George Hotz: 3 Problems of Autonomous Driving: Static, Dynamic, Counterfactual | AI Podcast Clips
00:00:00.000 | [guitar music]
00:00:02.580 | So the way you leapfrog, right,
00:00:10.320 | is you come up with an idea
00:00:13.200 | or you take a direction, perhaps secretly,
00:00:15.680 | that the other people aren't taking.
00:00:17.760 | And so the Cruz, Waymo, even Aurora.
00:00:23.760 | - I don't know Aurora, Zooks is the same stack as well.
00:00:27.200 | They're all the same code base even.
00:00:28.840 | They're all the same DARPA Urban Challenge code base.
00:00:31.440 | - So the question is, do you think there's a room
00:00:34.880 | for brilliance and innovation there
00:00:36.240 | that will change everything?
00:00:37.600 | Like say, okay, so I'll give you examples.
00:00:41.020 | It could be if revolution and mapping, for example,
00:00:46.760 | that allow you to map things,
00:00:50.160 | do HD maps of the whole world,
00:00:52.960 | all weather conditions somehow really well,
00:00:55.200 | or revolution and simulation
00:01:00.200 | to where all the way you said before becomes incorrect,
00:01:05.940 | that kind of thing.
00:01:08.440 | Any room for breakthrough innovation?
00:01:10.300 | - What I said before about,
00:01:13.080 | oh, they actually get the whole thing.
00:01:15.400 | I'll say this about, we divide driving into three problems
00:01:19.760 | and I actually haven't solved the third yet,
00:01:20.920 | but I have an idea how to do it.
00:01:21.920 | So there's the static.
00:01:23.220 | The static driving problem is assuming
00:01:25.120 | you are the only car on the road.
00:01:27.240 | And this problem can be solved 100%
00:01:29.080 | with mapping and localization.
00:01:31.120 | This is why farms work the way they do.
00:01:32.840 | If all you have to deal with is the static problem
00:01:35.560 | and you can statically schedule your machines,
00:01:37.280 | it's the same as like statically scheduling processes.
00:01:39.800 | You can statically schedule your tractors
00:01:41.160 | to never hit each other on their paths
00:01:43.240 | 'cause you know the speed they go at.
00:01:44.600 | So that's the static driving problem.
00:01:47.260 | Maps only helps you with the static driving problem.
00:01:51.040 | - Yeah, the question about static driving.
00:01:54.100 | You've just made it sound like it's really easy.
00:01:55.900 | - Static driving is really easy.
00:01:57.440 | - How easy?
00:02:00.140 | 'Cause the whole drifting out of lane,
00:02:04.140 | when Tesla drifts out of lane,
00:02:05.900 | it's failing on the fundamental static driving problem.
00:02:09.140 | - Tesla is drifting out of lane?
00:02:11.600 | The static driving problem is not easy for the world.
00:02:14.860 | The static driving problem is easy for one route.
00:02:18.980 | - One route and one weather condition
00:02:21.080 | with one state of lane markings
00:02:25.080 | and like no deterioration, no cracks in the road.
00:02:28.080 | - I'm assuming you have a perfect localizer.
00:02:29.760 | So that's solved for the weather condition
00:02:31.340 | and the lane marking condition.
00:02:32.720 | - But that's the problem is how do you have a perfect--
00:02:34.840 | - You can build, perfect localizers
00:02:36.320 | are not that hard to build.
00:02:37.720 | - Okay, come on now.
00:02:38.560 | With LIDAR?
00:02:40.520 | - With LIDAR, yeah.
00:02:41.360 | - Oh, with LIDAR, okay.
00:02:42.180 | - With LIDAR, yeah, but you use LIDAR, right?
00:02:43.560 | Like use LIDAR to build a perfect localizer.
00:02:45.760 | Building a perfect localizer without LIDAR,
00:02:49.040 | (sighs)
00:02:50.120 | it's gonna be hard.
00:02:51.440 | You can get 10 centimeters without LIDAR,
00:02:52.880 | you can get one centimeter with LIDAR.
00:02:54.360 | - I'm not even concerned about the one or 10 centimeters.
00:02:56.400 | I'm concerned if every once in a while you're just way off.
00:02:59.800 | - Yeah, so this is why you have to
00:03:03.080 | carefully make sure you're always tracking your position.
00:03:07.120 | You wanna use LIDAR camera fusion,
00:03:08.840 | but you can get the reliability of that system
00:03:11.560 | up to 100,000 miles
00:03:15.120 | and then you write some fallback condition
00:03:16.800 | where it's not that bad if you're way off, right?
00:03:19.260 | I think that you can get it to the point,
00:03:20.860 | it's like as will be,
00:03:21.900 | that you're never in a case where you're way off
00:03:24.700 | and you don't know it.
00:03:25.540 | - Yeah, okay, so this is brilliant.
00:03:27.300 | So that's the static.
00:03:28.220 | - Static.
00:03:29.360 | - We can, especially with LIDAR and good HD maps,
00:03:33.060 | you can solve that problem.
00:03:34.140 | - Easy.
00:03:34.980 | - No, I just disagree with your word easy.
00:03:37.580 | - The static problem's so easy.
00:03:39.020 | - Very typical for you to say something's easy.
00:03:41.100 | I got it.
00:03:41.940 | It's not as challenging as the other ones, okay.
00:03:44.020 | - Well, it's, okay, maybe it's obvious how to solve it.
00:03:45.860 | The third one's the hardest.
00:03:46.940 | So where do we get,
00:03:47.780 | and a lot of people don't even think about the third one
00:03:49.020 | and even see it as different from the second one.
00:03:50.740 | So the second one is dynamic.
00:03:52.840 | The second one is like,
00:03:54.060 | say there's an obvious examples,
00:03:55.720 | like a car stopped at a red light, right?
00:03:57.480 | You can't have that car in your map
00:03:59.640 | because you don't know whether that car
00:04:00.840 | is gonna be there or not.
00:04:02.000 | So you have to detect that car in real time
00:04:05.080 | and then you have to do the appropriate action, right?
00:04:07.800 | Also, that car is not a fixed object.
00:04:11.920 | That car may move
00:04:12.840 | and you have to predict what that car will do, right?
00:04:15.840 | So this is the dynamic problem.
00:04:17.960 | - Yeah.
00:04:18.800 | - So you have to deal with this.
00:04:19.920 | This involves, again,
00:04:22.000 | like you're gonna need models of other people's behavior.
00:04:24.900 | - Do you, are you including in that,
00:04:27.280 | I don't wanna step on the third one.
00:04:29.440 | - Oh.
00:04:30.280 | - But are you including in that your influence on people?
00:04:34.080 | - Ah, that's the third one.
00:04:35.320 | - Okay.
00:04:36.160 | - That's the third one.
00:04:36.980 | We call it the counterfactual.
00:04:38.960 | - Yeah, brilliant. - And that.
00:04:40.200 | - I just talked to Judea Pearl
00:04:41.440 | who's obsessed with counterfactuals.
00:04:42.960 | - The counterfactual, oh yeah, yeah, I read his books.
00:04:45.720 | - So the static and the dynamic.
00:04:47.880 | - Yeah.
00:04:49.080 | - Our approach right now for lateral
00:04:51.840 | will scale completely to the static and dynamic.
00:04:54.680 | The counterfactual, the only way I have to do it yet,
00:04:58.440 | the thing that I wanna do once we have all of these cars
00:05:01.080 | is I wanna do reinforcement learning on the world.
00:05:03.880 | I'm always gonna turn the exploiter up to max.
00:05:05.980 | I'm not gonna have them explore.
00:05:07.540 | But the only real way to get at the counterfactual
00:05:09.880 | is to do reinforcement learning
00:05:11.200 | because the other agents are humans.
00:05:13.440 | (whooshing)
00:05:15.760 | (upbeat music)
00:05:18.340 | (whooshing)
00:05:20.660 | (whooshing)
00:05:22.980 | (whooshing)
00:05:25.300 | (whooshing)
00:05:27.620 | (whooshing)
00:05:29.940 | [BLANK_AUDIO]