Vijay Kumar: Flying Robots #37

Transcript

00:00:00 The following is a conversation with Vijay Kumar.

00:00:03 He’s one of the top roboticists in the world,

00:00:05 a professor at the University of Pennsylvania,

00:00:08 a dean of pen engineering, former director of Grasp Lab,

00:00:12 or the General Robotics Automation Sensing

00:00:15 and Perception Laboratory at Penn,

00:00:17 that was established back in 1979, that’s 40 years ago.

00:00:22 Vijay is perhaps best known for his work

00:00:25 in multi robot systems, robot swarms,

00:00:28 and micro aerial vehicles,

00:00:30 robots that elegantly cooperate in flight

00:00:34 under all the uncertainty and challenges

00:00:36 that the real world conditions present.

00:00:38 This is the Artificial Intelligence Podcast.

00:00:41 If you enjoy it, subscribe on YouTube,

00:00:44 give it five stars on iTunes, support on Patreon,

00:00:47 or simply connect with me on Twitter

00:00:49 at Lex Friedman, spelled F R I D M A N.

00:00:53 And now, here’s my conversation with Vijay Kumar.

00:00:58 What is the first robot you’ve ever built

00:01:01 or were a part of building?

00:01:02 Way back when I was in graduate school,

00:01:04 I was part of a fairly big project

00:01:06 that involved building a very large hexapod.

00:01:12 It’s weighed close to 7,000 pounds,

00:01:17 and it was powered by hydraulic actuation,

00:01:21 or it was actuated by hydraulics with 18 motors,

00:01:27 hydraulic motors, each controlled by an Intel 8085 processor

00:01:34 and an 8086 co processor.

00:01:38 And so imagine this huge monster that had 18 joints,

00:01:44 each controlled by an independent computer,

00:01:46 and there was a 19th computer that actually did

00:01:49 the coordination between these 18 joints.

00:01:52 So I was part of this project,

00:01:53 and my thesis work was how do you coordinate the 18 legs?

00:02:02 And in particular, the pressures in the hydraulic cylinders

00:02:06 to get efficient locomotion.

00:02:09 It sounds like a giant mess.

00:02:11 So how difficult is it to make all the motors communicate?

00:02:14 Presumably, you have to send signals hundreds of times

00:02:17 a second, or at least.

00:02:18 So this was not my work,

00:02:19 but the folks who worked on this wrote what I believe

00:02:23 to be the first multiprocessor operating system.

00:02:26 This was in the 80s, and you had to make sure

00:02:30 that obviously messages got across

00:02:32 from one joint to another.

00:02:34 You have to remember the clock speeds on those computers

00:02:37 were about half a megahertz.

00:02:39 Right, the 80s.

00:02:42 So not to romanticize the notion,

00:02:45 but how did it make you feel to see that robot move?

00:02:51 It was amazing.

00:02:52 In hindsight, it looks like, well, we built this thing

00:02:55 which really should have been much smaller.

00:02:57 And of course, today’s robots are much smaller.

00:02:59 You look at Boston Dynamics or Ghost Robotics,

00:03:03 a spinoff from Penn.

00:03:06 But back then, you were stuck with the substrate you had,

00:03:10 the compute you had, so things were unnecessarily big.

00:03:13 But at the same time, and this is just human psychology,

00:03:18 somehow bigger means grander.

00:03:21 People never had the same appreciation

00:03:23 for nanotechnology or nanodevices

00:03:26 as they do for the Space Shuttle or the Boeing 747.

00:03:30 Yeah, you’ve actually done quite a good job

00:03:32 at illustrating that small is beautiful

00:03:36 in terms of robotics.

00:03:37 So what is on that topic is the most beautiful

00:03:42 or elegant robot in motion that you’ve ever seen?

00:03:46 Not to pick favorites or whatever,

00:03:47 but something that just inspires you that you remember.

00:03:51 Well, I think the thing that I’m most proud of

00:03:54 that my students have done is really think about

00:03:57 small UAVs that can maneuver in constrained spaces

00:04:00 and in particular, their ability to coordinate

00:04:03 with each other and form three dimensional patterns.

00:04:06 So once you can do that,

00:04:08 you can essentially create 3D objects in the sky

00:04:14 and you can deform these objects on the fly.

00:04:17 So in some sense, your toolbox of what you can create

00:04:21 has suddenly got enhanced.

00:04:25 And before that, we did the two dimensional version of this.

00:04:27 So we had ground robots forming patterns and so on.

00:04:31 So that was not as impressive, that was not as beautiful.

00:04:34 But if you do it in 3D,

00:04:36 suspended in midair, and you’ve got to go back to 2011

00:04:40 when we did this, now it’s actually pretty standard

00:04:43 to do these things eight years later.

00:04:45 But back then it was a big accomplishment.

00:04:47 So the distributed cooperation

00:04:50 is where beauty emerges in your eyes?

00:04:53 Well, I think beauty to an engineer is very different

00:04:55 from beauty to someone who’s looking at robots

00:04:59 from the outside, if you will.

00:05:01 But what I meant there, so before we said that grand,

00:05:04 so before we said that grand is associated with size.

00:05:10 And another way of thinking about this

00:05:13 is just the physical shape

00:05:15 and the idea that you can get physical shapes in midair

00:05:18 and have them deform, that’s beautiful.

00:05:21 But the individual components,

00:05:23 the agility is beautiful too, right?

00:05:24 That is true too.

00:05:25 So then how quickly can you actually manipulate

00:05:28 these three dimensional shapes

00:05:29 and the individual components?

00:05:31 Yes, you’re right.

00:05:32 But by the way, you said UAV, unmanned aerial vehicle.

00:05:36 What’s a good term for drones, UAVs, quad copters?

00:05:41 Is there a term that’s being standardized?

00:05:44 I don’t know if there is.

00:05:45 Everybody wants to use the word drones.

00:05:47 And I’ve often said this, drones to me is a pejorative word.

00:05:51 It signifies something that’s dumb,

00:05:53 that’s pre programmed, that does one little thing

00:05:56 and robots are anything but drones.

00:05:58 So I actually don’t like that word,

00:06:00 but that’s what everybody uses.

00:06:02 You could call it unpiloted.

00:06:04 Unpiloted.

00:06:05 But even unpiloted could be radio controlled,

00:06:08 could be remotely controlled in many different ways.

00:06:11 And I think the right word is,

00:06:12 thinking about it as an aerial robot.

00:06:15 You also say agile, autonomous, aerial robot, right?

00:06:19 Yeah, so agility is an attribute, but they don’t have to be.

00:06:23 So what biological system,

00:06:24 because you’ve also drawn a lot of inspiration with those.

00:06:27 I’ve seen bees and ants that you’ve talked about.

00:06:30 What living creatures have you found to be most inspiring

00:06:35 as an engineer, instructive in your work in robotics?

00:06:38 To me, so ants are really quite incredible creatures, right?

00:06:43 So you, I mean, the individuals arguably are very simple

00:06:47 in how they’re built and yet they’re incredibly resilient

00:06:52 as a population.

00:06:53 And as individuals, they’re incredibly robust.

00:06:56 So, if you take an ant, it’s six legs,

00:07:00 you remove one leg, it still works just fine.

00:07:04 And it moves along.

00:07:05 And I don’t know that he even realizes it’s lost a leg.

00:07:09 So that’s the robustness at the individual ant level.

00:07:13 But then you look about this instinct

00:07:15 for self preservation of the colonies

00:07:17 and they adapt in so many amazing ways.

00:07:20 You know, transcending gaps by just chaining themselves

00:07:26 together when you have a flood,

00:07:29 being able to recruit other teammates

00:07:32 to carry big morsels of food,

00:07:35 and then going out in different directions looking for food,

00:07:38 and then being able to demonstrate consensus,

00:07:43 even though they don’t communicate directly with each other

00:07:47 the way we communicate with each other.

00:07:49 In some sense, they also know how to do democracy,

00:07:51 probably better than what we do.

00:07:53 Yeah, somehow it’s even democracy is emergent.

00:07:57 It seems like all of the phenomena that we see

00:07:59 is all emergent.

00:08:00 It seems like there’s no centralized communicator.

00:08:03 There is, so I think a lot is made about that word,

00:08:06 emergent, and it means lots of things to different people.

00:08:09 But you’re absolutely right.

00:08:10 I think as an engineer, you think about

00:08:13 what element, elemental behaviors

00:08:17 were primitives you could synthesize

00:08:21 so that the whole looks incredibly powerful,

00:08:25 incredibly synergistic,

00:08:26 the whole definitely being greater than some of the parts,

00:08:29 and ants are living proof of that.

00:08:32 So when you see these beautiful swarms

00:08:34 where there’s biological systems of robots,

00:08:38 do you sometimes think of them

00:08:40 as a single individual living intelligent organism?

00:08:44 So it’s the same as thinking of our human beings

00:08:47 are human civilization as one organism,

00:08:51 or do you still, as an engineer,

00:08:52 think about the individual components

00:08:54 and all the engineering

00:08:55 that went into the individual components?

00:08:57 Well, that’s very interesting.

00:08:58 So again, philosophically as engineers,

00:09:01 what we wanna do is to go beyond

00:09:05 the individual components, the individual units,

00:09:08 and think about it as a unit, as a cohesive unit,

00:09:11 without worrying about the individual components.

00:09:15 If you start obsessing about

00:09:17 the individual building blocks and what they do,

00:09:23 you inevitably will find it hard to scale up.

00:09:27 Just mathematically,

00:09:29 just think about individual things you wanna model,

00:09:31 and if you want to have 10 of those,

00:09:34 then you essentially are taking Cartesian products

00:09:36 of 10 things, and that makes it really complicated.

00:09:39 Then to do any kind of synthesis or design

00:09:41 in that high dimension space is really hard.

00:09:44 So the right way to do this

00:09:45 is to think about the individuals in a clever way

00:09:49 so that at the higher level,

00:09:51 when you look at lots and lots of them,

00:09:53 abstractly, you can think of them

00:09:55 in some low dimensional space.

00:09:57 So what does that involve?

00:09:58 For the individual, do you have to try to make

00:10:02 the way they see the world as local as possible?

00:10:05 And the other thing,

00:10:06 do you just have to make them robust to collisions?

00:10:09 Like you said with the ants,

00:10:10 if something fails, the whole swarm doesn’t fail.

00:10:15 Right, I think as engineers, we do this.

00:10:17 I mean, you think about, we build planes,

00:10:19 or we build iPhones,

00:10:22 and we know that by taking individual components,

00:10:26 well engineered components with well specified interfaces

00:10:30 that behave in a predictable way,

00:10:31 you can build complex systems.

00:10:34 So that’s ingrained, I would claim,

00:10:36 in most engineers thinking,

00:10:39 and it’s true for computer scientists as well.

00:10:41 I think what’s different here is that you want

00:10:44 the individuals to be robust in some sense,

00:10:49 as we do in these other settings,

00:10:52 but you also want some degree of resiliency

00:10:54 for the population.

00:10:56 And so you really want them to be able to reestablish

00:11:02 communication with their neighbors.

00:11:03 You want them to rethink their strategy for group behavior.

00:11:08 You want them to reorganize.

00:11:12 And that’s where I think a lot of the challenges lie.

00:11:15 So just at a high level,

00:11:18 what does it take for a bunch of,

00:11:22 what should we call them, flying robots,

00:11:24 to create a formation?

00:11:26 Just for people who are not familiar

00:11:28 with robotics in general, how much information is needed?

00:11:32 How do you even make it happen

00:11:35 without a centralized controller?

00:11:39 So, I mean, there are a couple of different ways

00:11:41 of looking at this.

00:11:43 If you are a purist,

00:11:45 you think of it as a way of recreating what nature does.

00:11:53 So nature forms groups for several reasons,

00:11:58 but mostly it’s because of this instinct

00:12:02 that organisms have of preserving their colonies,

00:12:05 their population, which means what?

00:12:09 You need shelter, you need food, you need to procreate,

00:12:12 and that’s basically it.

00:12:14 So the kinds of interactions you see are all organic.

00:12:18 They’re all local.

00:12:20 And the only information that they share,

00:12:24 and mostly it’s indirectly, is to, again,

00:12:27 preserve the herd or the flock,

00:12:30 or the swarm, and either by looking for new sources of food

00:12:37 or looking for new shelters, right?

00:12:39 Right.

00:12:41 As engineers, when we build swarms, we have a mission.

00:12:46 And when you think of a mission, and it involves mobility,

00:12:52 most often it’s described in some kind

00:12:55 of a global coordinate system.

00:12:56 As a human, as an operator, as a commander,

00:12:59 or as a collaborator, I have my coordinate system,

00:13:03 and I want the robots to be consistent with that.

00:13:07 So I might think of it slightly differently.

00:13:11 I might want the robots to recognize that coordinate system,

00:13:15 which means not only do they have to think locally

00:13:17 in terms of who their immediate neighbors are,

00:13:19 but they have to be cognizant

00:13:20 of what the global environment is.

00:13:24 They have to be cognizant of what the global environment

00:13:27 looks like.

00:13:28 So if I say, surround this building

00:13:31 and protect this from intruders,

00:13:33 well, they’re immediately in a building centered

00:13:35 coordinate system, and I have to tell them

00:13:37 where the building is.

00:13:38 And they’re globally collaborating

00:13:40 on the map of that building.

00:13:41 They’re maintaining some kind of global,

00:13:44 not just in the frame of the building,

00:13:45 but there’s information that’s ultimately being built up

00:13:49 explicitly as opposed to kind of implicitly,

00:13:53 like nature might.

00:13:54 Correct, correct.

00:13:55 So in some sense, nature is very, very sophisticated,

00:13:57 but the tasks that nature solves or needs to solve

00:14:01 are very different from the kind of engineered tasks,

00:14:05 artificial tasks that we are forced to address.

00:14:09 And again, there’s nothing preventing us

00:14:12 from solving these other problems,

00:14:15 but ultimately it’s about impact.

00:14:16 You want these swarms to do something useful.

00:14:19 And so you’re kind of driven into this very unnatural,

00:14:24 if you will.

00:14:25 Unnatural, meaning not like how nature does, setting.

00:14:29 And it’s probably a little bit more expensive

00:14:31 to do it the way nature does,

00:14:33 because nature is less sensitive

00:14:37 to the loss of the individual.

00:14:39 And cost wise in robotics,

00:14:42 I think you’re more sensitive to losing individuals.

00:14:45 I think that’s true, although if you look at the price

00:14:49 to performance ratio of robotic components,

00:14:51 it’s coming down dramatically, right?

00:14:54 It continues to come down.

00:14:56 So I think we’re asymptotically approaching the point

00:14:58 where we would get, yeah,

00:14:59 the cost of individuals would really become insignificant.

00:15:05 So let’s step back at a high level view,

00:15:07 the impossible question of what kind of, as an overview,

00:15:12 what kind of autonomous flying vehicles

00:15:14 are there in general?

00:15:16 I think the ones that receive a lot of notoriety

00:15:19 are obviously the military vehicles.

00:15:22 Military vehicles are controlled by a base station,

00:15:26 but have a lot of human supervision.

00:15:29 But they have limited autonomy,

00:15:31 which is the ability to go from point A to point B.

00:15:34 And even the more sophisticated now,

00:15:37 sophisticated vehicles can do autonomous takeoff

00:15:40 and landing.

00:15:41 And those usually have wings and they’re heavy.

00:15:44 Usually they’re wings,

00:15:45 but then there’s nothing preventing us from doing this

00:15:47 for helicopters as well.

00:15:49 There are many military organizations

00:15:52 that have autonomous helicopters in the same vein.

00:15:56 And by the way, you look at autopilots and airplanes

00:16:00 and it’s actually very similar.

00:16:02 In fact, one interesting question we can ask is,

00:16:07 if you look at all the air safety violations,

00:16:12 all the crashes that occurred,

00:16:14 would they have happened if the plane were truly autonomous?

00:16:18 And I think you’ll find that in many of the cases,

00:16:21 because of pilot error, we made silly decisions.

00:16:24 And so in some sense, even in air traffic,

00:16:26 commercial air traffic, there’s a lot of applications,

00:16:29 although we only see autonomy being enabled

00:16:33 at very high altitudes when the plane is an autopilot.

00:16:38 The plane is an autopilot.

00:16:41 There’s still a role for the human

00:16:42 and that kind of autonomy is, you’re kind of implying,

00:16:47 I don’t know what the right word is,

00:16:48 but it’s a little dumber than it could be.

00:16:53 Right, so in the lab, of course,

00:16:55 we can afford to be a lot more aggressive.

00:16:59 And the question we try to ask is,

00:17:04 can we make robots that will be able to make decisions

00:17:10 without any kind of external infrastructure?

00:17:13 So what does that mean?

00:17:14 So the most common piece of infrastructure

00:17:16 that airplanes use today is GPS.

00:17:20 GPS is also the most brittle form of information.

00:17:26 If you have driven in a city, try to use GPS navigation,

00:17:30 in tall buildings, you immediately lose GPS.

00:17:32 And so that’s not a very sophisticated way

00:17:36 of building autonomy.

00:17:37 I think the second piece of infrastructure

00:17:39 they rely on is communications.

00:17:41 Again, it’s very easy to jam communications.

00:17:47 In fact, if you use wifi, you know that wifi signals

00:17:51 drop out, cell signals drop out.

00:17:53 So to rely on something like that is not good.

00:17:58 The third form of infrastructure we use,

00:18:01 and I hate to call it infrastructure,

00:18:02 but it is that, in the sense of robots, is people.

00:18:06 So you could rely on somebody to pilot you.

00:18:09 And so the question you wanna ask is,

00:18:11 if there are no pilots, there’s no communications

00:18:14 with any base station, if there’s no knowledge of position,

00:18:18 and if there’s no a priori map,

00:18:21 a priori knowledge of what the environment looks like,

00:18:24 a priori model of what might happen in the future,

00:18:28 can robots navigate?

00:18:29 So that is true autonomy.

00:18:31 So that’s true autonomy, and we’re talking about,

00:18:34 you mentioned like military application of drones.

00:18:36 Okay, so what else is there?

00:18:38 You talk about agile, autonomous flying robots,

00:18:42 aerial robots, so that’s a different kind of,

00:18:45 it’s not winged, it’s not big, at least it’s small.

00:18:48 So I use the word agility mostly,

00:18:50 or at least we’re motivated to do agile robots,

00:18:53 mostly because robots can operate

00:18:58 and should be operating in constrained environments.

00:19:02 And if you want to operate the way a global hawk operates,

00:19:06 I mean, the kinds of conditions in which you operate

00:19:09 are very, very restrictive.

00:19:11 If you wanna go inside a building,

00:19:13 for example, for search and rescue,

00:19:15 or to locate an active shooter,

00:19:18 or you wanna navigate under the canopy in an orchard

00:19:22 to look at health of plants,

00:19:23 or to look for, to count fruits,

00:19:28 to measure the tree trunks.

00:19:31 These are things we do, by the way.

00:19:33 There’s some cool agriculture stuff you’ve shown

00:19:35 in the past, it’s really awesome.

00:19:37 So in those kinds of settings, you do need that agility.

00:19:40 Agility does not necessarily mean

00:19:42 you break records for the 100 meters dash.

00:19:45 What it really means is you see the unexpected

00:19:48 and you’re able to maneuver in a safe way,

00:19:51 and in a way that gets you the most information

00:19:55 about the thing you’re trying to do.

00:19:57 By the way, you may be the only person

00:20:00 who, in a TED Talk, has used a math equation,

00:20:04 which is amazing, people should go see one of your TED Talks.

00:20:07 Actually, it’s very interesting,

00:20:08 because the TED curator, Chris Anderson,

00:20:12 told me, you can’t show math.

00:20:15 And I thought about it, but that’s who I am.

00:20:18 I mean, that’s our work.

00:20:20 And so I felt compelled to give the audience a taste

00:20:25 for at least some math.

00:20:27 So on that point, simply, what does it take

00:20:32 to make a thing with four motors fly, a quadcopter,

00:20:37 one of these little flying robots?

00:20:41 How hard is it to make it fly?

00:20:43 How do you coordinate the four motors?

00:20:46 How do you convert those motors into actual movement?

00:20:52 So this is an interesting question.

00:20:54 We’ve been trying to do this since 2000.

00:20:58 It is a commentary on the sensors

00:21:00 that were available back then,

00:21:02 the computers that were available back then.

00:21:05 And a number of things happened between 2000 and 2007.

00:21:11 One is the advances in computing,

00:21:14 which is, so we all know about Moore’s Law,

00:21:16 but I think 2007 was a tipping point,

00:21:19 the year of the iPhone, the year of the cloud.

00:21:22 Lots of things happened in 2007.

00:21:25 But going back even further,

00:21:27 inertial measurement units as a sensor really matured.

00:21:31 Again, lots of reasons for that.

00:21:33 Certainly, there’s a lot of federal funding,

00:21:35 particularly DARPA in the US,

00:21:38 but they didn’t anticipate this boom in IMUs.

00:21:42 But if you look, subsequently what happened

00:21:46 is that every car manufacturer had to put an airbag in,

00:21:50 which meant you had to have an accelerometer on board.

00:21:52 And so that drove down the price to performance ratio.

00:21:55 Wow, I should know this.

00:21:56 That’s very interesting.

00:21:57 That’s very interesting, the connection there.

00:21:59 And that’s why research is very,

00:22:01 it’s very hard to predict the outcomes.

00:22:04 And again, the federal government spent a ton of money

00:22:07 on things that they thought were useful for resonators,

00:22:12 but it ended up enabling these small UAVs, which is great,

00:22:16 because I could have never raised that much money

00:22:18 and sold this project,

00:22:20 hey, we want to build these small UAVs.

00:22:22 Can you actually fund the development of low cost IMUs?

00:22:25 So why do you need an IMU on an IMU?

00:22:27 So I’ll come back to that.

00:22:31 So in 2007, 2008, we were able to build these.

00:22:33 And then the question you’re asking was a good one.

00:22:35 How do you coordinate the motors to develop this?

00:22:40 But over the last 10 years, everything is commoditized.

00:22:43 A high school kid today can pick up

00:22:46 a Raspberry Pi kit and build this.

00:22:50 All the low levels functionality is all automated.

00:22:54 But basically at some level,

00:22:56 you have to drive the motors at the right RPMs,

00:23:01 the right velocity,

00:23:04 in order to generate the right amount of thrust,

00:23:07 in order to position it and orient it in a way

00:23:10 that you need to in order to fly.

00:23:13 The feedback that you get is from onboard sensors,

00:23:16 and the IMU is an important part of it.

00:23:18 The IMU tells you what the acceleration is,

00:23:23 as well as what the angular velocity is.

00:23:26 And those are important pieces of information.

00:23:30 In addition to that, you need some kind of local position

00:23:34 or velocity information.

00:23:37 For example, when we walk,

00:23:39 we implicitly have this information

00:23:41 because we kind of know what our stride length is.

00:23:46 We also are looking at images fly past our retina,

00:23:51 if you will, and so we can estimate velocity.

00:23:54 We also have accelerometers in our head,

00:23:56 and we’re able to integrate all these pieces of information

00:23:59 to determine where we are as we walk.

00:24:02 And so robots have to do something very similar.

00:24:04 You need an IMU, you need some kind of a camera

00:24:08 or other sensor that’s measuring velocity,

00:24:12 and then you need some kind of a global reference frame

00:24:15 if you really want to think about doing something

00:24:19 in a world coordinate system.

00:24:21 And so how do you estimate your position

00:24:23 with respect to that global reference frame?

00:24:25 That’s important as well.

00:24:26 So coordinating the RPMs of the four motors

00:24:29 is what allows you to, first of all, fly and hover,

00:24:32 and then you can change the orientation

00:24:35 and the velocity and so on.

00:24:37 Exactly, exactly.

00:24:38 So it’s a bunch of degrees of freedom

00:24:40 that you’re complaining about.

00:24:41 There’s six degrees of freedom,

00:24:42 but you only have four inputs, the four motors.

00:24:44 And it turns out to be a remarkably versatile configuration.

00:24:50 You think at first, well, I only have four motors,

00:24:53 how do I go sideways?

00:24:55 But it’s not too hard to say, well, if I tilt myself,

00:24:57 I can go sideways, and then you have four motors

00:25:00 pointing up, how do I rotate in place

00:25:03 about a vertical axis?

00:25:05 Well, you rotate them at different speeds

00:25:07 and that generates reaction moments

00:25:09 and that allows you to turn.

00:25:11 So it’s actually a pretty, it’s an optimal configuration

00:25:14 from an engineer standpoint.

00:25:18 It’s very simple, very cleverly done, and very versatile.

00:25:23 So if you could step back to a time,

00:25:27 so I’ve always known flying robots as,

00:25:31 to me, it was natural that a quadcopter should fly.

00:25:35 But when you first started working with it,

00:25:38 how surprised are you that you can make,

00:25:42 do so much with the four motors?

00:25:45 How surprising is it that you can make this thing fly,

00:25:47 first of all, that you can make it hover,

00:25:49 that you can add control to it?

00:25:52 Firstly, this is not, the four motor configuration

00:25:55 is not ours.

00:25:56 You can, it has at least a hundred year history.

00:26:00 And various people, various people try to get quadrotors

00:26:04 to fly without much success.

00:26:08 As I said, we’ve been working on this since 2000.

00:26:10 Our first designs were, well, this is way too complicated.

00:26:14 Why not we try to get an omnidirectional flying robot?

00:26:18 So our early designs, we had eight rotors.

00:26:21 And so these eight rotors were arranged uniformly

00:26:26 on a sphere, if you will.

00:26:28 So you can imagine a symmetric configuration.

00:26:30 And so you should be able to fly anywhere.

00:26:33 But the real challenge we had is the strength to weight ratio

00:26:36 is not enough.

00:26:37 And of course, we didn’t have the sensors and so on.

00:26:40 So everybody knew, or at least the people

00:26:43 who worked with rotorcrafts knew,

00:26:44 four rotors will get it done.

00:26:47 So that was not our idea.

00:26:49 But it took a while before we could actually do

00:26:52 the onboard sensing and the computation that was needed

00:26:56 for the kinds of agile maneuvering that we wanted to do

00:27:01 in our little aerial robots.

00:27:03 And that only happened between 2007 and 2009 in our lab.

00:27:07 Yeah, and you have to send the signal

00:27:09 maybe a hundred times a second.

00:27:12 So the compute there, everything has to come down in price.

00:27:15 And what are the steps of getting from point A to point B?

00:27:21 So we just talked about like local control.

00:27:25 But if all the kind of cool dancing in the air

00:27:30 that I’ve seen you show, how do you make it happen?

00:27:34 How do you make a trajectory?

00:27:37 First of all, okay, figure out a trajectory.

00:27:40 So plan a trajectory.

00:27:41 And then how do you make that trajectory happen?

00:27:44 Yeah, I think planning is a very fundamental problem

00:27:47 in robotics.

00:27:48 I think 10 years ago it was an esoteric thing,

00:27:50 but today with self driving cars,

00:27:53 everybody can understand this basic idea

00:27:55 that a car sees a whole bunch of things

00:27:57 and it has to keep a lane or maybe make a right turn

00:28:00 or switch lanes.

00:28:01 It has to plan a trajectory.

00:28:02 It has to be safe.

00:28:03 It has to be efficient.

00:28:04 So everybody’s familiar with that.

00:28:06 That’s kind of the first step that you have to think about

00:28:10 when you say autonomy.

00:28:14 And so for us, it’s about finding smooth motions,

00:28:19 motions that are safe.

00:28:21 So we think about these two things.

00:28:22 One is optimality, one is safety.

00:28:24 Clearly you cannot compromise safety.

00:28:28 So you’re looking for safe, optimal motions.

00:28:31 The other thing you have to think about is

00:28:34 can you actually compute a reasonable trajectory

00:28:38 in a small amount of time?

00:28:40 Cause you have a time budget.

00:28:42 So the optimal becomes suboptimal,

00:28:45 but in our lab we focus on synthesizing smooth trajectory

00:28:51 that satisfy all the constraints.

00:28:53 In other words, don’t violate any safety constraints

00:28:58 and is as efficient as possible.

00:29:02 And when I say efficient,

00:29:04 it could mean I want to get from point A to point B

00:29:06 as quickly as possible,

00:29:08 or I want to get to it as gracefully as possible,

00:29:12 or I want to consume as little energy as possible.

00:29:15 But always staying within the safety constraints.

00:29:18 But yes, always finding a safe trajectory.

00:29:22 So there’s a lot of excitement and progress

00:29:25 in the field of machine learning

00:29:27 and reinforcement learning

00:29:29 and the neural network variant of that

00:29:32 with deep reinforcement learning.

00:29:33 Do you see a role of machine learning

00:29:36 in, so a lot of the success of flying robots

00:29:40 did not rely on machine learning,

00:29:42 except for maybe a little bit of the perception

00:29:45 on the computer vision side.

00:29:46 On the control side and the planning,

00:29:48 do you see there’s a role in the future

00:29:50 for machine learning?

00:29:51 So let me disagree a little bit with you.

00:29:53 I think we never perhaps called out in my work,

00:29:56 called out learning,

00:29:57 but even this very simple idea of being able to fly

00:30:00 through a constrained space.

00:30:02 The first time you try it, you’ll invariably,

00:30:05 you might get it wrong if the task is challenging.

00:30:08 And the reason is to get it perfectly right,

00:30:12 you have to model everything in the environment.

00:30:15 And flying is notoriously hard to model.

00:30:19 There are aerodynamic effects that we constantly discover.

00:30:26 Even just before I was talking to you,

00:30:29 I was talking to a student about how blades flap

00:30:33 when they fly.

00:30:35 And that ends up changing how a rotorcraft

00:30:40 is accelerated in the angular direction.

00:30:43 Does he use like micro flaps or something?

00:30:46 It’s not micro flaps.

00:30:47 So we assume that each blade is rigid,

00:30:49 but actually it flaps a little bit.

00:30:51 It bends.

00:30:52 Interesting, yeah.

00:30:53 And so the models rely on the fact,

00:30:56 on the assumption that they’re not rigid.

00:30:58 On the assumption that they’re actually rigid,

00:31:00 but that’s not true.

00:31:02 If you’re flying really quickly,

00:31:03 these effects become significant.

00:31:06 If you’re flying close to the ground,

00:31:09 you get pushed off by the ground, right?

00:31:12 Something which every pilot knows when he tries to land

00:31:14 or she tries to land, this is called a ground effect.

00:31:18 Something very few pilots think about

00:31:21 is what happens when you go close to a ceiling

00:31:23 or you get sucked into a ceiling.

00:31:25 There are very few aircrafts

00:31:26 that fly close to any kind of ceiling.

00:31:29 Likewise, when you go close to a wall,

00:31:33 there are these wall effects.

00:31:35 And if you’ve gone on a train

00:31:37 and you pass another train that’s traveling

00:31:39 in the opposite direction, you feel the buffeting.

00:31:42 And so these kinds of microclimates

00:31:45 affect our UAV significantly.

00:31:47 So if you want…

00:31:48 And they’re impossible to model, essentially.

00:31:50 I wouldn’t say they’re impossible to model,

00:31:52 but the level of sophistication you would need

00:31:54 in the model and the software would be tremendous.

00:32:00 Plus, to get everything right would be awfully tedious.

00:32:02 So the way we do this is over time,

00:32:05 we figure out how to adapt to these conditions.

00:32:10 So early on, we use the form of learning

00:32:13 that we call iterative learning.

00:32:15 So this idea, if you want to perform a task,

00:32:18 there are a few things that you need to change

00:32:22 and iterate over a few parameters

00:32:24 that over time you can figure out.

00:32:29 So I could call it policy gradient reinforcement learning,

00:32:33 but actually it was just iterative learning.

00:32:34 Iterative learning.

00:32:36 And so this was there way back.

00:32:37 I think what’s interesting is,

00:32:39 if you look at autonomous vehicles today,

00:32:43 learning occurs, could occur in two pieces.

00:32:45 One is perception, understanding the world.

00:32:47 Second is action, taking actions.

00:32:50 Everything that I’ve seen that is successful

00:32:52 is on the perception side of things.

00:32:54 So in computer vision,

00:32:55 we’ve made amazing strides in the last 10 years.

00:32:57 So recognizing objects, actually detecting objects,

00:33:01 classifying them and tagging them in some sense,

00:33:06 annotating them.

00:33:07 This is all done through machine learning.

00:33:09 On the action side, on the other hand,

00:33:12 I don’t know of any examples

00:33:13 where there are fielded systems

00:33:15 where we actually learn

00:33:17 the right behavior.

00:33:20 Outside of single demonstration is successful.

00:33:22 In the laboratory, this is the holy grail.

00:33:24 Can you do end to end learning?

00:33:26 Can you go from pixels to motor currents?

00:33:30 This is really, really hard.

00:33:32 And I think if you go forward,

00:33:35 the right way to think about these things

00:33:37 is data driven approaches,

00:33:40 learning based approaches,

00:33:42 in concert with model based approaches,

00:33:45 which is the traditional way of doing things.

00:33:47 So I think there’s a piece,

00:33:48 there’s a role for each of these methodologies.

00:33:51 So what do you think,

00:33:52 just jumping out on topic

00:33:53 since you mentioned autonomous vehicles,

00:33:56 what do you think are the limits on the perception side?

00:33:58 So I’ve talked to Elon Musk

00:34:01 and there on the perception side,

00:34:03 they’re using primarily computer vision

00:34:05 to perceive the environment.

00:34:08 In your work with,

00:34:09 because you work with the real world a lot

00:34:12 and the physical world,

00:34:13 what are the limits of computer vision?

00:34:15 Do you think we can solve autonomous vehicles

00:34:19 on the perception side,

00:34:20 focusing on vision alone and machine learning?

00:34:24 So, we also have a spinoff company,

00:34:27 Exxon Technologies that works underground in mines.

00:34:31 So you go into mines, they’re dark, they’re dirty.

00:34:36 You fly in a dirty area,

00:34:38 there’s stuff you kick up from by the propellers,

00:34:41 the downwash kicks up dust.

00:34:42 I challenge you to get a computer vision algorithm

00:34:45 to work there.

00:34:46 So we use LIDARs in that setting.

00:34:51 Indoors and even outdoors when we fly through fields,

00:34:55 I think there’s a lot of potential

00:34:57 for just solving the problem using computer vision alone.

00:35:01 But I think the bigger question is,

00:35:02 can you actually solve

00:35:06 or can you actually identify all the corner cases

00:35:09 using a single sensing modality and using learning alone?

00:35:13 So what’s your intuition there?

00:35:15 So look, if you have a corner case

00:35:17 and your algorithm doesn’t work,

00:35:20 your instinct is to go get data about the corner case

00:35:23 and patch it up, learn how to deal with that corner case.

00:35:27 But at some point, this is gonna saturate,

00:35:32 this approach is not viable.

00:35:34 So today, computer vision algorithms can detect

00:35:38 90% of the objects or can detect objects 90% of the time,

00:35:41 classify them 90% of the time.

00:35:43 Cats on the internet probably can do 95%, I don’t know.

00:35:47 But to get from 90% to 99%, you need a lot more data.

00:35:52 And then I tell you, well, that’s not enough

00:35:54 because I have a safety critical application,

00:35:56 I wanna go from 99% to 99.9%.

00:36:00 That’s even more data.

00:36:01 So I think if you look at wanting accuracy on the X axis

00:36:09 and look at the amount of data on the Y axis,

00:36:14 I believe that curve is an exponential curve.

00:36:16 Wow, okay, it’s even hard if it’s linear.

00:36:19 It’s hard if it’s linear, totally,

00:36:20 but I think it’s exponential.

00:36:22 And the other thing you have to think about

00:36:24 is that this process is a very, very power hungry process

00:36:29 to run data farms or servers.

00:36:32 Power, do you mean literally power?

00:36:34 Literally power, literally power.

00:36:36 So in 2014, five years ago, and I don’t have more recent data,

00:36:41 2% of US electricity consumption was from data farms.

00:36:48 So we think about this as an information science

00:36:52 and information processing problem.

00:36:54 Actually, it is an energy processing problem.

00:36:57 And so unless we figured out better ways of doing this,

00:37:00 I don’t think this is viable.

00:37:02 So talking about driving, which is a safety critical application

00:37:06 and some aspect of flight is safety critical,

00:37:10 maybe philosophical question, maybe an engineering one,

00:37:12 what problem do you think is harder to solve,

00:37:15 autonomous driving or autonomous flight?

00:37:18 That’s a really interesting question.

00:37:19 I think autonomous flight has several advantages

00:37:25 that autonomous driving doesn’t have.

00:37:29 So look, if I want to go from point A to point B,

00:37:32 I have a very, very safe trajectory.

00:37:34 Go vertically up to a maximum altitude,

00:37:36 fly horizontally to just about the destination,

00:37:39 and then come down vertically.

00:37:42 This is preprogrammed.

00:37:45 The equivalent of that is very hard to find

00:37:48 in the self driving car world because you’re on the ground,

00:37:51 you’re in a two dimensional surface,

00:37:53 and the trajectories on the two dimensional surface

00:37:56 are more likely to encounter obstacles.

00:38:00 I mean this in an intuitive sense, but mathematically true.

00:38:03 That’s mathematically as well, that’s true.

00:38:06 There’s other option on the 2G space of platooning,

00:38:10 or because there’s so many obstacles,

00:38:11 you can connect with those obstacles

00:38:13 and all these kind of options.

00:38:14 Sure, but those exist in the three dimensional space as well.

00:38:16 So they do.

00:38:17 So the question also implies how difficult are obstacles

00:38:21 in the three dimensional space in flight?

00:38:23 So that’s the downside.

00:38:25 I think in three dimensional space,

00:38:26 you’re modeling three dimensional world,

00:38:29 not just because you want to avoid it,

00:38:31 but you want to reason about it,

00:38:33 and you want to work in the three dimensional environment,

00:38:35 and that’s significantly harder.

00:38:37 So that’s one disadvantage.

00:38:38 I think the second disadvantage is of course,

00:38:41 anytime you fly, you have to put up

00:38:43 with the peculiarities of aerodynamics

00:38:46 and their complicated environments.

00:38:48 How do you negotiate that?

00:38:49 So that’s always a problem.

00:38:51 Do you see a time in the future where there is,

00:38:55 you mentioned there’s agriculture applications.

00:38:58 So there’s a lot of applications of flying robots,

00:39:01 but do you see a time in the future

00:39:03 where there’s tens of thousands,

00:39:05 or maybe hundreds of thousands of delivery drones

00:39:08 that fill the sky, delivery flying robots?

00:39:12 I think there’s a lot of potential

00:39:14 for the last mile delivery.

00:39:15 And so in crowded cities, I don’t know,

00:39:19 if you go to a place like Hong Kong,

00:39:21 just crossing the river can take half an hour,

00:39:24 and while a drone can just do it in five minutes at most.

00:39:29 I think you look at delivery of supplies to remote villages.

00:39:35 I work with a nonprofit called Weave Robotics.

00:39:38 So they work in the Peruvian Amazon,

00:39:40 where the only highways that are available

00:39:44 are the only highways or rivers.

00:39:47 And to get from point A to point B may take five hours,

00:39:52 while with a drone, you can get there in 30 minutes.

00:39:56 So just delivering drugs,

00:39:59 retrieving samples for testing vaccines,

00:40:05 I think there’s huge potential here.

00:40:07 So I think the challenges are not technological,

00:40:09 but the challenge is economical.

00:40:12 The one thing I’ll tell you that nobody thinks about

00:40:15 is the fact that we’ve not made huge strides

00:40:18 in battery technology.

00:40:20 Yes, it’s true, batteries are becoming less expensive

00:40:23 because we have these mega factories that are coming up,

00:40:26 but they’re all based on lithium based technologies.

00:40:28 And if you look at the energy density

00:40:31 and the power density,

00:40:33 those are two fundamentally limiting numbers.

00:40:38 So power density is important

00:40:39 because for a UAV to take off vertically into the air,

00:40:42 which most drones do, they don’t have a runway,

00:40:46 you consume roughly 200 watts per kilo at the small size.

00:40:51 That’s a lot, right?

00:40:53 In contrast, the human brain consumes less than 80 watts,

00:40:57 the whole of the human brain.

00:40:59 So just imagine just lifting yourself into the air

00:41:03 is like two or three light bulbs,

00:41:06 which makes no sense to me.

00:41:07 Yeah, so you’re going to have to at scale

00:41:10 solve the energy problem then,

00:41:12 charging the batteries, storing the energy and so on.

00:41:18 And then the storage is the second problem,

00:41:20 but storage limits the range.

00:41:22 But you have to remember that you have to burn

00:41:28 a lot of it per given time.

00:41:31 So the burning is another problem.

00:41:32 Which is a power question.

00:41:34 Yes, and do you think just your intuition,

00:41:38 there are breakthroughs in batteries on the horizon?

00:41:44 How hard is that problem?

00:41:46 Look, there are a lot of companies

00:41:47 that are promising flying cars that are autonomous

00:41:53 and that are clean.

00:41:59 I think they’re over promising.

00:42:01 The autonomy piece is doable.

00:42:04 The clean piece, I don’t think so.

00:42:08 There’s another company that I work with called JetOptra.

00:42:11 They make small jet engines.

00:42:15 And they can get up to 50 miles an hour very easily

00:42:18 and lift 50 kilos.

00:42:19 But they’re jet engines, they’re efficient,

00:42:23 they’re a little louder than electric vehicles,

00:42:26 but they can build flying cars.

00:42:28 So your sense is that there’s a lot of pieces

00:42:32 that have come together.

00:42:33 So on this crazy question,

00:42:37 if you look at companies like Kitty Hawk,

00:42:39 working on electric, so the clean,

00:42:43 talking to Sebastian Thrun, right?

00:42:45 It’s a crazy dream, you know?

00:42:48 But you work with flight a lot.

00:42:52 You’ve mentioned before that manned flights

00:42:55 or carrying a human body is very difficult to do.

00:43:01 So how crazy is flying cars?

00:43:04 Do you think there’ll be a day

00:43:05 when we have vertical takeoff and landing vehicles

00:43:11 that are sufficiently affordable

00:43:14 that we’re going to see a huge amount of them?

00:43:17 And they would look like something like we dream of

00:43:19 when we think about flying cars.

00:43:21 Yeah, like the Jetsons.

00:43:22 The Jetsons, yeah.

00:43:23 So look, there are a lot of smart people working on this

00:43:25 and you never say something is not possible

00:43:29 when you have people like Sebastian Thrun working on it.

00:43:32 So I totally think it’s viable.

00:43:35 I question, again, the electric piece.

00:43:38 The electric piece, yeah.

00:43:39 And again, for short distances, you can do it.

00:43:41 And there’s no reason to suggest

00:43:43 that these all just have to be rotorcrafts.

00:43:45 You take off vertically,

00:43:46 but then you morph into a forward flight.

00:43:49 I think there are a lot of interesting designs.

00:43:51 The question to me is, are these economically viable?

00:43:56 And if you agree to do this with fossil fuels,

00:43:59 it instantly immediately becomes viable.

00:44:01 That’s a real challenge.

00:44:03 Do you think it’s possible for robots and humans

00:44:06 to collaborate successfully on tasks?

00:44:08 So a lot of robotics folks that I talk to and work with,

00:44:13 I mean, humans just add a giant mess to the picture.

00:44:18 So it’s best to remove them from consideration

00:44:20 when solving specific tasks.

00:44:22 It’s very difficult to model.

00:44:23 There’s just a source of uncertainty.

00:44:26 In your work with these agile flying robots,

00:44:32 do you think there’s a role for collaboration with humans?

00:44:35 Or is it best to model tasks in a way

00:44:38 that doesn’t have a human in the picture?

00:44:43 Well, I don’t think we should ever think about robots

00:44:46 without human in the picture.

00:44:48 Ultimately, robots are there because we want them

00:44:50 to solve problems for humans.

00:44:54 But there’s no general solution to this problem.

00:44:58 I think if you look at human interaction

00:45:00 and how humans interact with robots,

00:45:02 you know, we think of these in sort of three different ways.

00:45:05 One is the human commanding the robot.

00:45:08 The second is the human collaborating with the robot.

00:45:12 So for example, we work on how a robot

00:45:15 can actually pick up things with a human and carry things.

00:45:18 That’s like true collaboration.

00:45:20 And third, we think about humans as bystanders,

00:45:25 self driving cars, what’s the human’s role

00:45:27 and how do self driving cars

00:45:30 acknowledge the presence of humans?

00:45:32 So I think all of these things are different scenarios.

00:45:35 It depends on what kind of humans, what kind of task.

00:45:39 And I think it’s very difficult to say

00:45:41 that there’s a general theory that we all have for this.

00:45:45 But at the same time, it’s also silly to say

00:45:48 that we should think about robots independent of humans.

00:45:52 So to me, human robot interaction

00:45:55 is almost a mandatory aspect of everything we do.

00:45:59 Yes, but to which degree, so your thoughts,

00:46:02 if we jump to autonomous vehicles, for example,

00:46:05 there’s a big debate between what’s called

00:46:08 level two and level four.

00:46:10 So semi autonomous and autonomous vehicles.

00:46:13 And so the Tesla approach currently at least

00:46:16 has a lot of collaboration between human and machine.

00:46:18 So the human is supposed to actively supervise

00:46:22 the operation of the robot.

00:46:23 Part of the safety definition of how safe a robot is

00:46:29 in that case is how effective is the human in monitoring it.

00:46:32 Do you think that’s ultimately not a good approach

00:46:37 in sort of having a human in the picture,

00:46:42 not as a bystander or part of the infrastructure,

00:46:47 but really as part of what’s required

00:46:50 to make the system safe?

00:46:51 This is harder than it sounds.

00:46:53 I think, you know, if you, I mean,

00:46:58 I’m sure you’ve driven before in highways and so on.

00:47:01 It’s really very hard to have to relinquish control

00:47:06 to a machine and then take over when needed.

00:47:10 So I think Tesla’s approach is interesting

00:47:12 because it allows you to periodically establish

00:47:14 some kind of contact with the car.

00:47:18 Toyota, on the other hand, is thinking about

00:47:20 shared autonomy or collaborative autonomy as a paradigm.

00:47:24 If I may argue, these are very, very simple ways

00:47:27 of human robot collaboration,

00:47:29 because the task is pretty boring.

00:47:31 You sit in a vehicle, you go from point A to point B.

00:47:35 I think the more interesting thing to me is,

00:47:37 for example, search and rescue.

00:47:38 I’ve got a human first responder, robot first responders.

00:47:43 I gotta do something.

00:47:45 It’s important.

00:47:46 I have to do it in two minutes.

00:47:47 The building is burning.

00:47:49 There’s been an explosion.

00:47:50 It’s collapsed.

00:47:51 How do I do it?

00:47:52 I think to me, those are the interesting things

00:47:54 where it’s very, very unstructured.

00:47:57 And what’s the role of the human?

00:47:58 What’s the role of the robot?

00:48:00 Clearly, there’s lots of interesting challenges

00:48:02 and there’s a field.

00:48:03 I think we’re gonna make a lot of progress in this area.

00:48:05 Yeah, it’s an exciting form of collaboration.

00:48:07 You’re right.

00:48:08 In autonomous driving, the main enemy

00:48:11 is just boredom of the human.

00:48:13 Yes.

00:48:13 As opposed to in rescue operations,

00:48:15 it’s literally life and death.

00:48:18 And the collaboration enables

00:48:22 the effective completion of the mission.

00:48:23 So it’s exciting.

00:48:24 In some sense, we’re also doing this.

00:48:27 You think about the human driving a car

00:48:30 and almost invariably, the human’s trying

00:48:33 to estimate the state of the car,

00:48:35 they estimate the state of the environment and so on.

00:48:37 But what if the car were to estimate the state of the human?

00:48:40 So for example, I’m sure you have a smartphone

00:48:41 and the smartphone tries to figure out what you’re doing

00:48:44 and send you reminders and oftentimes telling you

00:48:48 to drive to a certain place,

00:48:49 although you have no intention of going there

00:48:51 because it thinks that that’s where you should be

00:48:53 because of some Gmail calendar entry

00:48:57 or something like that.

00:48:58 And it’s trying to constantly figure out who you are,

00:49:01 what you’re doing.

00:49:02 If a car were to do that,

00:49:04 maybe that would make the driver safer

00:49:06 because the car is trying to figure out

00:49:08 is the driver paying attention,

00:49:09 looking at his or her eyes,

00:49:12 looking at circadian movements.

00:49:14 So I think the potential is there,

00:49:16 but from the reverse side,

00:49:18 it’s not robot modeling, but it’s human modeling.

00:49:21 It’s more on the human, right.

00:49:22 And I think the robots can do a very good job

00:49:25 of modeling humans if you really think about the framework

00:49:29 that you have a human sitting in a cockpit,

00:49:32 surrounded by sensors, all staring at him,

00:49:35 in addition to be staring outside,

00:49:37 but also staring at him.

00:49:39 I think there’s a real synergy there.

00:49:40 Yeah, I love that problem

00:49:42 because it’s the new 21st century form of psychology,

00:49:45 actually AI enabled psychology.

00:49:48 A lot of people have sci fi inspired fears

00:49:51 of walking robots like those from Boston Dynamics.

00:49:54 If you just look at shows on Netflix and so on,

00:49:56 or flying robots like those you work with,

00:49:59 how would you, how do you think about those fears?

00:50:03 How would you alleviate those fears?

00:50:05 Do you have inklings, echoes of those same concerns?

00:50:09 You know, anytime we develop a technology

00:50:11 meaning to have positive impact in the world,

00:50:14 there’s always the worry that,

00:50:17 you know, somebody could subvert those technologies

00:50:21 and use it in an adversarial setting.

00:50:23 And robotics is no exception, right?

00:50:25 So I think it’s very easy to weaponize robots.

00:50:29 I think we talk about swarms.

00:50:31 One thing I worry a lot about is,

00:50:33 so, you know, for us to get swarms to work

00:50:35 and do something reliably, it’s really hard.

00:50:38 But suppose I have this challenge

00:50:42 of trying to destroy something,

00:50:44 and I have a swarm of robots,

00:50:45 where only one out of the swarm

00:50:47 needs to get to its destination.

00:50:48 So that suddenly becomes a lot more doable.

00:50:52 And so I worry about, you know,

00:50:54 this general idea of using autonomy

00:50:56 with lots and lots of agents.

00:51:00 I mean, having said that, look,

00:51:01 a lot of this technology is not very mature.

00:51:03 My favorite saying is that

00:51:06 if somebody had to develop this technology,

00:51:10 wouldn’t you rather the good guys do it?

00:51:12 So the good guys have a good understanding

00:51:13 of the technology, so they can figure out

00:51:15 how this technology is being used in a bad way,

00:51:18 or could be used in a bad way and try to defend against it.

00:51:21 So we think a lot about that.

00:51:22 So we have, we’re doing research

00:51:25 on how to defend against swarms, for example.

00:51:28 That’s interesting.

00:51:29 There’s in fact a report by the National Academies

00:51:32 on counter UAS technologies.

00:51:36 This is a real threat,

00:51:38 but we’re also thinking about how to defend against this

00:51:40 and knowing how swarms work.

00:51:42 Knowing how autonomy works is, I think, very important.

00:51:47 So it’s not just politicians?

00:51:49 Do you think engineers have a role in this discussion?

00:51:51 Absolutely.

00:51:52 I think the days where politicians

00:51:55 can be agnostic to technology are gone.

00:51:59 I think every politician needs to be

00:52:03 literate in technology.

00:52:05 And I often say technology is the new liberal art.

00:52:09 Understanding how technology will change your life,

00:52:12 I think is important.

00:52:14 And every human being needs to understand that.

00:52:18 And maybe we can elect some engineers

00:52:20 to office as well on the other side.

00:52:22 What are the biggest open problems in robotics?

00:52:24 And you said we’re in the early days in some sense.

00:52:27 What are the problems we would like to solve in robotics?

00:52:31 I think there are lots of problems, right?

00:52:32 But I would phrase it in the following way.

00:52:36 If you look at the robots we’re building,

00:52:39 they’re still very much tailored towards

00:52:43 doing specific tasks and specific settings.

00:52:46 I think the question of how do you get them to operate

00:52:49 in much broader settings

00:52:53 where things can change in unstructured environments

00:52:58 is up in the air.

00:52:59 So think of self driving cars.

00:53:02 Today, we can build a self driving car in a parking lot.

00:53:05 We can do level five autonomy in a parking lot.

00:53:10 But can you do a level five autonomy

00:53:13 in the streets of Napoli in Italy or Mumbai in India?

00:53:16 No.

00:53:17 So in some sense, when we think about robotics,

00:53:22 we have to think about where they’re functioning,

00:53:25 what kind of environment, what kind of a task.

00:53:27 We have no understanding

00:53:29 of how to put both those things together.

00:53:32 So we’re in the very early days

00:53:34 of applying it to the physical world.

00:53:35 And I was just in Naples actually.

00:53:38 And there’s levels of difficulty and complexity

00:53:42 depending on which area you’re applying it to.

00:53:45 I think so.

00:53:46 And we don’t have a systematic way of understanding that.

00:53:51 Everybody says, just because a computer

00:53:53 can now beat a human at any board game,

00:53:56 we certainly know something about intelligence.

00:53:59 That’s not true.

00:54:01 A computer board game is very, very structured.

00:54:04 It is the equivalent of working in a Henry Ford factory

00:54:08 where things, parts come, you assemble, move on.

00:54:11 It’s a very, very, very structured setting.

00:54:14 That’s the easiest thing.

00:54:15 And we know how to do that.

00:54:18 So you’ve done a lot of incredible work

00:54:20 at the UPenn, University of Pennsylvania, GraspLab.

00:54:23 You’re now Dean of Engineering at UPenn.

00:54:26 What advice do you have for a new bright eyed undergrad

00:54:31 interested in robotics or AI or engineering?

00:54:34 Well, I think there’s really three things.

00:54:36 One is you have to get used to the idea

00:54:40 that the world will not be the same in five years

00:54:42 or four years whenever you graduate, right?

00:54:45 Which is really hard to do.

00:54:46 So this thing about predicting the future,

00:54:48 every one of us needs to be trying

00:54:50 to predict the future always.

00:54:53 Not because you’ll be any good at it,

00:54:54 but by thinking about it,

00:54:56 I think you sharpen your senses and you become smarter.

00:55:00 So that’s number one.

00:55:02 Number two, it’s a corollary of the first piece,

00:55:05 which is you really don’t know what’s gonna be important.

00:55:09 So this idea that I’m gonna specialize in something

00:55:12 which will allow me to go in a particular direction,

00:55:15 it may be interesting,

00:55:16 but it’s important also to have this breadth

00:55:18 so you have this jumping off point.

00:55:22 I think the third thing,

00:55:23 and this is where I think Penn excels.

00:55:25 I mean, we teach engineering,

00:55:27 but it’s always in the context of the liberal arts.

00:55:29 It’s always in the context of society.

00:55:32 As engineers, we cannot afford to lose sight of that.

00:55:35 So I think that’s important.

00:55:37 But I think one thing that people underestimate

00:55:39 when they do robotics

00:55:40 is the importance of mathematical foundations,

00:55:43 the importance of representations.

00:55:47 Not everything can just be solved

00:55:50 by looking for Ross packages on the internet

00:55:52 or to find a deep neural network that works.

00:55:56 I think the representation question is key,

00:55:59 even to machine learning,

00:56:00 where if you ever hope to achieve or get to explainable AI,

00:56:05 somehow there need to be representations

00:56:07 that you can understand.

00:56:09 So if you wanna do robotics,

00:56:11 you should also do mathematics.

00:56:12 And you said liberal arts, a little literature.

00:56:16 If you wanna build a robot,

00:56:17 it should be reading Dostoyevsky.

00:56:19 I agree with that.

00:56:20 Very good.

00:56:21 So Vijay, thank you so much for talking today.

00:56:23 It was an honor.

00:56:24 Thank you.

00:56:25 It was just a very exciting conversation.

00:56:26 Thank you.