WEBVTT 00:00.000 --> 00:03.160 The following is a conversation with Vijay Kumar. 00:03.160 --> 00:05.800 He's one of the top roboticists in the world, 00:05.800 --> 00:08.760 a professor at the University of Pennsylvania, 00:08.760 --> 00:10.680 a Dean of Penn Engineering, 00:10.680 --> 00:12.880 former director of Grasp Lab, 00:12.880 --> 00:15.320 or the General Robotics Automation Sensing 00:15.320 --> 00:17.560 and Perception Laboratory at Penn, 00:17.560 --> 00:22.560 that was established back in 1979, that's 40 years ago. 00:22.600 --> 00:24.720 Vijay is perhaps best known 00:24.720 --> 00:28.520 for his work in multi robot systems, robot swarms, 00:28.520 --> 00:30.880 and micro aerial vehicles. 00:30.880 --> 00:34.040 Robots that elegantly cooperate in flight 00:34.040 --> 00:36.200 under all the uncertainty and challenges 00:36.200 --> 00:38.760 that the real world conditions present. 00:38.760 --> 00:41.960 This is the Artificial Intelligence Podcast. 00:41.960 --> 00:44.320 If you enjoy it, subscribe on YouTube, 00:44.320 --> 00:46.080 give it five stars on iTunes, 00:46.080 --> 00:47.560 support it on Patreon, 00:47.560 --> 00:49.480 or simply connect with me on Twitter 00:49.480 --> 00:53.280 at Lex Freedman spelled FRID MAN. 00:53.280 --> 00:57.560 And now here's my conversation with Vijay Kumar. 00:58.680 --> 01:01.080 What is the first robot you've ever built 01:01.080 --> 01:02.840 or were a part of building? 01:02.840 --> 01:04.760 Way back when I was in graduate school, 01:04.760 --> 01:06.760 I was part of a fairly big project 01:06.760 --> 01:11.760 that involved building a very large hexapod. 01:12.000 --> 01:17.000 This weighed close to 7,000 pounds. 01:17.520 --> 01:21.600 And it was powered by hydraulic actuation, 01:21.600 --> 01:26.600 or was actuated by hydraulics with 18 motors, 01:27.760 --> 01:32.760 hydraulic motors, each controlled by an Intel 8085 processor 01:34.200 --> 01:36.680 and an 8086 co processor. 01:38.160 --> 01:43.160 And so imagine this huge monster that had 18 joints, 01:44.840 --> 01:47.000 each controlled by an independent computer. 01:47.000 --> 01:48.560 And there was a 19th computer 01:48.560 --> 01:50.160 that actually did the coordination 01:50.160 --> 01:52.360 between these 18 joints. 01:52.360 --> 01:53.760 So as part of this project, 01:53.760 --> 01:57.960 and my thesis work was, 01:57.960 --> 02:01.080 how do you coordinate the 18 legs? 02:02.120 --> 02:06.360 And in particular, the pressures in the hydraulic cylinders 02:06.360 --> 02:09.240 to get efficient locomotion. 02:09.240 --> 02:11.680 It sounds like a giant mess. 02:11.680 --> 02:14.480 So how difficult is it to make all the motors communicate? 02:14.480 --> 02:16.880 Presumably you have to send signals 02:16.880 --> 02:18.720 hundreds of times a second, or at least... 02:18.720 --> 02:22.800 This was not my work, but the folks who worked on this 02:22.800 --> 02:24.200 wrote what I believe to be 02:24.200 --> 02:26.640 the first multiprocessor operating system. 02:26.640 --> 02:27.960 This was in the 80s. 02:29.080 --> 02:32.240 And you had to make sure that obviously messages 02:32.240 --> 02:34.640 got across from one joint to another. 02:34.640 --> 02:37.960 You have to remember the clock speeds on those computers 02:37.960 --> 02:39.640 were about half a megahertz. 02:39.640 --> 02:40.480 Right. 02:40.480 --> 02:42.200 So the 80s. 02:42.200 --> 02:45.320 So not to romanticize the notion, 02:45.320 --> 02:47.960 but how did it make you feel to make, 02:47.960 --> 02:49.680 to see that robot move? 02:51.040 --> 02:52.240 It was amazing. 02:52.240 --> 02:54.160 In hindsight, it looks like, well, 02:54.160 --> 02:57.240 we built the thing which really should have been much smaller. 02:57.240 --> 02:59.080 And of course, today's robots are much smaller. 02:59.080 --> 03:02.200 You look at, you know, Boston Dynamics, 03:02.200 --> 03:04.720 our ghost robotics has been off from pen. 03:06.000 --> 03:08.640 But back then, you were stuck 03:08.640 --> 03:11.120 with the substrate you had, the compute you had, 03:11.120 --> 03:13.640 so things were unnecessarily big. 03:13.640 --> 03:18.000 But at the same time, and this is just human psychology, 03:18.000 --> 03:20.380 somehow bigger means grander. 03:21.280 --> 03:23.600 You know, people never have the same appreciation 03:23.600 --> 03:26.320 for nanotechnology or nano devices 03:26.320 --> 03:30.120 as they do for the space shuttle or the Boeing 747. 03:30.120 --> 03:32.720 Yeah, you've actually done quite a good job 03:32.720 --> 03:35.960 at illustrating that small is beautiful 03:35.960 --> 03:37.680 in terms of robotics. 03:37.680 --> 03:42.520 So what is on that topic is the most beautiful 03:42.520 --> 03:46.200 or elegant robot emotion that you've ever seen. 03:46.200 --> 03:47.880 Not to pick favorites or whatever, 03:47.880 --> 03:51.000 but something that just inspires you that you remember. 03:51.000 --> 03:54.000 Well, I think the thing that I'm most proud of 03:54.000 --> 03:57.200 that my students have done is really think about 03:57.200 --> 04:00.360 small UAVs that can maneuver and constrain spaces 04:00.360 --> 04:03.640 and in particular, their ability to coordinate 04:03.640 --> 04:06.760 with each other and form three dimensional patterns. 04:06.760 --> 04:08.920 So once you can do that, 04:08.920 --> 04:13.920 you can essentially create 3D objects in the sky 04:17.000 --> 04:19.800 and you can deform these objects on the fly. 04:19.800 --> 04:23.520 So in some sense, your toolbox of what you can create 04:23.520 --> 04:25.300 has suddenly got enhanced. 04:27.400 --> 04:29.920 And before that, we did the two dimensional version of this. 04:29.920 --> 04:33.760 So we had ground robots forming patterns and so on. 04:33.760 --> 04:37.080 So that was not as impressive, that was not as beautiful. 04:37.080 --> 04:40.480 But if you do it in 3D, suspend it in midair 04:40.480 --> 04:43.640 and you've got to go back to 2011 when we did this. 04:43.640 --> 04:45.960 Now it's actually pretty standard to do these things 04:45.960 --> 04:49.800 eight years later, but back then it was a big accomplishment. 04:49.800 --> 04:52.440 So the distributed cooperation 04:52.440 --> 04:55.640 is where beauty emerges in your eyes? 04:55.640 --> 04:57.960 Well, I think beauty to an engineer is very different 04:57.960 --> 05:01.520 from beauty to someone who's looking at robots 05:01.520 --> 05:03.400 from the outside, if you will. 05:03.400 --> 05:07.920 But what I meant there, so before we said that grand 05:07.920 --> 05:10.480 is associated with size. 05:10.480 --> 05:13.640 And another way of thinking about this 05:13.640 --> 05:16.480 is just the physical shape and the idea 05:16.480 --> 05:18.320 that you can create physical shapes in midair 05:18.320 --> 05:21.520 and have them deform, that's beautiful. 05:21.520 --> 05:23.000 But the individual components, 05:23.000 --> 05:24.840 the agility is beautiful too, right? 05:24.840 --> 05:25.680 That is true too. 05:25.680 --> 05:28.400 So then how quickly can you actually manipulate 05:28.400 --> 05:29.560 these three dimensional shapes 05:29.560 --> 05:31.200 and the individual components? 05:31.200 --> 05:32.200 Yes, you're right. 05:32.200 --> 05:36.760 Oh, by the way, said UAV, unmanned aerial vehicle. 05:36.760 --> 05:41.760 What's a good term for drones, UAVs, quadcopters? 05:41.840 --> 05:44.520 Is there a term that's being standardized? 05:44.520 --> 05:45.440 I don't know if there is. 05:45.440 --> 05:47.880 Everybody wants to use the word drones. 05:47.880 --> 05:49.760 And I've often said there's drones to me 05:49.760 --> 05:51.000 is a pejorative word. 05:51.000 --> 05:53.960 It signifies something that's dumb, 05:53.960 --> 05:56.320 a pre program that does one little thing 05:56.320 --> 05:58.600 and robots are anything but drones. 05:58.600 --> 06:00.680 So I actually don't like that word, 06:00.680 --> 06:02.960 but that's what everybody uses. 06:02.960 --> 06:04.880 You could call it unpiloted. 06:04.880 --> 06:05.800 Unpiloted. 06:05.800 --> 06:08.080 But even unpiloted could be radio controlled, 06:08.080 --> 06:10.480 could be remotely controlled in many different ways. 06:11.560 --> 06:12.680 And I think the right word 06:12.680 --> 06:15.040 is thinking about it as an aerial robot. 06:15.040 --> 06:19.080 You also say agile, autonomous aerial robot, right? 06:19.080 --> 06:20.600 Yeah, so agility is an attribute, 06:20.600 --> 06:22.200 but they don't have to be. 06:23.080 --> 06:24.800 So what biological system, 06:24.800 --> 06:26.880 because you've also drawn a lot of inspiration 06:26.880 --> 06:28.640 with those I've seen bees and ants 06:28.640 --> 06:32.400 that you've talked about, what living creatures 06:32.400 --> 06:35.280 have you found to be most inspiring 06:35.280 --> 06:38.560 as an engineer, instructive in your work in robotics? 06:38.560 --> 06:43.480 To me, so ants are really quite incredible creatures, right? 06:43.480 --> 06:47.920 So you, I mean, the individuals arguably are very simple 06:47.920 --> 06:52.400 in how they're built, and yet they're incredibly resilient 06:52.400 --> 06:54.000 as a population. 06:54.000 --> 06:56.800 And as individuals, they're incredibly robust. 06:56.800 --> 07:01.800 So, if you take an ant with six legs, you remove one leg, 07:02.080 --> 07:04.160 it still works just fine. 07:04.160 --> 07:05.800 And it moves along, 07:05.800 --> 07:08.800 and I don't know that he even realizes it's lost a leg. 07:09.800 --> 07:13.480 So that's the robustness at the individual ant level. 07:13.480 --> 07:15.400 But then you look about this instinct 07:15.400 --> 07:17.760 for self preservation of the colonies, 07:17.760 --> 07:20.440 and they adapt in so many amazing ways, 07:20.440 --> 07:25.440 you know, transcending gaps by just chaining themselves 07:27.680 --> 07:30.400 together when you have a flood, 07:30.400 --> 07:33.160 being able to recruit other teammates 07:33.160 --> 07:35.200 to carry big morsels of food, 07:36.560 --> 07:38.240 and then going out in different directions, 07:38.240 --> 07:39.560 looking for food, 07:39.560 --> 07:43.040 and then being able to demonstrate consensus, 07:43.880 --> 07:47.240 even though they don't communicate directly 07:47.240 --> 07:50.400 with each other the way we communicate with each other, 07:50.400 --> 07:53.840 in some sense, they also know how to do democracy, 07:53.840 --> 07:55.480 probably better than what we do. 07:55.480 --> 07:59.000 Yeah, somehow, even democracy is emergent. 07:59.000 --> 08:00.640 It seems like all of the phenomena 08:00.640 --> 08:02.480 that we see is all emergent. 08:02.480 --> 08:05.600 It seems like there's no centralized communicator. 08:05.600 --> 08:09.840 There is, so I think a lot is made about that word, emergent, 08:09.840 --> 08:11.560 and it means lots of things to different people, 08:11.560 --> 08:12.600 but you're absolutely right. 08:12.600 --> 08:17.600 I think as an engineer, you think about what element, 08:17.600 --> 08:22.600 elemental behaviors, what primitives you could synthesize 08:22.720 --> 08:26.640 so that the whole looks incredibly powerful, 08:26.640 --> 08:27.920 incredibly synergistic, 08:27.920 --> 08:30.960 the whole definitely being greater than some of the parts, 08:30.960 --> 08:33.760 and ants are living proof of that. 08:33.760 --> 08:36.280 So when you see these beautiful swarms 08:36.280 --> 08:38.800 where there's biological systems of robots, 08:39.920 --> 08:41.560 do you sometimes think of them 08:41.560 --> 08:45.920 as a single individual living intelligent organism? 08:45.920 --> 08:49.400 So it's the same as thinking of our human civilization 08:49.400 --> 08:52.920 as one organism, or do you still, as an engineer, 08:52.920 --> 08:54.560 think about the individual components 08:54.560 --> 08:55.840 and all the engineering that went into 08:55.840 --> 08:57.280 the individual components? 08:57.280 --> 08:58.600 Well, that's very interesting. 08:58.600 --> 09:01.440 So again, philosophically, as engineers, 09:01.440 --> 09:06.440 what we want to do is to go beyond the individual components, 09:06.800 --> 09:10.240 the individual units, and think about it as a unit, 09:10.240 --> 09:11.480 as a cohesive unit, 09:11.480 --> 09:14.240 without worrying about the individual components. 09:14.240 --> 09:19.240 If you start obsessing about the individual building blocks 09:19.680 --> 09:24.680 and what they do, you inevitably will find it hard 09:26.400 --> 09:27.920 to scale up. 09:27.920 --> 09:28.960 Just mathematically, 09:28.960 --> 09:31.560 just think about individual things you want to model, 09:31.560 --> 09:34.000 and if you want to have 10 of those, 09:34.000 --> 09:36.400 then you essentially are taking Cartesian products 09:36.400 --> 09:39.280 of 10 things, and that makes it really complicated 09:39.280 --> 09:41.800 than to do any kind of synthesis or design 09:41.800 --> 09:44.160 in that high dimension space is really hard. 09:44.160 --> 09:45.840 So the right way to do this 09:45.840 --> 09:49.080 is to think about the individuals in a clever way 09:49.080 --> 09:51.160 so that at the higher level, 09:51.160 --> 09:53.400 when you look at lots and lots of them, 09:53.400 --> 09:55.320 abstractly, you can think of them 09:55.320 --> 09:57.120 in some low dimensional space. 09:57.120 --> 09:58.680 So what does that involve? 09:58.680 --> 10:02.160 For the individual, you have to try to make 10:02.160 --> 10:05.160 the way they see the world as local as possible, 10:05.160 --> 10:06.440 and the other thing, 10:06.440 --> 10:09.560 do you just have to make them robust to collisions? 10:09.560 --> 10:10.880 Like you said, with the ants, 10:10.880 --> 10:15.320 if something fails, the whole swarm doesn't fail. 10:15.320 --> 10:17.760 Right, I think as engineers, we do this. 10:17.760 --> 10:18.840 I mean, you know, think about, 10:18.840 --> 10:21.280 we build planes or we build iPhones, 10:22.240 --> 10:26.280 and we know that by taking individual components, 10:26.280 --> 10:27.600 well engineered components, 10:27.600 --> 10:30.080 with well specified interfaces 10:30.080 --> 10:31.680 that behave in a predictable way, 10:31.680 --> 10:33.560 you can build complex systems. 10:34.400 --> 10:36.840 So that's ingrained I would claim 10:36.840 --> 10:39.400 in most engineers thinking, 10:39.400 --> 10:41.600 and it's true for computer scientists as well. 10:41.600 --> 10:44.760 I think what's different here is that you want 10:44.760 --> 10:49.480 the individuals to be robust in some sense, 10:49.480 --> 10:52.000 as we do in these other settings, 10:52.000 --> 10:54.480 but you also want some degree of resiliency 10:54.480 --> 10:56.320 for the population. 10:56.320 --> 10:58.720 And so you really want them to be able 10:58.720 --> 11:03.720 to reestablish communication with their neighbors. 11:03.840 --> 11:07.320 You want them to rethink their strategy 11:07.320 --> 11:09.040 for group behavior. 11:09.040 --> 11:11.000 You want them to reorganize. 11:12.440 --> 11:16.120 And that's where I think a lot of the challenges lie. 11:16.120 --> 11:18.400 So just at a high level, 11:18.400 --> 11:21.120 what does it take for a bunch of, 11:22.400 --> 11:23.560 what should we call them, 11:23.560 --> 11:26.920 flying robots to create a formation? 11:26.920 --> 11:30.400 Just for people who are not familiar with robotics 11:30.400 --> 11:33.000 in general, how much information is needed? 11:33.000 --> 11:36.040 How do you even make it happen 11:36.040 --> 11:39.720 without a centralized controller? 11:39.720 --> 11:41.320 So I mean, there are a couple of different ways 11:41.320 --> 11:43.400 of looking at this. 11:43.400 --> 11:48.400 If you are a purist, you think of it as a way 11:50.040 --> 11:52.160 of recreating what nature does. 11:53.800 --> 11:58.680 So nature forms groups for several reasons, 11:58.680 --> 12:02.200 but mostly it's because of this instinct 12:02.200 --> 12:07.200 that organisms have of preserving their colonies, 12:07.280 --> 12:11.200 their population, which means what? 12:11.200 --> 12:14.640 You need shelter, you need food, you need to procreate, 12:14.640 --> 12:16.480 and that's basically it. 12:16.480 --> 12:20.120 So the kinds of interactions you see are all organic. 12:20.120 --> 12:21.320 They're all local. 12:22.320 --> 12:25.760 And the only information that they share, 12:25.760 --> 12:27.800 and mostly it's indirectly, 12:27.800 --> 12:31.040 is to again preserve the herd or the flock 12:31.040 --> 12:36.040 or the swarm and either by looking for new sources of food 12:39.400 --> 12:41.240 or looking for new shelters, right? 12:42.960 --> 12:47.200 As engineers, when we build swarms, we have a mission. 12:48.280 --> 12:50.760 And when you think of a mission, 12:52.080 --> 12:54.360 and it involves mobility, 12:54.360 --> 12:56.840 most often it's described in some kind 12:56.840 --> 12:58.800 of a global coordinate system. 12:58.800 --> 13:03.080 As a human, as an operator, as a commander, 13:03.080 --> 13:07.120 or as a collaborator, I have my coordinate system 13:07.120 --> 13:10.160 and I want the robots to be consistent with that. 13:11.120 --> 13:14.720 So I might think of it slightly differently. 13:14.720 --> 13:18.960 I might want the robots to recognize that coordinate system, 13:18.960 --> 13:21.360 which means not only do they have to think locally 13:21.360 --> 13:23.160 in terms of who their immediate neighbors are, 13:23.160 --> 13:24.640 but they have to be cognizant 13:24.640 --> 13:28.320 of what the global environment looks like. 13:28.320 --> 13:31.080 So if I go, if I say surround this building 13:31.080 --> 13:33.280 and protect this from intruders, 13:33.280 --> 13:35.160 well, they're immediately in a building 13:35.160 --> 13:36.520 centered coordinate system 13:36.520 --> 13:38.720 and I have to tell them where the building is. 13:38.720 --> 13:40.080 And they're globally collaborating 13:40.080 --> 13:41.360 on the map of that building. 13:41.360 --> 13:44.240 They're maintaining some kind of global, 13:44.240 --> 13:45.560 not just in the frame of the building, 13:45.560 --> 13:49.040 but there's information that's ultimately being built up 13:49.040 --> 13:53.320 explicitly as opposed to kind of implicitly, 13:53.320 --> 13:54.400 like nature might. 13:54.400 --> 13:55.240 Correct, correct. 13:55.240 --> 13:57.720 So in some sense, nature is very, very sophisticated, 13:57.720 --> 14:00.480 but the tasks that nature solves 14:00.480 --> 14:03.040 or needs to solve are very different 14:03.040 --> 14:05.160 from the kind of engineered tasks, 14:05.160 --> 14:09.800 artificial tasks that we are forced to address. 14:09.800 --> 14:12.560 And again, there's nothing preventing us 14:12.560 --> 14:15.200 from solving these other problems, 14:15.200 --> 14:16.640 but ultimately it's about impact. 14:16.640 --> 14:19.400 You want these swarms to do something useful. 14:19.400 --> 14:24.400 And so you're kind of driven into this very unnatural, 14:24.400 --> 14:27.840 if you will, unnatural meaning, not like how nature does, 14:27.840 --> 14:29.000 setting. 14:29.000 --> 14:31.720 And it's probably a little bit more expensive 14:31.720 --> 14:33.560 to do it the way nature does, 14:33.560 --> 14:38.560 because nature is less sensitive to the loss of the individual 14:39.280 --> 14:42.080 and cost wise in robotics, 14:42.080 --> 14:45.280 I think you're more sensitive to losing individuals. 14:45.280 --> 14:48.800 I think that's true, although if you look at the price 14:48.800 --> 14:51.320 to performance ratio of robotic components, 14:51.320 --> 14:53.640 it's coming down dramatically. 14:53.640 --> 14:54.480 I'm interested. 14:54.480 --> 14:56.040 Right, it continues to come down. 14:56.040 --> 14:58.920 So I think we're asymptotically approaching the point 14:58.920 --> 14:59.960 where we would get, yeah, 14:59.960 --> 15:05.080 the cost of individuals would really become insignificant. 15:05.080 --> 15:07.600 So let's step back at a high level view, 15:07.600 --> 15:11.680 the impossible question of what kind of, 15:11.680 --> 15:14.400 as an overview, what kind of autonomous flying vehicles 15:14.400 --> 15:16.200 are there in general? 15:16.200 --> 15:19.720 I think the ones that receive a lot of notoriety 15:19.720 --> 15:22.560 are obviously the military vehicles. 15:22.560 --> 15:26.280 Military vehicles are controlled by a base station, 15:26.280 --> 15:29.640 but have a lot of human supervision, 15:29.640 --> 15:31.800 but have limited autonomy, 15:31.800 --> 15:34.760 which is the ability to go from point A to point B, 15:34.760 --> 15:38.320 and even the more sophisticated vehicles 15:38.320 --> 15:41.760 can do autonomous takeoff and landing. 15:41.760 --> 15:44.400 And those usually have wings and they're heavy? 15:44.400 --> 15:45.360 Usually they're wings, 15:45.360 --> 15:47.440 but there's nothing preventing us from doing this 15:47.440 --> 15:49.000 for helicopters as well. 15:49.000 --> 15:53.440 There are many military organizations that have 15:53.440 --> 15:56.560 autonomous helicopters in the same vein. 15:56.560 --> 16:00.080 And by the way, you look at autopilots and airplanes, 16:00.080 --> 16:02.800 and it's actually very similar. 16:02.800 --> 16:07.160 In fact, one interesting question we can ask is, 16:07.160 --> 16:12.120 if you look at all the air safety violations, 16:12.120 --> 16:14.080 all the crashes that occurred, 16:14.080 --> 16:16.880 would they have happened if the plane 16:16.880 --> 16:20.200 were truly autonomous, and I think you'll find 16:20.200 --> 16:21.960 that in many of the cases, 16:21.960 --> 16:24.600 because of pilot error, we made silly decisions. 16:24.600 --> 16:26.920 And so in some sense, even in air traffic, 16:26.920 --> 16:29.760 commercial air traffic, there's a lot of applications, 16:29.760 --> 16:33.920 although we only see autonomy being enabled 16:33.920 --> 16:38.920 at very high altitudes when the plane is an autopilot. 16:41.160 --> 16:42.520 There's still a role for the human, 16:42.520 --> 16:47.520 and that kind of autonomy is, you're kind of implying, 16:47.640 --> 16:48.680 I don't know what the right word is, 16:48.680 --> 16:52.600 but it's a little dumber than it could be. 16:53.480 --> 16:55.720 Right, so in the lab, of course, 16:55.720 --> 16:59.200 we can afford to be a lot more aggressive. 16:59.200 --> 17:04.200 And the question we try to ask is, 17:04.600 --> 17:09.600 can we make robots that will be able to make decisions 17:09.600 --> 17:13.680 without any kind of external infrastructure? 17:13.680 --> 17:14.880 So what does that mean? 17:14.880 --> 17:16.960 So the most common piece of infrastructure 17:16.960 --> 17:19.640 that airplanes use today is GPS. 17:20.560 --> 17:25.160 GPS is also the most brittle form of information. 17:26.680 --> 17:30.480 If you've driven in a city, try to use GPS navigation, 17:30.480 --> 17:32.760 tall buildings, you immediately lose GPS. 17:33.720 --> 17:36.320 And so that's not a very sophisticated way 17:36.320 --> 17:37.880 of building autonomy. 17:37.880 --> 17:39.560 I think the second piece of infrastructure 17:39.560 --> 17:41.920 that I rely on is communications. 17:41.920 --> 17:46.200 Again, it's very easy to jam communications. 17:47.400 --> 17:49.680 In fact, if you use Wi Fi, 17:49.680 --> 17:51.880 you know that Wi Fi signals drop out, 17:51.880 --> 17:53.560 cell signals drop out. 17:53.560 --> 17:56.840 So to rely on something like that is not good. 17:58.600 --> 18:01.240 The third form of infrastructure we use, 18:01.240 --> 18:02.960 and I hate to call it infrastructure, 18:02.960 --> 18:06.400 but it is that in the sense of robots, it's people. 18:06.400 --> 18:08.760 So you could rely on somebody to pilot you. 18:08.760 --> 18:09.960 Right. 18:09.960 --> 18:11.600 And so the question you wanna ask is 18:11.600 --> 18:13.400 if there are no pilots, 18:13.400 --> 18:16.200 if there's no communications with any base station, 18:16.200 --> 18:18.720 if there's no knowledge of position, 18:18.720 --> 18:21.640 and if there's no a priori map, 18:21.640 --> 18:24.880 a priori knowledge of what the environment looks like, 18:24.880 --> 18:28.280 a priori model of what might happen in the future. 18:28.280 --> 18:29.560 Can robots navigate? 18:29.560 --> 18:31.440 So that is true autonomy. 18:31.440 --> 18:33.240 So that's true autonomy. 18:33.240 --> 18:35.040 And we're talking about, you mentioned 18:35.040 --> 18:36.880 like military applications and drones. 18:36.880 --> 18:38.280 Okay, so what else is there? 18:38.280 --> 18:43.280 You talk about agile autonomous flying robots, aerial robots. 18:43.480 --> 18:46.320 So that's a different kind of, it's not winged, 18:46.320 --> 18:48.120 it's not big, at least it's small. 18:48.120 --> 18:50.800 So I use the word agility mostly, 18:50.800 --> 18:53.480 or at least we're motivated to do agile robots, 18:53.480 --> 18:57.960 mostly because robots can operate 18:57.960 --> 19:01.120 and should be operating in constrained environments. 19:02.120 --> 19:06.960 And if you want to operate the way a global hawk operates, 19:06.960 --> 19:09.120 I mean, the kinds of conditions in which you operate 19:09.120 --> 19:10.760 are very, very restrictive. 19:11.760 --> 19:13.720 If you wanna go inside a building, 19:13.720 --> 19:15.600 for example, for search and rescue, 19:15.600 --> 19:18.120 or to locate an active shooter, 19:18.120 --> 19:22.120 or you wanna navigate under the canopy in an orchard 19:22.120 --> 19:23.880 to look at health of plants, 19:23.880 --> 19:28.880 or to count fruits to measure the tree trunks. 19:31.240 --> 19:33.240 These are things we do, by the way. 19:33.240 --> 19:35.920 Yeah, some cool agriculture stuff you've shown in the past, 19:35.920 --> 19:36.760 it's really awesome. 19:36.760 --> 19:40.400 So in those kinds of settings, you do need that agility. 19:40.400 --> 19:42.560 Agility does not necessarily mean 19:42.560 --> 19:45.440 you break records for the 100 meters dash. 19:45.440 --> 19:48.040 What it really means is you see the unexpected 19:48.040 --> 19:51.520 and you're able to maneuver in a safe way, 19:51.520 --> 19:55.440 and in a way that gets you the most information 19:55.440 --> 19:57.720 about the thing you're trying to do. 19:57.720 --> 20:00.520 By the way, you may be the only person 20:00.520 --> 20:04.280 who in a TED talk has used a math equation, 20:04.280 --> 20:07.720 which is amazing, people should go see one of your TED talks. 20:07.720 --> 20:08.840 Actually, it's very interesting 20:08.840 --> 20:13.560 because the TED curator, Chris Anderson, told me, 20:13.560 --> 20:15.400 you can't show math. 20:15.400 --> 20:18.240 And I thought about it, but that's who I am. 20:18.240 --> 20:20.800 I mean, that's our work. 20:20.800 --> 20:25.800 And so I felt compelled to give the audience a taste 20:25.800 --> 20:27.680 for at least some math. 20:27.680 --> 20:32.680 So on that point, simply, what does it take 20:32.680 --> 20:37.120 to make a thing with four motors fly, a quadcopter, 20:37.120 --> 20:40.400 one of these little flying robots? 20:41.560 --> 20:43.800 How hard is it to make it fly? 20:43.800 --> 20:46.360 How do you coordinate the four motors? 20:47.360 --> 20:52.360 How do you convert those motors into actual movement? 20:52.400 --> 20:54.600 So this is an interesting question. 20:54.600 --> 20:57.840 We've been trying to do this since 2000. 20:57.840 --> 21:00.360 It is a commentary on the sensors 21:00.360 --> 21:01.880 that were available back then, 21:01.880 --> 21:04.320 and the computers that were available back then. 21:05.640 --> 21:08.080 And a number of things happened 21:08.080 --> 21:10.320 between 2000 and 2007. 21:11.640 --> 21:15.560 One is the advances in computing, which is, 21:15.560 --> 21:16.840 so we all know about Moore's Law, 21:16.840 --> 21:19.760 but I think 2007 was a tipping point, 21:19.760 --> 21:22.800 the year of the iPhone, the year of the cloud. 21:22.800 --> 21:24.720 Lots of things happened in 2007. 21:25.680 --> 21:27.640 But going back even further, 21:27.640 --> 21:31.440 inertial measurement units as a sensor really matured. 21:31.440 --> 21:33.040 Again, lots of reasons for that. 21:34.000 --> 21:35.480 Certainly there's a lot of federal funding, 21:35.480 --> 21:37.440 particularly DARPA in the US, 21:38.360 --> 21:42.840 but they didn't anticipate this boom in IMUs. 21:43.800 --> 21:46.080 But if you look subsequently, 21:46.080 --> 21:49.000 what happened is that every car manufacturer 21:49.000 --> 21:50.120 had to put an airbag in, 21:50.120 --> 21:52.720 which meant you had to have an accelerometer on board. 21:52.720 --> 21:54.080 And so that drove down the price 21:54.080 --> 21:56.280 to performance ratio of the sensors. 21:56.280 --> 21:57.960 I should know this, that's very interesting. 21:57.960 --> 21:59.480 It's very interesting, the connection there. 21:59.480 --> 22:03.160 And that's why research is very hard to predict the outcomes. 22:04.920 --> 22:07.760 And again, the federal government spent a ton of money 22:07.760 --> 22:12.360 on things that they thought were useful for resonators, 22:12.360 --> 22:16.920 but it ended up enabling these small UAVs, which is great, 22:16.920 --> 22:18.600 because I could have never raised that much money 22:18.600 --> 22:20.800 and told, sold this project, 22:20.800 --> 22:22.280 hey, we want to build these small UAVs. 22:22.280 --> 22:25.520 Can you actually fund the development of low cost IMUs? 22:25.520 --> 22:27.720 So why do you need an IMU on an IMU? 22:27.720 --> 22:30.440 So I'll come back to that, 22:30.440 --> 22:33.400 but so in 2007, 2008, we were able to build these. 22:33.400 --> 22:35.280 And then the question you're asking was a good one, 22:35.280 --> 22:40.280 how do you coordinate the motors to develop this? 22:40.320 --> 22:43.920 But over the last 10 years, everything is commoditized. 22:43.920 --> 22:47.920 A high school kid today can pick up a Raspberry Pi kit 22:49.520 --> 22:50.600 and build this, 22:50.600 --> 22:53.240 all the low levels functionality is all automated. 22:53.240 --> 22:58.240 But basically at some level, you have to drive the motors 22:59.160 --> 23:03.660 at the right RPMs, the right velocity, 23:04.560 --> 23:07.480 in order to generate the right amount of thrust 23:07.480 --> 23:09.960 in order to position it and orient it 23:09.960 --> 23:12.840 in a way that you need to in order to fly. 23:13.800 --> 23:16.680 The feedback that you get is from onboard sensors 23:16.680 --> 23:18.400 and the IMU is an important part of it. 23:18.400 --> 23:23.400 The IMU tells you what the acceleration is 23:23.840 --> 23:26.400 as well as what the angular velocity is. 23:26.400 --> 23:29.200 And those are important pieces of information. 23:30.440 --> 23:34.200 In addition to that, you need some kind of local position 23:34.200 --> 23:36.480 or velocity information. 23:37.440 --> 23:39.320 For example, when we walk, 23:39.320 --> 23:41.520 we implicitly have this information 23:41.520 --> 23:45.800 because we kind of know how, what our stride length is. 23:45.800 --> 23:50.800 We also are looking at images fly past our retina, 23:51.440 --> 23:54.240 if you will, and so we can estimate velocity. 23:54.240 --> 23:56.280 We also have accelerometers in our head 23:56.280 --> 23:59.120 and we're able to integrate all these pieces of information 23:59.120 --> 24:02.320 to determine where we are as we walk. 24:02.320 --> 24:04.320 And so robots have to do something very similar. 24:04.320 --> 24:08.160 You need an IMU, you need some kind of a camera 24:08.160 --> 24:11.600 or other sensor that's measuring velocity. 24:11.600 --> 24:15.800 And then you need some kind of a global reference frame 24:15.800 --> 24:19.480 if you really want to think about doing something 24:19.480 --> 24:21.280 in a world coordinate system. 24:21.280 --> 24:23.680 And so how do you estimate your position 24:23.680 --> 24:25.160 with respect to that global reference frame? 24:25.160 --> 24:26.560 That's important as well. 24:26.560 --> 24:29.520 So coordinating the RPMs of the four motors 24:29.520 --> 24:32.680 is what allows you to, first of all, fly and hover 24:32.680 --> 24:35.640 and then you can change the orientation 24:35.640 --> 24:37.640 and the velocity and so on. 24:37.640 --> 24:38.480 Exactly, exactly. 24:38.480 --> 24:40.360 So there's a bunch of degrees of freedom 24:40.360 --> 24:42.240 or there's six degrees of freedom 24:42.240 --> 24:44.960 but you only have four inputs, the four motors. 24:44.960 --> 24:49.960 And it turns out to be a remarkably versatile configuration. 24:50.960 --> 24:53.120 You think at first, well, I only have four motors, 24:53.120 --> 24:55.040 how do I go sideways? 24:55.040 --> 24:56.360 But it's not too hard to say, well, 24:56.360 --> 24:59.200 if I tilt myself, I can go sideways. 24:59.200 --> 25:01.200 And then you have four motors pointing up, 25:01.200 --> 25:05.400 how do I rotate in place about a vertical axis? 25:05.400 --> 25:07.840 Well, you rotate them at different speeds 25:07.840 --> 25:09.760 and that generates reaction moments 25:09.760 --> 25:11.560 and that allows you to turn. 25:11.560 --> 25:13.400 So it's actually a pretty, 25:13.400 --> 25:17.040 it's an optimal configuration from an engineer standpoint. 25:17.960 --> 25:22.960 It's very simple, very cleverly done and very versatile. 25:23.800 --> 25:26.520 So if you could step back to a time, 25:27.320 --> 25:30.120 so I've always known flying robots as, 25:31.120 --> 25:35.840 to me it was natural that the quadcopters should fly. 25:35.840 --> 25:38.040 But when you first started working with it, 25:38.040 --> 25:42.040 how surprised are you that you can make, 25:42.040 --> 25:45.560 do so much with the four motors? 25:45.560 --> 25:47.640 How surprising is that you can make this thing fly, 25:47.640 --> 25:49.800 first of all, that you can make it hover, 25:49.800 --> 25:52.040 then you can add control to it? 25:52.920 --> 25:55.800 Firstly, this is not, the four motor configuration 25:55.800 --> 26:00.120 is not ours, it has at least a hundred year history. 26:01.080 --> 26:02.480 And various people, 26:02.480 --> 26:06.280 various people try to get quadrotors to fly 26:06.280 --> 26:08.120 without much success. 26:09.240 --> 26:11.560 As I said, we've been working on this since 2000. 26:11.560 --> 26:13.360 Our first designs were, 26:13.360 --> 26:15.160 well, this is way too complicated. 26:15.160 --> 26:19.160 Why not we try to get an omnidirectional flying robot? 26:19.160 --> 26:22.760 So our early designs, we had eight rotors. 26:22.760 --> 26:26.080 And so these eight rotors were arranged uniformly 26:27.520 --> 26:28.880 on a sphere, if you will. 26:28.880 --> 26:31.360 So you can imagine a symmetric configuration 26:31.360 --> 26:34.160 and so you should be able to fly anywhere. 26:34.160 --> 26:36.280 But the real challenge we had is the strength 26:36.280 --> 26:37.880 to weight ratio is not enough, 26:37.880 --> 26:41.240 and of course we didn't have the sensors and so on. 26:41.240 --> 26:43.840 So everybody knew, or at least the people 26:43.840 --> 26:45.680 who worked with rotor crafts knew, 26:45.680 --> 26:47.320 four rotors would get it done. 26:48.280 --> 26:50.200 So that was not our idea. 26:50.200 --> 26:53.480 But it took a while before we could actually do 26:53.480 --> 26:56.520 the onboard sensing and the computation 26:56.520 --> 27:00.400 that was needed for the kinds of agile maneuvering 27:00.400 --> 27:03.800 that we wanted to do in our little aerial robots. 27:03.800 --> 27:08.320 And that only happened between 2007 and 2009 in our lab. 27:08.320 --> 27:10.680 Yeah, and you have to send the signal 27:10.680 --> 27:13.200 maybe a hundred times a second. 27:13.200 --> 27:15.400 So the compute there is everything 27:15.400 --> 27:16.720 has to come down in price. 27:16.720 --> 27:21.720 And what are the steps of getting from point A to point B? 27:22.320 --> 27:25.840 So we just talked about like local control, 27:25.840 --> 27:30.840 but if all the kind of cool dancing in the air 27:30.840 --> 27:34.480 that I've seen you show, how do you make it happen? 27:34.480 --> 27:38.840 Make it trajectory, first of all, okay, 27:38.840 --> 27:41.600 figure out a trajectory, so plan a trajectory, 27:41.600 --> 27:44.320 and then how do you make that trajectory happen? 27:44.320 --> 27:47.320 I think planning is a very fundamental problem in robotics. 27:47.320 --> 27:50.120 I think 10 years ago it was an esoteric thing, 27:50.120 --> 27:52.360 but today with self driving cars, 27:52.360 --> 27:55.160 everybody can understand this basic idea 27:55.160 --> 27:57.320 that a car sees a whole bunch of things 27:57.320 --> 27:59.720 and it has to keep a lane or maybe make a right turn 27:59.720 --> 28:02.160 or switch lanes, it has to plan a trajectory, 28:02.160 --> 28:04.320 it has to be safe, it has to be efficient. 28:04.320 --> 28:06.120 So everybody's familiar with that. 28:06.120 --> 28:07.400 That's kind of the first step 28:07.400 --> 28:12.400 that you have to think about when you say autonomy. 28:14.320 --> 28:18.600 And so for us, it's about finding smooth motions, 28:18.600 --> 28:20.760 motions that are safe. 28:20.760 --> 28:22.360 So we think about these two things. 28:22.360 --> 28:24.160 One is optimality, one is safety. 28:24.160 --> 28:26.720 Clearly you cannot compromise safety. 28:26.720 --> 28:30.160 So you're looking for safe, optimal motions. 28:30.160 --> 28:33.160 The other thing you have to think about 28:33.160 --> 28:37.360 is can you actually compute a reasonable trajectory 28:37.360 --> 28:41.560 in a small amount of time, because you have a time budget. 28:41.560 --> 28:44.360 So the optimal becomes suboptimal, 28:44.360 --> 28:49.360 but in our lab we focus on synthesizing smooth trajectory 28:50.560 --> 28:52.360 that satisfy all the constraints. 28:52.360 --> 28:57.200 In other words, don't violate any safety constraints 28:57.200 --> 29:02.200 and is as efficient as possible. 29:02.200 --> 29:04.600 And when I say efficient, it could mean 29:04.600 --> 29:07.760 I want to get from point A to point B as quickly as possible 29:07.760 --> 29:11.200 or I want to get to it as gracefully as possible 29:11.200 --> 29:15.360 or I want to consume as little energy as possible. 29:15.360 --> 29:17.600 But always staying within the safety constraints. 29:17.600 --> 29:22.440 But yes, always finding a safe trajectory. 29:22.440 --> 29:24.440 So there's a lot of excitement and progress 29:24.440 --> 29:26.440 in the field of machine learning. 29:26.440 --> 29:31.440 And reinforcement learning and the neural network variant 29:31.440 --> 29:33.440 of that with deeper reinforcement learning. 29:33.440 --> 29:37.440 Do you see a role of machine learning in... 29:37.440 --> 29:40.040 So a lot of the success with flying robots 29:40.040 --> 29:41.840 did not rely on machine learning, 29:41.840 --> 29:44.440 except for maybe a little bit of the perception 29:44.440 --> 29:46.040 on the computer vision side. 29:46.040 --> 29:48.040 On the control side and the planning, 29:48.040 --> 29:50.040 do you see there's a role in the future 29:50.040 --> 29:51.040 for machine learning? 29:51.040 --> 29:53.040 So let me disagree a little bit with you. 29:53.040 --> 29:56.040 I think we never perhaps called out in my work 29:56.040 --> 29:59.040 called out learning, but even this very simple idea 29:59.040 --> 30:04.040 of being able to fly through a constrained space. 30:04.040 --> 30:07.040 The first time you try it, you'll invariably... 30:07.040 --> 30:10.040 You might get it wrong if the task is challenging. 30:10.040 --> 30:14.040 And the reason is to get it perfectly right, 30:14.040 --> 30:17.040 you have to model everything in the environment. 30:17.040 --> 30:22.040 And flying is notoriously hard to model. 30:22.040 --> 30:28.040 There are aerodynamic effects that we constantly discover, 30:28.040 --> 30:31.040 even just before I was talking to you, 30:31.040 --> 30:37.040 I was talking to a student about how blades flap when they fly. 30:37.040 --> 30:43.040 And that ends up changing how a rotorcraft 30:43.040 --> 30:46.040 is accelerated in the angular direction. 30:46.040 --> 30:48.040 Does it use like microflaps or something? 30:48.040 --> 30:49.040 It's not microflaps. 30:49.040 --> 30:52.040 We assume that each blade is rigid, 30:52.040 --> 30:54.040 but actually it flaps a little bit. 30:54.040 --> 30:55.040 It bends. 30:55.040 --> 30:56.040 Interesting, yeah. 30:56.040 --> 30:58.040 And so the models rely on the fact, 30:58.040 --> 31:01.040 on the assumption that they're actually rigid. 31:01.040 --> 31:02.040 But that's not true. 31:02.040 --> 31:04.040 If you're flying really quickly, 31:04.040 --> 31:07.040 these effects become significant. 31:07.040 --> 31:09.040 If you're flying close to the ground, 31:09.040 --> 31:12.040 you get pushed off by the ground. 31:12.040 --> 31:15.040 Something which every pilot knows when he tries to land 31:15.040 --> 31:19.040 or she tries to land, this is called a ground effect. 31:19.040 --> 31:21.040 Something very few pilots think about 31:21.040 --> 31:23.040 is what happens when you go close to a ceiling, 31:23.040 --> 31:25.040 or you get sucked into a ceiling. 31:25.040 --> 31:29.040 There are very few aircraft that fly close to any kind of ceiling. 31:29.040 --> 31:33.040 Likewise, when you go close to a wall, 31:33.040 --> 31:36.040 there are these wall effects. 31:36.040 --> 31:39.040 And if you've gone on a train and you pass another train 31:39.040 --> 31:41.040 that's traveling the opposite direction, 31:41.040 --> 31:43.040 you can feel the buffeting. 31:43.040 --> 31:46.040 And so these kinds of microclimates 31:46.040 --> 31:48.040 affect our UAVs significantly. 31:48.040 --> 31:51.040 And they're impossible to model, essentially. 31:51.040 --> 31:53.040 I wouldn't say they're impossible to model, 31:53.040 --> 31:55.040 but the level of sophistication you would need 31:55.040 --> 32:00.040 in the model and the software would be tremendous. 32:00.040 --> 32:03.040 Plus, to get everything right would be awfully tedious. 32:03.040 --> 32:05.040 So the way we do this is over time, 32:05.040 --> 32:10.040 we figure out how to adapt to these conditions. 32:10.040 --> 32:13.040 So early on, we use the form of learning 32:13.040 --> 32:15.040 that we call iterative learning. 32:15.040 --> 32:18.040 So this idea, if you want to perform a task, 32:18.040 --> 32:23.040 there are a few things that you need to change 32:23.040 --> 32:25.040 and iterate over a few parameters 32:25.040 --> 32:29.040 that over time you can figure out. 32:29.040 --> 32:34.040 So I could call it policy gradient reinforcement learning, 32:34.040 --> 32:36.040 but actually it was just iterative learning. 32:36.040 --> 32:38.040 And so this was there way back. 32:38.040 --> 32:40.040 I think what's interesting is, 32:40.040 --> 32:43.040 if you look at autonomous vehicles today, 32:43.040 --> 32:46.040 learning occurs, could occur in two pieces. 32:46.040 --> 32:48.040 One is perception, understanding the world. 32:48.040 --> 32:51.040 Second is action, taking actions. 32:51.040 --> 32:54.040 Everything that I've seen that is successful 32:54.040 --> 32:56.040 is on the perception side of things. 32:56.040 --> 32:59.040 So in computer vision, we've made amazing strides 32:59.040 --> 33:00.040 in the last 10 years. 33:00.040 --> 33:03.040 So recognizing objects, actually detecting objects, 33:03.040 --> 33:08.040 classifying them and tagging them in some sense, 33:08.040 --> 33:11.040 annotating them, this is all done through machine learning. 33:11.040 --> 33:13.040 On the action side, on the other hand, 33:13.040 --> 33:17.040 I don't know if any examples where there are fielded systems 33:17.040 --> 33:21.040 where we actually learn the right behavior. 33:21.040 --> 33:23.040 Outside of single demonstration is successful. 33:23.040 --> 33:25.040 On the laboratory, this is the Holy Grail. 33:25.040 --> 33:27.040 Can you do end to end learning? 33:27.040 --> 33:31.040 Can you go from pixels to motor currents? 33:31.040 --> 33:33.040 This is really, really hard. 33:33.040 --> 33:36.040 And I think if you go forward, 33:36.040 --> 33:38.040 the right way to think about these things 33:38.040 --> 33:43.040 is data driven approaches, learning based approaches, 33:43.040 --> 33:46.040 in concert with model based approaches, 33:46.040 --> 33:48.040 which is the traditional way of doing things. 33:48.040 --> 33:50.040 So I think there's a piece, 33:50.040 --> 33:52.040 there's a role for each of these methodologies. 33:52.040 --> 33:55.040 So what do you think, just jumping out on topics, 33:55.040 --> 33:57.040 since you mentioned autonomous vehicles, 33:57.040 --> 33:59.040 what do you think are the limits on the perception side? 33:59.040 --> 34:02.040 So I've talked to Elon Musk, 34:02.040 --> 34:04.040 and there on the perception side, 34:04.040 --> 34:07.040 they're using primarily computer vision 34:07.040 --> 34:09.040 to perceive the environment. 34:09.040 --> 34:13.040 In your work with, because you work with the real world a lot, 34:13.040 --> 34:15.040 and the physical world, 34:15.040 --> 34:17.040 what are the limits of computer vision? 34:17.040 --> 34:20.040 Do you think you can solve autonomous vehicles, 34:20.040 --> 34:22.040 focusing on the perception side, 34:22.040 --> 34:25.040 focusing on vision alone and machine learning? 34:25.040 --> 34:29.040 So we also have a spin off company, Exxon Technologies, 34:29.040 --> 34:32.040 that works underground in mines. 34:32.040 --> 34:35.040 So you go into mines, they're dark. 34:35.040 --> 34:37.040 They're dirty. 34:37.040 --> 34:39.040 You fly in a dirty area, 34:39.040 --> 34:42.040 there's stuff you kick up by the propellers, 34:42.040 --> 34:44.040 the downwash kicks up dust. 34:44.040 --> 34:48.040 I challenge you to get a computer vision algorithm to work there. 34:48.040 --> 34:53.040 So we use LIDARS in that setting. 34:53.040 --> 34:57.040 Indoors, and even outdoors when we fly through fields, 34:57.040 --> 34:59.040 I think there's a lot of potential 34:59.040 --> 35:03.040 for just solving the problem using computer vision alone. 35:03.040 --> 35:05.040 But I think the bigger question is, 35:05.040 --> 35:08.040 can you actually solve, 35:08.040 --> 35:11.040 or can you actually identify all the corner cases 35:11.040 --> 35:14.040 using a single sensing modality 35:14.040 --> 35:16.040 and using learning alone? 35:16.040 --> 35:18.040 What's your intuition there? 35:18.040 --> 35:20.040 So look, if you have a corner case 35:20.040 --> 35:22.040 and your algorithm doesn't work, 35:22.040 --> 35:25.040 your instinct is to go get data about the corner case 35:25.040 --> 35:29.040 and patch it up, learn how to deal with that corner case. 35:29.040 --> 35:32.040 But at some point, 35:32.040 --> 35:36.040 this is going to saturate, this approach is not viable. 35:36.040 --> 35:39.040 So today, computer vision algorithms 35:39.040 --> 35:43.040 can detect objects 90% of the time, 35:43.040 --> 35:45.040 classify them 90% of the time. 35:45.040 --> 35:49.040 Cats on the internet probably can do 95%, I don't know. 35:49.040 --> 35:54.040 But to get from 90% to 99%, you need a lot more data. 35:54.040 --> 35:56.040 And then I tell you, well, that's not enough 35:56.040 --> 35:58.040 because I have a safety critical application 35:58.040 --> 36:01.040 that want to go from 99% to 99.9%, 36:01.040 --> 36:03.040 well, that's even more data. 36:03.040 --> 36:07.040 So I think if you look at 36:07.040 --> 36:11.040 wanting accuracy on the x axis 36:11.040 --> 36:15.040 and look at the amount of data on the y axis, 36:15.040 --> 36:18.040 I believe that curve is an exponential curve. 36:18.040 --> 36:21.040 Wow, okay, it's even hard if it's linear. 36:21.040 --> 36:24.040 It's hard if it's linear, totally, but I think it's exponential. 36:24.040 --> 36:26.040 And the other thing you have to think about 36:26.040 --> 36:31.040 is that this process is a very, very power hungry process 36:31.040 --> 36:34.040 to run data farms or servers. 36:34.040 --> 36:36.040 Power, do you mean literally power? 36:36.040 --> 36:38.040 Literally power, literally power. 36:38.040 --> 36:43.040 So in 2014, five years ago, and I don't have more recent data, 36:43.040 --> 36:50.040 2% of US electricity consumption was from data farms. 36:50.040 --> 36:54.040 So we think about this as an information science 36:54.040 --> 36:56.040 and information processing problem. 36:56.040 --> 36:59.040 Actually, it is an energy processing problem. 36:59.040 --> 37:02.040 And so unless we've figured out better ways of doing this, 37:02.040 --> 37:04.040 I don't think this is viable. 37:04.040 --> 37:08.040 So talking about driving, which is a safety critical application 37:08.040 --> 37:11.040 and some aspect of the flight is safety critical, 37:11.040 --> 37:14.040 maybe philosophical question, maybe an engineering one. 37:14.040 --> 37:16.040 What problem do you think is harder to solve? 37:16.040 --> 37:19.040 Autonomous driving or autonomous flight? 37:19.040 --> 37:21.040 That's a really interesting question. 37:21.040 --> 37:26.040 I think autonomous flight has several advantages 37:26.040 --> 37:30.040 that autonomous driving doesn't have. 37:30.040 --> 37:33.040 So look, if I want to go from point A to point B, 37:33.040 --> 37:35.040 I have a very, very safe trajectory. 37:35.040 --> 37:38.040 Go vertically up to a maximum altitude, 37:38.040 --> 37:41.040 fly horizontally to just about the destination 37:41.040 --> 37:43.040 and then come down vertically. 37:43.040 --> 37:46.040 This is preprogrammed. 37:46.040 --> 37:49.040 The equivalent of that is very hard to find 37:49.040 --> 37:53.040 in a self driving car world because you're on the ground, 37:53.040 --> 37:55.040 you're in a two dimensional surface, 37:55.040 --> 37:58.040 and the trajectories on the two dimensional surface 37:58.040 --> 38:01.040 are more likely to encounter obstacles. 38:01.040 --> 38:03.040 I mean this in an intuitive sense, 38:03.040 --> 38:05.040 but mathematically true, that's... 38:05.040 --> 38:08.040 Mathematically as well, that's true. 38:08.040 --> 38:11.040 There's other option on the 2G space of platooning 38:11.040 --> 38:13.040 or because there's so many obstacles, 38:13.040 --> 38:15.040 you can connect with those obstacles 38:15.040 --> 38:16.040 and all these kinds of problems. 38:16.040 --> 38:18.040 But those exist in the three dimensional space as well. 38:18.040 --> 38:19.040 So they do. 38:19.040 --> 38:23.040 So the question also implies how difficult are obstacles 38:23.040 --> 38:25.040 in the three dimensional space in flight? 38:25.040 --> 38:27.040 So that's the downside. 38:27.040 --> 38:29.040 I think in three dimensional space, 38:29.040 --> 38:31.040 you're modeling three dimensional world, 38:31.040 --> 38:33.040 not just because you want to avoid it, 38:33.040 --> 38:35.040 but you want to reason about it 38:35.040 --> 38:37.040 and you want to work in that three dimensional environment. 38:37.040 --> 38:39.040 And that's significantly harder. 38:39.040 --> 38:41.040 So that's one disadvantage. 38:41.040 --> 38:43.040 I think the second disadvantage is of course, 38:43.040 --> 38:45.040 anytime you fly, you have to put up 38:45.040 --> 38:49.040 with the peculiarities of aerodynamics 38:49.040 --> 38:51.040 and their complicated environments. 38:51.040 --> 38:52.040 How do you negotiate that? 38:52.040 --> 38:54.040 So that's always a problem. 38:54.040 --> 38:57.040 Do you see a time in the future where there is... 38:57.040 --> 39:00.040 You mentioned there's agriculture applications. 39:00.040 --> 39:03.040 So there's a lot of applications of flying robots. 39:03.040 --> 39:07.040 But do you see a time in the future where there is tens of thousands 39:07.040 --> 39:10.040 or maybe hundreds of thousands of delivery drones 39:10.040 --> 39:14.040 that fill the sky, a delivery of flying robots? 39:14.040 --> 39:18.040 I think there's a lot of potential for the last mile delivery. 39:18.040 --> 39:21.040 And so in crowded cities, 39:21.040 --> 39:24.040 I don't know if you go to a place like Hong Kong, 39:24.040 --> 39:27.040 just crossing the river can take half an hour. 39:27.040 --> 39:32.040 And while a drone can just do it in five minutes at most. 39:32.040 --> 39:38.040 I think you look at delivery of supplies to remote villages. 39:38.040 --> 39:41.040 I work with a nonprofit called Weave Robotics. 39:41.040 --> 39:43.040 So they work in the Peruvian Amazon, 39:43.040 --> 39:47.040 where the only highways are rivers. 39:47.040 --> 39:49.040 And to get from point A to point B 39:49.040 --> 39:52.040 may take five hours. 39:52.040 --> 39:56.040 While with a drone, you can get there in 30 minutes. 39:56.040 --> 39:59.040 So just delivering drugs, 39:59.040 --> 40:04.040 retrieving samples for testing vaccines. 40:04.040 --> 40:06.040 I think there's huge potential here. 40:06.040 --> 40:09.040 So I think the challenges are not technological. 40:09.040 --> 40:12.040 The challenge is economical. 40:12.040 --> 40:16.040 The one thing I'll tell you that nobody thinks about 40:16.040 --> 40:21.040 is the fact that we've not made huge strides in battery technology. 40:21.040 --> 40:22.040 Yes, it's true. 40:22.040 --> 40:24.040 Batteries are becoming less expensive 40:24.040 --> 40:27.040 because we have these mega factories that are coming up. 40:27.040 --> 40:29.040 But they're all based on lithium based technologies. 40:29.040 --> 40:34.040 And if you look at the energy density and the power density, 40:34.040 --> 40:39.040 those are two fundamentally limiting numbers. 40:39.040 --> 40:41.040 So power density is important because for a UAV 40:41.040 --> 40:43.040 to take off vertically into the air, 40:43.040 --> 40:47.040 which most drones do, they don't have a runway, 40:47.040 --> 40:52.040 you consume roughly 200 watts per kilo at the small size. 40:52.040 --> 40:54.040 That's a lot. 40:54.040 --> 40:58.040 In contrast, the human brain consumes less than 80 watts, 40:58.040 --> 41:00.040 the whole of the human brain. 41:00.040 --> 41:04.040 So just imagine just lifting yourself into the air 41:04.040 --> 41:08.040 is like two or three light bulbs, which makes no sense to me. 41:08.040 --> 41:12.040 Yeah, so you're going to have to at scale solve the energy problem 41:12.040 --> 41:19.040 then charging the batteries, storing the energy and so on. 41:19.040 --> 41:21.040 And then the storage is the second problem. 41:21.040 --> 41:23.040 But storage limits the range. 41:23.040 --> 41:30.040 But you have to remember that you have to burn a lot of it 41:30.040 --> 41:32.040 for a given time. 41:32.040 --> 41:33.040 So the burning is another problem. 41:33.040 --> 41:35.040 Which is a power question. 41:35.040 --> 41:36.040 Yes. 41:36.040 --> 41:39.040 And do you think just your intuition, 41:39.040 --> 41:45.040 there are breakthroughs in batteries on the horizon? 41:45.040 --> 41:47.040 How hard is that problem? 41:47.040 --> 41:52.040 Look, there are a lot of companies that are promising flying cars, 41:52.040 --> 42:00.040 that are autonomous, and that are clean. 42:00.040 --> 42:02.040 I think they're over promising. 42:02.040 --> 42:05.040 The autonomy piece is doable. 42:05.040 --> 42:08.040 The clean piece, I don't think so. 42:08.040 --> 42:12.040 There's another company that I work with called Jatatra. 42:12.040 --> 42:16.040 They make small jet engines. 42:16.040 --> 42:20.040 And they can get up to 50 miles an hour very easily and lift 50 kilos. 42:20.040 --> 42:22.040 But they're jet engines. 42:22.040 --> 42:24.040 They're efficient. 42:24.040 --> 42:26.040 They're a little louder than electric vehicles. 42:26.040 --> 42:29.040 But they can build flying cars. 42:29.040 --> 42:33.040 So your sense is that there's a lot of pieces that have come together. 42:33.040 --> 42:39.040 So on this crazy question, if you look at companies like Kitty Hawk, 42:39.040 --> 42:45.040 working on electric, so the clean, talking as the bashing through. 42:45.040 --> 42:52.040 It's a crazy dream, but you work with flight a lot. 42:52.040 --> 42:58.040 You've mentioned before that manned flights or carrying a human body 42:58.040 --> 43:01.040 is very difficult to do. 43:01.040 --> 43:04.040 So how crazy is flying cars? 43:04.040 --> 43:11.040 Do you think there will be a day when we have vertical takeoff and landing vehicles 43:11.040 --> 43:17.040 that are sufficiently affordable that we're going to see a huge amount of them? 43:17.040 --> 43:21.040 And they would look like something like we dream of when we think about flying cars. 43:21.040 --> 43:23.040 Yeah, like the Jetsons. 43:23.040 --> 43:26.040 So look, there are a lot of smart people working on this. 43:26.040 --> 43:32.040 And you never say something is not possible when you're people like Sebastian Thrun working on it. 43:32.040 --> 43:35.040 So I totally think it's viable. 43:35.040 --> 43:38.040 I question, again, the electric piece. 43:38.040 --> 43:40.040 The electric piece, yeah. 43:40.040 --> 43:42.040 For short distances, you can do it. 43:42.040 --> 43:46.040 And there's no reason to suggest that these all just have to be rotor crafts. 43:46.040 --> 43:50.040 You take off vertically, but then you morph into a forward flight. 43:50.040 --> 43:52.040 I think there are a lot of interesting designs. 43:52.040 --> 43:56.040 The question to me is, are these economically viable? 43:56.040 --> 44:02.040 And if you agree to do this with fossil fuels, it instantly immediately becomes viable. 44:02.040 --> 44:04.040 That's a real challenge. 44:04.040 --> 44:09.040 Do you think it's possible for robots and humans to collaborate successfully on tasks? 44:09.040 --> 44:18.040 So a lot of robotics folks that I talk to and work with, I mean, humans just add a giant mess to the picture. 44:18.040 --> 44:22.040 So it's best to remove them from consideration when solving specific tasks. 44:22.040 --> 44:24.040 It's very difficult to model. 44:24.040 --> 44:26.040 There's just a source of uncertainty. 44:26.040 --> 44:36.040 In your work with these agile flying robots, do you think there's a role for collaboration with humans? 44:36.040 --> 44:43.040 Is it best to model tasks in a way that doesn't have a human in the picture? 44:43.040 --> 44:48.040 I don't think we should ever think about robots without human in the picture. 44:48.040 --> 44:54.040 Ultimately, robots are there because we want them to solve problems for humans. 44:54.040 --> 44:58.040 But there's no general solution to this problem. 44:58.040 --> 45:02.040 I think if you look at human interaction and how humans interact with robots, 45:02.040 --> 45:06.040 you know, we think of these in sort of three different ways. 45:06.040 --> 45:09.040 One is the human commanding the robot. 45:09.040 --> 45:13.040 The second is the human collaborating with the robot. 45:13.040 --> 45:19.040 So for example, we work on how a robot can actually pick up things with a human and carry things. 45:19.040 --> 45:21.040 That's like true collaboration. 45:21.040 --> 45:26.040 And third, we think about humans as bystanders, self driving cars. 45:26.040 --> 45:33.040 What's the human's role and how do self driving cars acknowledge the presence of humans? 45:33.040 --> 45:36.040 So I think all of these things are different scenarios. 45:36.040 --> 45:39.040 It depends on what kind of humans, what kind of tasks. 45:39.040 --> 45:45.040 And I think it's very difficult to say that there's a general theory that we all have for this. 45:45.040 --> 45:52.040 But at the same time, it's also silly to say that we should think about robots independent of humans. 45:52.040 --> 45:59.040 So to me, human robot interaction is almost a mandatory aspect of everything we do. 45:59.040 --> 46:00.040 Yes. 46:00.040 --> 46:05.040 But to wish to agree, so your thoughts, if we jump to autonomous vehicles, for example, 46:05.040 --> 46:10.040 there's a big debate between what's called level two and level four. 46:10.040 --> 46:13.040 So semi autonomous and autonomous vehicles. 46:13.040 --> 46:19.040 And sort of the Tesla approach currently at least has a lot of collaboration between human and machine. 46:19.040 --> 46:24.040 So the human is supposed to actively supervise the operation of the robot. 46:24.040 --> 46:33.040 Part of the safety definition of how safe a robot is in that case is how effective is the human in monitoring it. 46:33.040 --> 46:43.040 Do you think that's ultimately not a good approach in sort of having a human in the picture, 46:43.040 --> 46:51.040 not as a bystander or part of the infrastructure, but really as part of what's required to make the system safe? 46:51.040 --> 46:53.040 This is harder than it sounds. 46:53.040 --> 47:01.040 I think, you know, if you, I mean, I'm sure you've driven before in highways and so on, 47:01.040 --> 47:10.040 it's really very hard to relinquish controls to a machine and then take over when needed. 47:10.040 --> 47:18.040 So I think Tesla's approach is interesting because it allows you to periodically establish some kind of contact with the car. 47:18.040 --> 47:24.040 Toyota, on the other hand, is thinking about shared autonomy or collaborative autonomy as a paradigm. 47:24.040 --> 47:31.040 If I may argue, these are very, very simple ways of human robot collaboration because the task is pretty boring. 47:31.040 --> 47:34.040 You sit in a vehicle, you go from point A to point B. 47:34.040 --> 47:42.040 I think the more interesting thing to me is, for example, search and rescue, I've got a human first responder, robot first responders. 47:42.040 --> 47:44.040 I got to do something. 47:44.040 --> 47:45.040 It's important. 47:45.040 --> 47:47.040 I have to do it in two minutes. 47:47.040 --> 47:48.040 The building is burning. 47:48.040 --> 47:50.040 There's been an explosion. 47:50.040 --> 47:51.040 It's collapsed. 47:51.040 --> 47:52.040 How do I do it? 47:52.040 --> 47:58.040 I think to me, those are the interesting things where it's very, very unstructured and what's the role of the human? 47:58.040 --> 47:59.040 What's the role of the robot? 47:59.040 --> 48:02.040 Clearly, there's lots of interesting challenges. 48:02.040 --> 48:05.040 As a field, I think we're going to make a lot of progress in this area. 48:05.040 --> 48:07.040 Yeah, it's an exciting form of collaboration. 48:07.040 --> 48:08.040 You're right. 48:08.040 --> 48:15.040 In the autonomous driving, the main enemy is just boredom of the human as opposed to the rescue operations. 48:15.040 --> 48:23.040 It's literally life and death and the collaboration enables the effective completion of the mission. 48:23.040 --> 48:24.040 So it's exciting. 48:24.040 --> 48:27.040 Well, in some sense, we're also doing this. 48:27.040 --> 48:37.040 You think about the human driving a car and almost invariably the human is trying to estimate the state of the car, the state of the environment, and so on. 48:37.040 --> 48:40.040 But what is the car where to estimate the state of the human? 48:40.040 --> 48:47.040 So for example, I'm sure you have a smartphone and the smartphone tries to figure out what you're doing and send you reminders. 48:47.040 --> 48:53.040 And oftentimes telling you to drive to a certain place, although you have no intention of going there, because it thinks that that's where you should be. 48:53.040 --> 48:59.040 Because of some Gmail calendar entry or something like that. 48:59.040 --> 49:02.040 And it's trying to constantly figure out who you are, what you're doing. 49:02.040 --> 49:06.040 If a car were to do that, maybe that would make the driver safer. 49:06.040 --> 49:14.040 Because the car is trying to figure out there's a driver paying attention, looking at his or her eyes, looking at circadian movements. 49:14.040 --> 49:16.040 So I think the potential is there. 49:16.040 --> 49:21.040 But from the reverse side, it's not robot modeling, but it's human modeling. 49:21.040 --> 49:23.040 It's more in the human, right? 49:23.040 --> 49:30.040 And I think the robots can do a very good job of modeling humans if you really think about the framework that you have. 49:30.040 --> 49:39.040 A human sitting in a cockpit surrounded by sensors, all staring at him, in addition to be staring outside, but also staring at him. 49:39.040 --> 49:41.040 I think there's a real synergy there. 49:41.040 --> 49:48.040 Yeah, I love that problem because it's the new 21st century form of psychology actually, AI enabled psychology. 49:48.040 --> 49:54.040 A lot of people have sci fi inspired fears of walking robots like those from Boston Dynamics. 49:54.040 --> 49:59.040 If you just look at shows on Netflix and so on, or flying robots like those you work with. 49:59.040 --> 50:03.040 How would you, how do you think about those fears? 50:03.040 --> 50:05.040 How would you alleviate those fears? 50:05.040 --> 50:09.040 Do you have inklings, echoes of those same concerns? 50:09.040 --> 50:23.040 Any time we develop a technology meaning to have positive impact in the world, there's always a worry that somebody could subvert those technologies and use it in an adversarial setting. 50:23.040 --> 50:25.040 And robotics is no exception, right? 50:25.040 --> 50:29.040 So I think it's very easy to weaponize robots. 50:29.040 --> 50:31.040 I think we talk about swarms. 50:31.040 --> 50:38.040 One thing I worry a lot about is, for us to get swarms to work and do something reliably, it's really hard. 50:38.040 --> 50:44.040 But suppose I have this challenge of trying to destroy something. 50:44.040 --> 50:49.040 And I have a swarm of robots where only one out of the swarm needs to get to its destination. 50:49.040 --> 50:53.040 So that suddenly becomes a lot more doable. 50:53.040 --> 51:00.040 And so I worry about this general idea of using autonomy with lots and lots of agents. 51:00.040 --> 51:04.040 I mean, having said that, look, a lot of this technology is not very mature. 51:04.040 --> 51:12.040 My favorite saying is that if somebody had to develop this technology, wouldn't you rather the good guys do it? 51:12.040 --> 51:21.040 So the good guys have a good understanding of the technology so they can figure out how this technology is being used in a bad way or could be used in a bad way and try to defend against it? 51:21.040 --> 51:23.040 So we think a lot about that. 51:23.040 --> 51:28.040 So we're doing research on how to defend against swarms, for example. 51:28.040 --> 51:29.040 That's interesting. 51:29.040 --> 51:36.040 There is, in fact, a report by the National Academies on counter UAS technologies. 51:36.040 --> 51:38.040 This is a real threat. 51:38.040 --> 51:47.040 But we're also thinking about how to defend against this and knowing how swarms work, knowing how autonomy works is, I think, very important. 51:47.040 --> 51:49.040 So it's not just politicians? 51:49.040 --> 51:51.040 You think engineers have a role in this discussion? 51:51.040 --> 51:52.040 Absolutely. 51:52.040 --> 51:59.040 I think the days where politicians can be agnostic to technology are gone. 51:59.040 --> 52:05.040 I think every politician needs to be literate in technology. 52:05.040 --> 52:09.040 And I often say technology is the new liberal art. 52:09.040 --> 52:18.040 Understanding how technology will change your life, I think, is important and every human being needs to understand that. 52:18.040 --> 52:22.040 And maybe we can elect some engineers to office as well on the other side. 52:22.040 --> 52:25.040 What are the biggest open problems in robotics in UV? 52:25.040 --> 52:27.040 You said we're in the early days in some sense. 52:27.040 --> 52:30.040 What are the problems we would like to solve in robotics? 52:30.040 --> 52:32.040 I think there are lots of problems, right? 52:32.040 --> 52:36.040 But I would phrase it in the following way. 52:36.040 --> 52:46.040 If you look at the robots we're building, they're still very much tailored towards doing specific tasks in specific settings. 52:46.040 --> 52:59.040 I think the question of how do you get them to operate in much broader settings where things can change in unstructured environments is up in the air. 52:59.040 --> 53:02.040 So think of the self driving cars. 53:02.040 --> 53:05.040 Today, we can build a self driving car in a parking lot. 53:05.040 --> 53:09.040 We can do level five autonomy in a parking lot. 53:09.040 --> 53:17.040 But can you do a level five autonomy in the streets of Napoli in Italy or Mumbai in India? 53:17.040 --> 53:18.040 No. 53:18.040 --> 53:27.040 So in some sense, when we think about robotics, we have to think about where they're functioning, what kind of environment, what kind of a task. 53:27.040 --> 53:32.040 We have no understanding of how to put both those things together. 53:32.040 --> 53:36.040 So we're in the very early days of applying it to the physical world. 53:36.040 --> 53:39.040 And I was just in Naples, actually. 53:39.040 --> 53:46.040 And there's levels of difficulty and complexity depending on which area you're applying it to. 53:46.040 --> 53:47.040 I think so. 53:47.040 --> 53:51.040 And we don't have a systematic way of understanding that. 53:51.040 --> 54:00.040 Everybody says just because a computer can now beat a human at any board game, we suddenly know something about intelligence. 54:00.040 --> 54:01.040 That's not true. 54:01.040 --> 54:04.040 A computer board game is very, very structured. 54:04.040 --> 54:11.040 It is the equivalent of working in a Henry Ford factory where things, parts come, you assemble, move on. 54:11.040 --> 54:14.040 It's a very, very, very structured setting. 54:14.040 --> 54:15.040 That's the easiest thing. 54:15.040 --> 54:18.040 And we know how to do that. 54:18.040 --> 54:23.040 So you've done a lot of incredible work at the UPenn University of Pennsylvania Grass Club. 54:23.040 --> 54:26.040 You're now Dean of Engineering at UPenn. 54:26.040 --> 54:34.040 What advice do you have for a new bright eyed undergrad interested in robotics or AI or engineering? 54:34.040 --> 54:37.040 Well, I think there's really three things. 54:37.040 --> 54:45.040 One is you have to get used to the idea that the world will not be the same in five years or four years whenever you graduate, right? 54:45.040 --> 54:46.040 Which is really hard to do. 54:46.040 --> 54:53.040 So this thing about predicting the future, every one of us needs to be trying to predict the future always. 54:53.040 --> 55:01.040 Not because you'll be any good at it, but by thinking about it, I think you sharpen your senses and you become smarter. 55:01.040 --> 55:02.040 So that's number one. 55:02.040 --> 55:09.040 Number two, it's a callery of the first piece, which is you really don't know what's going to be important. 55:09.040 --> 55:15.040 So this idea that I'm going to specialize in something which will allow me to go in a particular direction. 55:15.040 --> 55:22.040 It may be interesting, but it's important also to have this breadth so you have this jumping off point. 55:22.040 --> 55:25.040 I think the third thing, and this is where I think Penn excels. 55:25.040 --> 55:30.040 I mean, we teach engineering, but it's always in the context of the liberal arts. 55:30.040 --> 55:32.040 It's always in the context of society. 55:32.040 --> 55:35.040 As engineers, we cannot afford to lose sight of that. 55:35.040 --> 55:37.040 So I think that's important. 55:37.040 --> 55:43.040 But I think one thing that people underestimate when they do robotics is the importance of mathematical foundations, 55:43.040 --> 55:47.040 the importance of representations. 55:47.040 --> 55:56.040 Not everything can just be solved by looking for ROS packages on the Internet or to find a deep neural network that works. 55:56.040 --> 56:06.040 I think the representation question is key, even to machine learning, where if you ever hope to achieve or get to explainable AI, 56:06.040 --> 56:09.040 somehow there need to be representations that you can understand. 56:09.040 --> 56:16.040 So if you want to do robotics, you should also do mathematics, and you said liberal arts, a little literature. 56:16.040 --> 56:19.040 If you want to build a robot, you should be reading Dostoyevsky. 56:19.040 --> 56:20.040 I agree with that. 56:20.040 --> 56:21.040 Very good. 56:21.040 --> 56:24.040 So Vijay, thank you so much for talking today. It was an honor. 56:24.040 --> 56:47.040 Thank you. It was just a very exciting conversation. Thank you.