WEBVTT 00:00:01.680 --> 00:00:04.080 The following content is provided under a Creative 00:00:04.080 --> 00:00:05.620 Commons license. 00:00:05.620 --> 00:00:07.920 Your support will help MIT OpenCourseWare 00:00:07.920 --> 00:00:12.280 continue to offer high quality educational resources for free. 00:00:12.280 --> 00:00:14.910 To make a donation or view additional materials 00:00:14.910 --> 00:00:18.820 from hundreds of MIT courses, visit MIT OpenCourseWare 00:00:18.820 --> 00:00:20.010 at ocw.mit.edu. 00:00:22.497 --> 00:00:24.580 PATRICK WINSTON: Well, I suppose my first question 00:00:24.580 --> 00:00:28.600 has to do with some remarks that Tony made about Rod Brooks. 00:00:28.600 --> 00:00:34.150 I remember Rod Brooks' work for the one great idea he had, 00:00:34.150 --> 00:00:36.640 which was the idea of subsumption. 00:00:36.640 --> 00:00:38.410 And the idea of subsumption was to take 00:00:38.410 --> 00:00:40.870 the notion of procedural and data 00:00:40.870 --> 00:00:42.760 abstraction from ordinary programming 00:00:42.760 --> 00:00:45.890 and elevate it to a behavior level. 00:00:45.890 --> 00:00:48.550 And the reason for doing that was 00:00:48.550 --> 00:00:52.180 that if you weren't working so well at one level, 00:00:52.180 --> 00:00:53.860 you would appeal to another level 00:00:53.860 --> 00:00:56.250 to get you out of trouble. 00:00:56.250 --> 00:00:58.030 So that sounds like a powerful idea to me. 00:00:58.030 --> 00:01:02.260 And I'm just very interested in what the panelists construe 00:01:02.260 --> 00:01:04.810 to be the great principles of robotics 00:01:04.810 --> 00:01:06.540 that have emerged since then. 00:01:06.540 --> 00:01:08.200 Are there great principles that we 00:01:08.200 --> 00:01:13.150 can talk about in a classroom without just describing 00:01:13.150 --> 00:01:15.322 how a particular robot works? 00:01:15.322 --> 00:01:16.780 STEFANIE TELLEX: So we were talking 00:01:16.780 --> 00:01:18.180 about this ourselves a little bit 00:01:18.180 --> 00:01:22.330 and sort of asking ourselves what makes a systems paper? 00:01:22.330 --> 00:01:23.770 And what do you write down in one 00:01:23.770 --> 00:01:26.950 of those papers as these general principles 00:01:26.950 --> 00:01:28.630 that you extract from building a system? 00:01:28.630 --> 00:01:30.671 Because it seems like there's two kinds of papers 00:01:30.671 --> 00:01:33.130 in robotics-- the systems paper, where you said, 00:01:33.130 --> 00:01:35.690 I built this thing and here's kind of how it works. 00:01:35.690 --> 00:01:38.390 Where it's hard to extract, I think, general principles 00:01:38.390 --> 00:01:39.041 from that. 00:01:39.041 --> 00:01:40.790 It's like, I built this and this and this, 00:01:40.790 --> 00:01:41.790 and this is what it did. 00:01:41.790 --> 00:01:43.780 Here's a video. 00:01:43.780 --> 00:01:46.109 But it does something amazing, so it's cool. 00:01:46.109 --> 00:01:48.400 And then there's, like, kind of algorithm papers, which 00:01:48.400 --> 00:01:53.130 tend to get more citations and I don't know. 00:01:53.130 --> 00:01:54.580 And they usually work because they 00:01:54.580 --> 00:01:57.970 have this chunk of knowledge, subsumption architectures, 00:01:57.970 --> 00:01:59.730 RRT Star's one of my favorite examples. 00:01:59.730 --> 00:02:01.150 It's the kind of paper, there's an algorithm, 00:02:01.150 --> 00:02:03.525 and then there's math that shows how the algorithm works, 00:02:03.525 --> 00:02:04.900 some results that they show. 00:02:04.900 --> 00:02:08.509 And there's this nugget that transfers from your brain-- 00:02:08.509 --> 00:02:10.460 to the author's brain to your brain. 00:02:10.460 --> 00:02:14.500 And I think it's hard to know what that nugget is when 00:02:14.500 --> 00:02:16.624 you've built a giant system. 00:02:16.624 --> 00:02:18.790 One of the things that I've been thinking about that 00:02:18.790 --> 00:02:20.206 might be what that might look like 00:02:20.206 --> 00:02:22.450 is kind of design patterns for robots. 00:02:22.450 --> 00:02:25.180 This is a concept from software engineering. 00:02:25.180 --> 00:02:27.430 It's at a higher level abstraction than a library 00:02:27.430 --> 00:02:30.790 or something that you share, but it's things about the ways 00:02:30.790 --> 00:02:32.510 that you put software together. 00:02:32.510 --> 00:02:34.030 So if you're a hacker, you probably 00:02:34.030 --> 00:02:37.539 heard of some of these patterns like singleton and facade 00:02:37.539 --> 00:02:38.080 and strategy. 00:02:38.080 --> 00:02:40.600 They have these sort of evocative names from this book, 00:02:40.600 --> 00:02:42.175 Gang of Four is the nickname. 00:02:42.175 --> 00:02:45.280 And I think there's a set of design patterns for robotics 00:02:45.280 --> 00:02:47.560 that we are slowly discovering. 00:02:47.560 --> 00:02:50.710 So when I was hanging out in Seth Teller and Nick 00:02:50.710 --> 00:02:53.740 Roy's for my post-doc, there was one that I really got 00:02:53.740 --> 00:02:56.994 in my head, which was this idea of pub/sub-- 00:02:56.994 --> 00:02:57.910 publish and subscribe. 00:02:57.910 --> 00:03:00.970 You're talking about YARP and LCM and Russ. 00:03:00.970 --> 00:03:03.370 They all had this idea that you don't 00:03:03.370 --> 00:03:07.191 want to write a function to call to get the autodetection 00:03:07.191 --> 00:03:07.690 results. 00:03:07.690 --> 00:03:09.910 You just want you detector blasting out the results 00:03:09.910 --> 00:03:12.090 as fast as it can all the time, and then break 00:03:12.090 --> 00:03:13.150 that abstraction. 00:03:13.150 --> 00:03:15.649 And you get a lot of robustness in exchange for that. 00:03:15.649 --> 00:03:17.440 I think that's a design pattern for robots. 00:03:17.440 --> 00:03:20.650 I think there's probably about 30 more of them. 00:03:20.650 --> 00:03:22.680 And I bet Russ knows a lot of them. 00:03:22.680 --> 00:03:25.510 And Albert and Ed-- there's people who know them maybe, 00:03:25.510 --> 00:03:26.719 but they're not written down. 00:03:26.719 --> 00:03:29.176 I think one thing I'd like to do is write some more of them 00:03:29.176 --> 00:03:29.770 down. 00:03:29.770 --> 00:03:31.255 Sorry I didn't. 00:03:31.255 --> 00:03:33.340 PATRICK WINSTON: Russ, what do you think? 00:03:33.340 --> 00:03:40.799 Your talk seemed to focus on optimization as the answer. 00:03:40.799 --> 00:03:42.340 RUSS TEDRAKE: It's a framework that I 00:03:42.340 --> 00:03:46.670 think you can guess a lot of the problems and get clear results. 00:03:46.670 --> 00:03:51.760 I think looking across you can point to clearer sort of ideas 00:03:51.760 --> 00:03:52.880 that worked very well. 00:03:52.880 --> 00:03:55.690 So I think for estimation Bayes Rule works really well. 00:03:55.690 --> 00:03:58.240 And you know, Monte Carlo estimation 00:03:58.240 --> 00:04:00.656 has worked really well for planning 00:04:00.656 --> 00:04:01.780 in high dimensional spaces. 00:04:01.780 --> 00:04:04.570 Somehow randomization was a magical bullet 00:04:04.570 --> 00:04:07.180 where people start doing RRT-type things as well 00:04:07.180 --> 00:04:10.870 as trajectory optimization-type things. 00:04:10.870 --> 00:04:13.930 I do think that the open source movement and the ability 00:04:13.930 --> 00:04:15.790 to write software and components and modules 00:04:15.790 --> 00:04:18.290 has been a huge, huge thing. 00:04:18.290 --> 00:04:21.910 And I do think that at the low-level control level 00:04:21.910 --> 00:04:24.070 is optimization-based controllers 00:04:24.070 --> 00:04:25.670 have been a magical thing. 00:04:25.670 --> 00:04:27.040 I think any one of these-- 00:04:27.040 --> 00:04:28.660 in any one of these sub-disciplines, 00:04:28.660 --> 00:04:31.120 you can point to a few real go-ahead 00:04:31.120 --> 00:04:35.050 ideas that have rocked our world, 00:04:35.050 --> 00:04:37.210 and everybody gets behind them. 00:04:37.210 --> 00:04:40.000 You know, I think maybe the biggest one of all, 00:04:40.000 --> 00:04:43.880 actually, has been LIDAR. 00:04:43.880 --> 00:04:45.220 And then the ability-- 00:04:45.220 --> 00:04:50.560 I think sensing has really come online and been so enabling 00:04:50.560 --> 00:04:53.440 in the last few years that I think if you look back 00:04:53.440 --> 00:04:55.390 at the last 15, 20 years of robotics, 00:04:55.390 --> 00:04:56.930 the biggest point changes I think 00:04:56.930 --> 00:05:00.480 has been with the sensors-- sensors upped their frame rate 00:05:00.480 --> 00:05:01.710 resolution, gave depth. 00:05:01.710 --> 00:05:07.410 When LIDAR and Kinect came out, those just changed everything. 00:05:07.410 --> 00:05:09.660 PATRICK WINSTON: Before we leave the subject of Brooks 00:05:09.660 --> 00:05:12.640 and subsumption, Tony, you brought it up, 00:05:12.640 --> 00:05:15.280 and there was a little exchange between you 00:05:15.280 --> 00:05:18.720 and Russ about why it might be useful. 00:05:18.720 --> 00:05:23.850 I noted that both Russ and John talked about two-- 00:05:23.850 --> 00:05:27.340 one respect-- one each major blunders 00:05:27.340 --> 00:05:29.956 that their machines made. 00:05:29.956 --> 00:05:32.260 Do you construe that any of Brooks's stuff 00:05:32.260 --> 00:05:33.940 might have been useful in avoiding 00:05:33.940 --> 00:05:36.756 those kinds of blunders? 00:05:36.756 --> 00:05:37.880 TONY PRESCOTT: Potentially. 00:05:37.880 --> 00:05:41.240 But I think these are really challenging things that we're 00:05:41.240 --> 00:05:42.610 trying to do with these robots. 00:05:42.610 --> 00:05:45.280 So the biomimetic approach that I take, I think, 00:05:45.280 --> 00:05:50.260 is partly I want to mine biology for insights 00:05:50.260 --> 00:05:52.840 about how to solve problems. 00:05:52.840 --> 00:05:56.590 And the problems that we're realizing 00:05:56.590 --> 00:05:58.630 are hard in robotics are the problems that 00:05:58.630 --> 00:06:00.230 are hard in biology as well. 00:06:00.230 --> 00:06:03.250 So I think we underestimated the problem 00:06:03.250 --> 00:06:05.390 of manipulating objects. 00:06:05.390 --> 00:06:08.830 But if you look in biology, what other species apart from us 00:06:08.830 --> 00:06:12.670 has really solved that problem to the level of dexterity? 00:06:12.670 --> 00:06:16.146 An octopus trunk, maybe but-- 00:06:16.146 --> 00:06:18.530 sorry, elephant trunk. 00:06:18.530 --> 00:06:23.324 So I think that these challenges take a-- 00:06:23.324 --> 00:06:25.490 prove to be much more difficult than we might think. 00:06:25.490 --> 00:06:28.690 And then other things that intuitively seem hard 00:06:28.690 --> 00:06:30.040 are actually quite easy. 00:06:30.040 --> 00:06:33.550 So the path that we're taking in some of our biomimetic robots, 00:06:33.550 --> 00:06:38.530 like the MiRo robot toy is to do stuff which actually people 00:06:38.530 --> 00:06:42.300 think looks hard, but it's relatively easy to do, 00:06:42.300 --> 00:06:44.420 and these are solvable problems. 00:06:44.420 --> 00:06:48.100 And then we can progress towards what are obviously the harder 00:06:48.100 --> 00:06:50.650 problems, and where the brain has dedicated 00:06:50.650 --> 00:06:52.270 a lot of processing power. 00:06:52.270 --> 00:06:55.780 And I think object manipulation, if you look in the brain, 00:06:55.780 --> 00:06:59.600 is massive representation for the hand. 00:06:59.600 --> 00:07:02.860 And there's all these extra motor systems in cortex that 00:07:02.860 --> 00:07:05.230 aren't there in non-mammals. 00:07:05.230 --> 00:07:08.570 And even in simpler mammals they don't have motor cortex. 00:07:08.570 --> 00:07:10.720 We've developed all these extra motor 00:07:10.720 --> 00:07:13.960 cortical areas, direct corticospinal projections. 00:07:13.960 --> 00:07:17.530 All of this is dedicated to the manipulation problem. 00:07:17.530 --> 00:07:21.910 So I think we're finding out what the hard problems are 00:07:21.910 --> 00:07:23.830 by trying to build the robots. 00:07:23.830 --> 00:07:26.290 And then we can maybe find out what the solutions 00:07:26.290 --> 00:07:27.880 are by looking at the biology. 00:07:27.880 --> 00:07:30.340 Because in that case, particularly, you 00:07:30.340 --> 00:07:32.470 have low level systems for-- 00:07:32.470 --> 00:07:35.560 that can do grasp, and that are there in reptiles. 00:07:35.560 --> 00:07:38.950 And then you have these additional systems in mammals, 00:07:38.950 --> 00:07:40.510 and particularly in primates, that 00:07:40.510 --> 00:07:43.250 can override the low-level systems to do 00:07:43.250 --> 00:07:45.009 dexterous control. 00:07:45.009 --> 00:07:46.800 PATRICK WINSTON: John, you look like you're 00:07:46.800 --> 00:07:47.590 eager to say something. 00:07:47.590 --> 00:07:49.756 JOHN LEONARD: I just want to talk about subsumption, 00:07:49.756 --> 00:07:51.790 because it had such a huge effect on my life 00:07:51.790 --> 00:07:52.600 as a grad student. 00:07:52.600 --> 00:07:54.580 It was sort of like the big thing in 1990 00:07:54.580 --> 00:07:56.380 when I was finishing up. 00:07:56.380 --> 00:07:59.650 But I really developed a strong aversion to it, 00:07:59.650 --> 00:08:02.800 and so I tried to argue with Rod back then, but not very 00:08:02.800 --> 00:08:03.930 successfully. 00:08:03.930 --> 00:08:06.070 I would say, when will you build a robot that 00:08:06.070 --> 00:08:07.390 knows its position? 00:08:07.390 --> 00:08:09.580 And he would say, I don't know my position. 00:08:09.580 --> 00:08:11.740 But I think some of the biological evidence, 00:08:11.740 --> 00:08:14.200 like the grid cells and things that maybe 00:08:14.200 --> 00:08:16.180 at an autonomic subconscious level 00:08:16.180 --> 00:08:20.490 there is sort of position information in the brain. 00:08:20.490 --> 00:08:22.690 But subsumption, I think, as a way 00:08:22.690 --> 00:08:24.760 of trying to strive for robustness where 00:08:24.760 --> 00:08:27.010 you build on layers, that's a great inspiration. 00:08:27.010 --> 00:08:31.750 But I feel that the intelligence without representation 00:08:31.750 --> 00:08:34.929 sort of work that Rod did, I just don't buy it. 00:08:34.929 --> 00:08:36.419 I think we need representation. 00:08:36.419 --> 00:08:38.289 PATRICK WINSTON: I guess there are two separable ideas. 00:08:38.289 --> 00:08:38.570 JOHN LEONARD: Yes. 00:08:38.570 --> 00:08:40.361 PATRICK WINSTON: The no representation idea 00:08:40.361 --> 00:08:41.380 and the layering idea. 00:08:41.380 --> 00:08:41.580 JOHN LEONARD: Yeah. 00:08:41.580 --> 00:08:43.030 I like the layering, and I'm less 00:08:43.030 --> 00:08:45.234 keen on the no representation. 00:08:45.234 --> 00:08:47.110 RUSS TEDRAKE: I'm curious if Giorgio-- 00:08:47.110 --> 00:08:49.419 I mean, you showed complicated system diagrams. 00:08:49.419 --> 00:08:51.460 And you're obviously doing very complicated tasks 00:08:51.460 --> 00:08:52.730 with a very complicated robot. 00:08:52.730 --> 00:08:54.855 Do you think-- do you see subsumption when you look 00:08:54.855 --> 00:08:55.570 at those-- 00:08:55.570 --> 00:08:57.320 GIORGIO METTA: Well, it's actually there. 00:08:57.320 --> 00:08:59.470 I don't have time to enter into the details. 00:08:59.470 --> 00:09:04.850 But the way it's implemented, ER, 00:09:04.850 --> 00:09:08.480 allows you to do subsumption or other things. 00:09:08.480 --> 00:09:11.710 There's a way to physically take the modules 00:09:11.710 --> 00:09:14.560 and, without modifying the modules, 00:09:14.560 --> 00:09:20.680 connect them through scripts that can insert logic 00:09:20.680 --> 00:09:24.430 on each module to sort of preprocess the messages 00:09:24.430 --> 00:09:27.970 and decide whether to subsume one of the module or not, 00:09:27.970 --> 00:09:30.800 or activate the various behaviors. 00:09:30.800 --> 00:09:33.580 So in practice, can be a subsumption architecture, 00:09:33.580 --> 00:09:37.480 a purer one or whatever other combination you have in mind 00:09:37.480 --> 00:09:38.920 for the particular task. 00:09:38.920 --> 00:09:40.961 RUSS TEDRAKE: Maybe a slightly different question 00:09:40.961 --> 00:09:42.700 is did the subsumption view of the world 00:09:42.700 --> 00:09:44.325 shape the way you designed your system? 00:09:47.640 --> 00:09:51.260 GIORGIO METTA: It may have happened without knowing, 00:09:51.260 --> 00:09:56.720 because that piece of software started at MIT while I 00:09:56.720 --> 00:09:57.650 was working with Rod. 00:09:57.650 --> 00:10:02.010 But we didn't try to build the subsumption architecture 00:10:02.010 --> 00:10:03.530 in that specific case. 00:10:03.530 --> 00:10:07.255 But the style, maybe of the publish-subscribe 00:10:07.255 --> 00:10:10.410 that we ended up doing was derived 00:10:10.410 --> 00:10:12.800 from subsumption in a sense. 00:10:12.800 --> 00:10:17.660 Was in spirit, although not in the software itself. 00:10:17.660 --> 00:10:23.900 But going back to whether there's a clear message that we 00:10:23.900 --> 00:10:26.480 can take, certainly I will subscribe 00:10:26.480 --> 00:10:31.730 to Stefanie message of the recyclable software, the fact 00:10:31.730 --> 00:10:34.190 that we can now build these modules 00:10:34.190 --> 00:10:36.020 and build large architectures. 00:10:36.020 --> 00:10:38.780 This allows doing experiments that we never dreamed 00:10:38.780 --> 00:10:41.130 of until a few years ago. 00:10:41.130 --> 00:10:46.420 So we can connect many powerful computers 00:10:46.420 --> 00:10:49.580 and run vision and control optimisation 00:10:49.580 --> 00:10:53.340 very efficiently, and especially if recycling the software. 00:10:53.340 --> 00:10:58.227 So you don't have to implement inverse kinematics everyday. 00:10:58.227 --> 00:11:00.310 PATRICK WINSTON: Well, I think the casual observer 00:11:00.310 --> 00:11:04.470 this afternoon, one good example being me, 00:11:04.470 --> 00:11:08.540 would get the sense that, with the exception of Tony, 00:11:08.540 --> 00:11:12.200 the other four of you are interested in the behavior, 00:11:12.200 --> 00:11:15.890 but not necessarily interested in understanding the biology. 00:11:15.890 --> 00:11:18.195 Is that a misimpression or is that correct? 00:11:18.195 --> 00:11:20.570 And when you mention LIDAR, for example, that's something 00:11:20.570 --> 00:11:21.560 I don't think I use. 00:11:21.560 --> 00:11:23.780 And it's doing something-- 00:11:23.780 --> 00:11:25.550 it's enabling a robot with a mechanism 00:11:25.550 --> 00:11:27.260 that is not biological. 00:11:27.260 --> 00:11:31.160 So to what degree are any of you interested in understanding 00:11:31.160 --> 00:11:35.329 the nature of biological computation? 00:11:35.329 --> 00:11:36.870 JOHN LEONARD: I care deeply about it. 00:11:36.870 --> 00:11:40.050 I just don't feel I have the tools or the bandwidth 00:11:40.050 --> 00:11:41.040 to really dive into it. 00:11:41.040 --> 00:11:46.080 But every time I talk to Matt Wilson I leave feeling in awe, 00:11:46.080 --> 00:11:49.470 like I wish I could clone myself and hang out across the street. 00:11:52.080 --> 00:11:53.180 GIORGIO METTA: Well-- 00:11:53.180 --> 00:11:53.680 Yeah. 00:11:53.680 --> 00:11:54.977 Go ahead. 00:11:54.977 --> 00:11:56.060 STEFANIE TELLEX: I get it. 00:11:56.060 --> 00:11:57.140 I kind of feel the same. 00:11:57.140 --> 00:11:58.140 I mean, I'm an engineer. 00:11:58.140 --> 00:11:58.780 I'm a hacker. 00:11:58.780 --> 00:11:59.650 I build things. 00:11:59.650 --> 00:12:02.381 And I think that's the way that I 00:12:02.381 --> 00:12:04.380 can make the most progress towards understanding 00:12:04.380 --> 00:12:07.140 intelligence is by trying to build things. 00:12:07.140 --> 00:12:10.180 But every time I talk to Josh, you know, 00:12:10.180 --> 00:12:14.277 I learn something new, and Noah and Vikash. 00:12:14.277 --> 00:12:15.610 PATRICK WINSTON: Say that again? 00:12:15.610 --> 00:12:18.068 STEFANIE TELLEX: Every time I talk to Josh and Noah Goodman 00:12:18.068 --> 00:12:21.280 and Bertram Malle, people from the psychology and cognitive 00:12:21.280 --> 00:12:23.890 science [INAUDIBLE],, I learn something 00:12:23.890 --> 00:12:25.980 and I take things away. 00:12:25.980 --> 00:12:29.502 But I don't get excited about trying to build a faithful 00:12:29.502 --> 00:12:31.710 model that incorporates everything that we know about 00:12:31.710 --> 00:12:33.084 the brain, because I just don't-- 00:12:33.084 --> 00:12:36.180 I can't put all that in my brain. 00:12:36.180 --> 00:12:38.604 And I feel that I'm better guided by my engineering 00:12:38.604 --> 00:12:40.020 intuition and the things that I've 00:12:40.020 --> 00:12:42.840 learned by trying to build systems 00:12:42.840 --> 00:12:44.185 and seeing how that plays out. 00:12:44.185 --> 00:12:45.893 PATRICK WINSTON: On the other hand, Tony, 00:12:45.893 --> 00:12:50.220 you are interested in biology and you do build stuff. 00:12:50.220 --> 00:12:51.120 TONY PRESCOTT: Yeah. 00:12:51.120 --> 00:12:53.203 I'm interested in it, because I trained originally 00:12:53.203 --> 00:12:53.970 a psychologist. 00:12:53.970 --> 00:12:59.250 So I came into robotics in order to build physical models. 00:12:59.250 --> 00:13:01.110 PATRICK WINSTON: So why do you build stuff? 00:13:01.110 --> 00:13:06.810 TONY PRESCOTT: Because I think the theory is the machine. 00:13:06.810 --> 00:13:09.920 Our theories and psychology and neuroscience 00:13:09.920 --> 00:13:12.560 are never going to be like theories are in physics. 00:13:12.560 --> 00:13:14.310 We're not going to be able to express them 00:13:14.310 --> 00:13:18.270 concisely and convince people of them in a short paper. 00:13:18.270 --> 00:13:20.190 We're going to be able to, though, build them 00:13:20.190 --> 00:13:24.300 into machines like robots and show people 00:13:24.300 --> 00:13:27.270 that they do behavior, and hopefully 00:13:27.270 --> 00:13:29.604 convince people that way that we have a complete theory. 00:13:29.604 --> 00:13:31.645 PATRICK WINSTON: So it's a demonstration purpose? 00:13:31.645 --> 00:13:32.750 Or a convincing purpose? 00:13:32.750 --> 00:13:35.280 TONY PRESCOTT: It's partly to demonstrate 00:13:35.280 --> 00:13:37.140 the sufficiency of the theory. 00:13:37.140 --> 00:13:39.420 I think that's the big reason. 00:13:39.420 --> 00:13:44.160 But another motivation that has grown more important for me 00:13:44.160 --> 00:13:49.310 is to be able to ask questions of biologists 00:13:49.310 --> 00:13:50.790 that that wouldn't occur to them. 00:13:50.790 --> 00:13:53.920 Because I think the engineering approach-- 00:13:53.920 --> 00:13:55.560 you're actually building something-- 00:13:55.560 --> 00:13:58.890 raises a lot of different questions. 00:13:58.890 --> 00:14:01.680 And those are then interesting questions 00:14:01.680 --> 00:14:05.580 to pursue in biological studies and questions that might not 00:14:05.580 --> 00:14:08.660 occur to you otherwise. 00:14:08.660 --> 00:14:10.950 So I go back to Brightenburg's comment 00:14:10.950 --> 00:14:13.320 that when you try to understand a system, 00:14:13.320 --> 00:14:16.970 there's a tendency to overestimate its complexity, 00:14:16.970 --> 00:14:19.350 and that when you do synthesis that's 00:14:19.350 --> 00:14:21.810 a whole lot different from analysis. 00:14:21.810 --> 00:14:24.090 And actually, with the brain, we tend 00:14:24.090 --> 00:14:27.540 to either underestimate or overestimate its complexity 00:14:27.540 --> 00:14:29.200 and we rarely get it right. 00:14:29.200 --> 00:14:32.130 So the things that we think are complex sometimes 00:14:32.130 --> 00:14:33.690 turn out to be easy. 00:14:33.690 --> 00:14:36.910 So an example would be in our whisker system, 00:14:36.910 --> 00:14:41.792 it's really quite easy to measure texture with a whisker. 00:14:41.792 --> 00:14:44.250 And there's lots of different ways of doing that that work. 00:14:44.250 --> 00:14:47.640 But intuitively you might not have thought that. 00:14:47.640 --> 00:14:49.440 But getting shape out of whiskers 00:14:49.440 --> 00:14:51.660 is harder, because it's an integration 00:14:51.660 --> 00:14:55.170 problem across time, and you have 00:14:55.170 --> 00:14:57.300 to track position and all these other things. 00:14:57.300 --> 00:15:00.340 So these things turn out to be hard. 00:15:00.340 --> 00:15:04.980 So I think trying to build synthetic systems helps 00:15:04.980 --> 00:15:07.560 us understand what are the real challenges the brain has 00:15:07.560 --> 00:15:09.437 to solve, and that's interesting for me. 00:15:09.437 --> 00:15:11.520 PATRICK WINSTON: Is there an example of a question 00:15:11.520 --> 00:15:13.615 that you didn't know is there when you started? 00:15:13.615 --> 00:15:13.940 TONY PRESCOTT: Yeah. 00:15:13.940 --> 00:15:15.315 PATRICK WINSTON: And you wouldn't 00:15:15.315 --> 00:15:17.160 have found if you hadn't attempted to build? 00:15:17.160 --> 00:15:18.360 TONY PRESCOTT: So when we started 00:15:18.360 --> 00:15:19.920 trying to build artificial whiskers, 00:15:19.920 --> 00:15:22.710 the engineers that were building the robot 00:15:22.710 --> 00:15:26.250 said, well, how much power does the motor have to have 00:15:26.250 --> 00:15:28.110 that drives the whisker? 00:15:28.110 --> 00:15:30.910 I mean, what happens when the whisker touches something? 00:15:30.910 --> 00:15:34.700 Does it continue to move and bend against the surface? 00:15:34.700 --> 00:15:36.190 Or does it stop? 00:15:36.190 --> 00:15:38.822 And well, we said we'll look that up in the literature. 00:15:38.822 --> 00:15:40.530 And of course, there wasn't an experiment 00:15:40.530 --> 00:15:42.180 that answered that question. 00:15:42.180 --> 00:15:44.760 So at that point we said, OK, we'll get a high speed camera 00:15:44.760 --> 00:15:46.920 and we'll start watching rats. 00:15:46.920 --> 00:15:49.350 And we found that when the whiskers touch, 00:15:49.350 --> 00:15:51.220 they stopped moving very quickly. 00:15:51.220 --> 00:15:52.840 So they make a light touch. 00:15:52.840 --> 00:15:56.371 And intuitively, yeah, maybe-- 00:15:56.371 --> 00:15:57.620 because we make a light touch. 00:15:57.620 --> 00:16:00.840 Obviously, we don't bash our hands against surfaces. 00:16:00.840 --> 00:16:02.340 But it's not obvious, necessarily, 00:16:02.340 --> 00:16:05.310 when you have a flexible sensor that that's what you would do. 00:16:05.310 --> 00:16:07.000 And in some circumstances, the rats 00:16:07.000 --> 00:16:09.420 allow their whiskers to bend against objects. 00:16:09.420 --> 00:16:13.200 So understanding when you make a light touch and when you bend 00:16:13.200 --> 00:16:16.080 was really a question that became important to us 00:16:16.080 --> 00:16:17.640 after we'd started thinking about how 00:16:17.640 --> 00:16:18.810 to engineer the system. 00:16:18.810 --> 00:16:20.675 How powerful do the motors need to be? 00:16:20.675 --> 00:16:22.800 PATRICK WINSTON: I'm very sympathetic to that view, 00:16:22.800 --> 00:16:24.420 being an engineer myself. 00:16:24.420 --> 00:16:26.729 I always say if you can't build it, 00:16:26.729 --> 00:16:28.020 you don't really understand it. 00:16:28.020 --> 00:16:29.020 TONY PRESCOTT: Yeah. 00:16:29.020 --> 00:16:30.280 PATRICK WINSTON: So many of you-- all of you 00:16:30.280 --> 00:16:32.140 have talked about impressive systems today. 00:16:32.140 --> 00:16:36.270 And I wonder if any of you would like 00:16:36.270 --> 00:16:39.625 to comment on some problem you didn't know that was there 00:16:39.625 --> 00:16:41.250 and you wouldn't have discovered if you 00:16:41.250 --> 00:16:49.769 hadn't been building the kinds of stuff that you have built. 00:16:49.769 --> 00:16:51.060 RUSS TEDRAKE: It's a long list. 00:16:51.060 --> 00:16:53.720 I mean, I think we learn a lot every day. 00:16:56.360 --> 00:16:57.280 Let me be specific. 00:16:57.280 --> 00:17:02.520 So with Atlas, we took a robot to a level of maturity 00:17:02.520 --> 00:17:05.760 that I've never taken before. 00:17:05.760 --> 00:17:09.119 I see videos from companies like Boston Dynamics 00:17:09.119 --> 00:17:10.650 that are extremely impressive. 00:17:10.650 --> 00:17:14.430 I think one of the things that separates a company like that 00:17:14.430 --> 00:17:17.790 from the results you get in a research lab 00:17:17.790 --> 00:17:21.540 is incredible amounts of hours, sort 00:17:21.540 --> 00:17:25.560 of a religion to data logging and analysis, and sort 00:17:25.560 --> 00:17:31.760 of finding corner cases, logging them, addressing them, 00:17:31.760 --> 00:17:32.760 incremental improvement. 00:17:32.760 --> 00:17:34.500 And researchers don't often do that. 00:17:34.500 --> 00:17:36.474 And actually, I think a theme that, 00:17:36.474 --> 00:17:37.890 at least in a couple of the talks, 00:17:37.890 --> 00:17:41.730 was that maybe this is actually a central requirement. 00:17:41.730 --> 00:17:43.650 And in some sense, our autonomy really 00:17:43.650 --> 00:17:47.110 should be well suited to doing that, 00:17:47.110 --> 00:17:49.500 to maybe automatically finding corner cases 00:17:49.500 --> 00:17:51.880 and proving robustness and all these things. 00:17:51.880 --> 00:18:01.080 But the places that broke our theory were weird. 00:18:01.080 --> 00:18:04.620 I mean, so the stiction in the joints of Atlas 00:18:04.620 --> 00:18:06.000 is just dominant. 00:18:06.000 --> 00:18:07.500 So we do torque control, but we have 00:18:07.500 --> 00:18:09.958 to send a feedforward velocity signal to get over friction. 00:18:09.958 --> 00:18:13.170 If we're started at zero and we have to start moving, 00:18:13.170 --> 00:18:15.390 if we don't send a feedforward velocity signal, 00:18:15.390 --> 00:18:18.780 our model is just completely wrong. 00:18:18.780 --> 00:18:20.570 When you're walking on cinder blocks 00:18:20.570 --> 00:18:25.600 and you go near the ankle limit in pitch, 00:18:25.600 --> 00:18:27.600 there's a strange coupling between the mechanism 00:18:27.600 --> 00:18:28.980 which causes the ankle to roll. 00:18:28.980 --> 00:18:30.896 And it'll kick your robot over just like that, 00:18:30.896 --> 00:18:32.120 if you don't watch for it. 00:18:32.120 --> 00:18:34.380 And we thought about putting that into our model, 00:18:34.380 --> 00:18:36.420 addressing it with sophisticated things. 00:18:36.420 --> 00:18:38.050 It's hard and ugly and gross. 00:18:38.050 --> 00:18:42.636 And it's possible, but we do things to just-- 00:18:42.636 --> 00:18:44.010 you know, Band-Aid solution that. 00:18:44.010 --> 00:18:48.570 And I think there's all this stuff, all these details that 00:18:48.570 --> 00:18:49.920 come in. 00:18:49.920 --> 00:18:52.140 I think the theory should address it all. 00:18:52.140 --> 00:18:53.580 I think we did pretty well. 00:18:53.580 --> 00:18:55.170 I'd say we got 80% of the way there, 00:18:55.170 --> 00:18:57.810 70%, 80% of the way there with our theory this time. 00:18:57.810 --> 00:18:59.970 And then we just decided that there was a deadline 00:18:59.970 --> 00:19:02.077 and we had to cover some stuff up with band-aids. 00:19:02.077 --> 00:19:03.160 But that's the good stuff. 00:19:03.160 --> 00:19:04.985 That's the stuff we should be focused on. 00:19:04.985 --> 00:19:07.110 That's the stuff we should be devoting our research 00:19:07.110 --> 00:19:07.890 efforts on. 00:19:07.890 --> 00:19:10.890 PATRICK WINSTON: If you were to go into a log cabin today 00:19:10.890 --> 00:19:14.451 and write a book on Atlas and your work on it, 00:19:14.451 --> 00:19:16.700 what fraction of that book would be about corner cases 00:19:16.700 --> 00:19:19.800 and which fraction would be about principles? 00:19:19.800 --> 00:19:24.992 RUSS TEDRAKE: We really stuck to principles until last November. 00:19:24.992 --> 00:19:25.950 That was our threshold. 00:19:25.950 --> 00:19:27.870 November, we had to send the robot back for upgrades. 00:19:27.870 --> 00:19:30.120 We said, until then, we're going to do research. 00:19:30.120 --> 00:19:32.580 The code base is going to be clean. 00:19:32.580 --> 00:19:36.060 And then when we got the robot back in January, 00:19:36.060 --> 00:19:38.430 we did everything we needed to to make the robot compete 00:19:38.430 --> 00:19:40.710 in the challenge. 00:19:40.710 --> 00:19:43.030 So I think 70% or 80% of the way, we got there. 00:19:43.030 --> 00:19:45.210 And that it was just putting hours on the robot, 00:19:45.210 --> 00:19:47.370 finding those screw cases. 00:19:47.370 --> 00:19:49.860 And then, if I were to write a book in five years, 00:19:49.860 --> 00:19:51.582 I hope it would be-- 00:19:51.582 --> 00:19:54.329 PATRICK WINSTON: Is there a principle on that ankle roll? 00:19:54.329 --> 00:19:54.870 Or was that-- 00:19:54.870 --> 00:19:55.350 RUSS TEDRAKE: Oh, absolutely. 00:19:55.350 --> 00:19:57.240 We could have thrown that into the model. 00:19:57.240 --> 00:19:59.240 It just would have increased the dimensionality. 00:19:59.240 --> 00:20:01.180 It would have been a non-linear term. 00:20:01.180 --> 00:20:03.625 We could have done it, we just didn't have time 00:20:03.625 --> 00:20:04.500 to do it at the time. 00:20:04.500 --> 00:20:06.300 And it was going to be one of many things 00:20:06.300 --> 00:20:08.700 we would have had to do if we had taken the principle 00:20:08.700 --> 00:20:10.260 approach throughout. 00:20:10.260 --> 00:20:12.480 There's other things that we couldn't have put nicely 00:20:12.480 --> 00:20:16.690 into the model, that we would have needed to address. 00:20:16.690 --> 00:20:18.970 And that should be our agenda. 00:20:18.970 --> 00:20:21.060 TONY PRESCOTT: If that was your own robot, 00:20:21.060 --> 00:20:22.990 would you have just re-engineered the ankle 00:20:22.990 --> 00:20:25.830 to make that problem less of an issue? 00:20:25.830 --> 00:20:27.510 RUSS TEDRAKE: It wasn't about the ankle. 00:20:27.510 --> 00:20:29.010 It was about the fact that there's always 00:20:29.010 --> 00:20:30.690 going to be something unmodeled that's 00:20:30.690 --> 00:20:32.940 going to come up and get you. 00:20:32.940 --> 00:20:35.700 And with robots, I think we're data starved. 00:20:35.700 --> 00:20:38.670 We don't have the big data problem in robotics yet. 00:20:38.670 --> 00:20:40.890 And I think you're limited by the hours you 00:20:40.890 --> 00:20:41.970 can put on your robot. 00:20:41.970 --> 00:20:46.050 We need to think about how do you aggressively 00:20:46.050 --> 00:20:48.180 search for the cases that are going to get you? 00:20:48.180 --> 00:20:52.350 How do you prove robustness to unmodeled things? 00:20:52.350 --> 00:20:55.650 I think this is fundamental. 00:20:55.650 --> 00:20:58.350 It's not a theme I would have prioritized if I hadn't 00:20:58.350 --> 00:20:59.569 gotten this far with a robot. 00:20:59.569 --> 00:21:01.860 PATRICK WINSTON: But Giorgio, what about building iCub? 00:21:01.860 --> 00:21:04.410 Were there big problems that emerged 00:21:04.410 --> 00:21:07.650 in the course of building that you hadn't anticipated? 00:21:07.650 --> 00:21:10.160 GIORGIO METTA: Well first of all, 00:21:10.160 --> 00:21:13.020 there's a problem of power. 00:21:13.020 --> 00:21:14.640 I guess for Atlas it's very different. 00:21:14.640 --> 00:21:19.710 But for electric motors, you're always short of space, 00:21:19.710 --> 00:21:23.340 short of space where to put the actuators. 00:21:23.340 --> 00:21:26.790 And you start filling the available-- if you 00:21:26.790 --> 00:21:29.860 have a target size, you start filling it very quickly. 00:21:29.860 --> 00:21:32.890 And there's no room for anything else. 00:21:32.890 --> 00:21:36.270 And then you start sticking the electronics wherever you can, 00:21:36.270 --> 00:21:40.020 and cables and everything. 00:21:40.020 --> 00:21:46.500 Certainly if-- I mean, a difference 00:21:46.500 --> 00:21:52.270 in design from the biological actuators and the artificial 00:21:52.270 --> 00:21:55.060 actuators to make life very difficult. 00:21:55.060 --> 00:21:58.570 And especially when you have to go through something 00:21:58.570 --> 00:22:00.370 like a hand, where you like to have 00:22:00.370 --> 00:22:01.930 a lot of degrees of freedom. 00:22:01.930 --> 00:22:05.340 But there's no way you could actually build it, 00:22:05.340 --> 00:22:07.990 so you have to take shortcuts here and there. 00:22:07.990 --> 00:22:12.480 And I guess the same is true, then, for computation. 00:22:12.480 --> 00:22:16.960 And you resort to putting as many micro-controllers 00:22:16.960 --> 00:22:18.940 as you can inside the robot, because you 00:22:18.940 --> 00:22:21.130 want to have efficient control loops. 00:22:21.130 --> 00:22:24.730 And then you say, well, maybe have a cable 00:22:24.730 --> 00:22:28.480 for a proper image processing because there's 00:22:28.480 --> 00:22:31.740 no way you can squeeze that into the robot itself. 00:22:31.740 --> 00:22:33.220 It's not surprising. 00:22:33.220 --> 00:22:36.260 It's just a matter of when you're doing to design, 00:22:36.260 --> 00:22:39.670 you soon discover that there are limitations 00:22:39.670 --> 00:22:41.400 you have to take into account. 00:22:41.400 --> 00:22:43.280 I don't know whether it is surprising. 00:22:43.280 --> 00:22:46.030 I mean, I guess we learn the lesson 00:22:46.030 --> 00:22:48.610 across many years of design. 00:22:48.610 --> 00:22:50.960 We designed other robots before the iCub. 00:22:50.960 --> 00:22:53.290 We sort of-- 00:22:53.290 --> 00:22:56.730 I thought we knew where the limits where 00:22:56.730 --> 00:22:58.024 with the current technology. 00:22:58.024 --> 00:22:59.440 PATRICK WINSTON: I wonder if the-- 00:22:59.440 --> 00:23:01.570 you say you learned a lot building iCub. 00:23:01.570 --> 00:23:06.790 I wonder if this knowledge is accessible. 00:23:06.790 --> 00:23:08.670 It's knowledge that you discussed 00:23:08.670 --> 00:23:12.450 in meetings and seminars, and thought about at night 00:23:12.450 --> 00:23:13.760 and fixed it the next day. 00:23:13.760 --> 00:23:15.400 Is any of it-- 00:23:15.400 --> 00:23:18.970 if I wanted to build iCub and couldn't talk to you, 00:23:18.970 --> 00:23:22.505 would I have to start from scratch? 00:23:22.505 --> 00:23:24.880 I know you've got the stuff on the web and whatnot, but-- 00:23:24.880 --> 00:23:25.120 GIORGIO METTA: Yeah. 00:23:25.120 --> 00:23:26.304 That's probably enough for-- 00:23:26.304 --> 00:23:27.720 PATRICK WINSTON: --reasons in them 00:23:27.720 --> 00:23:28.594 GIORGIO METTA: Sorry? 00:23:28.594 --> 00:23:30.880 PATRICK WINSTON: Your web material has designs, 00:23:30.880 --> 00:23:32.290 but it doesn't have reasons. 00:23:32.290 --> 00:23:33.123 GIORGIO METTA: Yeah. 00:23:33.123 --> 00:23:34.360 Yeah, that's-- that's right. 00:23:34.360 --> 00:23:39.190 No, the other thing we documented the process itself. 00:23:39.190 --> 00:23:42.910 So that information, I don't know, 00:23:42.910 --> 00:23:46.810 resides in the people that actually made the choices when 00:23:46.810 --> 00:23:48.310 we were doing the design. 00:23:48.310 --> 00:23:52.650 There's one other thing that maybe is important, is that-- 00:23:52.650 --> 00:23:56.290 so the iCub is about 5,000 parts. 00:23:56.290 --> 00:23:58.240 And that's not good, because there 00:23:58.240 --> 00:24:00.730 are 5,000 parts that can break. 00:24:00.730 --> 00:24:06.610 And that may be something interesting for design 00:24:06.610 --> 00:24:11.635 of materials for robots, or new ways of building the robots. 00:24:14.620 --> 00:24:16.390 And at the moment, basically everything 00:24:16.390 --> 00:24:20.000 that could potentially break has happened, 00:24:20.000 --> 00:24:26.150 that it failed on the iCub over many years. 00:24:26.150 --> 00:24:30.310 Even parts that theoretically we didn't think could break, 00:24:30.310 --> 00:24:31.720 well, they could. 00:24:31.720 --> 00:24:35.440 But we estimated maximum torques, 00:24:35.440 --> 00:24:37.210 whatever, and then it happened. 00:24:37.210 --> 00:24:41.020 Somebody did something silly and we broke a shoulder. 00:24:41.020 --> 00:24:44.520 It's a steel part that we never thought 00:24:44.520 --> 00:24:46.550 it could actually break. 00:24:46.550 --> 00:24:50.080 And it completely failed. 00:24:50.080 --> 00:24:54.790 I mean, those type of things are maybe interesting 00:24:54.790 --> 00:24:58.250 for future designs, or for either simplifying 00:24:58.250 --> 00:25:03.490 the number of parts or figuring out ways of doing less parts 00:25:03.490 --> 00:25:08.170 or, let's say, different ways of actually building 00:25:08.170 --> 00:25:09.834 the mechanics of the robot. 00:25:09.834 --> 00:25:11.500 PATRICK WINSTON: I suppose I bring it up 00:25:11.500 --> 00:25:15.850 because some of us in CSAIL are addressing-- 00:25:15.850 --> 00:25:17.380 not me, but some people in CSAIL are 00:25:17.380 --> 00:25:20.560 interested in how you capture design rationale, 00:25:20.560 --> 00:25:22.195 how you capture those conversations, 00:25:22.195 --> 00:25:24.580 those whiteboard sketches and all of that sort of thing 00:25:24.580 --> 00:25:28.900 so that the next generation can learn by some mechanism other 00:25:28.900 --> 00:25:32.260 than apprenticeship. 00:25:32.260 --> 00:25:33.010 But let's see. 00:25:33.010 --> 00:25:35.570 Where to go from here? 00:25:35.570 --> 00:25:37.900 iCub is obviously a major undertaking 00:25:37.900 --> 00:25:39.760 and Russ had been working like a slave 00:25:39.760 --> 00:25:41.590 for three years on Atlas robot. 00:25:44.110 --> 00:25:47.006 Do you-- I don't know quite how to phrase this 00:25:47.006 --> 00:25:48.130 without being too trumpish. 00:25:48.130 --> 00:25:51.400 But the soldering time to thinking time 00:25:51.400 --> 00:25:56.950 must be very high on projects like this. 00:25:56.950 --> 00:25:58.060 Is that your sense? 00:25:58.060 --> 00:26:00.910 Or do you think that the building of these things 00:26:00.910 --> 00:26:04.690 is actually essential to working out the ideas? 00:26:04.690 --> 00:26:07.420 Maybe that's not quite the question I'm going to ask. 00:26:07.420 --> 00:26:09.880 Maybe a sharper question is given the high ratio 00:26:09.880 --> 00:26:11.442 of soldering time to thinking time, 00:26:11.442 --> 00:26:13.150 is it something that a student should do? 00:26:16.450 --> 00:26:19.150 RUSS TEDRAKE: I'm lucky that someone else built 00:26:19.150 --> 00:26:20.280 the robot for us. 00:26:20.280 --> 00:26:23.685 Giorgio has done much more than we have in this regard. 00:26:23.685 --> 00:26:25.810 PATRICK WINSTON: Well, by soldering time you know-- 00:26:25.810 --> 00:26:26.110 RUSS TEDRAKE: I know. 00:26:26.110 --> 00:26:26.901 Yeah, yeah, sure. 00:26:26.901 --> 00:26:27.400 We-- 00:26:27.400 --> 00:26:28.240 PATRICK WINSTON: It's a metaphor. 00:26:28.240 --> 00:26:29.300 RUSS TEDRAKE: Yeah. 00:26:29.300 --> 00:26:34.570 but we got pretty far into it with the good graces 00:26:34.570 --> 00:26:38.530 of DARPA and Google slash Boston Dynamics. 00:26:38.530 --> 00:26:41.570 The software is where we've invested our solder time. 00:26:41.570 --> 00:26:43.950 A huge amount of software engineering effort. 00:26:43.950 --> 00:26:47.150 I spent countless hours on setting up 00:26:47.150 --> 00:26:48.300 build servers and stuff. 00:26:51.495 --> 00:26:54.510 Am I stronger, you know, am I better for it? 00:26:54.510 --> 00:26:59.230 I think having invested, we can do research very fast now. 00:26:59.230 --> 00:27:04.050 So I'm in a new position to be able to try 00:27:04.050 --> 00:27:06.270 really complicated ideas very quickly 00:27:06.270 --> 00:27:08.190 because of that investment. 00:27:08.190 --> 00:27:10.050 It was actually-- I knew going in what 00:27:10.050 --> 00:27:11.340 I was going to be doing. 00:27:11.340 --> 00:27:13.980 I saw what John and other people got out 00:27:13.980 --> 00:27:16.972 of being in the Urban Challenge, including in especially 00:27:16.972 --> 00:27:19.180 the tools, like the LCM that we've been talking about 00:27:19.180 --> 00:27:20.340 and stuff today. 00:27:20.340 --> 00:27:22.500 And I wanted that for my group. 00:27:22.500 --> 00:27:24.480 So it was a very conscious decision. 00:27:24.480 --> 00:27:28.260 I'm at a place now that we can do fantastic research. 00:27:28.260 --> 00:27:32.130 Every one of the students involved in great research work 00:27:32.130 --> 00:27:33.887 on the project. 00:27:33.887 --> 00:27:35.970 We hired a few staff programmers to help with some 00:27:35.970 --> 00:27:38.410 of the non-research stuff. 00:27:38.410 --> 00:27:40.860 And I think the hardware is important. 00:27:40.860 --> 00:27:43.455 It's hard to balance, but I do think it's important. 00:27:43.455 --> 00:27:45.830 PATRICK WINSTON: Just one short follow-up question there. 00:27:45.830 --> 00:27:48.000 Earlier you said that some of your students 00:27:48.000 --> 00:27:49.430 didn't want to work on it. 00:27:49.430 --> 00:27:50.370 And why was that? 00:27:50.370 --> 00:27:53.899 Was that a principal reason? 00:27:53.899 --> 00:27:56.190 RUSS TEDRAKE: People knew how much soldering time there 00:27:56.190 --> 00:27:56.891 was going to be. 00:27:56.891 --> 00:27:57.390 Right? 00:27:57.390 --> 00:28:01.830 And the people that had their research agenda and it 00:28:01.830 --> 00:28:03.960 was more theoretical, and they didn't 00:28:03.960 --> 00:28:05.610 want that soldering time. 00:28:05.610 --> 00:28:08.070 Other people said, I'm still looking for ideas. 00:28:08.070 --> 00:28:10.230 This is going to motivate me for my future work. 00:28:10.230 --> 00:28:11.140 They jumped right in. 00:28:11.140 --> 00:28:14.537 And super strong students made different decisions on that. 00:28:14.537 --> 00:28:16.870 PATRICK WINSTON: And they both made the right decisions. 00:28:16.870 --> 00:28:17.360 RUSS TEDRAKE: I think so. 00:28:17.360 --> 00:28:17.875 PATRICK WINSTON: Yeah. 00:28:17.875 --> 00:28:18.250 RUSS TEDRAKE: Yeah. 00:28:18.250 --> 00:28:18.960 PATRICK WINSTON: But John, you've also 00:28:18.960 --> 00:28:21.120 been involved in-- well, the just driving car thing 00:28:21.120 --> 00:28:24.330 was a major DARPA grand challenge. 00:28:24.330 --> 00:28:26.790 Some people have been critical of these grand challenges 00:28:26.790 --> 00:28:29.460 because they say that, well, they 00:28:29.460 --> 00:28:31.290 drive the technology up the closest hill, 00:28:31.290 --> 00:28:33.207 but they don't get you on a different hill. 00:28:33.207 --> 00:28:35.040 Do you have any feelings about these things, 00:28:35.040 --> 00:28:37.740 if they're good idea in retrospect, 00:28:37.740 --> 00:28:39.360 having participated in them? 00:28:39.360 --> 00:28:40.360 JOHN LEONARD: Let's see. 00:28:40.360 --> 00:28:43.110 I'm really torn on that one because I 00:28:43.110 --> 00:28:46.152 see how the short term benefits of the community-- and you 00:28:46.152 --> 00:28:47.860 can point to things like the Google car-- 00:28:47.860 --> 00:28:50.150 that there's a clear impact. 00:28:50.150 --> 00:28:53.580 But DARPA does have a mindset that once they've 00:28:53.580 --> 00:28:55.750 done something, they declare victory and move on. 00:28:55.750 --> 00:28:58.740 So now if you work, say, on legged locomotion, which 00:28:58.740 --> 00:29:01.030 one of my junior colleagues does, 00:29:01.030 --> 00:29:02.590 DARPA won't answer his emails. 00:29:02.590 --> 00:29:04.910 It's like, OK, we did legged locomotion. 00:29:04.910 --> 00:29:07.530 And so I think that the challenge 00:29:07.530 --> 00:29:10.830 is to be mindful of where we are in terms 00:29:10.830 --> 00:29:13.300 of the real long-term progress. 00:29:13.300 --> 00:29:15.810 And it's not an easy conversation 00:29:15.810 --> 00:29:18.745 to have with the funding agencies, but-- 00:29:18.745 --> 00:29:20.370 PATRICK WINSTON: But what about a brand 00:29:20.370 --> 00:29:22.140 new way of doing something that is not 00:29:22.140 --> 00:29:24.990 going to be competitive in terms of demonstration for a while? 00:29:27.910 --> 00:29:30.540 Is that a problem that's amplified 00:29:30.540 --> 00:29:32.040 by these DARPA grand challenges? 00:29:32.040 --> 00:29:33.670 I mean, take chess, for example. 00:29:33.670 --> 00:29:36.900 If you had a great idea about how humans play chess, 00:29:36.900 --> 00:29:40.220 you would never be competitive with Deep Blue, 00:29:40.220 --> 00:29:41.680 or not for a long time. 00:29:41.680 --> 00:29:43.830 So you wouldn't be in a DARPA program 00:29:43.830 --> 00:29:47.580 that was aimed at doing chess. 00:29:47.580 --> 00:29:48.777 So is that a-- 00:29:48.777 --> 00:29:49.985 do you see that as a problem? 00:29:52.821 --> 00:29:54.570 RUSS TEDRAKE: I think it's a huge problem. 00:29:54.570 --> 00:29:56.610 But I still see a role for these kind 00:29:56.610 --> 00:29:59.880 of competitions as benchmarks. 00:29:59.880 --> 00:30:01.860 And I wouldn't do another one today. 00:30:01.860 --> 00:30:04.890 I mean, for me it was the right time 00:30:04.890 --> 00:30:07.900 to sort of see how our theory had gotten, 00:30:07.900 --> 00:30:10.860 try it on a much more complicated robot, benchmark 00:30:10.860 --> 00:30:12.820 where we are, get some new ideas going forward. 00:30:12.820 --> 00:30:15.540 It was perfect for me. 00:30:15.540 --> 00:30:17.190 But you can't set a research agenda. 00:30:17.190 --> 00:30:18.290 JOHN LEONARD: And they're dangerous for students. 00:30:18.290 --> 00:30:20.190 So one of our strongest students never 00:30:20.190 --> 00:30:24.180 got his PhD, because his wife is in program PhD 00:30:24.180 --> 00:30:26.970 program in biology and he did the DARPA challenge. 00:30:26.970 --> 00:30:28.376 And she finished her thesis. 00:30:28.376 --> 00:30:30.750 And he said, I don't want to live alone on the east coast 00:30:30.750 --> 00:30:33.010 while she starts her faculty position in California. 00:30:33.010 --> 00:30:34.155 So I'm out of here. 00:30:34.155 --> 00:30:37.446 And that's the sort of thing. 00:30:37.446 --> 00:30:39.570 STEFANIE TELLEX: I kind of made different decisions 00:30:39.570 --> 00:30:40.980 about that over my career. 00:30:40.980 --> 00:30:44.790 So when I was a post-doc at MIT, I really, 00:30:44.790 --> 00:30:47.094 really, really worked to avoid soldering time. 00:30:47.094 --> 00:30:47.760 I was fortunate. 00:30:47.760 --> 00:30:48.974 I kind of walked around. 00:30:48.974 --> 00:30:50.890 There's all these great robot robotic systems. 00:30:50.890 --> 00:30:52.910 And I would bolt language on, and get one paper. 00:30:52.910 --> 00:30:55.380 And bolt language on another way and get another paper. 00:30:55.380 --> 00:30:58.340 And you look to get a faculty position, 00:30:58.340 --> 00:31:00.990 you have to have this focused research agenda. 00:31:00.990 --> 00:31:03.200 And so I was focused on that. 00:31:03.200 --> 00:31:03.880 And it worked. 00:31:03.880 --> 00:31:06.180 I think it was a very productive time for me. 00:31:06.180 --> 00:31:08.700 But I really valued the past two years at Brown, 00:31:08.700 --> 00:31:12.610 where there's not as much other roboticists around there. 00:31:12.610 --> 00:31:16.560 So I've really been forced to broaden myself as a roboticist 00:31:16.560 --> 00:31:19.230 and spend a lot more time soldering, 00:31:19.230 --> 00:31:23.420 making this system for pick and place on Baxter. 00:31:23.420 --> 00:31:24.920 The first year I didn't hack at all, 00:31:24.920 --> 00:31:26.040 and the second year I started hacking 00:31:26.040 --> 00:31:27.440 on that system, the one that was doing 00:31:27.440 --> 00:31:28.648 the grasping with my student. 00:31:28.648 --> 00:31:30.570 It was the best decision I ever made. 00:31:30.570 --> 00:31:33.850 I learned so much about the abstractions. 00:31:33.850 --> 00:31:36.210 Because the problems that we needed 00:31:36.210 --> 00:31:39.210 to solve at the beginning, before I started hacking, 00:31:39.210 --> 00:31:40.562 I just didn't understand. 00:31:40.562 --> 00:31:42.270 The problems I thought we needed to solve 00:31:42.270 --> 00:31:43.853 were not the problems that we actually 00:31:43.853 --> 00:31:46.326 needed to solve to make the robot do something useful. 00:31:46.326 --> 00:31:47.700 And I don't think there's any way 00:31:47.700 --> 00:31:51.180 we could have gotten to that knowledge without hacking 00:31:51.180 --> 00:31:54.090 and trying to build it. 00:31:54.090 --> 00:31:57.030 GIORGIO METTA: In our case, we've 00:31:57.030 --> 00:32:00.120 been lucky, in a sense, that we had 00:32:00.120 --> 00:32:07.230 resources in terms of engineers that could do the soldering. 00:32:07.230 --> 00:32:10.500 So at the moment we still have about 25 people that 00:32:10.500 --> 00:32:12.930 are just doing the soldering. 00:32:12.930 --> 00:32:15.070 So it's just a large number. 00:32:15.070 --> 00:32:15.730 Yeah. 00:32:15.730 --> 00:32:18.540 PATRICK WINSTON: That would look like a battalion, something 00:32:18.540 --> 00:32:19.363 like that. 00:32:19.363 --> 00:32:21.430 JOHN LEONARD: Can I say something more generally? 00:32:21.430 --> 00:32:26.190 So there's a lot of claims in the media and sort 00:32:26.190 --> 00:32:29.350 of hyped fears about robots that take over the world, 00:32:29.350 --> 00:32:31.320 or sort of very strong AI. 00:32:31.320 --> 00:32:34.020 And partly, they sometimes they point to Moore's law 00:32:34.020 --> 00:32:36.390 as this evidence of great progress. 00:32:36.390 --> 00:32:39.570 But I would say that in robotics we're 00:32:39.570 --> 00:32:43.380 lacking high-performance commodity robot 00:32:43.380 --> 00:32:46.466 hardware that lets us make tremendous progress. 00:32:46.466 --> 00:32:48.840 And so things like Baxter are great because they're cheap 00:32:48.840 --> 00:32:51.131 and they're safe, and they're a step in that direction. 00:32:51.131 --> 00:32:54.030 But I think we're going to look back 20 years from now 00:32:54.030 --> 00:32:57.870 and say, how did we make any progress with the robots we 00:32:57.870 --> 00:32:58.680 had at the time? 00:32:58.680 --> 00:33:02.790 Like, we really need better robots that just get massively 00:33:02.790 --> 00:33:04.210 out there in the labs. 00:33:04.210 --> 00:33:05.414 RUSS TEDRAKE: But-- 00:33:05.414 --> 00:33:07.080 TONY PRESCOTT: I was going to echo that, 00:33:07.080 --> 00:33:12.501 because I think robotics is massively interdisciplinary. 00:33:12.501 --> 00:33:15.840 And you've maybe got people more towards control here slightly. 00:33:18.946 --> 00:33:20.820 What we're trying to do in Sheffield robotics 00:33:20.820 --> 00:33:23.670 is actually bringing in more of the other disciplines 00:33:23.670 --> 00:33:26.760 in engineering, but also science, social science. 00:33:26.760 --> 00:33:29.710 Everybody has a potential contribution to make. 00:33:29.710 --> 00:33:31.680 Certainly electronic engineering, 00:33:31.680 --> 00:33:33.600 mechanical engineering. 00:33:33.600 --> 00:33:37.800 Soft robotics, I think, depends very much on new materials, 00:33:37.800 --> 00:33:39.290 materials science. 00:33:39.290 --> 00:33:41.970 And then these things have different control challenges. 00:33:41.970 --> 00:33:44.580 But sometimes the control problem 00:33:44.580 --> 00:33:48.540 is really simplified if you have the right material substrates. 00:33:48.540 --> 00:33:51.060 So if you can solve Giorgio's problem, having 00:33:51.060 --> 00:33:54.605 a powerful actuator, then his problem of building iCub 00:33:54.605 --> 00:33:55.890 is much simplified. 00:33:55.890 --> 00:33:58.200 So I think we have to think of robotics 00:33:58.200 --> 00:34:01.190 as this large, multi-disciplinary enterprise. 00:34:01.190 --> 00:34:05.040 And if we're going to build robots that are useful, 00:34:05.040 --> 00:34:06.944 you have to pull in all this expertise. 00:34:06.944 --> 00:34:08.819 And we're interested in pulling the expertise 00:34:08.819 --> 00:34:11.760 in from social science as well. 00:34:11.760 --> 00:34:13.775 Because I think one of the major problems 00:34:13.775 --> 00:34:16.710 that we will face in AI and in robotics 00:34:16.710 --> 00:34:20.350 is kind of backlash, which is already happening. 00:34:20.350 --> 00:34:22.070 Do we really want these machines? 00:34:22.070 --> 00:34:24.300 And how are they going to change the world? 00:34:24.300 --> 00:34:26.425 Understanding what the impacts will be 00:34:26.425 --> 00:34:29.520 and trying to build in safeguards 00:34:29.520 --> 00:34:32.471 against the negative impact is something we should work on. 00:34:32.471 --> 00:34:34.679 PATRICK WINSTON: But Giorgio had on one of the slides 00:34:34.679 --> 00:34:39.630 that one of the reasons for doing all this was fun. 00:34:39.630 --> 00:34:42.000 And I wonder to what degree that is the motivation? 00:34:42.000 --> 00:34:45.580 Because all of you talked about how difficult the problems are, 00:34:45.580 --> 00:34:48.420 and some of them are like the ones you talked about, 00:34:48.420 --> 00:34:50.579 John, watching that policeman say, 00:34:50.579 --> 00:34:51.870 go ahead through the red light. 00:34:51.870 --> 00:34:56.880 Those seem not insurmountable, but very tough and sound 00:34:56.880 --> 00:34:59.350 like they would take five decades. 00:34:59.350 --> 00:35:02.917 So is the motivation largely that it's fun? 00:35:02.917 --> 00:35:04.500 RUSS TEDRAKE: That's a big part of it. 00:35:04.500 --> 00:35:07.920 I mean, so we've done some work on steady aerodynamics 00:35:07.920 --> 00:35:12.300 and the like, too, and I like-- so we made robotic birds. 00:35:12.300 --> 00:35:15.150 And I tried to make robotic birds like on a perch. 00:35:15.150 --> 00:35:18.180 And then we had a small side project 00:35:18.180 --> 00:35:21.450 where we tried to show that the exact same technology could 00:35:21.450 --> 00:35:23.126 help a wind turbine be more efficient. 00:35:23.126 --> 00:35:24.000 PATRICK WINSTON: Yeah 00:35:24.000 --> 00:35:26.860 RUSS TEDRAKE: And that's the important problem. 00:35:26.860 --> 00:35:30.612 I could have easily started off and done some of the same work 00:35:30.612 --> 00:35:33.070 by saying I was going to make wind turbines more efficient. 00:35:33.070 --> 00:35:34.528 I was going to study pitch control. 00:35:34.528 --> 00:35:37.420 I'd be very serious about that. 00:35:37.420 --> 00:35:38.920 But I did it the other way around. 00:35:38.920 --> 00:35:40.806 I wanted to try to make a robot bird. 00:35:40.806 --> 00:35:42.180 And I think the win-- not only do 00:35:42.180 --> 00:35:45.989 I get excited going in, try to make a bird 00:35:45.989 --> 00:35:48.030 fly for the first time instead of getting 2% more 00:35:48.030 --> 00:35:49.590 efficient on a wind turbine. 00:35:49.590 --> 00:35:52.286 I mean, I go in more excited, but I also-- 00:35:52.286 --> 00:35:53.910 I get to recruit the very best students 00:35:53.910 --> 00:35:56.120 in the world because of it. 00:35:56.120 --> 00:35:59.780 There's just so many good reasons to do that. 00:35:59.780 --> 00:36:02.280 Sometimes it makes me feel a little shallow because the wind 00:36:02.280 --> 00:36:05.670 turbine's way more important than a robotic bird. 00:36:05.670 --> 00:36:09.274 But that is-- the fun is the choice. 00:36:09.274 --> 00:36:10.940 PATRICK WINSTON: What about it, Giorgio? 00:36:10.940 --> 00:36:11.670 Do you do it-- 00:36:11.670 --> 00:36:14.340 or you have a huge group there. 00:36:14.340 --> 00:36:17.790 Somebody must be paid for all those people. 00:36:17.790 --> 00:36:23.005 Are they in expectation of applications in the near term? 00:36:23.005 --> 00:36:23.880 GIORGIO METTA: Sorry. 00:36:23.880 --> 00:36:25.921 PATRICK WINSTON: You have a huge group of people. 00:36:25.921 --> 00:36:27.030 GIORGIO METTA: Yeah. 00:36:27.030 --> 00:36:31.050 I mean, the group is mainly funded internally 00:36:31.050 --> 00:36:39.440 by IIT, which is public funding for large groups basically. 00:36:39.440 --> 00:36:45.240 And actually, the robotics program at IIT is even larger-- 00:36:45.240 --> 00:36:47.790 the group on the iCub is actually-- 00:36:47.790 --> 00:36:51.200 there are four PIs working on it, 00:36:51.200 --> 00:36:53.970 and collaborations with other people 00:36:53.970 --> 00:36:58.370 like with the IIT-MIT group also. 00:36:58.370 --> 00:37:03.710 But the overall robotics program at IIT's about 250 people, 00:37:03.710 --> 00:37:04.710 I would say. 00:37:04.710 --> 00:37:09.120 So they're certainly also part of the reason 00:37:09.120 --> 00:37:13.785 why we've been able to go for a complicated platform. 00:37:13.785 --> 00:37:17.330 There was one that actually participated in the DARPA 00:37:17.330 --> 00:37:18.950 robotics challenge. 00:37:18.950 --> 00:37:24.000 There's people doing quadrupeds and people doing robotics 00:37:24.000 --> 00:37:25.257 for rehabilitation. 00:37:25.257 --> 00:37:26.340 So there's various things. 00:37:26.340 --> 00:37:28.673 PATRICK WINSTON: So there must be princes and princesses 00:37:28.673 --> 00:37:31.680 of science back there somewhere who view this as a long-term 00:37:31.680 --> 00:37:33.330 investment that will have some-- 00:37:33.330 --> 00:37:35.288 GIORGIO METTA: It was in the scientific program 00:37:35.288 --> 00:37:36.970 in the Institute to invest in robotics. 00:37:36.970 --> 00:37:41.670 And while one day they may be looking at the results 00:37:41.670 --> 00:37:45.020 and see whether we've done a good job, 00:37:45.020 --> 00:37:48.140 desire to fire us all, whatever. 00:37:48.140 --> 00:37:51.790 Hey man, that might be the case. 00:37:51.790 --> 00:37:54.930 And I think it was-- 00:37:54.930 --> 00:37:58.770 unexpectedly IIT started in 2006. 00:37:58.770 --> 00:38:00.900 And the centrifuge program include 00:38:00.900 --> 00:38:05.550 robotics and all the hype about robotics 00:38:05.550 --> 00:38:09.180 that started in recent years, Google acquiring companies, 00:38:09.180 --> 00:38:13.290 this and that, I think in hindsight 00:38:13.290 --> 00:38:17.257 has been a good choice to be in robotics at that time. 00:38:17.257 --> 00:38:18.465 Just by sheer luck, probably. 00:38:21.174 --> 00:38:22.590 RUSS TEDRAKE: To be clear, I think 00:38:22.590 --> 00:38:25.290 we're having fun but solving all the right problems while-- 00:38:25.290 --> 00:38:26.510 I think we just sort of-- 00:38:26.510 --> 00:38:27.160 yeah. 00:38:27.160 --> 00:38:28.620 We lucked out, maybe, a little bit. 00:38:28.620 --> 00:38:31.120 But we found a way to have fun and solve the right problems. 00:38:31.120 --> 00:38:32.530 So I don't feel that we're-- 00:38:32.530 --> 00:38:34.530 GIORGIO METTA: I think it's a combination of fun 00:38:34.530 --> 00:38:40.050 and the challenge, so not solving trivial things 00:38:40.050 --> 00:38:43.320 just because it's fun, but a combination of the two, 00:38:43.320 --> 00:38:45.975 seeing something as an unsolved problem. 00:38:45.975 --> 00:38:47.850 STEFANIE TELLEX: So I try really hard to only 00:38:47.850 --> 00:38:49.560 work on things that are fun and to spend 00:38:49.560 --> 00:38:51.909 as little time as possible on things that are not fun. 00:38:51.909 --> 00:38:53.700 And I don't think of it as a shallow thing. 00:38:53.700 --> 00:38:55.991 I think of it as a kind of resource optimization thing, 00:38:55.991 --> 00:38:58.140 because I'm about 1,000 times more productive when 00:38:58.140 --> 00:39:00.780 I'm having fun than when I'm not having fun. 00:39:00.780 --> 00:39:04.320 So even if it was more serious or something, 00:39:04.320 --> 00:39:07.260 I would get so much less done that it's just not worth it. 00:39:07.260 --> 00:39:11.430 It's better to do the fun thing and work the long hours 00:39:11.430 --> 00:39:13.170 because it's fun. 00:39:13.170 --> 00:39:16.590 So for me it's just-- it's still obviously the right thing 00:39:16.590 --> 00:39:20.530 because so much more gets done that way, for me. 00:39:20.530 --> 00:39:23.430 PATRICK WINSTON: Well, to put another twist on this, 00:39:23.430 --> 00:39:27.090 if you were a DARPA program manager, 00:39:27.090 --> 00:39:35.270 what would you do for the next round of progress in robotics? 00:39:35.270 --> 00:39:37.490 Do have a sense of what ought to be next? 00:39:37.490 --> 00:39:40.910 Or maybe what the flaws in previous programs have been? 00:39:44.786 --> 00:39:47.160 STEFANIE TELLEX: So we've been talking to a DARPA program 00:39:47.160 --> 00:39:49.140 manager about what they should do next. 00:39:49.140 --> 00:39:51.420 And we got a seedling for a program 00:39:51.420 --> 00:39:54.900 to think about planning in really large state action 00:39:54.900 --> 00:39:58.242 spaces to enable-- 00:39:58.242 --> 00:39:59.700 the sort of middle part of my talk, 00:39:59.700 --> 00:40:01.616 we were talking about the dime problem, right? 00:40:01.616 --> 00:40:04.830 So we wanted a planner that could find actions 00:40:04.830 --> 00:40:09.090 like picking up a like small scale actions, 00:40:09.090 --> 00:40:11.460 but also large scale things like unload the truck 00:40:11.460 --> 00:40:13.270 or clean up the warehouse. 00:40:13.270 --> 00:40:14.220 And so we were-- 00:40:14.220 --> 00:40:15.780 because we thought that is what's 00:40:15.780 --> 00:40:17.730 needed to interpret natural language commands 00:40:17.730 --> 00:40:20.870 and interact with a person at the level of abstraction. 00:40:20.870 --> 00:40:23.560 So we have a seedling to work on that. 00:40:23.560 --> 00:40:25.770 JOHN LEONARD: So if I could clone myself, 00:40:25.770 --> 00:40:27.600 say I made four or five of my selves. 00:40:27.600 --> 00:40:30.660 One of them I would, if I were DARPA program manager, 00:40:30.660 --> 00:40:32.740 to do Google for the physical world. 00:40:32.740 --> 00:40:35.190 So think about having like an object-based understanding 00:40:35.190 --> 00:40:39.060 of things and people in the environment and places, 00:40:39.060 --> 00:40:43.080 and being able to do the equivalent of internet search, 00:40:43.080 --> 00:40:46.440 physical world search, just combining perception 00:40:46.440 --> 00:40:48.280 and then being able to go get objects. 00:40:48.280 --> 00:40:51.510 So the physical-- like, w get for the physical world. 00:40:51.510 --> 00:40:56.350 That's what I would like to do. 00:40:56.350 --> 00:40:58.110 TONY PRESCOTT: In the UK, I think-- 00:40:58.110 --> 00:40:59.820 so I don't know about DARPA. 00:40:59.820 --> 00:41:02.380 But the government made it one of their eight 00:41:02.380 --> 00:41:04.230 great technologies a few years ago, 00:41:04.230 --> 00:41:06.470 robotics and autonomous systems. 00:41:06.470 --> 00:41:09.845 And looking again, now, at the priorities and they're now 00:41:09.845 --> 00:41:11.970 looking, well, what are the disruptive technologies 00:41:11.970 --> 00:41:15.720 in robotics, again, is coming out as one of the things 00:41:15.720 --> 00:41:17.230 that they think is important. 00:41:17.230 --> 00:41:21.160 So in terms of potential economic and societal impact, 00:41:21.160 --> 00:41:22.590 I think it's huge. 00:41:22.590 --> 00:41:25.500 And so if US funding agencies aren't doing it-- 00:41:25.500 --> 00:41:29.640 PATRICK WINSTON: What do you see those applications as being? 00:41:29.640 --> 00:41:31.110 TONY PRESCOTT: I think-- 00:41:31.110 --> 00:41:35.500 well, the big one that interests me is assistive technology. 00:41:35.500 --> 00:41:37.620 In Europe, Japan, I think the US, 00:41:37.620 --> 00:41:40.140 we're faced with aging society issues. 00:41:40.140 --> 00:41:43.752 And I think assistive robotics in all sorts of ways 00:41:43.752 --> 00:41:44.460 are going to be-- 00:41:44.460 --> 00:41:46.830 PATRICK WINSTON: So you mean for home health care 00:41:46.830 --> 00:41:48.030 type of applications? 00:41:48.030 --> 00:41:50.010 TONY PRESCOTT: Home health care-- 00:41:50.010 --> 00:41:52.880 prosthetics is already a massive growth area. 00:41:52.880 --> 00:41:57.310 But robots-- I mean, my generation, 00:41:57.310 --> 00:42:01.800 I've looked at the statistics and the number of people that 00:42:01.800 --> 00:42:05.640 are going to be in the age group 80 plus 00:42:05.640 --> 00:42:10.410 is going to be 50% higher when I reach that age. 00:42:10.410 --> 00:42:14.320 And it's a huge burden on younger people to care for us. 00:42:14.320 --> 00:42:18.420 So I think independence in my own age, I think in my old age 00:42:18.420 --> 00:42:21.122 I would love to be supported by technology. 00:42:21.122 --> 00:42:22.830 You can do what you like with a computer, 00:42:22.830 --> 00:42:24.905 but you can't physically help somebody. 00:42:24.905 --> 00:42:27.080 And that's where robots are different. 00:42:27.080 --> 00:42:29.880 So that would be one of the things that excites me, 00:42:29.880 --> 00:42:33.690 and one of the reasons I'm interested in applications. 00:42:33.690 --> 00:42:35.940 I'm driven, I think, by the excitement of the research 00:42:35.940 --> 00:42:37.160 and building stuff. 00:42:37.160 --> 00:42:40.260 But I'm also motivated by the potential benefits 00:42:40.260 --> 00:42:42.077 of the applications we can make. 00:42:42.077 --> 00:42:44.160 PATRICK WINSTON: I suppose if they're good enough, 00:42:44.160 --> 00:42:45.900 we won't need dishwashers because they 00:42:45.900 --> 00:42:48.221 can do the dishes themselves. 00:42:48.221 --> 00:42:48.720 OK. 00:42:48.720 --> 00:42:51.640 So now we have a question from the audience, 00:42:51.640 --> 00:42:54.570 which if I may paraphrase, there have been a-- 00:42:57.930 --> 00:43:00.060 have there been examples-- 00:43:00.060 --> 00:43:02.190 I know you think of them all the time, Tony. 00:43:02.190 --> 00:43:03.810 But have there been examples where 00:43:03.810 --> 00:43:07.260 work on robotics in your respective activities 00:43:07.260 --> 00:43:11.340 have shed new light on a biological problem 00:43:11.340 --> 00:43:13.290 or inspired a biological inquiry that 00:43:13.290 --> 00:43:16.680 wouldn't have happened without the kind of stuff you do? 00:43:16.680 --> 00:43:19.229 RUSS TEDRAKE: I started off more as a biologist, I guess. 00:43:19.229 --> 00:43:20.770 I was in a computational neuroscience 00:43:20.770 --> 00:43:22.260 lab with Sebastian Seung. 00:43:22.260 --> 00:43:24.700 I tried to study a lot about how the brain works, 00:43:24.700 --> 00:43:27.270 how the motor system works, in the hopes 00:43:27.270 --> 00:43:29.310 that it would help me make better robots. 00:43:29.310 --> 00:43:30.935 PATRICK WINSTON: Oh, maybe pause there. 00:43:30.935 --> 00:43:31.620 Did it? 00:43:31.620 --> 00:43:33.150 RUSS TEDRAKE: Yeah, it didn't. 00:43:33.150 --> 00:43:35.460 So I don't use that stuff right now. 00:43:35.460 --> 00:43:37.360 I mean, maybe one day again. 00:43:37.360 --> 00:43:40.930 But our hardware is very different, 00:43:40.930 --> 00:43:44.694 our computational hardware is very different right now. 00:43:44.694 --> 00:43:47.110 I think there's sort of a race to understand intelligence, 00:43:47.110 --> 00:43:49.170 and maybe we'll converge again right now. 00:43:49.170 --> 00:43:52.650 But the things I write down for the robots today 00:43:52.650 --> 00:43:56.270 don't look anything like what I think the brain-- 00:43:56.270 --> 00:44:01.640 what I was learning about what the brain did back then. 00:44:01.640 --> 00:44:05.250 But that doesn't mean there's not tons of cross-pollination. 00:44:05.250 --> 00:44:10.250 So we have a great project with a biologist at Harvard, Andy 00:44:10.250 --> 00:44:12.290 Biewener. 00:44:12.290 --> 00:44:15.800 Andy has been studying maneuvering flight in birds. 00:44:15.800 --> 00:44:19.110 He's instrumenting birds flying through dense obstacles. 00:44:19.110 --> 00:44:21.690 We're trying to make UAVs fly through dense obstacles. 00:44:21.690 --> 00:44:24.649 We're exchanging capabilities and ideas 00:44:24.649 --> 00:44:25.690 and going back and forth. 00:44:25.690 --> 00:44:27.630 Just the algorithms that we have written 00:44:27.630 --> 00:44:30.610 helps him understand what birds are doing and vice versa. 00:44:30.610 --> 00:44:32.560 So there's tons of exchanges. 00:44:32.560 --> 00:44:35.340 But the code that I write to power of the robots today, 00:44:35.340 --> 00:44:38.800 I think, is not quite what the brain is doing, 00:44:38.800 --> 00:44:41.100 and nor should it be. 00:44:41.100 --> 00:44:44.110 PATRICK WINSTON: Any other thoughts on that? 00:44:44.110 --> 00:44:47.350 GIORGIO METTA: Well, we have experiments 00:44:47.350 --> 00:44:50.560 that I meant to actually present today, 00:44:50.560 --> 00:44:56.620 where we've been working with neuroscientists on trying 00:44:56.620 --> 00:45:01.770 to bring some of the principles from neuroscience to the robot 00:45:01.770 --> 00:45:02.640 construction. 00:45:02.640 --> 00:45:06.700 Let's say, not the physical robot, but the software. 00:45:06.700 --> 00:45:15.260 And I find always difficult to find the level of abstraction 00:45:15.260 --> 00:45:19.550 that actually motivate something from neuroscience 00:45:19.550 --> 00:45:22.980 and manages to show something important for computation. 00:45:22.980 --> 00:45:29.670 I think I only have one example, or two maybe overall. 00:45:29.670 --> 00:45:35.520 And it always happen not copying in details brain structure, 00:45:35.520 --> 00:45:38.120 but just taking an idea what information 00:45:38.120 --> 00:45:40.550 may be relevant for a certain task 00:45:40.550 --> 00:45:42.290 and trying to figure out solutions 00:45:42.290 --> 00:45:44.690 that use that information. 00:45:44.690 --> 00:45:48.880 In particular, a couple of things we've done 00:45:48.880 --> 00:45:55.010 had to do with the involvement of motor controlling 00:45:55.010 --> 00:45:57.140 information in perception. 00:45:57.140 --> 00:46:00.570 And that's something the sort of paid off, 00:46:00.570 --> 00:46:04.280 at least in the smallest experiments. 00:46:04.280 --> 00:46:08.540 Still, we can't compare with full-blown systems. 00:46:08.540 --> 00:46:11.320 Like we've done experiments in speech perception 00:46:11.320 --> 00:46:15.470 and that show to be over-performing systems 00:46:15.470 --> 00:46:17.510 that don't use motor information, 00:46:17.510 --> 00:46:19.430 but on limited settings. 00:46:19.430 --> 00:46:22.490 We don't know if we build the full speech recognition 00:46:22.490 --> 00:46:25.880 system what happens, whether we better or worse existing 00:46:25.880 --> 00:46:30.200 commercial methods or commercial systems. 00:46:30.200 --> 00:46:35.300 So it's still a long way to actually show 00:46:35.300 --> 00:46:41.580 that we managed to get something from the biological 00:46:41.580 --> 00:46:42.940 counter-part. 00:46:42.940 --> 00:46:45.220 Although maybe for the neuroscientists 00:46:45.220 --> 00:46:46.510 this explains something. 00:46:46.510 --> 00:46:51.760 Because where they didn't have a specific theory, 00:46:51.760 --> 00:46:53.290 at least we showed the advantages 00:46:53.290 --> 00:46:54.940 of that particular solution, that it's 00:46:54.940 --> 00:46:58.780 being used by the brain. 00:46:58.780 --> 00:47:00.570 TONY PRESCOTT: So I think there's-- 00:47:00.570 --> 00:47:04.830 we tend to forget in our history where our ideas came from. 00:47:04.830 --> 00:47:08.460 So for instance, reinforcement learning-- 00:47:08.460 --> 00:47:10.290 Demis Hassabis explained last night 00:47:10.290 --> 00:47:13.740 how he's using this to play Atari computer 00:47:13.740 --> 00:47:16.530 games and these amazing system he's developing. 00:47:16.530 --> 00:47:19.560 If you go back in the history of reinforcement learning, 00:47:19.560 --> 00:47:23.400 the key idea there came from two psychologists, 00:47:23.400 --> 00:47:25.680 Rescorla Wagner developing a theory 00:47:25.680 --> 00:47:27.470 of classical conditioning. 00:47:27.470 --> 00:47:34.170 And then that got picked up in machine learning in the 1980s, 00:47:34.170 --> 00:47:40.120 and it got really developed and hugely accelerated. 00:47:40.120 --> 00:47:42.900 But then there was crossover back into neuroscience 00:47:42.900 --> 00:47:44.900 with dopamine theory and so on. 00:47:44.900 --> 00:47:47.880 And ideas about hierarchical reinforcement learning 00:47:47.880 --> 00:47:51.600 have been developed that are partly brain inspired. 00:47:51.600 --> 00:47:53.630 So I think it's-- 00:47:53.630 --> 00:47:56.370 there is crossover, and sometimes we 00:47:56.370 --> 00:47:58.764 may lose track of how much crossover there is. 00:47:58.764 --> 00:48:01.180 PATRICK WINSTON: I have another comment from the audience, 00:48:01.180 --> 00:48:03.870 I see we are under some pressure to not drone on 00:48:03.870 --> 00:48:05.530 for the rest of the evening. 00:48:05.530 --> 00:48:10.170 So the comment is I think relevant to the last topic I 00:48:10.170 --> 00:48:13.980 wanted to bring up, which is the question of ethics 00:48:13.980 --> 00:48:14.670 in all of this. 00:48:14.670 --> 00:48:19.260 And the comment is why should we make 00:48:19.260 --> 00:48:20.939 robots that are good at operating, doing 00:48:20.939 --> 00:48:22.980 things in the household, take care of the elderly 00:48:22.980 --> 00:48:26.140 and so on, when the rest of AI is going hell-bent 00:48:26.140 --> 00:48:28.570 to put a lot of people out of work and people 00:48:28.570 --> 00:48:30.910 who could perhaps use those jobs. 00:48:30.910 --> 00:48:35.700 But in any event, there's been a lot of concern, 00:48:35.700 --> 00:48:39.420 perhaps spawned by some of the films like Ex Machina 00:48:39.420 --> 00:48:42.140 and so on, that robots will take over. 00:48:42.140 --> 00:48:44.820 And I don't think are going to take over in that sense 00:48:44.820 --> 00:48:45.380 very soon. 00:48:45.380 --> 00:48:46.570 But do you see-- 00:48:46.570 --> 00:48:49.470 do you worry-- do you think about any dangers of the kinds 00:48:49.470 --> 00:48:51.210 of technology you're working on in terms 00:48:51.210 --> 00:48:54.522 of economic dislocation or battlefield robots 00:48:54.522 --> 00:48:55.980 or anything of that sort that might 00:48:55.980 --> 00:48:59.807 come about as a consequence of what you do? 00:48:59.807 --> 00:49:01.390 RUSS TEDRAKE: I think it's inevitable. 00:49:01.390 --> 00:49:03.870 I think we shouldn't fear it, but we 00:49:03.870 --> 00:49:05.860 have to be conscious of it. 00:49:05.860 --> 00:49:08.640 So I mean, would you look back to the 1980s 00:49:08.640 --> 00:49:12.300 and avoid the invent of the personal computer 00:49:12.300 --> 00:49:15.030 because it was going to change the way people had to do work? 00:49:15.030 --> 00:49:17.160 I mean, of course you wouldn't. 00:49:17.160 --> 00:49:19.934 But at the same time that changed the way 00:49:19.934 --> 00:49:20.850 people had to do work. 00:49:20.850 --> 00:49:25.015 And it was painful for a big portion of the population, 00:49:25.015 --> 00:49:26.640 but ultimately it was good for society. 00:49:26.640 --> 00:49:29.990 I think robots will have the same sort of effect. 00:49:29.990 --> 00:49:32.520 It's going to raise the bar on what 00:49:32.520 --> 00:49:33.800 people are capable of doing. 00:49:33.800 --> 00:49:36.540 It's going to raise the bar on what people have 00:49:36.540 --> 00:49:38.790 to be successful in their jobs. 00:49:38.790 --> 00:49:41.130 And it might be painful, but I think 00:49:41.130 --> 00:49:45.350 it's super important for society to keep moving on it. 00:49:45.350 --> 00:49:47.641 PATRICK WINSTON: Why, again, is it super important? 00:49:47.641 --> 00:49:49.640 RUSS TEDRAKE: Because it's going to advance what 00:49:49.640 --> 00:49:50.890 we're capable of as a society. 00:49:50.890 --> 00:49:54.352 It's going to make us ultimately more productive. 00:49:54.352 --> 00:49:56.392 PATRICK WINSTON: Other thoughts? 00:49:56.392 --> 00:49:57.350 TONY PRESCOTT: I agree. 00:49:57.350 --> 00:50:00.560 I mean, I think the people that are worrying about jobs being 00:50:00.560 --> 00:50:02.460 taken by robots aren't the people that 00:50:02.460 --> 00:50:03.980 want to do those jobs. 00:50:03.980 --> 00:50:06.590 Because most of the jobs are ones that it's 00:50:06.590 --> 00:50:08.780 very hard to get anyone to do. 00:50:08.780 --> 00:50:13.080 They're low paid, they're unpleasant. 00:50:13.080 --> 00:50:17.870 And we're automating the dull and dreary aspects 00:50:17.870 --> 00:50:19.610 of human existence. 00:50:19.610 --> 00:50:21.800 And that gives the opportunity for people 00:50:21.800 --> 00:50:23.750 to have more fulfilling lives. 00:50:23.750 --> 00:50:26.570 Now, the problem isn't that we're 00:50:26.570 --> 00:50:29.510 doing this great work to get robots or machines 00:50:29.510 --> 00:50:30.940 to do these things for us. 00:50:30.940 --> 00:50:33.410 It's that, as a society, we're not 00:50:33.410 --> 00:50:36.230 thinking about how we adjust to that, 00:50:36.230 --> 00:50:39.160 how we make sure people will have fulfilling lives 00:50:39.160 --> 00:50:43.280 and will be supported materially to enjoy that prosperity. 00:50:43.280 --> 00:50:45.620 So I think it's disruptive in many ways, 00:50:45.620 --> 00:50:47.710 and it's going to be disruptive politically. 00:50:47.710 --> 00:50:49.460 And we're going to have to adapt. 00:50:49.460 --> 00:50:52.940 Because if you're not working, then you 00:50:52.940 --> 00:50:55.030 have to be supported to enjoy your life. 00:50:55.030 --> 00:50:58.650 And maybe that means a change in the political system. 00:50:58.650 --> 00:51:01.790 So those are questions perhaps not for us. 00:51:01.790 --> 00:51:03.560 But I think we maybe-- 00:51:03.560 --> 00:51:05.960 as the technologists, we have to be prepared 00:51:05.960 --> 00:51:09.800 to admit that what we're working on are really disrupted systems 00:51:09.800 --> 00:51:11.780 and they are going to have these large impacts. 00:51:11.780 --> 00:51:13.790 And people are waking up to that. 00:51:13.790 --> 00:51:16.620 And if we wave our hands and say, don't worry, 00:51:16.620 --> 00:51:20.810 I think we're not going to be taken seriously. 00:51:20.810 --> 00:51:23.030 PATRICK WINSTON: Other thoughts? 00:51:23.030 --> 00:51:26.420 JOHN LEONARD: I see how these are really important questions. 00:51:26.420 --> 00:51:28.070 And I see-- 00:51:28.070 --> 00:51:29.000 I have mixed emotions. 00:51:29.000 --> 00:51:30.680 I'm really torn. 00:51:30.680 --> 00:51:32.360 I came from a family that was affected 00:51:32.360 --> 00:51:33.920 by unemployment in the 1970s. 00:51:33.920 --> 00:51:39.140 So I feel like I'm very sympathetic to the potential 00:51:39.140 --> 00:51:40.440 for losing jobs. 00:51:40.440 --> 00:51:43.580 At CSAIL we've had this wonderful discussion 00:51:43.580 --> 00:51:46.100 with some economists at MIT the last few years, 00:51:46.100 --> 00:51:53.780 Frank Levy, David Autor, and Eric Brynjolfsson, Andy McAfee, 00:51:53.780 --> 00:51:55.630 and I've learned a lot from them. 00:51:55.630 --> 00:51:58.100 And I think that they vary in their views. 00:51:58.100 --> 00:52:00.994 I think I am more along the lines of someone 00:52:00.994 --> 00:52:02.660 like David Autor, who's an economist who 00:52:02.660 --> 00:52:05.780 thinks that we shouldn't fear too 00:52:05.780 --> 00:52:07.320 rapid a replacement of robots. 00:52:07.320 --> 00:52:10.310 If you look at the data, that the things that are-- 00:52:10.310 --> 00:52:13.920 I would say the things that are hard for robots are still hard. 00:52:13.920 --> 00:52:17.030 But on the other hand, I think, longer term, we 00:52:17.030 --> 00:52:19.370 do have to be mindful of as a society, that like, 00:52:19.370 --> 00:52:22.046 as Russ said, things like this are going to happen. 00:52:22.046 --> 00:52:24.420 I think that the short term introduction, if you look at, 00:52:24.420 --> 00:52:27.060 for example, Kiva and how they they've changed the way 00:52:27.060 --> 00:52:27.920 a warehouse works. 00:52:27.920 --> 00:52:31.460 I think replacing humans just completely with robots, 00:52:31.460 --> 00:52:33.950 like, say, for gardening or agriculture, some really 00:52:33.950 --> 00:52:37.310 hard things to do, because the problems are so hard. 00:52:37.310 --> 00:52:41.000 But if you rethink the task to have humans and robots working 00:52:41.000 --> 00:52:43.940 together, Kiva's a good example of how you actually 00:52:43.940 --> 00:52:44.976 can change things. 00:52:44.976 --> 00:52:46.850 And so that's where I think the short term is 00:52:46.850 --> 00:52:49.490 going to come from, is humans and robots working together. 00:52:49.490 --> 00:52:52.744 That's why I think HRI is such an important topic. 00:52:52.744 --> 00:52:54.160 PATRICK WINSTON: Well I don't know 00:52:54.160 --> 00:52:57.560 if you running for president, but be that is it may, 00:52:57.560 --> 00:52:59.740 do any of you have a one-minute closing 00:52:59.740 --> 00:53:02.412 statement you like to make? 00:53:02.412 --> 00:53:04.870 JOHN LEONARD: Well I'll go sort of the deep learning thing. 00:53:04.870 --> 00:53:10.900 I think in robotics we have this, potentially, a coming 00:53:10.900 --> 00:53:12.760 divide between the folks that believe more 00:53:12.760 --> 00:53:16.070 in that data-driven learning methods and more models. 00:53:16.070 --> 00:53:19.580 And I'm a believer more on the model-based side, 00:53:19.580 --> 00:53:21.895 that we don't have enough data and enough systems. 00:53:25.690 --> 00:53:29.530 But I do fear that we could be in a society-- 00:53:29.530 --> 00:53:33.010 for certain classes of problems, he or she who has the data 00:53:33.010 --> 00:53:35.830 may win, in terms of if Google or Facebook have 00:53:35.830 --> 00:53:38.740 just such massive amounts of data for certain problems 00:53:38.740 --> 00:53:40.561 that academics can't compete. 00:53:40.561 --> 00:53:43.060 So I do feel there's a place for the professor and the seven 00:53:43.060 --> 00:53:45.340 grad students and a couple of post-docs. 00:53:45.340 --> 00:53:46.964 But if you do you have to be careful 00:53:46.964 --> 00:53:48.880 in terms of problem selection, that you're not 00:53:48.880 --> 00:53:52.030 going right up against the sort of one of these data machine 00:53:52.030 --> 00:53:52.655 companies. 00:53:52.655 --> 00:53:54.280 RUSS TEDRAKE: I was going to say that I 00:53:54.280 --> 00:53:56.030 was looking at humans right now and trying 00:53:56.030 --> 00:53:59.050 to inform the robots, I wouldn't look 00:53:59.050 --> 00:54:01.577 at center-out reaching movements or nominal walking 00:54:01.577 --> 00:54:02.410 or things like this. 00:54:02.410 --> 00:54:04.120 I'd be pushing for the corner cases. 00:54:04.120 --> 00:54:06.460 I've been trying to really understand 00:54:06.460 --> 00:54:09.920 the performance of biological intelligence 00:54:09.920 --> 00:54:13.720 in the screw cases, in the cases where they didn't 00:54:13.720 --> 00:54:15.250 have a lot of prior data. 00:54:15.250 --> 00:54:18.070 They were once in a lifetime experiences. 00:54:18.070 --> 00:54:20.980 How did natural intelligence respond? 00:54:20.980 --> 00:54:23.740 That's, I think, a grand challenge 00:54:23.740 --> 00:54:26.840 for us on the computational intelligence side. 00:54:26.840 --> 00:54:29.964 And maybe there's a lot to learn. 00:54:29.964 --> 00:54:31.630 PATRICK WINSTON: And the grand challenge 00:54:31.630 --> 00:54:33.820 for me, to conclude all this, has 00:54:33.820 --> 00:54:38.230 to do with what it would really take to make a robot humanoid. 00:54:38.230 --> 00:54:41.050 And I've been thinking a lot about that 00:54:41.050 --> 00:54:44.050 recently in connection with self-awareness, 00:54:44.050 --> 00:54:47.290 understanding the story, having the robot understand 00:54:47.290 --> 00:54:50.080 the story of what's going on throughout the day, 00:54:50.080 --> 00:54:53.530 having it able to use previous experiences to guide 00:54:53.530 --> 00:54:55.880 its future experiences, and so on. 00:54:55.880 --> 00:54:58.090 So there's a lot to be done, that's for sure. 00:54:58.090 --> 00:55:01.064 And I'm sure we'll be working together as time goes on. 00:55:01.064 --> 00:55:02.730 Now I'd just like to thank the panelists 00:55:02.730 --> 00:55:05.480 and conclude the evening.