, Jr.
43:03
SPEAKER_00: Yeah.
43:04
SPEAKER_01: Uh Dexterous manipulation this is the thing, like this is like the hard problem.
43:10
SPEAKER_00: Trevor Burrus, Jr.: Dexterous manipulation is is if we had dexterous manipulation, it would change everything.
43:16
SPEAKER_00: Yeah.
43:17
SPEAKER_00: Are we going to get there anytime soon?
43:20
SPEAKER_00: I don't know.
43:21
SPEAKER_00: It's gonna be a complex thing.
43:22
SPEAKER_00: So a a lot of a lot of manipulation of packages that's done is uh with suction cups.
43:30
SPEAKER_00: And what has happened, if you look carefully, over the last 20 years has been a co-evolution of the s material that's put around stuff so that it doesn't get broken by a suction cup sucking too hard so that that technology can move the packages around.
43:47
SPEAKER_00: So we're likely to see coevolutions rather than take the world exactly as is, put this humanoid robot which can do anything because it looks like a human, and then everything's gonna be the same.
43:58
SPEAKER_00: It's gonna be much more complex.
44:00
SPEAKER_01: Yeah.
44:00
SPEAKER_01: I I I mean your your solution is it is a uh a brownfield solution, but i is there a sense in which um uh you know factory or warehouse owners are also kind of adapting the warehouses to your robots?
44:16
SPEAKER_00: Yeah.
44:16
SPEAKER_00: Well we we we actually have incredible data now on where the hotspots are, where there's um too many, too many people moving at once, and we can recommend how to move product around so that things will be much smoother and just more efficient.
44:34
SPEAKER_00: So that's a a great a great thing, I think, for for people having having that data that they can query.
44:41
SPEAKER_00: Aaron Powell You are you're achieving ROI, I assume, with with the systems at this point.
44:45
SPEAKER_00: Oh yeah, the way we the way we sell, actually is we get data from people if they're willing to give it to us, and then we we give them back in a day or two uh a recommendation for how many robots, and we say what it's how that is going to um increase their productivity and what the return on investment will be.
45:06
SPEAKER_00: And we don't get paid until we meet the agreed-upon numbers.
45:10
SPEAKER_01: Aaron Powell Is Is there invariably always gonna be, though, if you're kind of uh like a robotics company uh or a company that's early into the space there there's always gonna be a little bit of the customer being that uh as as you were sort of alluding to around humanoids, like they want to be the first to something, they want to be a part of this exciting technology.
45:30
SPEAKER_00: That's one aspect of it.
45:32
SPEAKER_00: Another aspect right now is um most big companies can't afford to be to not say, oh, we're looking at humanoids.
45:38
SPEAKER_00: Yeah, yeah, we're looking at them.
45:39
SPEAKER_00: Um we we we we we we trialed one uh at uh at our facility in South Carolina.
45:45
SPEAKER_00: Um there's a they don't want to look like they're technically backwards, uh even though they might think that that particular form is not gonna work out for them.
45:54
SPEAKER_00: So there's two aspects.
45:55
SPEAKER_01: Yeah, I'm I'm hearing some kind of alarm, and I'm just gonna take that to mean that we're like close close to time.
46:00
SPEAKER_01: Um so we we can we can wrap up on this, but um by the way, I do want to say I'm completely positive about robots.
46:08
SPEAKER_00: It's a gold it's a golden age for robots.
46:10
SPEAKER_00: It's fantastic.
46:11
SPEAKER_00: The sensors we have, the the onboard compute, the GPUs, it's fantastic.
46:16
SPEAKER_00: I just think people have gotten b beyond themselves with with hype and and uh promising stuff they're not gonna deliver anytime soon.
46:23
SPEAKER_01: Aaron Powell I mean, yeah, I again uh to to bring us back to the beginning, you you know, you are you are bullish on and have been bullish on the humanoid form factor.
46:33
SPEAKER_01: You know, you have these systems, you are ship it's important to say that you are shipping right now.
46:38
SPEAKER_01: And a lot of that is because you haven't overextended yourself.
46:41
SPEAKER_01: You know, you you're you haven't uh you're not one of these companies that that's had to build full stack and and you know it's gonna take take uh however many years.
46:51
SPEAKER_01: But that that said, uh, even though you are pragmatic and you are approaching the market right now, uh, you know, you still must be looking forward to those potential solutions.
47:04
SPEAKER_01: Again, I mean, one of the things I've I've said a lot is it seems like we really skipped a bunch of steps between AMRs and humanoids.
47:11
SPEAKER_01: There's a lot of spots in in the middle.
47:13
SPEAKER_01: It seems like a given that that manipulation, that bidexterous manipulation is going to play a role in the industrial setting.
47:21
SPEAKER_00: Oh, yeah.
47:22
SPEAKER_01: Yeah.
47:22
SPEAKER_01: So so um I assume that some of the work that you're doing, probably that you can't talk about right now, is sort of continuing to evolve along those lines.
47:30
SPEAKER_01: If I was doing it, I couldn't talk about it.
47:33
SPEAKER_01: Okay.
47:34
SPEAKER_01: Blink twice if you're doing it.
47:37
SPEAKER_01: All right, Rod.
47:38
SPEAKER_01: Well, thank you so much for taking time.
47:40
SPEAKER_01: Thank you.
47:41
SPEAKER_01: Thanks, everybody.
47:46
SPEAKER_01: There you go.
47:46
SPEAKER_01: Thank you so much to Rodney Brooks, uh, Robust AI, Andrew Kay at Silicon Valley Robotics, uh Circuit Launch, everyone who showed up.
47:56
SPEAKER_01: Uh thank you so much to them.
47:59
SPEAKER_01: Thank you so much for tuning in.
48:01
SPEAKER_01: If you're enjoying the show, please like and subscribe.
48:04
SPEAKER_01: Both help a lot.
48:05
SPEAKER_01: Please tell a friend.
48:06
SPEAKER_01: And don't forget to subscribe to our newsletter over at automated.fm.
48:10
SPEAKER_01: And we will see you soon for another episode of Automated.