vtt/episode_037_large.vtt

WEBVTT

00:00.000 --> 00:03.080
 The following is a conversation with Vijay Kumar.

00:03.080 --> 00:05.760
 He's one of the top roboticists in the world,

00:05.760 --> 00:08.760
 a professor at the University of Pennsylvania,

00:08.760 --> 00:12.880
 a dean of pen engineering, former director of Grasp Lab,

00:12.880 --> 00:15.300
 or the General Robotics Automation Sensing

00:15.300 --> 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 --> 00:25.280
 Vijay is perhaps best known for his work

00:25.280 --> 00:28.520
 in multi robot systems, robot swarms,

00:28.520 --> 00:30.880
 and micro aerial vehicles,

00:30.880 --> 00:34.020
 robots that elegantly cooperate in flight

00:34.020 --> 00:36.200
 under all the uncertainty and challenges

00:36.200 --> 00:38.760
 that the real world conditions present.

00:38.760 --> 00:41.920
 This is the Artificial Intelligence Podcast.

00:41.920 --> 00:44.320
 If you enjoy it, subscribe on YouTube,

00:44.320 --> 00:47.560
 give it five stars on iTunes, support on Patreon,

00:47.560 --> 00:49.500
 or simply connect with me on Twitter

00:49.500 --> 00:53.280
 at Lex Friedman, spelled F R I D M A N.

00:53.280 --> 00:58.280
 And now, here's my conversation with Vijay Kumar.

00:58.700 --> 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.

01:12.040 --> 01:16.700
 It's weighed close to 7,000 pounds,

01:17.520 --> 01:21.620
 and it was powered by hydraulic actuation,

01:21.620 --> 01:26.620
 or it was actuated by hydraulics with 18 motors,

01:27.720 --> 01:32.720
 hydraulic motors, each controlled by an Intel 8085 processor

01:34.160 --> 01:36.680
 and an 8086 co processor.

01:38.120 --> 01:43.120
 And so imagine this huge monster that had 18 joints,

01:44.800 --> 01:46.960
 each controlled by an independent computer,

01:46.960 --> 01:49.320
 and there was a 19th computer that actually did

01:49.320 --> 01:52.320
 the coordination between these 18 joints.

01:52.320 --> 01:53.720
 So I was part of this project,

01:53.720 --> 01:58.720
 and my thesis work was how do you coordinate the 18 legs?

02:02.080 --> 02:06.320
 And in particular, the pressures in the hydraulic cylinders

02:06.320 --> 02:09.200
 to get efficient locomotion.

02:09.200 --> 02:11.640
 It sounds like a giant mess.

02:11.640 --> 02:14.440
 So how difficult is it to make all the motors communicate?

02:14.440 --> 02:17.600
 Presumably, you have to send signals hundreds of times

02:17.600 --> 02:18.440
 a second, or at least.

02:18.440 --> 02:19.880
 So this was not my work,

02:19.880 --> 02:23.960
 but the folks who worked on this wrote what I believe

02:23.960 --> 02:26.640
 to be the first multiprocessor operating system.

02:26.640 --> 02:30.320
 This was in the 80s, and you had to make sure

02:30.320 --> 02:32.800
 that obviously messages got across

02:32.800 --> 02:34.640
 from one joint to another.

02:34.640 --> 02:37.960
 You have to remember the clock speeds on those computers

02:37.960 --> 02:39.660
 were about half a megahertz.

02:39.660 --> 02:42.180
 Right, the 80s.

02:42.180 --> 02:45.320
 So not to romanticize the notion,

02:45.320 --> 02:49.700
 but how did it make you feel to see that robot move?

02:51.080 --> 02:52.280
 It was amazing.

02:52.280 --> 02:55.280
 In hindsight, it looks like, well, we built this thing

02:55.280 --> 02:57.320
 which really should have been much smaller.

02:57.320 --> 02:59.160
 And of course, today's robots are much smaller.

02:59.160 --> 03:03.120
 You look at Boston Dynamics or Ghost Robotics,

03:03.120 --> 03:04.780
 a spinoff from Penn.

03:06.080 --> 03:10.080
 But back then, you were stuck with the substrate you had,

03:10.080 --> 03:13.720
 the compute you had, so things were unnecessarily big.

03:13.720 --> 03:18.040
 But at the same time, and this is just human psychology,

03:18.040 --> 03:20.400
 somehow bigger means grander.

03:21.600 --> 03:23.640
 People never had the same appreciation

03:23.640 --> 03:26.360
 for nanotechnology or nanodevices

03:26.360 --> 03:30.160
 as they do for the Space Shuttle or the Boeing 747.

03:30.160 --> 03:32.760
 Yeah, you've actually done quite a good job

03:32.760 --> 03:36.000
 at illustrating that small is beautiful

03:36.000 --> 03:37.760
 in terms of robotics.

03:37.760 --> 03:42.600
 So what is on that topic is the most beautiful

03:42.600 --> 03:46.200
 or elegant robot in motion 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 in constrained 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

04:14.960 --> 04:17.680
 and you can deform these objects on the fly.

04:17.680 --> 04:21.560
 So in some sense, your toolbox of what you can create

04:21.560 --> 04:23.400
 has suddenly got enhanced.

04:25.240 --> 04:27.800
 And before that, we did the two dimensional version of this.

04:27.800 --> 04:31.680
 So we had ground robots forming patterns and so on.

04:31.680 --> 04:34.960
 So that was not as impressive, that was not as beautiful.

04:34.960 --> 04:36.560
 But if you do it in 3D,

04:36.560 --> 04:40.240
 suspended in midair, and you've got to go back to 2011

04:40.240 --> 04:43.040
 when we did this, now it's actually pretty standard

04:43.040 --> 04:45.600
 to do these things eight years later.

04:45.600 --> 04:47.680
 But back then it was a big accomplishment.

04:47.680 --> 04:50.280
 So the distributed cooperation

04:50.280 --> 04:53.480
 is where beauty emerges in your eyes?

04:53.480 --> 04:55.800
 Well, I think beauty to an engineer is very different

04:55.800 --> 04:59.400
 from beauty to someone who's looking at robots

04:59.400 --> 05:01.240
 from the outside, if you will.

05:01.240 --> 05:04.800
 But what I meant there, so before we said that grand,

05:04.800 --> 05:09.800
 so before we said that grand is associated with size.

05:10.520 --> 05:13.720
 And another way of thinking about this

05:13.720 --> 05:15.600
 is just the physical shape

05:15.600 --> 05:18.400
 and the idea that you can get physical shapes in midair

05:18.400 --> 05:21.560
 and have them deform, that's beautiful.

05:21.560 --> 05:23.040
 But the individual components,

05:23.040 --> 05:24.880
 the agility is beautiful too, right?

05:24.880 --> 05:25.720
 That is true too.

05:25.720 --> 05:28.480
 So then how quickly can you actually manipulate

05:28.480 --> 05:29.560
 these three dimensional shapes

05:29.560 --> 05:31.280
 and the individual components?

05:31.280 --> 05:32.240
 Yes, you're right.

05:32.240 --> 05:36.760
 But by the way, you said UAV, unmanned aerial vehicle.

05:36.760 --> 05:41.760
 What's a good term for drones, UAVs, quad copters?

05:41.840 --> 05:44.560
 Is there a term that's being standardized?

05:44.560 --> 05:45.440
 I don't know if there is.

05:45.440 --> 05:47.920
 Everybody wants to use the word drones.

05:47.920 --> 05:51.080
 And I've often said this, drones to me is a pejorative word.

05:51.080 --> 05:53.960
 It signifies something that's dumb,

05:53.960 --> 05:56.360
 that's pre programmed, that does one little thing

05:56.360 --> 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 --> 06:08.120
 But even unpiloted could be radio controlled,

06:08.120 --> 06:11.560
 could be remotely controlled in many different ways.

06:11.560 --> 06:12.960
 And I think the right word is,

06:12.960 --> 06:15.040
 thinking about it as an aerial robot.

06:15.040 --> 06:19.080
 You also say agile, autonomous, aerial robot, right?

06:19.080 --> 06:22.160
 Yeah, so agility is an attribute, but they don't have to be.

06:23.080 --> 06:24.800
 So what biological system,

06:24.800 --> 06:27.200
 because you've also drawn a lot of inspiration with those.

06:27.200 --> 06:30.360
 I've seen bees and ants that you've talked about.

06:30.360 --> 06:35.240
 What living creatures have you found to be most inspiring

06:35.240 --> 06:38.520
 as an engineer, instructive in your work in robotics?

06:38.520 --> 06:43.440
 To me, so ants are really quite incredible creatures, right?

06:43.440 --> 06:47.880
 So you, I mean, the individuals arguably are very simple

06:47.880 --> 06:52.360
 in how they're built and yet they're incredibly resilient

06:52.360 --> 06:53.960
 as a population.

06:53.960 --> 06:56.760
 And as individuals, they're incredibly robust.

06:56.760 --> 07:00.600
 So, if you take an ant, it's six legs,

07:00.600 --> 07:04.120
 you remove one leg, it still works just fine.

07:04.120 --> 07:05.760
 And it moves along.

07:05.760 --> 07:08.720
 And I don't know that he even realizes it's lost a leg.

07:09.760 --> 07:12.520
 So that's the robustness at the individual ant level.

07:13.400 --> 07:15.360
 But then you look about this instinct

07:15.360 --> 07:17.680
 for self preservation of the colonies

07:17.680 --> 07:20.400
 and they adapt in so many amazing ways.

07:20.400 --> 07:25.400
 You know, transcending gaps by just chaining themselves

07:26.800 --> 07:29.600
 together when you have a flood,

07:29.600 --> 07:32.360
 being able to recruit other teammates

07:32.360 --> 07:34.320
 to carry big morsels of food,

07:35.760 --> 07:38.760
 and then going out in different directions looking for food,

07:38.760 --> 07:43.160
 and then being able to demonstrate consensus,

07:43.160 --> 07:47.040
 even though they don't communicate directly with each other

07:47.040 --> 07:49.080
 the way we communicate with each other.

07:49.080 --> 07:51.880
 In some sense, they also know how to do democracy,

07:51.880 --> 07:53.640
 probably better than what we do.

07:53.640 --> 07:57.000
 Yeah, somehow it's even democracy is emergent.

07:57.000 --> 07:59.120
 It seems like all of the phenomena that we see

07:59.120 --> 08:00.480
 is all emergent.

08:00.480 --> 08:03.560
 It seems like there's no centralized communicator.

08:03.560 --> 08:06.520
 There is, so I think a lot is made about that word,

08:06.520 --> 08:09.640
 emergent, and it means lots of things to different people.

08:09.640 --> 08:10.680
 But you're absolutely right.

08:10.680 --> 08:13.040
 I think as an engineer, you think about

08:13.040 --> 08:17.720
 what element, elemental behaviors

08:17.720 --> 08:21.320
 were primitives you could synthesize

08:21.320 --> 08:25.240
 so that the whole looks incredibly powerful,

08:25.240 --> 08:26.520
 incredibly synergistic,

08:26.520 --> 08:29.520
 the whole definitely being greater than some of the parts,

08:29.520 --> 08:31.480
 and ants are living proof of that.

08:32.480 --> 08:34.960
 So when you see these beautiful swarms

08:34.960 --> 08:37.520
 where there's biological systems of robots,

08:38.520 --> 08:40.200
 do you sometimes think of them

08:40.200 --> 08:44.640
 as a single individual living intelligent organism?

08:44.640 --> 08:47.400
 So it's the same as thinking of our human beings

08:47.400 --> 08:51.160
 are human civilization as one organism,

08:51.160 --> 08:52.960
 or do you still, as an engineer,

08:52.960 --> 08:54.600
 think about the individual components

08:54.600 --> 08:55.440
 and all the engineering

08:55.440 --> 08:57.320
 that went into the individual components?

08:57.320 --> 08:58.640
 Well, that's very interesting.

08:58.640 --> 09:01.480
 So again, philosophically as engineers,

09:01.480 --> 09:05.400
 what we wanna do is to go beyond

09:05.400 --> 09:08.280
 the individual components, the individual units,

09:08.280 --> 09:11.520
 and think about it as a unit, as a cohesive unit,

09:11.520 --> 09:15.120
 without worrying about the individual components.

09:15.120 --> 09:17.760
 If you start obsessing about

09:17.760 --> 09:22.120
 the individual building blocks and what they do,

09:23.320 --> 09:27.960
 you inevitably will find it hard to scale up.

09:27.960 --> 09:29.000
 Just mathematically,

09:29.000 --> 09:31.600
 just think about individual things you wanna model,

09:31.600 --> 09:34.040
 and if you want to have 10 of those,

09:34.040 --> 09:36.440
 then you essentially are taking Cartesian products

09:36.440 --> 09:39.320
 of 10 things, and that makes it really complicated.

09:39.320 --> 09:41.840
 Then to do any kind of synthesis or design

09:41.840 --> 09:44.200
 in that high dimension space is really hard.

09:44.200 --> 09:45.800
 So the right way to do this

09:45.800 --> 09:49.040
 is to think about the individuals in a clever way

09:49.040 --> 09:51.120
 so that at the higher level,

09:51.120 --> 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, do 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 --> 10:19.760
 I mean, you think about, we build planes,

10:19.760 --> 10:21.240
 or we build iPhones,

10:22.240 --> 10:26.280
 and we know that by taking individual components,

10:26.280 --> 10:30.080
 well engineered components 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.

10:34.440 --> 10:36.880
 So that's ingrained, I would claim,

10:36.880 --> 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 --> 11:00.560
 And so you really want them to be able to reestablish

11:02.040 --> 11:03.840
 communication with their neighbors.

11:03.840 --> 11:08.840
 You want them to rethink their strategy for group behavior.

11:08.840 --> 11:10.760
 You want them to reorganize.

11:12.200 --> 11:15.920
 And that's where I think a lot of the challenges lie.

11:15.920 --> 11:18.160
 So just at a high level,

11:18.160 --> 11:20.880
 what does it take for a bunch of,

11:22.200 --> 11:24.440
 what should we call them, flying robots,

11:24.440 --> 11:26.680
 to create a formation?

11:26.680 --> 11:28.680
 Just for people who are not familiar

11:28.680 --> 11:32.760
 with robotics in general, how much information is needed?

11:32.760 --> 11:35.840
 How do you even make it happen

11:35.840 --> 11:39.520
 without a centralized controller?

11:39.520 --> 11:41.080
 So, I mean, there are a couple of different ways

11:41.080 --> 11:43.160
 of looking at this.

11:43.160 --> 11:45.680
 If you are a purist,

11:45.680 --> 11:50.680
 you think of it as a way of recreating what nature does.

11:53.560 --> 11:58.440
 So nature forms groups for several reasons,

11:58.440 --> 12:02.000
 but mostly it's because of this instinct

12:02.000 --> 12:05.680
 that organisms have of preserving their colonies,

12:05.680 --> 12:09.520
 their population, which means what?

12:09.520 --> 12:12.920
 You need shelter, you need food, you need to procreate,

12:12.920 --> 12:14.760
 and that's basically it.

12:14.760 --> 12:18.440
 So the kinds of interactions you see are all organic.

12:18.440 --> 12:19.760
 They're all local.

12:20.760 --> 12:24.080
 And the only information that they share,

12:24.080 --> 12:27.520
 and mostly it's indirectly, is to, again,

12:27.520 --> 12:30.000
 preserve the herd or the flock,

12:30.000 --> 12:35.000
 or the swarm, and either by looking for new sources of food

12:37.480 --> 12:39.440
 or looking for new shelters, right?

12:39.440 --> 12:40.280
 Right.

12:41.240 --> 12:45.360
 As engineers, when we build swarms, we have a mission.

12:46.560 --> 12:51.560
 And when you think of a mission, and it involves mobility,

12:52.480 --> 12:55.000
 most often it's described in some kind

12:55.000 --> 12:56.880
 of a global coordinate system.

12:56.880 --> 12:59.440
 As a human, as an operator, as a commander,

12:59.440 --> 13:03.560
 or as a collaborator, I have my coordinate system,

13:03.560 --> 13:06.640
 and I want the robots to be consistent with that.

13:07.600 --> 13:11.240
 So I might think of it slightly differently.

13:11.240 --> 13:15.440
 I might want the robots to recognize that coordinate system,

13:15.440 --> 13:17.720
 which means not only do they have to think locally

13:17.720 --> 13:19.600
 in terms of who their immediate neighbors are,

13:19.600 --> 13:20.920
 but they have to be cognizant

13:20.920 --> 13:24.040
 of what the global environment is.

13:24.040 --> 13:27.040
 They have to be cognizant of what the global environment

13:27.040 --> 13:28.280
 looks like.

13:28.280 --> 13:31.040
 So if I say, surround this building

13:31.040 --> 13:33.240
 and protect this from intruders,

13:33.240 --> 13:35.600
 well, they're immediately in a building centered

13:35.600 --> 13:37.040
 coordinate system, and I have to tell them

13:37.040 --> 13:38.680
 where the building is.

13:38.680 --> 13:40.040
 And they're globally collaborating

13:40.040 --> 13:41.280
 on the map of that building.

13:41.280 --> 13:44.160
 They're maintaining some kind of global,

13:44.160 --> 13:45.480
 not just in the frame of the building,

13:45.480 --> 13:49.000
 but there's information that's ultimately being built up

13:49.000 --> 13:53.280
 explicitly as opposed to kind of implicitly,

13:53.280 --> 13:54.360
 like nature might.

13:54.360 --> 13:55.200
 Correct, correct.

13:55.200 --> 13:57.680
 So in some sense, nature is very, very sophisticated,

13:57.680 --> 14:01.880
 but the tasks that nature solves or needs to solve

14:01.880 --> 14:05.160
 are very different from the kind of engineered tasks,

14:05.160 --> 14:09.760
 artificial tasks that we are forced to address.

14:09.760 --> 14:12.520
 And again, there's nothing preventing us

14:12.520 --> 14:15.160
 from solving these other problems,

14:15.160 --> 14:16.600
 but ultimately it's about impact.

14:16.600 --> 14:19.360
 You want these swarms to do something useful.

14:19.360 --> 14:24.360
 And so you're kind of driven into this very unnatural,

14:24.640 --> 14:25.480
 if you will.

14:25.480 --> 14:29.160
 Unnatural, meaning not like how nature does, setting.

14:29.160 --> 14:31.920
 And it's probably a little bit more expensive

14:31.920 --> 14:33.760
 to do it the way nature does,

14:33.760 --> 14:37.560
 because nature is less sensitive

14:37.560 --> 14:39.480
 to the loss of the individual.

14:39.480 --> 14:42.280
 And cost wise in robotics,

14:42.280 --> 14:45.480
 I think you're more sensitive to losing individuals.

14:45.480 --> 14:49.000
 I think that's true, although if you look at the price

14:49.000 --> 14:51.520
 to performance ratio of robotic components,

14:51.520 --> 14:54.720
 it's coming down dramatically, right?

14:54.720 --> 14:56.040
 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 --> 15:05.040
 the cost of individuals would really become insignificant.

15:05.040 --> 15:07.640
 So let's step back at a high level view,

15:07.640 --> 15:12.480
 the impossible question of what kind of, as an overview,

15:12.480 --> 15:14.400
 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 they have limited autonomy,

15:31.800 --> 15:34.760
 which is the ability to go from point A to point B.

15:34.760 --> 15:37.080
 And even the more sophisticated now,

15:37.080 --> 15:40.400
 sophisticated vehicles can do autonomous takeoff

15:40.400 --> 15:41.760
 and landing.

15:41.760 --> 15:44.360
 And those usually have wings and they're heavy.

15:44.360 --> 15:45.360
 Usually they're wings,

15:45.360 --> 15:47.440
 but then there's nothing preventing us from doing this

15:47.440 --> 15:49.000
 for helicopters as well.

15:49.000 --> 15:52.480
 There are many military organizations

15:52.480 --> 15:56.560
 that have 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 --> 16:02.840
 and it's actually very similar.

16:02.840 --> 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 --> 16:18.640
 would they have happened if the plane were truly autonomous?

16:18.640 --> 16:21.960
 And I think you'll find that in many of the cases,

16:21.960 --> 16:24.600
 because of pilot error, we made silly decisions.

16:24.600 --> 16:26.960
 And so in some sense, even in air traffic,

16:26.960 --> 16:29.800
 commercial air traffic, there's a lot of applications,

16:29.800 --> 16:33.960
 although we only see autonomy being enabled

16:33.960 --> 16:38.960
 at very high altitudes when the plane is an autopilot.

16:38.960 --> 16:41.960
 The plane is an autopilot.

16:41.960 --> 16:42.800
 There's still a role for the human

16:42.800 --> 16:47.640
 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 --> 16:53.480
 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,

17:04.200 --> 17:09.200
 can we make robots that will be able to make decisions

17:10.360 --> 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 have driven in a city, try to use GPS navigation,

17:30.480 --> 17:32.760
 in tall buildings, you immediately lose GPS.

17:32.760 --> 17:36.280
 And so that's not a very sophisticated way

17:36.280 --> 17:37.840
 of building autonomy.

17:37.840 --> 17:39.560
 I think the second piece of infrastructure

17:39.560 --> 17:41.920
 they rely on is communications.

17:41.920 --> 17:46.200
 Again, it's very easy to jam communications.

17:47.360 --> 17:51.320
 In fact, if you use wifi, you know that wifi signals

17:51.320 --> 17:53.520
 drop out, cell signals drop out.

17:53.520 --> 17:56.800
 So to rely on something like that is not good.

17:58.560 --> 18:01.200
 The third form of infrastructure we use,

18:01.200 --> 18:02.920
 and I hate to call it infrastructure,

18:02.920 --> 18:06.360
 but it is that, in the sense of robots, is people.

18:06.360 --> 18:08.640
 So you could rely on somebody to pilot you.

18:09.960 --> 18:11.600
 And so the question you wanna ask is,

18:11.600 --> 18:14.760
 if there are no pilots, there's no communications

18:14.760 --> 18:18.720
 with any base station, 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 --> 18:28.240
 a priori model of what might happen in the future,

18:28.240 --> 18:29.560
 can robots navigate?

18:29.560 --> 18:31.480
 So that is true autonomy.

18:31.480 --> 18:34.160
 So that's true autonomy, and we're talking about,

18:34.160 --> 18:36.880
 you mentioned like military application of drones.

18:36.880 --> 18:38.320
 Okay, so what else is there?

18:38.320 --> 18:42.080
 You talk about agile, autonomous flying robots,

18:42.080 --> 18:45.680
 aerial robots, so that's a different kind of,

18:45.680 --> 18:48.160
 it's not winged, it's not big, at least it's small.

18:48.160 --> 18:50.840
 So I use the word agility mostly,

18:50.840 --> 18:53.520
 or at least we're motivated to do agile robots,

18:53.520 --> 18:58.000
 mostly because robots can operate

18:58.000 --> 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 --> 19:28.240
 or to look for, to count fruits,

19:28.240 --> 19:31.240
 to measure the tree trunks.

19:31.240 --> 19:33.240
 These are things we do, by the way.

19:33.240 --> 19:35.400
 There's some cool agriculture stuff you've shown

19:35.400 --> 19:37.080
 in the past, it's really awesome.

19:37.080 --> 19:40.360
 So in those kinds of settings, you do need that agility.

19:40.360 --> 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 --> 19:48.000
 What it really means is you see the unexpected

19:48.000 --> 19:51.480
 and you're able to maneuver in a safe way,

19:51.480 --> 19:55.400
 and in a way that gets you the most information

19:55.400 --> 19:57.640
 about the thing you're trying to do.

19:57.640 --> 20:00.440
 By the way, you may be the only person

20:00.440 --> 20:04.200
 who, in a TED Talk, has used a math equation,

20:04.200 --> 20:07.600
 which is amazing, people should go see one of your TED Talks.

20:07.600 --> 20:08.800
 Actually, it's very interesting,

20:08.800 --> 20:12.400
 because the TED curator, Chris Anderson,

20:12.400 --> 20:15.360
 told me, you can't show math.

20:15.360 --> 20:18.200
 And I thought about it, but that's who I am.

20:18.200 --> 20:20.760
 I mean, that's our work.

20:20.760 --> 20:25.760
 And so I felt compelled to give the audience a taste

20:25.760 --> 20:27.640
 for at least some math.

20:27.640 --> 20:32.640
 So on that point, simply, what does it take

20:32.880 --> 20:37.360
 to make a thing with four motors fly, a quadcopter,

20:37.360 --> 20:40.640
 one of these little flying robots?

20:41.760 --> 20:43.960
 How hard is it to make it fly?

20:43.960 --> 20:46.560
 How do you coordinate the four motors?

20:46.560 --> 20:51.560
 How do you convert those motors into actual movement?

20:52.600 --> 20:54.800
 So this is an interesting question.

20:54.800 --> 20:58.080
 We've been trying to do this since 2000.

20:58.080 --> 21:00.560
 It is a commentary on the sensors

21:00.560 --> 21:02.080
 that were available back then,

21:02.080 --> 21:04.280
 the computers that were available back then.

21:05.560 --> 21:10.280
 And a number of things happened between 2000 and 2007.

21:11.520 --> 21:14.120
 One is the advances in computing,

21:14.120 --> 21:16.760
 which is, so we all know about Moore's Law,

21:16.760 --> 21:19.680
 but I think 2007 was a tipping point,

21:19.680 --> 21:22.720
 the year of the iPhone, the year of the cloud.

21:22.720 --> 21:24.640
 Lots of things happened in 2007.

21:25.600 --> 21:27.600
 But going back even further,

21:27.600 --> 21:31.360
 inertial measurement units as a sensor really matured.

21:31.360 --> 21:33.040
 Again, lots of reasons for that.

21:33.920 --> 21:35.400
 Certainly, there's a lot of federal funding,

21:35.400 --> 21:37.360
 particularly DARPA in the US,

21:38.320 --> 21:42.760
 but they didn't anticipate this boom in IMUs.

21:42.760 --> 21:46.560
 But if you look, subsequently what happened

21:46.560 --> 21:50.040
 is that every car manufacturer had to put an airbag in,

21:50.040 --> 21:52.600
 which meant you had to have an accelerometer on board.

21:52.600 --> 21:55.000
 And so that drove down the price to performance ratio.

21:55.000 --> 21:56.880
 Wow, I should know this.

21:56.880 --> 21:57.960
 That's very interesting.

21:57.960 --> 21:59.360
 That's very interesting, the connection there.

21:59.360 --> 22:01.320
 And that's why research is very,

22:01.320 --> 22:03.280
 it's very hard to predict the outcomes.

22:04.840 --> 22:07.640
 And again, the federal government spent a ton of money

22:07.640 --> 22:12.280
 on things that they thought were useful for resonators,

22:12.280 --> 22:16.840
 but it ended up enabling these small UAVs, which is great,

22:16.840 --> 22:18.520
 because I could have never raised that much money

22:18.520 --> 22:20.760
 and sold this project,

22:20.760 --> 22:22.200
 hey, we want to build these small UAVs.

22:22.200 --> 22:25.440
 Can you actually fund the development of low cost IMUs?

22:25.440 --> 22:27.600
 So why do you need an IMU on an IMU?

22:27.600 --> 22:31.000
 So I'll come back to that.

22:31.000 --> 22:33.320
 So in 2007, 2008, we were able to build these.

22:33.320 --> 22:35.200
 And then the question you're asking was a good one.

22:35.200 --> 22:40.240
 How do you coordinate the motors to develop this?

22:40.240 --> 22:43.880
 But over the last 10 years, everything is commoditized.

22:43.880 --> 22:46.240
 A high school kid today can pick up

22:46.240 --> 22:50.560
 a Raspberry Pi kit and build this.

22:50.560 --> 22:53.200
 All the low levels functionality is all automated.

22:54.160 --> 22:56.360
 But basically at some level,

22:56.360 --> 23:01.360
 you have to drive the motors at the right RPMs,

23:01.360 --> 23:03.680
 the right velocity,

23:04.560 --> 23:07.480
 in order to generate the right amount of thrust,

23:07.480 --> 23:10.360
 in order to position it and orient it in a way

23:10.360 --> 23:12.840
 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 --> 23:37.480
 or velocity information.

23:37.480 --> 23:39.360
 For example, when we walk,

23:39.360 --> 23:41.560
 we implicitly have this information

23:41.560 --> 23:45.840
 because we kind of know what our stride length is.

23:46.720 --> 23:51.480
 We also are looking at images fly past our retina,

23:51.480 --> 23:54.280
 if you will, and so we can estimate velocity.

23:54.280 --> 23:56.360
 We also have accelerometers in our head,

23:56.360 --> 23:59.160
 and we're able to integrate all these pieces of information

23:59.160 --> 24:02.360
 to determine where we are as we walk.

24:02.360 --> 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 --> 24:11.640
 or other sensor that's measuring velocity,

24:12.560 --> 24:15.800
 and then you need some kind of a global reference frame

24:15.800 --> 24:19.520
 if you really want to think about doing something

24:19.520 --> 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 --> 24:32.640
 is what allows you to, first of all, fly and hover,

24:32.640 --> 24:35.600
 and then you can change the orientation

24:35.600 --> 24:37.600
 and the velocity and so on.

24:37.600 --> 24:38.440
 Exactly, exactly.

24:38.440 --> 24:40.320
 So it's a bunch of degrees of freedom

24:40.320 --> 24:41.160
 that you're complaining about.

24:41.160 --> 24:42.200
 There's six degrees of freedom,

24:42.200 --> 24:44.920
 but you only have four inputs, the four motors.

24:44.920 --> 24:49.920
 And it turns out to be a remarkably versatile configuration.

24:50.920 --> 24:53.080
 You think at first, well, I only have four motors,

24:53.080 --> 24:55.000
 how do I go sideways?

24:55.000 --> 24:57.280
 But it's not too hard to say, well, if I tilt myself,

24:57.280 --> 25:00.440
 I can go sideways, and then you have four motors

25:00.440 --> 25:03.320
 pointing up, how do I rotate in place

25:03.320 --> 25:05.360
 about a vertical axis?

25:05.360 --> 25:07.800
 Well, you rotate them at different speeds

25:07.800 --> 25:09.720
 and that generates reaction moments

25:09.720 --> 25:11.520
 and that allows you to turn.

25:11.520 --> 25:14.960
 So it's actually a pretty, it's an optimal configuration

25:14.960 --> 25:17.040
 from an engineer standpoint.

25:18.360 --> 25:23.360
 It's very simple, very cleverly done, and very versatile.

25:23.360 --> 25:27.240
 So if you could step back to a time,

25:27.240 --> 25:30.000
 so I've always known flying robots as,

25:31.040 --> 25:35.760
 to me, it was natural that a quadcopter should fly.

25:35.760 --> 25:37.880
 But when you first started working with it,

25:38.800 --> 25:42.000
 how surprised are you that you can make,

25:42.000 --> 25:45.520
 do so much with the four motors?

25:45.520 --> 25:47.600
 How surprising is it that you can make this thing fly,

25:47.600 --> 25:49.760
 first of all, that you can make it hover,

25:49.760 --> 25:52.000
 that you can add control to it?

25:52.000 --> 25:55.080
 Firstly, this is not, the four motor configuration

25:55.080 --> 25:56.400
 is not ours.

25:56.400 --> 25:59.320
 You can, it has at least a hundred year history.

26:00.320 --> 26:04.160
 And various people, various people try to get quadrotors

26:04.160 --> 26:06.840
 to fly without much success.

26:08.480 --> 26:10.760
 As I said, we've been working on this since 2000.

26:10.760 --> 26:14.400
 Our first designs were, well, this is way too complicated.

26:14.400 --> 26:18.480
 Why not we try to get an omnidirectional flying robot?

26:18.480 --> 26:21.760
 So our early designs, we had eight rotors.

26:21.760 --> 26:25.200
 And so these eight rotors were arranged uniformly

26:26.600 --> 26:28.000
 on a sphere, if you will.

26:28.000 --> 26:30.440
 So you can imagine a symmetric configuration.

26:30.440 --> 26:33.280
 And so you should be able to fly anywhere.

26:33.280 --> 26:36.240
 But the real challenge we had is the strength to weight ratio

26:36.240 --> 26:37.080
 is not enough.

26:37.080 --> 26:39.680
 And of course, we didn't have the sensors and so on.

26:40.520 --> 26:43.040
 So everybody knew, or at least the people

26:43.040 --> 26:44.800
 who worked with rotorcrafts knew,

26:44.800 --> 26:46.520
 four rotors will get it done.

26:47.520 --> 26:49.400
 So that was not our idea.

26:49.400 --> 26:52.800
 But it took a while before we could actually do

26:52.800 --> 26:56.920
 the onboard sensing and the computation that was needed

26:56.920 --> 27:01.000
 for the kinds of agile maneuvering that we wanted to do

27:01.000 --> 27:03.000
 in our little aerial robots.

27:03.000 --> 27:07.560
 And that only happened between 2007 and 2009 in our lab.

27:07.560 --> 27:09.960
 Yeah, and you have to send the signal

27:09.960 --> 27:12.480
 maybe a hundred times a second.

27:12.480 --> 27:15.960
 So the compute there, everything has to come down in price.

27:15.960 --> 27:20.960
 And what are the steps of getting from point A to point B?

27:21.720 --> 27:25.200
 So we just talked about like local control.

27:25.200 --> 27:30.200
 But if all the kind of cool dancing in the air

27:30.840 --> 27:34.520
 that I've seen you show, how do you make it happen?

27:34.520 --> 27:37.360
 How do you make a trajectory?

27:37.360 --> 27:40.520
 First of all, okay, figure out a trajectory.

27:40.520 --> 27:41.680
 So plan a trajectory.

27:41.680 --> 27:44.400
 And then how do you make that trajectory happen?

27:44.400 --> 27:47.280
 Yeah, I think planning is a very fundamental problem

27:47.280 --> 27:48.120
 in robotics.

27:48.120 --> 27:50.800
 I think 10 years ago it was an esoteric thing,

27:50.800 --> 27:53.040
 but today with self driving cars,

27:53.040 --> 27:55.840
 everybody can understand this basic idea

27:55.840 --> 27:57.920
 that a car sees a whole bunch of things

27:57.920 --> 28:00.320
 and it has to keep a lane or maybe make a right turn

28:00.320 --> 28:01.280
 or switch lanes.

28:01.280 --> 28:02.680
 It has to plan a trajectory.

28:02.680 --> 28:03.560
 It has to be safe.

28:03.560 --> 28:04.840
 It has to be efficient.

28:04.840 --> 28:06.640
 So everybody's familiar with that.

28:06.640 --> 28:10.240
 That's kind of the first step that you have to think about

28:10.240 --> 28:14.800
 when you say autonomy.

28:14.800 --> 28:19.120
 And so for us, it's about finding smooth motions,

28:19.120 --> 28:21.320
 motions that are safe.

28:21.320 --> 28:22.880
 So we think about these two things.

28:22.880 --> 28:24.680
 One is optimality, one is safety.

28:24.680 --> 28:27.200
 Clearly you cannot compromise safety.

28:28.440 --> 28:31.360
 So you're looking for safe, optimal motions.

28:31.360 --> 28:34.480
 The other thing you have to think about is

28:34.480 --> 28:38.160
 can you actually compute a reasonable trajectory

28:38.160 --> 28:40.760
 in a small amount of time?

28:40.760 --> 28:42.280
 Cause you have a time budget.

28:42.280 --> 28:45.160
 So the optimal becomes suboptimal,

28:45.160 --> 28:50.160
 but in our lab we focus on synthesizing smooth trajectory

28:51.160 --> 28:53.000
 that satisfy all the constraints.

28:53.000 --> 28:57.120
 In other words, don't violate any safety constraints

28:58.440 --> 29:02.880
 and is as efficient as possible.

29:02.880 --> 29:04.360
 And when I say efficient,

29:04.360 --> 29:06.600
 it could mean I want to get from point A to point B

29:06.600 --> 29:08.360
 as quickly as possible,

29:08.360 --> 29:11.840
 or I want to get to it as gracefully as possible,

29:12.840 --> 29:15.960
 or I want to consume as little energy as possible.

29:15.960 --> 29:18.240
 But always staying within the safety constraints.

29:18.240 --> 29:22.800
 But yes, always finding a safe trajectory.

29:22.800 --> 29:25.040
 So there's a lot of excitement and progress

29:25.040 --> 29:27.360
 in the field of machine learning

29:27.360 --> 29:29.360
 and reinforcement learning

29:29.360 --> 29:32.200
 and the neural network variant of that

29:32.200 --> 29:33.920
 with deep reinforcement learning.

29:33.920 --> 29:36.360
 Do you see a role of machine learning

29:36.360 --> 29:40.560
 in, so a lot of the success of flying robots

29:40.560 --> 29:42.320
 did not rely on machine learning,

29:42.320 --> 29:45.040
 except for maybe a little bit of the perception

29:45.040 --> 29:46.600
 on the computer vision side.

29:46.600 --> 29:48.440
 On the control side and the planning,

29:48.440 --> 29:50.400
 do you see there's a role in the future

29:50.400 --> 29:51.680
 for machine learning?

29:51.680 --> 29:53.800
 So let me disagree a little bit with you.

29:53.800 --> 29:56.800
 I think we never perhaps called out in my work,

29:56.800 --> 29:57.720
 called out learning,

29:57.720 --> 30:00.600
 but even this very simple idea of being able to fly

30:00.600 --> 30:02.200
 through a constrained space.

30:02.200 --> 30:05.680
 The first time you try it, you'll invariably,

30:05.680 --> 30:08.440
 you might get it wrong if the task is challenging.

30:08.440 --> 30:12.200
 And the reason is to get it perfectly right,

30:12.200 --> 30:14.600
 you have to model everything in the environment.

30:15.600 --> 30:19.960
 And flying is notoriously hard to model.

30:19.960 --> 30:24.960
 There are aerodynamic effects that we constantly discover.

30:26.520 --> 30:29.440
 Even just before I was talking to you,

30:29.440 --> 30:33.440
 I was talking to a student about how blades flap

30:33.440 --> 30:35.320
 when they fly.

30:35.320 --> 30:40.320
 And that ends up changing how a rotorcraft

30:40.880 --> 30:43.960
 is accelerated in the angular direction.

30:43.960 --> 30:46.360
 Does he use like micro flaps or something?

30:46.360 --> 30:47.280
 It's not micro flaps.

30:47.280 --> 30:49.640
 So we assume that each blade is rigid,

30:49.640 --> 30:51.720
 but actually it flaps a little bit.

30:51.720 --> 30:52.880
 It bends.

30:52.880 --> 30:53.720
 Interesting, yeah.

30:53.720 --> 30:56.040
 And so the models rely on the fact,

30:56.040 --> 30:58.640
 on the assumption that they're not rigid.

30:58.640 --> 31:00.640
 On the assumption that they're actually rigid,

31:00.640 --> 31:02.240
 but that's not true.

31:02.240 --> 31:03.720
 If you're flying really quickly,

31:03.720 --> 31:06.920
 these effects become significant.

31:06.920 --> 31:09.240
 If you're flying close to the ground,

31:09.240 --> 31:12.160
 you get pushed off by the ground, right?

31:12.160 --> 31:14.920
 Something which every pilot knows when he tries to land

31:14.920 --> 31:18.000
 or she tries to land, this is called a ground effect.

31:18.920 --> 31:21.000
 Something very few pilots think about

31:21.000 --> 31:23.040
 is what happens when you go close to a ceiling

31:23.040 --> 31:25.320
 or you get sucked into a ceiling.

31:25.320 --> 31:26.880
 There are very few aircrafts

31:26.880 --> 31:29.520
 that fly close to any kind of ceiling.

31:29.520 --> 31:33.520
 Likewise, when you go close to a wall,

31:33.520 --> 31:35.720
 there are these wall effects.

31:35.720 --> 31:37.680
 And if you've gone on a train

31:37.680 --> 31:39.600
 and you pass another train that's traveling

31:39.600 --> 31:42.400
 in the opposite direction, you feel the buffeting.

31:42.400 --> 31:45.400
 And so these kinds of microclimates

31:45.400 --> 31:47.880
 affect our UAV significantly.

31:47.880 --> 31:48.720
 So if you want...

31:48.720 --> 31:50.640
 And they're impossible to model, essentially.

31:50.640 --> 31:52.480
 I wouldn't say they're impossible to model,

31:52.480 --> 31:54.880
 but the level of sophistication you would need

31:54.880 --> 31:58.600
 in the model and the software would be tremendous.

32:00.000 --> 32:02.920
 Plus, to get everything right would be awfully tedious.

32:02.920 --> 32:05.080
 So the way we do this is over time,

32:05.080 --> 32:09.000
 we figure out how to adapt to these conditions.

32:10.360 --> 32:13.160
 So early on, we use the form of learning

32:13.160 --> 32:15.760
 that we call iterative learning.

32:15.760 --> 32:18.600
 So this idea, if you want to perform a task,

32:18.600 --> 32:22.120
 there are a few things that you need to change

32:22.120 --> 32:24.960
 and iterate over a few parameters

32:24.960 --> 32:29.280
 that over time you can figure out.

32:29.280 --> 32:33.400
 So I could call it policy gradient reinforcement learning,

32:33.400 --> 32:34.920
 but actually it was just iterative learning.

32:34.920 --> 32:36.000
 Iterative learning.

32:36.000 --> 32:37.800
 And so this was there way back.

32:37.800 --> 32:39.440
 I think what's interesting is,

32:39.440 --> 32:41.640
 if you look at autonomous vehicles today,

32:43.120 --> 32:45.680
 learning occurs, could occur in two pieces.

32:45.680 --> 32:47.960
 One is perception, understanding the world.

32:47.960 --> 32:50.080
 Second is action, taking actions.

32:50.080 --> 32:52.240
 Everything that I've seen that is successful

32:52.240 --> 32:54.360
 is on the perception side of things.

32:54.360 --> 32:55.400
 So in computer vision,

32:55.400 --> 32:57.840
 we've made amazing strides in the last 10 years.

32:57.840 --> 33:01.640
 So recognizing objects, actually detecting objects,

33:01.640 --> 33:06.400
 classifying them and tagging them in some sense,

33:06.400 --> 33:07.440
 annotating them.

33:07.440 --> 33:09.640
 This is all done through machine learning.

33:09.640 --> 33:12.160
 On the action side, on the other hand,

33:12.160 --> 33:13.720
 I don't know of any examples

33:13.720 --> 33:15.560
 where there are fielded systems

33:15.560 --> 33:17.560
 where we actually learn

33:17.560 --> 33:20.560
 the right behavior.

33:20.560 --> 33:22.760
 Outside of single demonstration is successful.

33:22.760 --> 33:24.640
 In the laboratory, this is the holy grail.

33:24.640 --> 33:26.040
 Can you do end to end learning?

33:26.040 --> 33:28.800
 Can you go from pixels to motor currents?

33:30.200 --> 33:31.600
 This is really, really hard.

33:32.800 --> 33:35.080
 And I think if you go forward,

33:35.080 --> 33:37.600
 the right way to think about these things

33:37.600 --> 33:40.720
 is data driven approaches,

33:40.720 --> 33:42.400
 learning based approaches,

33:42.400 --> 33:45.280
 in concert with model based approaches,

33:45.280 --> 33:47.320
 which is the traditional way of doing things.

33:47.320 --> 33:48.720
 So I think there's a piece,

33:48.720 --> 33:51.400
 there's a role for each of these methodologies.

33:51.400 --> 33:52.440
 So what do you think,

33:52.440 --> 33:53.880
 just jumping out on topic

33:53.880 --> 33:56.200
 since you mentioned autonomous vehicles,

33:56.200 --> 33:58.480
 what do you think are the limits on the perception side?

33:58.480 --> 34:01.080
 So I've talked to Elon Musk

34:01.080 --> 34:03.320
 and there on the perception side,

34:03.320 --> 34:05.960
 they're using primarily computer vision

34:05.960 --> 34:08.080
 to perceive the environment.

34:08.080 --> 34:09.760
 In your work with,

34:09.760 --> 34:12.560
 because you work with the real world a lot

34:12.560 --> 34:13.720
 and the physical world,

34:13.720 --> 34:15.800
 what are the limits of computer vision?

34:15.800 --> 34:18.000
 Do you think we can solve autonomous vehicles

34:19.160 --> 34:20.880
 on the perception side,

34:20.880 --> 34:24.240
 focusing on vision alone and machine learning?

34:24.240 --> 34:27.480
 So, we also have a spinoff company,

34:27.480 --> 34:31.840
 Exxon Technologies that works underground in mines.

34:31.840 --> 34:35.600
 So you go into mines, they're dark, they're dirty.

34:36.480 --> 34:38.600
 You fly in a dirty area,

34:38.600 --> 34:41.120
 there's stuff you kick up from by the propellers,

34:41.120 --> 34:42.720
 the downwash kicks up dust.

34:42.720 --> 34:45.520
 I challenge you to get a computer vision algorithm

34:45.520 --> 34:46.680
 to work there.

34:46.680 --> 34:49.600
 So we use LIDARs in that setting.

34:51.200 --> 34:55.360
 Indoors and even outdoors when we fly through fields,

34:55.360 --> 34:57.120
 I think there's a lot of potential

34:57.120 --> 34:59.960
 for just solving the problem using computer vision alone.

35:01.240 --> 35:02.760
 But I think the bigger question is,

35:02.760 --> 35:06.160
 can you actually solve

35:06.160 --> 35:09.440
 or can you actually identify all the corner cases

35:09.440 --> 35:13.920
 using a single sensing modality and using learning alone?

35:13.920 --> 35:15.400
 So what's your intuition there?

35:15.400 --> 35:17.920
 So look, if you have a corner case

35:17.920 --> 35:20.000
 and your algorithm doesn't work,

35:20.000 --> 35:23.200
 your instinct is to go get data about the corner case

35:23.200 --> 35:26.640
 and patch it up, learn how to deal with that corner case.

35:27.640 --> 35:32.040
 But at some point, this is gonna saturate,

35:32.040 --> 35:34.200
 this approach is not viable.

35:34.200 --> 35:38.000
 So today, computer vision algorithms can detect

35:38.000 --> 35:41.360
 90% of the objects or can detect objects 90% of the time,

35:41.360 --> 35:43.920
 classify them 90% of the time.

35:43.920 --> 35:47.960
 Cats on the internet probably can do 95%, I don't know.

35:47.960 --> 35:52.520
 But to get from 90% to 99%, you need a lot more data.

35:52.520 --> 35:54.480
 And then I tell you, well, that's not enough

35:54.480 --> 35:56.680
 because I have a safety critical application,

35:56.680 --> 36:00.160
 I wanna go from 99% to 99.9%.

36:00.160 --> 36:01.600
 That's even more data.

36:01.600 --> 36:08.600
 So I think if you look at wanting accuracy on the X axis

36:09.600 --> 36:14.080
 and look at the amount of data on the Y axis,

36:14.080 --> 36:16.440
 I believe that curve is an exponential curve.

36:16.440 --> 36:19.480
 Wow, okay, it's even hard if it's linear.

36:19.480 --> 36:20.800
 It's hard if it's linear, totally,

36:20.800 --> 36:22.560
 but I think it's exponential.

36:22.560 --> 36:24.120
 And the other thing you have to think about

36:24.120 --> 36:29.600
 is that this process is a very, very power hungry process

36:29.600 --> 36:32.880
 to run data farms or servers.

36:32.880 --> 36:34.600
 Power, do you mean literally power?

36:34.600 --> 36:36.600
 Literally power, literally power.

36:36.600 --> 36:41.760
 So in 2014, five years ago, and I don't have more recent data,

36:41.760 --> 36:48.360
 2% of US electricity consumption was from data farms.

36:48.360 --> 36:52.080
 So we think about this as an information science

36:52.080 --> 36:54.240
 and information processing problem.

36:54.240 --> 36:57.840
 Actually, it is an energy processing problem.

36:57.840 --> 37:00.440
 And so unless we figured out better ways of doing this,

37:00.440 --> 37:02.440
 I don't think this is viable.

37:02.440 --> 37:06.600
 So talking about driving, which is a safety critical application

37:06.600 --> 37:10.440
 and some aspect of flight is safety critical,

37:10.440 --> 37:12.960
 maybe philosophical question, maybe an engineering one,

37:12.960 --> 37:15.000
 what problem do you think is harder to solve,

37:15.000 --> 37:18.120
 autonomous driving or autonomous flight?

37:18.120 --> 37:19.920
 That's a really interesting question.

37:19.920 --> 37:25.440
 I think autonomous flight has several advantages

37:25.440 --> 37:29.360
 that autonomous driving doesn't have.

37:29.360 --> 37:32.400
 So look, if I want to go from point A to point B,

37:32.400 --> 37:34.320
 I have a very, very safe trajectory.

37:34.320 --> 37:36.800
 Go vertically up to a maximum altitude,

37:36.800 --> 37:39.480
 fly horizontally to just about the destination,

37:39.480 --> 37:42.400
 and then come down vertically.

37:42.400 --> 37:45.400
 This is preprogrammed.

37:45.400 --> 37:48.040
 The equivalent of that is very hard to find

37:48.040 --> 37:51.560
 in the self driving car world because you're on the ground,

37:51.560 --> 37:53.560
 you're in a two dimensional surface,

37:53.560 --> 37:56.680
 and the trajectories on the two dimensional surface

37:56.680 --> 38:00.200
 are more likely to encounter obstacles.

38:00.200 --> 38:03.280
 I mean this in an intuitive sense, but mathematically true.

38:03.280 --> 38:06.360
 That's mathematically as well, that's true.

38:06.360 --> 38:10.040
 There's other option on the 2G space of platooning,

38:10.040 --> 38:11.640
 or because there's so many obstacles,

38:11.640 --> 38:13.280
 you can connect with those obstacles

38:13.280 --> 38:14.560
 and all these kind of options.

38:14.560 --> 38:16.560
 Sure, but those exist in the three dimensional space as well.

38:16.560 --> 38:17.560
 So they do.

38:17.560 --> 38:21.800
 So the question also implies how difficult are obstacles

38:21.800 --> 38:23.800
 in the three dimensional space in flight?

38:23.800 --> 38:25.600
 So that's the downside.

38:25.600 --> 38:26.920
 I think in three dimensional space,

38:26.920 --> 38:29.080
 you're modeling three dimensional world,

38:29.080 --> 38:31.280
 not just because you want to avoid it,

38:31.280 --> 38:33.040
 but you want to reason about it,

38:33.040 --> 38:35.360
 and you want to work in the three dimensional environment,

38:35.360 --> 38:37.480
 and that's significantly harder.

38:37.480 --> 38:38.920
 So that's one disadvantage.

38:38.920 --> 38:41.040
 I think the second disadvantage is of course,

38:41.040 --> 38:43.200
 anytime you fly, you have to put up

38:43.200 --> 38:46.560
 with the peculiarities of aerodynamics

38:46.560 --> 38:48.720
 and their complicated environments.

38:48.720 --> 38:49.800
 How do you negotiate that?

38:49.800 --> 38:51.880
 So that's always a problem.

38:51.880 --> 38:55.240
 Do you see a time in the future where there is,

38:55.240 --> 38:58.720
 you mentioned there's agriculture applications.

38:58.720 --> 39:01.680
 So there's a lot of applications of flying robots,

39:01.680 --> 39:03.040
 but do you see a time in the future

39:03.040 --> 39:05.360
 where there's tens of thousands,

39:05.360 --> 39:08.160
 or maybe hundreds of thousands of delivery drones

39:08.160 --> 39:12.160
 that fill the sky, delivery flying robots?

39:12.160 --> 39:14.200
 I think there's a lot of potential

39:14.200 --> 39:15.920
 for the last mile delivery.

39:15.920 --> 39:19.240
 And so in crowded cities, I don't know,

39:19.240 --> 39:21.400
 if you go to a place like Hong Kong,

39:21.400 --> 39:24.400
 just crossing the river can take half an hour,

39:24.400 --> 39:29.400
 and while a drone can just do it in five minutes at most.

39:29.400 --> 39:34.400
 I think you look at delivery of supplies to remote villages.

39:35.800 --> 39:38.680
 I work with a nonprofit called Weave Robotics.

39:38.680 --> 39:40.920
 So they work in the Peruvian Amazon,

39:40.920 --> 39:44.680
 where the only highways that are available

39:44.680 --> 39:47.440
 are the only highways or rivers.

39:47.440 --> 39:52.440
 And to get from point A to point B may take five hours,

39:52.960 --> 39:55.600
 while with a drone, you can get there in 30 minutes.

39:56.680 --> 39:59.880
 So just delivering drugs,

39:59.880 --> 40:04.880
 retrieving samples for testing vaccines,

40:05.160 --> 40:07.120
 I think there's huge potential here.

40:07.120 --> 40:09.960
 So I think the challenges are not technological,

40:09.960 --> 40:12.040
 but the challenge is economical.

40:12.040 --> 40:15.560
 The one thing I'll tell you that nobody thinks about

40:15.560 --> 40:18.920
 is the fact that we've not made huge strides

40:18.920 --> 40:20.840
 in battery technology.

40:20.840 --> 40:23.520
 Yes, it's true, batteries are becoming less expensive

40:23.520 --> 40:26.240
 because we have these mega factories that are coming up,

40:26.240 --> 40:28.800
 but they're all based on lithium based technologies.

40:28.800 --> 40:31.480
 And if you look at the energy density

40:31.480 --> 40:33.240
 and the power density,

40:33.240 --> 40:38.000
 those are two fundamentally limiting numbers.

40:38.000 --> 40:39.680
 So power density is important

40:39.680 --> 40:42.480
 because for a UAV to take off vertically into the air,

40:42.480 --> 40:46.360
 which most drones do, they don't have a runway,

40:46.360 --> 40:50.240
 you consume roughly 200 watts per kilo at the small size.

40:51.560 --> 40:53.920
 That's a lot, right?

40:53.920 --> 40:57.520
 In contrast, the human brain consumes less than 80 watts,

40:57.520 --> 40:58.920
 the whole of the human brain.

40:59.920 --> 41:03.600
 So just imagine just lifting yourself into the air

41:03.600 --> 41:06.000
 is like two or three light bulbs,

41:06.000 --> 41:07.840
 which makes no sense to me.

41:07.840 --> 41:10.440
 Yeah, so you're going to have to at scale

41:10.440 --> 41:12.880
 solve the energy problem then,

41:12.880 --> 41:17.880
 charging the batteries, storing the energy and so on.

41:18.920 --> 41:20.680
 And then the storage is the second problem,

41:20.680 --> 41:22.960
 but storage limits the range.

41:22.960 --> 41:27.960
 But you have to remember that you have to burn

41:28.680 --> 41:31.600
 a lot of it per given time.

41:31.600 --> 41:32.920
 So the burning is another problem.

41:32.920 --> 41:34.640
 Which is a power question.

41:34.640 --> 41:38.640
 Yes, and do you think just your intuition,

41:38.640 --> 41:43.640
 there are breakthroughs in batteries on the horizon?

41:44.960 --> 41:46.440
 How hard is that problem?

41:46.440 --> 41:47.600
 Look, there are a lot of companies

41:47.600 --> 41:52.600
 that are promising flying cars that are autonomous

41:53.880 --> 41:55.120
 and that are clean.

41:59.400 --> 42:01.680
 I think they're over promising.

42:01.680 --> 42:04.800
 The autonomy piece is doable.

42:04.800 --> 42:07.040
 The clean piece, I don't think so.

42:08.000 --> 42:11.840
 There's another company that I work with called JetOptra.

42:11.840 --> 42:14.360
 They make small jet engines.

42:15.760 --> 42:18.080
 And they can get up to 50 miles an hour very easily

42:18.080 --> 42:19.960
 and lift 50 kilos.

42:19.960 --> 42:22.840
 But they're jet engines, they're efficient,

42:23.920 --> 42:26.320
 they're a little louder than electric vehicles,

42:26.320 --> 42:28.960
 but they can build flying cars.

42:28.960 --> 42:32.440
 So your sense is that there's a lot of pieces

42:32.440 --> 42:33.520
 that have come together.

42:33.520 --> 42:37.360
 So on this crazy question,

42:37.360 --> 42:39.720
 if you look at companies like Kitty Hawk,

42:39.720 --> 42:42.080
 working on electric, so the clean,

42:43.880 --> 42:45.840
 talking to Sebastian Thrun, right?

42:45.840 --> 42:48.840
 It's a crazy dream, you know?

42:48.840 --> 42:52.080
 But you work with flight a lot.

42:52.080 --> 42:55.760
 You've mentioned before that manned flights

42:55.760 --> 43:00.760
 or carrying a human body is very difficult to do.

43:01.640 --> 43:04.240
 So how crazy is flying cars?

43:04.240 --> 43:05.400
 Do you think there'll be a day

43:05.400 --> 43:10.400
 when we have vertical takeoff and landing vehicles

43:11.080 --> 43:14.040
 that are sufficiently affordable

43:14.960 --> 43:17.440
 that we're going to see a huge amount of them?

43:17.440 --> 43:19.680
 And they would look like something like we dream of

43:19.680 --> 43:21.080
 when we think about flying cars.

43:21.080 --> 43:22.200
 Yeah, like the Jetsons.

43:22.200 --> 43:23.160
 The Jetsons, yeah.

43:23.160 --> 43:25.560
 So look, there are a lot of smart people working on this

43:25.560 --> 43:29.640
 and you never say something is not possible

43:29.640 --> 43:32.200
 when you have people like Sebastian Thrun working on it.

43:32.200 --> 43:35.160
 So I totally think it's viable.

43:35.160 --> 43:38.240
 I question, again, the electric piece.

43:38.240 --> 43:39.520
 The electric piece, yeah.

43:39.520 --> 43:41.680
 And again, for short distances, you can do it.

43:41.680 --> 43:43.640
 And there's no reason to suggest

43:43.640 --> 43:45.840
 that these all just have to be rotorcrafts.

43:45.840 --> 43:46.920
 You take off vertically,

43:46.920 --> 43:49.680
 but then you morph into a forward flight.

43:49.680 --> 43:51.600
 I think there are a lot of interesting designs.

43:51.600 --> 43:56.040
 The question to me is, are these economically viable?

43:56.040 --> 43:59.160
 And if you agree to do this with fossil fuels,

43:59.160 --> 44:01.960
 it instantly immediately becomes viable.

44:01.960 --> 44:03.480
 That's a real challenge.

44:03.480 --> 44:06.560
 Do you think it's possible for robots and humans

44:06.560 --> 44:08.840
 to collaborate successfully on tasks?

44:08.840 --> 44:13.640
 So a lot of robotics folks that I talk to and work with,

44:13.640 --> 44:18.000
 I mean, humans just add a giant mess to the picture.

44:18.000 --> 44:20.320
 So it's best to remove them from consideration

44:20.320 --> 44:22.400
 when solving specific tasks.

44:22.400 --> 44:23.600
 It's very difficult to model.

44:23.600 --> 44:26.000
 There's just a source of uncertainty.

44:26.000 --> 44:31.000
 In your work with these agile flying robots,

44:32.560 --> 44:35.680
 do you think there's a role for collaboration with humans?

44:35.680 --> 44:38.600
 Or is it best to model tasks in a way

44:38.600 --> 44:43.400
 that doesn't have a human in the picture?

44:43.400 --> 44:46.760
 Well, I don't think we should ever think about robots

44:46.760 --> 44:48.120
 without human in the picture.

44:48.120 --> 44:50.960
 Ultimately, robots are there because we want them

44:50.960 --> 44:54.360
 to solve problems for humans.

44:54.360 --> 44:58.280
 But there's no general solution to this problem.

44:58.280 --> 45:00.000
 I think if you look at human interaction

45:00.000 --> 45:02.400
 and how humans interact with robots,

45:02.400 --> 45:05.280
 you know, we think of these in sort of three different ways.

45:05.280 --> 45:07.600
 One is the human commanding the robot.

45:08.880 --> 45:12.880
 The second is the human collaborating with the robot.

45:12.880 --> 45:15.520
 So for example, we work on how a robot

45:15.520 --> 45:18.720
 can actually pick up things with a human and carry things.

45:18.720 --> 45:20.880
 That's like true collaboration.

45:20.880 --> 45:25.000
 And third, we think about humans as bystanders,

45:25.000 --> 45:27.240
 self driving cars, what's the human's role

45:27.240 --> 45:30.320
 and how do self driving cars

45:30.320 --> 45:32.920
 acknowledge the presence of humans?

45:32.920 --> 45:35.840
 So I think all of these things are different scenarios.

45:35.840 --> 45:38.480
 It depends on what kind of humans, what kind of task.

45:39.640 --> 45:41.840
 And I think it's very difficult to say

45:41.840 --> 45:45.520
 that there's a general theory that we all have for this.

45:45.520 --> 45:48.440
 But at the same time, it's also silly to say

45:48.440 --> 45:52.000
 that we should think about robots independent of humans.

45:52.000 --> 45:55.760
 So to me, human robot interaction

45:55.760 --> 45:59.760
 is almost a mandatory aspect of everything we do.

45:59.760 --> 46:02.440
 Yes, but to which degree, so your thoughts,

46:02.440 --> 46:05.240
 if we jump to autonomous vehicles, for example,

46:05.240 --> 46:08.680
 there's a big debate between what's called

46:08.680 --> 46:10.640
 level two and level four.

46:10.640 --> 46:13.680
 So semi autonomous and autonomous vehicles.

46:13.680 --> 46:16.440
 And so the Tesla approach currently at least

46:16.440 --> 46:18.960
 has a lot of collaboration between human and machine.

46:18.960 --> 46:22.040
 So the human is supposed to actively supervise

46:22.040 --> 46:23.880
 the operation of the robot.

46:23.880 --> 46:28.880
 Part of the safety definition of how safe a robot is

46:29.160 --> 46:32.880
 in that case is how effective is the human in monitoring it.

46:32.880 --> 46:37.880
 Do you think that's ultimately not a good approach

46:37.880 --> 46:42.360
 in sort of having a human in the picture,

46:42.360 --> 46:47.360
 not as a bystander or part of the infrastructure,

46:47.400 --> 46:50.000
 but really as part of what's required

46:50.000 --> 46:51.560
 to make the system safe?

46:51.560 --> 46:53.720
 This is harder than it sounds.

46:53.720 --> 46:58.200
 I think, you know, if you, I mean,

46:58.200 --> 47:01.360
 I'm sure you've driven before in highways and so on.

47:01.360 --> 47:06.120
 It's really very hard to have to relinquish control

47:06.120 --> 47:10.440
 to a machine and then take over when needed.

47:10.440 --> 47:12.280
 So I think Tesla's approach is interesting

47:12.280 --> 47:14.800
 because it allows you to periodically establish

47:14.800 --> 47:18.520
 some kind of contact with the car.

47:18.520 --> 47:20.640
 Toyota, on the other hand, is thinking about

47:20.640 --> 47:24.800
 shared autonomy or collaborative autonomy as a paradigm.

47:24.800 --> 47:27.480
 If I may argue, these are very, very simple ways

47:27.480 --> 47:29.680
 of human robot collaboration,

47:29.680 --> 47:31.880
 because the task is pretty boring.

47:31.880 --> 47:35.000
 You sit in a vehicle, you go from point A to point B.

47:35.000 --> 47:37.360
 I think the more interesting thing to me is,

47:37.360 --> 47:38.760
 for example, search and rescue.

47:38.760 --> 47:41.980
 I've got a human first responder, robot first responders.

47:43.160 --> 47:45.120
 I gotta do something.

47:45.120 --> 47:46.000
 It's important.

47:46.000 --> 47:47.800
 I have to do it in two minutes.

47:47.800 --> 47:49.240
 The building is burning.

47:49.240 --> 47:50.440
 There's been an explosion.

47:50.440 --> 47:51.360
 It's collapsed.

47:51.360 --> 47:52.800
 How do I do it?

47:52.800 --> 47:54.740
 I think to me, those are the interesting things

47:54.740 --> 47:57.160
 where it's very, very unstructured.

47:57.160 --> 47:58.480
 And what's the role of the human?

47:58.480 --> 48:00.200
 What's the role of the robot?

48:00.200 --> 48:02.440
 Clearly, there's lots of interesting challenges

48:02.440 --> 48:03.440
 and there's a field.

48:03.440 --> 48:05.760
 I think we're gonna make a lot of progress in this area.

48:05.760 --> 48:07.600
 Yeah, it's an exciting form of collaboration.

48:07.600 --> 48:08.440
 You're right.

48:08.440 --> 48:11.120
 In autonomous driving, the main enemy

48:11.120 --> 48:13.120
 is just boredom of the human.

48:13.120 --> 48:13.960
 Yes.

48:13.960 --> 48:15.680
 As opposed to in rescue operations,

48:15.680 --> 48:18.360
 it's literally life and death.

48:18.360 --> 48:22.080
 And the collaboration enables

48:22.080 --> 48:23.820
 the effective completion of the mission.

48:23.820 --> 48:24.760
 So it's exciting.

48:24.760 --> 48:27.400
 In some sense, we're also doing this.

48:27.400 --> 48:30.520
 You think about the human driving a car

48:30.520 --> 48:33.800
 and almost invariably, the human's trying

48:33.800 --> 48:35.000
 to estimate the state of the car,

48:35.000 --> 48:37.280
 they estimate the state of the environment and so on.

48:37.280 --> 48:40.120
 But what if the car were to estimate the state of the human?

48:40.120 --> 48:41.960
 So for example, I'm sure you have a smartphone

48:41.960 --> 48:44.580
 and the smartphone tries to figure out what you're doing

48:44.580 --> 48:48.320
 and send you reminders and oftentimes telling you

48:48.320 --> 48:49.540
 to drive to a certain place,

48:49.540 --> 48:51.400
 although you have no intention of going there

48:51.400 --> 48:53.880
 because it thinks that that's where you should be

48:53.880 --> 48:56.240
 because of some Gmail calendar entry

48:57.520 --> 48:58.960
 or something like that.

48:58.960 --> 49:01.600
 And it's trying to constantly figure out who you are,

49:01.600 --> 49:02.740
 what you're doing.

49:02.740 --> 49:04.200
 If a car were to do that,

49:04.200 --> 49:06.840
 maybe that would make the driver safer

49:06.840 --> 49:08.160
 because the car is trying to figure out

49:08.160 --> 49:09.760
 is the driver paying attention,

49:09.760 --> 49:11.600
 looking at his or her eyes,

49:12.480 --> 49:14.400
 looking at circadian movements.

49:14.400 --> 49:16.480
 So I think the potential is there,

49:16.480 --> 49:18.600
 but from the reverse side,

49:18.600 --> 49:21.640
 it's not robot modeling, but it's human modeling.

49:21.640 --> 49:22.880
 It's more on the human, right.

49:22.880 --> 49:25.320
 And I think the robots can do a very good job

49:25.320 --> 49:29.120
 of modeling humans if you really think about the framework

49:29.120 --> 49:32.640
 that you have a human sitting in a cockpit,

49:32.640 --> 49:35.820
 surrounded by sensors, all staring at him,

49:35.820 --> 49:37.860
 in addition to be staring outside,

49:37.860 --> 49:39.160
 but also staring at him.

49:39.160 --> 49:40.960
 I think there's a real synergy there.

49:40.960 --> 49:42.360
 Yeah, I love that problem

49:42.360 --> 49:45.560
 because it's the new 21st century form of psychology,

49:45.560 --> 49:48.520
 actually AI enabled psychology.

49:48.520 --> 49:51.280
 A lot of people have sci fi inspired fears

49:51.280 --> 49:54.080
 of walking robots like those from Boston Dynamics.

49:54.080 --> 49:56.480
 If you just look at shows on Netflix and so on,

49:56.480 --> 49:59.040
 or flying robots like those you work with,

49:59.920 --> 50:03.160
 how would you, how do you think about those fears?

50:03.160 --> 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 --> 50:11.760
 You know, anytime we develop a technology

50:11.760 --> 50:14.160
 meaning to have positive impact in the world,

50:14.160 --> 50:15.780
 there's always the worry that,

50:17.440 --> 50:21.000
 you know, somebody could subvert those technologies

50:21.000 --> 50:23.280
 and use it in an adversarial setting.

50:23.280 --> 50:25.280
 And robotics is no exception, right?

50:25.280 --> 50:29.280
 So I think it's very easy to weaponize robots.

50:29.280 --> 50:30.880
 I think we talk about swarms.

50:31.720 --> 50:33.960
 One thing I worry a lot about is,

50:33.960 --> 50:35.880
 so, you know, for us to get swarms to work

50:35.880 --> 50:38.280
 and do something reliably, it's really hard.

50:38.280 --> 50:42.040
 But suppose I have this challenge

50:42.040 --> 50:44.360
 of trying to destroy something,

50:44.360 --> 50:45.720
 and I have a swarm of robots,

50:45.720 --> 50:47.280
 where only one out of the swarm

50:47.280 --> 50:48.920
 needs to get to its destination.

50:48.920 --> 50:52.640
 So that suddenly becomes a lot more doable.

50:52.640 --> 50:54.720
 And so I worry about, you know,

50:54.720 --> 50:56.920
 this general idea of using autonomy

50:56.920 --> 50:58.600
 with lots and lots of agents.

51:00.040 --> 51:01.320
 I mean, having said that, look,

51:01.320 --> 51:03.760
 a lot of this technology is not very mature.

51:03.760 --> 51:05.520
 My favorite saying is that

51:06.560 --> 51:10.520
 if somebody had to develop this technology,

51:10.520 --> 51:12.320
 wouldn't you rather the good guys do it?

51:12.320 --> 51:13.880
 So the good guys have a good understanding

51:13.880 --> 51:15.560
 of the technology, so they can figure out

51:15.560 --> 51:18.320
 how this technology is being used in a bad way,

51:18.320 --> 51:21.360
 or could be used in a bad way and try to defend against it.

51:21.360 --> 51:22.760
 So we think a lot about that.

51:22.760 --> 51:25.400
 So we have, we're doing research

51:25.400 --> 51:28.240
 on how to defend against swarms, for example.

51:28.240 --> 51:29.600
 That's interesting.

51:29.600 --> 51:32.960
 There's in fact a report by the National Academies

51:32.960 --> 51:35.520
 on counter UAS technologies.

51:36.680 --> 51:38.200
 This is a real threat,

51:38.200 --> 51:40.320
 but we're also thinking about how to defend against this

51:40.320 --> 51:42.920
 and knowing how swarms work.

51:42.920 --> 51:47.160
 Knowing how autonomy works is, I think, very important.

51:47.160 --> 51:49.280
 So it's not just politicians?

51:49.280 --> 51:51.640
 Do you think engineers have a role in this discussion?

51:51.640 --> 51:52.480
 Absolutely.

51:52.480 --> 51:55.280
 I think the days where politicians

51:55.280 --> 51:57.680
 can be agnostic to technology are gone.

51:59.200 --> 52:02.640
 I think every politician needs to be

52:03.840 --> 52:05.680
 literate in technology.

52:05.680 --> 52:08.640
 And I often say technology is the new liberal art.

52:09.800 --> 52:12.920
 Understanding how technology will change your life,

52:12.920 --> 52:14.480
 I think is important.

52:14.480 --> 52:18.080
 And every human being needs to understand that.

52:18.080 --> 52:20.160
 And maybe we can elect some engineers

52:20.160 --> 52:22.720
 to office as well on the other side.

52:22.720 --> 52:24.840
 What are the biggest open problems in robotics?

52:24.840 --> 52:27.760
 And you said we're in the early days in some sense.

52:27.760 --> 52:31.040
 What are the problems we would like to solve in robotics?

52:31.040 --> 52:32.520
 I think there are lots of problems, right?

52:32.520 --> 52:36.440
 But I would phrase it in the following way.

52:36.440 --> 52:39.520
 If you look at the robots we're building,

52:39.520 --> 52:43.160
 they're still very much tailored towards

52:43.160 --> 52:46.520
 doing specific tasks and specific settings.

52:46.520 --> 52:49.480
 I think the question of how do you get them to operate

52:49.480 --> 52:51.080
 in much broader settings

52:53.560 --> 52:58.040
 where things can change in unstructured environments

52:58.040 --> 52:59.160
 is up in the air.

52:59.160 --> 53:01.200
 So think of self driving cars.

53:02.920 --> 53:05.680
 Today, we can build a self driving car in a parking lot.

53:05.680 --> 53:09.000
 We can do level five autonomy in a parking lot.

53:10.040 --> 53:13.240
 But can you do a level five autonomy

53:13.240 --> 53:16.840
 in the streets of Napoli in Italy or Mumbai in India?

53:16.840 --> 53:17.760
 No.

53:17.760 --> 53:22.400
 So in some sense, when we think about robotics,

53:22.400 --> 53:25.120
 we have to think about where they're functioning,

53:25.120 --> 53:27.760
 what kind of environment, what kind of a task.

53:27.760 --> 53:29.800
 We have no understanding

53:29.800 --> 53:32.800
 of how to put both those things together.

53:32.800 --> 53:34.000
 So we're in the very early days

53:34.000 --> 53:35.920
 of applying it to the physical world.

53:35.920 --> 53:38.800
 And I was just in Naples actually.

53:38.800 --> 53:42.200
 And there's levels of difficulty and complexity

53:42.200 --> 53:45.880
 depending on which area you're applying it to.

53:45.880 --> 53:46.720
 I think so.

53:46.720 --> 53:49.320
 And we don't have a systematic way of understanding that.

53:51.040 --> 53:53.800
 Everybody says, just because a computer

53:53.800 --> 53:56.520
 can now beat a human at any board game,

53:56.520 --> 53:59.920
 we certainly know something about intelligence.

53:59.920 --> 54:01.360
 That's not true.

54:01.360 --> 54:04.400
 A computer board game is very, very structured.

54:04.400 --> 54:08.480
 It is the equivalent of working in a Henry Ford factory

54:08.480 --> 54:11.680
 where things, parts come, you assemble, move on.

54:11.680 --> 54:14.120
 It's a very, very, very structured setting.

54:14.120 --> 54:15.680
 That's the easiest thing.

54:15.680 --> 54:17.040
 And we know how to do that.

54:18.400 --> 54:20.400
 So you've done a lot of incredible work

54:20.400 --> 54:23.720
 at the UPenn, University of Pennsylvania, GraspLab.

54:23.720 --> 54:26.560
 You're now Dean of Engineering at UPenn.

54:26.560 --> 54:31.320
 What advice do you have for a new bright eyed undergrad

54:31.320 --> 54:34.640
 interested in robotics or AI or engineering?

54:34.640 --> 54:36.560
 Well, I think there's really three things.

54:36.560 --> 54:40.600
 One is you have to get used to the idea

54:40.600 --> 54:42.840
 that the world will not be the same in five years

54:42.840 --> 54:45.160
 or four years whenever you graduate, right?

54:45.160 --> 54:46.120
 Which is really hard to do.

54:46.120 --> 54:48.960
 So this thing about predicting the future,

54:48.960 --> 54:50.520
 every one of us needs to be trying

54:50.520 --> 54:52.360
 to predict the future always.

54:53.280 --> 54:54.960
 Not because you'll be any good at it,

54:54.960 --> 54:56.440
 but by thinking about it,

54:56.440 --> 55:00.880
 I think you sharpen your senses and you become smarter.

55:00.880 --> 55:02.080
 So that's number one.

55:02.080 --> 55:05.760
 Number two, it's a corollary of the first piece,

55:05.760 --> 55:09.360
 which is you really don't know what's gonna be important.

55:09.360 --> 55:12.080
 So this idea that I'm gonna specialize in something

55:12.080 --> 55:15.320
 which will allow me to go in a particular direction,

55:15.320 --> 55:16.480
 it may be interesting,

55:16.480 --> 55:18.480
 but it's important also to have this breadth

55:18.480 --> 55:20.360
 so you have this jumping off point.

55:22.000 --> 55:23.000
 I think the third thing,

55:23.000 --> 55:25.360
 and this is where I think Penn excels.

55:25.360 --> 55:27.240
 I mean, we teach engineering,

55:27.240 --> 55:29.960
 but it's always in the context of the liberal arts.

55:29.960 --> 55:32.360
 It's always in the context of society.

55:32.360 --> 55:35.840
 As engineers, we cannot afford to lose sight of that.

55:35.840 --> 55:37.640
 So I think that's important.

55:37.640 --> 55:39.960
 But I think one thing that people underestimate

55:39.960 --> 55:40.920
 when they do robotics

55:40.920 --> 55:43.440
 is the importance of mathematical foundations,

55:43.440 --> 55:46.880
 the importance of representations.

55:47.720 --> 55:50.040
 Not everything can just be solved

55:50.040 --> 55:52.440
 by looking for Ross packages on the internet

55:52.440 --> 55:56.280
 or to find a deep neural network that works.

55:56.280 --> 55:59.080
 I think the representation question is key,

55:59.080 --> 56:00.400
 even to machine learning,

56:00.400 --> 56:05.400
 where if you ever hope to achieve or get to explainable AI,

56:05.400 --> 56:07.760
 somehow there need to be representations

56:07.760 --> 56:09.080
 that you can understand.

56:09.080 --> 56:11.120
 So if you wanna do robotics,

56:11.120 --> 56:12.680
 you should also do mathematics.

56:12.680 --> 56:15.080
 And you said liberal arts, a little literature.

56:16.160 --> 56:17.200
 If you wanna build a robot,

56:17.200 --> 56:19.320
 it should be reading Dostoyevsky.

56:19.320 --> 56:20.360
 I agree with that.

56:20.360 --> 56:21.200
 Very good.

56:21.200 --> 56:23.560
 So Vijay, thank you so much for talking today.

56:23.560 --> 56:24.400
 It was an honor.

56:24.400 --> 56:25.240
 Thank you.

56:25.240 --> 56:26.200
 It was just a very exciting conversation.

56:26.200 --> 56:46.200
 Thank you.