WEBVTT 1 00:00:00.000 --> 00:00:02.100 to the first session of what promises, I think, 2 00:00:02.140 --> 00:00:05.440 to be an absolutely fantastic festival of ideas here in Adelaide. 3 00:00:05.940 --> 00:00:07.980 We're very fortunate this morning 4 00:00:07.980 --> 00:00:10.800 to have the Director of the Artificial Intelligence Laboratory 5 00:00:10.800 --> 00:00:15.620 at MIT in Boston here with us, Professor Rodney Brooks. 6 00:00:16.140 --> 00:00:20.320 I visited the Artificial Intelligence Laboratory in November last year 7 00:00:20.320 --> 00:00:24.840 and walked away absolutely flabbergasted by what I saw. 8 00:00:25.960 --> 00:00:29.380 I saw demonstrations of artificial limbs 9 00:00:29.380 --> 00:00:33.440 which were using mouse muscle fibres to work them. 10 00:00:33.620 --> 00:00:38.820 I saw a machine that mirrored my facial expressions, 11 00:00:39.680 --> 00:00:41.980 the same machine that's been shown on television recently 12 00:00:41.980 --> 00:00:44.300 in the ABC series based on the face. 13 00:00:45.120 --> 00:00:47.720 And this is the work of Rodney Brooks and his team, 14 00:00:47.840 --> 00:00:51.340 some 230 people based at MIT, 15 00:00:52.680 --> 00:00:56.360 developing or rolling back the frontier, really, 16 00:00:56.380 --> 00:00:59.280 of artificial intelligence 17 00:00:59.380 --> 00:01:02.940 and robotics in what is one of the most exciting laboratories 18 00:01:02.940 --> 00:01:06.920 that I've ever had the privilege to visit. 19 00:01:08.140 --> 00:01:09.880 Rodney's originally from Adelaide. 20 00:01:10.000 --> 00:01:13.000 He did his first degree in mathematics at Flinders University 21 00:01:13.000 --> 00:01:15.640 and has gone on to found companies 22 00:01:15.640 --> 00:01:17.660 as well as the Artificial Intelligence Laboratory 23 00:01:17.660 --> 00:01:23.020 and really is one of the world leaders in this field. 24 00:01:23.120 --> 00:01:24.820 So we're very, very pleased to have him here today. 25 00:01:25.840 --> 00:01:27.980 Before we get underway, I'll introduce Rodney, though. 26 00:01:28.040 --> 00:01:29.360 I should just mention that, you know, 27 00:01:29.380 --> 00:01:32.080 if any of you are staying here for the next session, 28 00:01:32.260 --> 00:01:35.900 would you please mind clearing the venue entirely 29 00:01:35.900 --> 00:01:37.480 and taking your bags with you 30 00:01:37.480 --> 00:01:39.720 because there's no guarantee that you'll have the same seat 31 00:01:39.720 --> 00:01:41.040 for the next session. 32 00:01:41.240 --> 00:01:43.480 So the organisers have asked me just to make sure 33 00:01:43.480 --> 00:01:46.660 that we clear promptly just after 11 this morning 34 00:01:46.660 --> 00:01:49.560 in order to make room for the next session to get underway. 35 00:01:50.260 --> 00:01:51.780 So with no more ado, 36 00:01:51.880 --> 00:01:54.380 let's hear from Rodney about artificial intelligence. 37 00:01:55.580 --> 00:01:56.060 APPLAUSE 38 00:02:00.720 --> 00:02:01.600 Well, good morning. 39 00:02:01.960 --> 00:02:05.080 Thank you for coming to this session. 40 00:02:05.240 --> 00:02:07.800 I know it was pretty crowded over in Bishop Spong's session 41 00:02:07.800 --> 00:02:09.180 and some of you probably couldn't fit in 42 00:02:09.180 --> 00:02:10.300 and that's why you ended up here. 43 00:02:10.620 --> 00:02:12.480 I do appreciate the audience, though. 44 00:02:13.380 --> 00:02:16.280 The Artificial Intelligence Lab at MIT, as Tim said, 45 00:02:16.400 --> 00:02:18.860 is about 230 people, about 20 faculty members, 46 00:02:19.060 --> 00:02:21.000 and there's a wide range of things that are covered there. 47 00:02:21.120 --> 00:02:23.500 This morning, I'm just going to talk about my own work 48 00:02:23.500 --> 00:02:27.020 within that larger grouping, my own research group. 49 00:02:27.300 --> 00:02:28.580 And I'm going to talk about machines, 50 00:02:29.380 --> 00:02:30.200 like humans, 51 00:02:30.440 --> 00:02:34.280 and one can read that two different ways 52 00:02:34.280 --> 00:02:37.320 and I will be addressing both of those issues here. 53 00:02:37.700 --> 00:02:39.580 When I was growing up here in Adelaide, 54 00:02:40.280 --> 00:02:42.220 I was inspired by a couple of things. 55 00:02:42.420 --> 00:02:44.180 One was the work of... 56 00:02:47.080 --> 00:02:48.520 ..Greywalter, W. Greywalter, 57 00:02:48.620 --> 00:02:53.080 who was a doctor in Bristol 58 00:02:53.080 --> 00:02:58.200 who built small robots back in around 1949 and 1950. 59 00:02:58.400 --> 00:02:59.300 And these small robots, 60 00:02:59.380 --> 00:03:03.400 they had a couple of valves as their computational elements 61 00:03:03.400 --> 00:03:06.540 and acted rather like simple creatures. 62 00:03:06.640 --> 00:03:07.800 He called them tortoises, 63 00:03:08.340 --> 00:03:11.200 following Lewis Carroll, what they taught us. 64 00:03:12.480 --> 00:03:14.760 And I spent a lot of time when I was young 65 00:03:14.760 --> 00:03:17.760 trying to emulate these machines and build similar machines. 66 00:03:18.020 --> 00:03:19.400 And then in the late 60s, 67 00:03:19.400 --> 00:03:21.900 I was very inspired by the movie 2001 68 00:03:21.900 --> 00:03:24.620 with the robot HAL 9000, 69 00:03:24.740 --> 00:03:26.200 which lived on the Starship 70 00:03:26.200 --> 00:03:28.900 and controlled it to a large extent. 71 00:03:29.380 --> 00:03:32.640 And I was rather taken with it being completely intelligent. 72 00:03:32.960 --> 00:03:34.820 It turned out to be a murdering psychopath, 73 00:03:35.260 --> 00:03:37.200 but that was not a big deal to me. 74 00:03:37.240 --> 00:03:40.400 I was just interested in the fact of having a machine be intelligent. 75 00:03:41.220 --> 00:03:45.640 So by the time I got to MIT in the mid-80s, 76 00:03:45.640 --> 00:03:48.900 I started thinking about how to build intelligent robots. 77 00:03:49.400 --> 00:03:51.160 And I looked at the history of the world 78 00:03:51.160 --> 00:03:52.980 for the last 4.6 billion years 79 00:03:52.980 --> 00:03:55.600 and looked at what evolution had spent its time on. 80 00:03:56.400 --> 00:03:58.660 And most of its time had been spent 81 00:03:58.660 --> 00:04:02.120 on getting to the point of first single-cell animals 82 00:04:02.120 --> 00:04:05.940 and then mobility and moving around in the world. 83 00:04:06.100 --> 00:04:08.780 And if you look over on the right-hand side of this diagram, 84 00:04:09.060 --> 00:04:11.360 all the stuff that we consider intelligence 85 00:04:11.360 --> 00:04:14.280 has come along very late in the game. 86 00:04:14.480 --> 00:04:16.720 And artificial intelligence research, 87 00:04:16.920 --> 00:04:18.820 which had been around since about 1956, 88 00:04:19.760 --> 00:04:21.260 the AI lab that I now direct 89 00:04:21.260 --> 00:04:23.820 was founded in 1959 by Marvin Minsky, 90 00:04:24.500 --> 00:04:28.100 that enterprise had concentrated on the stuff 91 00:04:28.100 --> 00:04:28.640 that was being used to build robots. 92 00:04:28.640 --> 00:04:30.640 And it was that academics found difficult, 93 00:04:30.860 --> 00:04:33.400 like proving theorems, playing chess, etc. 94 00:04:33.660 --> 00:04:35.440 And it seemed to me, looking at evolution, 95 00:04:35.760 --> 00:04:37.740 that just moving around in the world 96 00:04:37.740 --> 00:04:39.880 and seeing the world and interacting with the world 97 00:04:39.880 --> 00:04:41.000 was probably the hard stuff 98 00:04:41.000 --> 00:04:43.620 because evolution had taken so long to get there. 99 00:04:44.200 --> 00:04:47.060 So with that, Marvin Minsky, 100 00:04:47.840 --> 00:04:49.540 who some of you may have heard of, 101 00:04:49.600 --> 00:04:51.020 had also come to a similar conclusion 102 00:04:51.020 --> 00:04:52.640 in his book, Society of Mind. 103 00:04:53.060 --> 00:04:55.140 So with that, I started thinking about insects 104 00:04:55.140 --> 00:04:56.360 as a good place to start, 105 00:04:56.480 --> 00:04:57.460 trying to build robots, 106 00:04:57.460 --> 00:04:59.900 that could emulate insect behavior. 107 00:05:01.140 --> 00:05:03.580 And you can see one robot 108 00:05:03.580 --> 00:05:04.740 and a whole bunch of insects 109 00:05:04.740 --> 00:05:06.680 from the Natural History Museum at Harvard here. 110 00:05:06.960 --> 00:05:08.600 And I'm going to show you a little bit 111 00:05:08.600 --> 00:05:11.160 about one of these insect robots, Genghis, 112 00:05:11.240 --> 00:05:12.380 and then I'm going to move on 113 00:05:12.380 --> 00:05:15.580 to higher level sorts of interactions. 114 00:05:16.120 --> 00:05:18.560 So Genghis, which was built in the late 80s, 115 00:05:20.200 --> 00:05:23.460 could walk around our lab over rough terrain. 116 00:05:24.020 --> 00:05:25.780 It was a six-legged robot, 117 00:05:26.000 --> 00:05:27.440 very much inspired by, 118 00:05:27.460 --> 00:05:27.840 by insects. 119 00:05:28.200 --> 00:05:31.500 And we built a control system for it 120 00:05:31.500 --> 00:05:33.800 out of very simple little computational elements, 121 00:05:34.000 --> 00:05:36.340 each computational element much, much simpler 122 00:05:36.340 --> 00:05:38.920 than the simplest web page you download 123 00:05:38.920 --> 00:05:40.660 with a JavaScript on it these days. 124 00:05:40.840 --> 00:05:44.040 And we added more and more layers of control to it, 125 00:05:44.080 --> 00:05:47.560 and over time, it was able to, 126 00:05:48.140 --> 00:05:49.520 as we added these layers of control, 127 00:05:49.660 --> 00:05:50.880 it was able to get better and better. 128 00:05:50.980 --> 00:05:52.160 We were sort of modeling evolution 129 00:05:52.160 --> 00:05:54.620 of accreting more neural matter 130 00:05:54.620 --> 00:05:56.320 onto the system over time. 131 00:05:57.220 --> 00:05:57.240 The, 132 00:05:57.460 --> 00:05:59.460 the computations were very simple. 133 00:05:59.500 --> 00:06:01.660 These wires were sending messages, 134 00:06:01.900 --> 00:06:02.840 very simple messages. 135 00:06:02.960 --> 00:06:04.500 For instance, the one that I've indicated there 136 00:06:04.500 --> 00:06:06.520 just sent a number between zero and 25. 137 00:06:07.000 --> 00:06:09.680 And these little boxes were very simple computations. 138 00:06:09.980 --> 00:06:11.840 This one happens to be the most complex 139 00:06:11.840 --> 00:06:13.220 of all of the computations. 140 00:06:13.580 --> 00:06:15.880 This one just adds up at six input numbers, 141 00:06:16.040 --> 00:06:17.780 divides by six, and negates, 142 00:06:17.780 --> 00:06:19.880 and has a little zero band around the, 143 00:06:19.920 --> 00:06:20.880 around zero to, 144 00:06:20.900 --> 00:06:22.740 so that the robot doesn't twitch too much. 145 00:06:22.940 --> 00:06:24.900 But all the boxes that I'm showing you here, 146 00:06:25.060 --> 00:06:26.840 most of them are copied six times, 147 00:06:26.860 --> 00:06:27.320 one for each. 148 00:06:27.340 --> 00:06:27.440 So, 149 00:06:27.460 --> 00:06:30.380 they're just simple little computational elements like that. 150 00:06:30.580 --> 00:06:33.700 And with that sort of control, 151 00:06:33.920 --> 00:06:35.000 we built this robot, 152 00:06:35.100 --> 00:06:36.880 which was able to wander around. 153 00:06:37.760 --> 00:06:39.920 This is the very first day it ever walked. 154 00:06:42.520 --> 00:06:43.300 We found, 155 00:06:43.400 --> 00:06:44.360 we discovered that night 156 00:06:44.360 --> 00:06:46.500 that the nitric acid batteries, 157 00:06:46.600 --> 00:06:48.220 when they were recharged on their side, 158 00:06:48.320 --> 00:06:50.380 sprayed nitric acid all over the computers. 159 00:06:50.480 --> 00:06:51.340 That wasn't very good. 160 00:06:52.220 --> 00:06:53.200 So, in the rest, 161 00:06:53.280 --> 00:06:54.200 the later video, 162 00:06:54.280 --> 00:06:55.920 you'll see a cable coming into the back, 163 00:06:55.940 --> 00:06:56.940 which provides, 164 00:06:56.940 --> 00:06:58.380 provides power to this robot. 165 00:06:58.580 --> 00:06:59.500 So, here it is walking 166 00:06:59.500 --> 00:07:01.380 without really sensing the world very much. 167 00:07:02.320 --> 00:07:05.080 This was our simulation of Mons Olympus on Mars. 168 00:07:06.880 --> 00:07:08.600 With some careful editing here. 169 00:07:12.540 --> 00:07:14.200 With that simple walking, 170 00:07:14.300 --> 00:07:15.960 it didn't do very well over rough terrain, 171 00:07:16.100 --> 00:07:18.280 so we emulated evolution 172 00:07:18.280 --> 00:07:19.940 and added more layers of control, 173 00:07:20.120 --> 00:07:21.640 more control boxes, 174 00:07:21.700 --> 00:07:24.260 which modulated the behavior of the lower level systems. 175 00:07:24.700 --> 00:07:25.700 And eventually, 176 00:07:25.700 --> 00:07:30.380 the robot was able to navigate over pretty rough terrain, 177 00:07:31.340 --> 00:07:33.860 sensing what was in front of it with its legs, 178 00:07:33.920 --> 00:07:34.940 reacting to forces, 179 00:07:35.180 --> 00:07:36.600 lifting its legs up higher. 180 00:07:36.700 --> 00:07:39.220 I apologize for the contrast here. 181 00:07:42.060 --> 00:07:45.760 And without having any sort of central goal 182 00:07:45.760 --> 00:07:46.800 or central model 183 00:07:46.800 --> 00:07:48.640 or three-dimensional understanding 184 00:07:48.640 --> 00:07:49.820 of what was in front of it, 185 00:07:49.900 --> 00:07:54.820 but like a blind person crawling over some obstacles, 186 00:07:54.820 --> 00:07:55.660 feeling the way 187 00:07:55.660 --> 00:07:58.020 and getting over them, 188 00:07:58.080 --> 00:07:59.000 reacting locally. 189 00:07:59.660 --> 00:08:02.860 Here it is with some of those higher levels of control, 190 00:08:03.620 --> 00:08:04.860 sensing infrared radiation 191 00:08:04.860 --> 00:08:06.560 that our human bodies emit 192 00:08:06.560 --> 00:08:08.040 in about the 10 micron range. 193 00:08:08.420 --> 00:08:12.300 And if it senses more radiation on the left, 194 00:08:12.360 --> 00:08:14.460 it takes shorter footsteps on its left side, 195 00:08:14.580 --> 00:08:15.940 so it tends to turn to the left. 196 00:08:15.960 --> 00:08:17.300 If it senses more on its right, 197 00:08:17.400 --> 00:08:18.140 as it's sensing now, 198 00:08:18.300 --> 00:08:20.320 it takes shorter steps on the right, 199 00:08:20.380 --> 00:08:21.680 so it tends to turn to the right. 200 00:08:21.780 --> 00:08:24.480 So, an observer sees it following, 201 00:08:25.260 --> 00:08:26.560 this hot body. 202 00:08:26.760 --> 00:08:29.180 It doesn't have any internal representation 203 00:08:29.180 --> 00:08:30.440 of the hot body. 204 00:08:30.500 --> 00:08:31.540 It doesn't have any knowledge 205 00:08:31.540 --> 00:08:32.640 that it's following something. 206 00:08:32.920 --> 00:08:34.860 But when you put it in the right circumstance, 207 00:08:35.260 --> 00:08:37.300 here it is, down the bottom left, 208 00:08:37.420 --> 00:08:39.100 waiting for some prey to come along, 209 00:08:39.680 --> 00:08:42.040 and then it attacks. 210 00:08:44.860 --> 00:08:45.840 So, from that, 211 00:08:46.060 --> 00:08:51.140 we compared the way our robot worked 212 00:08:51.140 --> 00:08:53.180 to the way other robots, 213 00:08:53.320 --> 00:08:54.420 walking robots were working, 214 00:08:54.500 --> 00:08:54.680 and, 215 00:08:54.820 --> 00:08:56.200 for instance, 216 00:08:56.380 --> 00:08:57.240 at Carnegie Mellon, 217 00:08:57.300 --> 00:08:59.060 they had a six-legged walking robot 218 00:08:59.060 --> 00:09:02.040 that constructed a three-dimensional model 219 00:09:02.040 --> 00:09:03.020 of what was in front of it, 220 00:09:03.060 --> 00:09:05.360 planned an optimal path for the robot to move. 221 00:09:05.480 --> 00:09:06.100 From that, 222 00:09:06.100 --> 00:09:09.060 derived through big, hairy mathematical computations 223 00:09:09.060 --> 00:09:10.720 where it should place each foot 224 00:09:10.720 --> 00:09:11.820 to keep its balance, 225 00:09:11.980 --> 00:09:14.200 and then resolved how much torque 226 00:09:14.200 --> 00:09:15.620 it should apply to each motor. 227 00:09:15.780 --> 00:09:17.880 And after it had thought for a very long time, 228 00:09:18.020 --> 00:09:19.820 then it would move one foot, 229 00:09:19.920 --> 00:09:21.440 and then it would think for a long time again 230 00:09:21.440 --> 00:09:22.220 and do it again. 231 00:09:22.480 --> 00:09:23.940 We took a very different approach 232 00:09:23.940 --> 00:09:24.640 where we started, 233 00:09:24.820 --> 00:09:26.420 off with the idea of everything moving, 234 00:09:27.020 --> 00:09:28.600 and we just coupled simple feedback 235 00:09:28.600 --> 00:09:30.220 between the legs. 236 00:09:30.940 --> 00:09:33.400 So this is a very simple way 237 00:09:33.400 --> 00:09:35.260 of controlling a robot to walk around. 238 00:09:35.520 --> 00:09:36.280 This, by the way, 239 00:09:37.120 --> 00:09:39.240 and I'm not going to go into this whole story right now, 240 00:09:39.360 --> 00:09:40.400 but this, by the way, 241 00:09:40.440 --> 00:09:41.800 ended up inspiring, 242 00:09:42.400 --> 00:09:44.680 through a whole bunch of different angles, 243 00:09:44.920 --> 00:09:48.400 the robot Sojourner that went to Mars in 1996, 244 00:09:49.040 --> 00:09:51.020 landed in July 4th, 1997, 245 00:09:51.880 --> 00:09:54.720 and after 28 days under ground control, 246 00:09:54.720 --> 00:09:58.580 was let go to operate under exactly the same sort of control 247 00:09:58.580 --> 00:09:59.480 as I've shown you here, 248 00:09:59.580 --> 00:10:03.000 and explore its local terrain by itself. 249 00:10:03.920 --> 00:10:09.680 So we were building these robots like insects, 250 00:10:09.740 --> 00:10:12.780 and I had this plan that after we got insects working well, 251 00:10:12.960 --> 00:10:14.620 then I'd build a reptile, 252 00:10:14.680 --> 00:10:16.480 and then I'd build a proto-mammal, 253 00:10:16.580 --> 00:10:17.580 and then maybe a cat, 254 00:10:17.700 --> 00:10:19.540 and I'd sort of work up the evolutionary chain, 255 00:10:19.760 --> 00:10:21.840 building more and more complex robots. 256 00:10:22.400 --> 00:10:24.700 But we spent 10 years on building insects, 257 00:10:24.720 --> 00:10:29.360 and I multiplied 10 years out by this number here, 258 00:10:29.520 --> 00:10:30.580 and how long I was going to live, 259 00:10:30.700 --> 00:10:33.220 and I realized that the best I was going to get in life 260 00:10:33.220 --> 00:10:34.200 was build a cat, 261 00:10:34.380 --> 00:10:36.920 and that didn't seem real exciting. 262 00:10:37.460 --> 00:10:42.200 So I decided to jump in around 1992 263 00:10:42.200 --> 00:10:44.700 and try to build a human robot. 264 00:10:46.160 --> 00:10:48.080 Now, people were surprised by this, 265 00:10:48.160 --> 00:10:50.420 because we'd been building insect-like robots, 266 00:10:50.580 --> 00:10:50.820 and they said, 267 00:10:50.860 --> 00:10:53.520 well, humans are a lot more complex than insects. 268 00:10:55.140 --> 00:10:56.920 And I'm not sure I agree. 269 00:10:57.960 --> 00:11:02.080 I think, and I got into trouble when I said this before. 270 00:11:02.200 --> 00:11:04.300 The Wall Street Journal gave me a really hard time, 271 00:11:04.400 --> 00:11:07.280 but I think we tend to over-anthropomorphize people. 272 00:11:08.860 --> 00:11:10.920 And that maybe we're not as complex 273 00:11:10.920 --> 00:11:12.240 as we'd like to think we are. 274 00:11:13.160 --> 00:11:15.280 People had been trying to build humanoid robots 275 00:11:15.280 --> 00:11:16.080 for quite a while. 276 00:11:16.180 --> 00:11:18.600 The one on the left is from 1932 in Japan. 277 00:11:18.940 --> 00:11:20.840 The one on the right is from the front page 278 00:11:20.840 --> 00:11:22.360 of the New York Times in 1947. 279 00:11:23.680 --> 00:11:24.620 When we look at them, 280 00:11:24.620 --> 00:11:24.880 and today, 281 00:11:26.300 --> 00:11:29.000 the computational power on board looks a little pitiful. 282 00:11:29.080 --> 00:11:30.960 The one on the left, you see, has five valves, 283 00:11:31.920 --> 00:11:33.400 which doesn't quite seem enough 284 00:11:33.400 --> 00:11:34.920 to emulate all of human behavior. 285 00:11:35.620 --> 00:11:37.380 And I do point out to my students 286 00:11:37.380 --> 00:11:40.940 that maybe the 24 PCs they have running our current robots 287 00:11:40.940 --> 00:11:43.440 aren't going to be enough to emulate all of human behavior. 288 00:11:43.640 --> 00:11:44.820 So we have to be a little careful. 289 00:11:45.740 --> 00:11:47.820 But in the 80s, in Japan, 290 00:11:48.060 --> 00:11:50.040 there was a push to build humanoid robots 291 00:11:50.040 --> 00:11:50.960 at Waseda University. 292 00:11:51.640 --> 00:11:54.600 And then people had been building animatronic robots, 293 00:11:54.620 --> 00:11:56.760 robots for entertainment for quite a while. 294 00:11:57.180 --> 00:11:59.840 One of those heads on the right is a real head. 295 00:12:00.000 --> 00:12:02.240 The others are, that's Steve Jacobson 296 00:12:02.240 --> 00:12:03.820 from University of Utah. 297 00:12:05.600 --> 00:12:07.280 And then in the early 90s, 298 00:12:07.280 --> 00:12:09.320 our own work on the left and other work in Japan, 299 00:12:09.440 --> 00:12:10.960 there was really a humanoid explosion 300 00:12:10.960 --> 00:12:12.860 of people trying to build humanoid robots. 301 00:12:14.020 --> 00:12:17.060 Honda Motor Corporation is perhaps one of the best 302 00:12:17.060 --> 00:12:19.280 or most well-known examples of this, 303 00:12:19.320 --> 00:12:21.760 where they've spent a lot of internal resources 304 00:12:21.760 --> 00:12:24.320 on building a humanoid robot that can walk around, 305 00:12:24.480 --> 00:12:24.600 and they've been doing a lot of work on it. 306 00:12:24.620 --> 00:12:25.840 They use it in their advertising, 307 00:12:26.060 --> 00:12:29.360 but they're also pushing it ultimately for home care 308 00:12:29.360 --> 00:12:31.720 for elderly in their homes in Japan. 309 00:12:32.280 --> 00:12:34.640 And NASA has humanoid projects 310 00:12:34.640 --> 00:12:37.560 where they're building robots with human form 311 00:12:37.560 --> 00:12:40.660 so that they can be used on spacewalks 312 00:12:40.660 --> 00:12:41.800 outside the space station 313 00:12:41.800 --> 00:12:43.940 with human teleoperation control, 314 00:12:44.160 --> 00:12:46.180 and they're making them with exactly human form 315 00:12:46.180 --> 00:12:47.680 so that they can use the same tools 316 00:12:47.680 --> 00:12:50.360 that the astronauts use, the same footholds, et cetera. 317 00:12:50.620 --> 00:12:54.380 And that the astronaut inside, 318 00:12:54.480 --> 00:12:54.600 the astronaut inside the space station, 319 00:12:54.600 --> 00:12:56.920 can intuitively map their body 320 00:12:56.920 --> 00:12:58.260 to the robot they're controlling outside. 321 00:13:00.000 --> 00:13:02.880 We started by building a robot called COG, 322 00:13:05.520 --> 00:13:08.180 and this is one version of COG. 323 00:13:08.180 --> 00:13:10.940 COG's had about four head transplants over time 324 00:13:10.940 --> 00:13:12.340 and a couple of arm transplants. 325 00:13:12.620 --> 00:13:16.180 It's an upper torso, or torso and head, no legs. 326 00:13:16.660 --> 00:13:19.620 And we looked at child development literature 327 00:13:19.620 --> 00:13:22.580 to figure out how children developed over time. 328 00:13:22.740 --> 00:13:23.840 We looked at... 329 00:13:24.600 --> 00:13:27.900 a lot about different animals 330 00:13:27.900 --> 00:13:29.880 and what sort of capabilities they had. 331 00:13:30.100 --> 00:13:33.640 So we looked at both phylogeny of different animals 332 00:13:33.640 --> 00:13:35.720 and the ontogeny of human development 333 00:13:35.720 --> 00:13:38.280 to figure out what sort of capabilities 334 00:13:38.280 --> 00:13:40.020 there were inside humans 335 00:13:40.600 --> 00:13:43.200 and what sort of modules we should try and put together. 336 00:13:43.380 --> 00:13:44.980 And we wanted to build each of those modules 337 00:13:44.980 --> 00:13:48.840 as a very simple sort of set of computations, 338 00:13:48.860 --> 00:13:50.900 much like the insect-like computations 339 00:13:50.900 --> 00:13:51.780 I was showing you before. 340 00:13:52.640 --> 00:13:53.500 We also... 341 00:13:55.160 --> 00:13:56.240 realized that... 342 00:13:57.420 --> 00:13:59.800 reports from people on their own introspection 343 00:13:59.800 --> 00:14:00.760 were not reliable. 344 00:14:01.120 --> 00:14:03.100 And let me just show you one little example of this. 345 00:14:03.180 --> 00:14:05.520 This is from split brain studies, 346 00:14:05.660 --> 00:14:08.400 where as a treatment for epilepsy at one time, 347 00:14:08.500 --> 00:14:11.900 it was quite common to cut the corpus callosum, 348 00:14:11.940 --> 00:14:15.340 the main connection between the two halves of the cerebellum 349 00:14:15.340 --> 00:14:16.640 in human patients. 350 00:14:16.840 --> 00:14:22.060 And that stopped the epileptic storms going through the brain, 351 00:14:22.200 --> 00:14:23.440 and so it was quite successful. 352 00:14:23.600 --> 00:14:24.580 But then people realized, 353 00:14:24.580 --> 00:14:27.340 people were left with two half brains without connection. 354 00:14:28.380 --> 00:14:29.740 Now, in the human system, 355 00:14:30.800 --> 00:14:33.660 the right half of your brain controls the left half of your body. 356 00:14:33.780 --> 00:14:36.440 The left half of your brain controls the right half of your body. 357 00:14:36.540 --> 00:14:38.420 And you've probably all seen people with strokes 358 00:14:38.420 --> 00:14:40.880 where half of their body is not very functional 359 00:14:40.880 --> 00:14:43.540 because they've had a stroke in one half of their brain, 360 00:14:43.620 --> 00:14:44.760 the other half of their brain is fine. 361 00:14:45.360 --> 00:14:47.920 So left half of the brain controls the right half of the body. 362 00:14:48.060 --> 00:14:50.480 Right half of the brain controls the left half of the body. 363 00:14:51.060 --> 00:14:52.760 And even in our visual field, 364 00:14:52.760 --> 00:14:55.920 we see stuff on the left of our visual field 365 00:14:55.920 --> 00:14:57.020 with our right half brain. 366 00:14:57.220 --> 00:15:00.100 We see stuff on the right half of our visual field 367 00:15:00.100 --> 00:15:02.540 from both eyes with the left half of our brain. 368 00:15:03.760 --> 00:15:07.280 But we talk, most of us, with our left half of our brain. 369 00:15:07.780 --> 00:15:09.760 So if we talk to these patients, 370 00:15:10.860 --> 00:15:13.800 both halves of them are hearing us, 371 00:15:13.920 --> 00:15:16.540 but only the left half brain is responding. 372 00:15:16.900 --> 00:15:18.380 And some interesting things happen. 373 00:15:18.520 --> 00:15:19.960 Here's some example experiments. 374 00:15:20.860 --> 00:15:22.740 Up in the number one, 375 00:15:22.740 --> 00:15:25.620 the patient, the woman, is shown, 376 00:15:25.780 --> 00:15:27.380 just flashed up a word, 377 00:15:27.500 --> 00:15:29.940 and she's focusing on a spot in the center of the screen, 378 00:15:30.140 --> 00:15:32.180 so only her right visual, 379 00:15:32.360 --> 00:15:37.220 sorry, only her left visual field sees the word laugh. 380 00:15:37.440 --> 00:15:39.440 So the right half brain, 381 00:15:39.700 --> 00:15:41.480 which is disconnected from the left half brain, 382 00:15:41.620 --> 00:15:42.820 sees the word laugh. 383 00:15:43.200 --> 00:15:46.360 She starts laughing under control of her right half brain. 384 00:15:46.600 --> 00:15:48.080 Then the experimenter says, 385 00:15:48.180 --> 00:15:49.040 why are you laughing? 386 00:15:49.400 --> 00:15:52.140 Now it's the left half brain which gets to answer. 387 00:15:52.840 --> 00:15:55.240 The left half brain didn't see that word laugh. 388 00:15:55.460 --> 00:15:57.040 The left half brain has only noticed 389 00:15:57.040 --> 00:15:58.640 that the person is laughing, 390 00:15:58.740 --> 00:15:59.160 ha ha ha. 391 00:15:59.800 --> 00:16:01.500 But immediately she says, 392 00:16:01.660 --> 00:16:03.060 oh, you guys are really something. 393 00:16:03.360 --> 00:16:05.820 She makes up an excuse that she's laughing. 394 00:16:06.180 --> 00:16:07.720 There's a couple more examples there. 395 00:16:08.060 --> 00:16:09.300 Here's another example. 396 00:16:09.660 --> 00:16:12.720 Here, this is actually a 15-year-old boy from New Hampshire 397 00:16:12.720 --> 00:16:13.540 with a split brain. 398 00:16:15.520 --> 00:16:17.660 He's flashed up the picture on the big screen, 399 00:16:17.760 --> 00:16:18.640 the two pictures, 400 00:16:18.760 --> 00:16:22.100 the Volkswagen in the snow and the chicken foot. 401 00:16:22.740 --> 00:16:25.840 Just flashed up while he's focusing on that spot in the middle. 402 00:16:25.900 --> 00:16:28.660 So his right brain sees the Volkswagen in the snow. 403 00:16:28.840 --> 00:16:30.920 His left brain sees the chicken foot, 404 00:16:31.060 --> 00:16:32.660 flashes up, goes away. 405 00:16:32.920 --> 00:16:36.620 Then he's asked to point to something, 406 00:16:36.780 --> 00:16:38.540 one of those images down below on the table, 407 00:16:38.580 --> 00:16:39.940 which corresponds to what he saw. 408 00:16:40.480 --> 00:16:42.160 Well, his right half of the brain, 409 00:16:42.280 --> 00:16:45.080 which saw the snowy scene with the Volkswagen, 410 00:16:45.160 --> 00:16:46.440 points to a show snubble. 411 00:16:47.520 --> 00:16:50.060 Show snow, snow shovel. 412 00:16:50.280 --> 00:16:51.420 Ha ha ha. 413 00:16:52.960 --> 00:16:54.580 His left half of the brain, 414 00:16:54.720 --> 00:16:56.260 which saw the chicken foot, 415 00:16:56.900 --> 00:16:59.700 points with his right hand to the chicken head. 416 00:17:00.280 --> 00:17:00.680 Okay? 417 00:17:01.180 --> 00:17:02.440 Now, in this particular case, 418 00:17:02.600 --> 00:17:04.240 and there's lots and lots of experiments like this, 419 00:17:04.320 --> 00:17:05.580 the experimenter said to him, 420 00:17:05.720 --> 00:17:07.960 why did you point to the shovel? 421 00:17:08.920 --> 00:17:09.860 Who's going to answer? 422 00:17:10.040 --> 00:17:11.440 The left half of the brain. 423 00:17:11.520 --> 00:17:12.480 That's the one that has speech. 424 00:17:12.820 --> 00:17:15.340 The left half of the brain saw the chicken foot 425 00:17:16.360 --> 00:17:18.620 pointed with the right hand to the chicken head, 426 00:17:18.760 --> 00:17:19.620 and this kid says, 427 00:17:19.760 --> 00:17:21.400 oh, to clean up after the chicken. 428 00:17:22.280 --> 00:17:22.760 Okay? 429 00:17:22.900 --> 00:17:23.700 He's lying. 430 00:17:23.880 --> 00:17:26.900 The left half of his brain is just plain lying, 431 00:17:27.120 --> 00:17:28.660 just making stuff up. 432 00:17:30.000 --> 00:17:32.760 But he doesn't worry about this. 433 00:17:32.840 --> 00:17:36.380 His two half brains are happily living in coexistence 434 00:17:36.380 --> 00:17:37.100 inside his head, 435 00:17:37.220 --> 00:17:39.440 and he's quite willing to make up stories. 436 00:17:39.940 --> 00:17:40.720 And I think, 437 00:17:42.100 --> 00:17:44.060 and if you look further into the literature, 438 00:17:44.140 --> 00:17:46.840 you see that we are quite willing to make up stories 439 00:17:46.840 --> 00:17:48.200 about what goes on. 440 00:17:48.280 --> 00:17:49.720 And I'm sure you've all had that experience 441 00:17:49.720 --> 00:17:51.400 where someone's asked you a question, 442 00:17:51.400 --> 00:17:52.560 which is a little embarrassing, 443 00:17:52.680 --> 00:17:53.520 and you give some answer, 444 00:17:53.620 --> 00:17:54.380 and then afterwards you think, 445 00:17:54.520 --> 00:17:55.600 why did I say that? 446 00:17:55.760 --> 00:17:57.100 Why did I make up that excuse? 447 00:17:58.180 --> 00:18:00.140 So we have this view of ourselves, 448 00:18:00.480 --> 00:18:02.800 some higher level layers are rationalizing 449 00:18:02.800 --> 00:18:04.460 what's going on at the lower level layers, 450 00:18:05.000 --> 00:18:06.460 but those lower level layers, 451 00:18:06.640 --> 00:18:07.380 I believe, 452 00:18:07.620 --> 00:18:11.100 from all my reading and experiments we've looked at, 453 00:18:11.780 --> 00:18:13.920 are really going on rather autonomously. 454 00:18:14.320 --> 00:18:16.660 And there's this whole layer of stuff 455 00:18:16.660 --> 00:18:18.500 that we don't have conscious control over. 456 00:18:18.580 --> 00:18:21.040 So we started out to start working on those, 457 00:18:21.040 --> 00:18:22.580 unconscious layers, 458 00:18:22.820 --> 00:18:24.560 and this was our original development plan 459 00:18:24.560 --> 00:18:26.540 for our humanoid robots. 460 00:18:26.900 --> 00:18:28.600 And starting over on the left here 461 00:18:30.420 --> 00:18:32.240 was a time scale, 462 00:18:32.960 --> 00:18:34.160 and this time scale 463 00:18:35.280 --> 00:18:37.280 corresponded to two different sorts of time scales. 464 00:18:37.380 --> 00:18:40.140 We started out by building these modules on the left, 465 00:18:40.240 --> 00:18:43.040 and over time we were going to build 466 00:18:43.040 --> 00:18:44.260 more and more of these modules, 467 00:18:44.360 --> 00:18:45.800 and are continuing to build these modules. 468 00:18:46.380 --> 00:18:49.260 Originally I did have years attached to them at the bottom, 469 00:18:49.320 --> 00:18:50.780 but I cleverly erased that. 470 00:18:51.040 --> 00:18:51.620 In this diagram. 471 00:18:52.700 --> 00:18:54.300 So, time goes this way, 472 00:18:54.340 --> 00:18:55.800 and this is the modules we're going to build, 473 00:18:55.900 --> 00:18:57.260 but then for our individual robots, 474 00:18:57.380 --> 00:18:58.200 when we switched them on, 475 00:18:58.940 --> 00:19:01.080 time would proceed from left to right 476 00:19:01.080 --> 00:19:03.100 with these modules starting up first, 477 00:19:04.500 --> 00:19:06.500 learning how to interact with each other, 478 00:19:06.560 --> 00:19:08.760 and a development cycle going through, 479 00:19:08.800 --> 00:19:12.020 much like a child developing over the first few years. 480 00:19:12.540 --> 00:19:13.680 So this was our plan, 481 00:19:14.920 --> 00:19:17.200 and after we'd been working on this a few years, 482 00:19:18.200 --> 00:19:20.640 we were about here, 483 00:19:21.040 --> 00:19:22.440 and that's the sort of, 484 00:19:22.520 --> 00:19:24.360 the pieces on the left we'd built, 485 00:19:25.040 --> 00:19:29.880 and we just took some videotape of Cynthia Brazeal, 486 00:19:30.020 --> 00:19:31.900 one of my graduate students at the time, 487 00:19:31.920 --> 00:19:33.020 interacting with the robot, 488 00:19:33.240 --> 00:19:35.000 and you can see there, 489 00:19:35.060 --> 00:19:36.660 she's sort of taking turns with the robot, 490 00:19:37.460 --> 00:19:39.960 playing this little game where she shakes the eraser, 491 00:19:40.220 --> 00:19:41.160 the robot sees it, 492 00:19:41.260 --> 00:19:42.780 it moves its eyes towards it, 493 00:19:42.840 --> 00:19:44.240 it reaches out for it, 494 00:19:45.040 --> 00:19:46.520 being attracted by the motion. 495 00:19:47.160 --> 00:19:49.400 And sometimes it gets surprised by things, 496 00:19:50.420 --> 00:19:51.020 and, you know, 497 00:19:51.020 --> 00:19:52.240 the robot just kind of looks at it. 498 00:19:52.540 --> 00:19:55.040 Well, that was all very fine, that interaction, 499 00:19:55.540 --> 00:19:57.500 but then we went back and looked at our diagram, 500 00:19:57.720 --> 00:20:00.500 and that turn-taking you saw there was over on the right. 501 00:20:00.580 --> 00:20:01.900 We hadn't implemented that yet. 502 00:20:03.140 --> 00:20:04.980 Now, nothing magic was happening, 503 00:20:05.100 --> 00:20:07.680 but we had to analyze what was going on, 504 00:20:07.740 --> 00:20:09.280 and we realized that Cynthia, 505 00:20:09.460 --> 00:20:11.120 although she was a designer of the robot, 506 00:20:11.300 --> 00:20:13.220 and knew exactly what the robot was capable of, 507 00:20:13.340 --> 00:20:16.640 had engaged in a play behavior with the robot 508 00:20:16.640 --> 00:20:19.280 where she was providing the turn-taking, 509 00:20:19.280 --> 00:20:21.540 the robot wasn't capable of turn-taking, 510 00:20:21.720 --> 00:20:25.180 but she was capitalizing on what it could do, 511 00:20:25.360 --> 00:20:26.840 see something and reach for it, 512 00:20:26.920 --> 00:20:28.460 and she was making a game out of it, 513 00:20:28.540 --> 00:20:29.460 and out of that, 514 00:20:29.460 --> 00:20:31.700 the observer sees the robot and the human 515 00:20:31.700 --> 00:20:33.080 taking turns with each other 516 00:20:33.080 --> 00:20:36.020 when it's all the human's doing, the turn-taking. 517 00:20:36.100 --> 00:20:38.040 The robot wasn't actually turn-taking 518 00:20:38.040 --> 00:20:39.360 and didn't know anything about it. 519 00:20:40.960 --> 00:20:42.860 That got us looking in more literature, 520 00:20:42.920 --> 00:20:44.380 and we realized that was going on, 521 00:20:44.420 --> 00:20:47.840 and is well known in the child development literature, 522 00:20:48.060 --> 00:20:49.260 about what mothers do, 523 00:20:49.260 --> 00:20:50.140 what mothers do with their children. 524 00:20:50.240 --> 00:20:52.200 The mothers always believe their children 525 00:20:52.200 --> 00:20:54.540 are somewhat more capable than they currently are, 526 00:20:54.780 --> 00:20:57.940 engage them in a behavior the child can't do by itself, 527 00:20:58.180 --> 00:21:01.300 and from that, the child is learning what's going on, 528 00:21:01.360 --> 00:21:03.560 even though the mother thinks the child is already doing 529 00:21:03.560 --> 00:21:05.240 what they're engaging it with. 530 00:21:05.420 --> 00:21:07.400 They're pumping the interaction. 531 00:21:07.920 --> 00:21:10.060 So that got us to thinking that we really had to look 532 00:21:10.060 --> 00:21:13.260 at the social interactions between humans and robots, 533 00:21:13.440 --> 00:21:15.080 which was not something we'd been looking at before. 534 00:21:15.500 --> 00:21:17.760 And so Cynthia, on the right there, 535 00:21:17.760 --> 00:21:18.940 built this robot, Kismet. 536 00:21:20.360 --> 00:21:24.040 And besides having the basic behaviors that we'd had before, 537 00:21:24.140 --> 00:21:26.280 we added a new level, a social level, 538 00:21:26.440 --> 00:21:30.020 where we wanted the robot to respond to the human, 539 00:21:30.760 --> 00:21:32.780 but also to get the human to respond to the robot. 540 00:21:33.400 --> 00:21:38.040 So this robot, Kismet, has two eyes, 541 00:21:38.080 --> 00:21:41.740 and behind its eyes it has high-resolution cameras, 542 00:21:41.900 --> 00:21:44.020 which are very narrow-angle, foveal cameras, 543 00:21:44.200 --> 00:21:46.540 just like you have a fovea, which you use to read, 544 00:21:46.580 --> 00:21:47.620 and you scan your eyes, 545 00:21:47.760 --> 00:21:49.100 and you turn it around to get the high resolution. 546 00:21:49.540 --> 00:21:51.600 In its nose it has two hidden cameras, 547 00:21:51.680 --> 00:21:53.240 which give wide-angle views. 548 00:21:53.500 --> 00:21:55.840 Your eyes have both wide-angle and narrow-angle. 549 00:21:55.920 --> 00:21:58.160 This only has narrow-angle behind its eyes. 550 00:21:58.860 --> 00:22:02.300 It has a mouth, which can move with the jaw and lips, 551 00:22:02.500 --> 00:22:05.160 so it can show... 552 00:22:05.940 --> 00:22:08.980 It can correlate the sound that it's making with its lip motion, 553 00:22:09.200 --> 00:22:11.540 so people can understand what's going on a little more. 554 00:22:11.720 --> 00:22:13.760 It has ears, those pink ears up there, 555 00:22:13.820 --> 00:22:15.320 so it's not exactly a human robot. 556 00:22:15.520 --> 00:22:17.040 It's got a little bit of dog in it. 557 00:22:17.760 --> 00:22:18.700 Those ears are steerable. 558 00:22:18.740 --> 00:22:21.360 We can understand dogs' ears and how they move, 559 00:22:21.500 --> 00:22:24.780 because we evolved to understand ear motion. 560 00:22:24.900 --> 00:22:27.100 We still have, in our superior colliculus, 561 00:22:27.240 --> 00:22:31.280 a layer of neurons for ear control, which are vestigal. 562 00:22:34.220 --> 00:22:37.480 And the robot has a neck, but it doesn't have any body. 563 00:22:37.580 --> 00:22:38.980 But it has a little more than a neck, 564 00:22:38.980 --> 00:22:41.300 so it can sort of do something that we do with its shoulders. 565 00:22:42.000 --> 00:22:44.760 And then this robot, we want to... 566 00:22:44.760 --> 00:22:47.320 We tried to get it to interact with people, 567 00:22:48.840 --> 00:22:49.300 and we did. 568 00:22:49.300 --> 00:22:51.560 People come with their evolution, development, culture. 569 00:22:52.420 --> 00:22:55.500 The stuff in red is provided for us. 570 00:22:55.560 --> 00:22:56.400 We can't mess with that. 571 00:22:56.540 --> 00:22:58.900 The stuff in blue is what we as engineers could mess with. 572 00:22:59.040 --> 00:23:02.280 The robot's perception, attention, motivation, behavior, expression, 573 00:23:02.520 --> 00:23:05.020 and to try and tune the robot to a human. 574 00:23:05.140 --> 00:23:07.600 So what we were looking at was how we could get the robot 575 00:23:07.600 --> 00:23:10.540 to get people to react to it, 576 00:23:10.620 --> 00:23:12.760 and get the robot to react to people. 577 00:23:12.980 --> 00:23:16.940 And so we've got Cynthia Brazeal on the right there, 578 00:23:17.360 --> 00:23:18.140 Kismet on the left. 579 00:23:18.220 --> 00:23:19.840 She has a book coming out with MIT Press 580 00:23:19.840 --> 00:23:21.300 in the next couple of months, by the way, 581 00:23:21.320 --> 00:23:24.680 with a lot of data taken from many experiments with this robot. 582 00:23:25.460 --> 00:23:26.620 Oh, let me back up. 583 00:23:28.260 --> 00:23:30.280 Tim, I need my robot here. 584 00:23:31.100 --> 00:23:32.560 Tim's a naive robot. 585 00:23:32.720 --> 00:23:33.580 He knows nothing here. 586 00:23:36.460 --> 00:23:37.520 So I'm gonna... 587 00:23:38.600 --> 00:23:39.580 This is in the future. 588 00:23:39.680 --> 00:23:41.740 I've got a home robot, and I want him to... 589 00:23:41.740 --> 00:23:42.940 Well, maybe he works at my factory. 590 00:23:43.160 --> 00:23:46.520 And I want to get him to do some task for me. 591 00:23:46.520 --> 00:23:47.960 You know, maybe I want him to install 592 00:23:48.520 --> 00:23:50.440 10,000 batteries and 10,000 cell phones. 593 00:23:50.740 --> 00:23:52.880 And I could write a big C++ program, 594 00:23:53.720 --> 00:23:56.680 or I could try and show him how to do it, okay? 595 00:23:56.960 --> 00:23:57.640 Now, he's... 596 00:23:58.460 --> 00:24:00.680 He's actually showing that he understands what I'm saying 597 00:24:00.680 --> 00:24:02.560 with those little nods, and he's making eye contact. 598 00:24:03.000 --> 00:24:04.240 And then I put this up here. 599 00:24:04.280 --> 00:24:06.280 I just pull that up there, and he immediately looked down. 600 00:24:06.340 --> 00:24:07.520 We didn't practice this beforehand. 601 00:24:07.760 --> 00:24:10.400 This is... He's doing this all fresh. 602 00:24:11.560 --> 00:24:13.060 He immediately looks down at this, 603 00:24:13.100 --> 00:24:15.120 and I put that in there, and I click it on. 604 00:24:16.120 --> 00:24:17.580 And he just nodded. 605 00:24:18.040 --> 00:24:20.840 He nodded, and actually his nostrils flared a little bit 606 00:24:20.840 --> 00:24:23.240 because he was showing that he understood. 607 00:24:23.360 --> 00:24:24.480 Now I give him the task, 608 00:24:26.680 --> 00:24:28.700 and he's able to follow what I did, map, 609 00:24:28.800 --> 00:24:29.620 and he does it. 610 00:24:29.640 --> 00:24:31.200 Good robot, okay? 611 00:24:32.060 --> 00:24:34.400 So what you saw there between my... 612 00:24:34.400 --> 00:24:35.440 You can sit down now. 613 00:24:35.820 --> 00:24:38.380 What you saw there between the robot and me 614 00:24:38.380 --> 00:24:39.840 was I directed his visual attention, 615 00:24:40.100 --> 00:24:42.200 and I did that just by moving my hands up, 616 00:24:42.240 --> 00:24:43.160 and he knew exactly... 617 00:24:43.160 --> 00:24:44.920 And I moved my hands up and looked down, 618 00:24:45.120 --> 00:24:46.420 and he knew exactly where to look. 619 00:24:46.600 --> 00:24:49.360 I didn't have to go through some complex thing 620 00:24:49.360 --> 00:24:52.580 because socially, that's how humans interact all the time. 621 00:24:52.660 --> 00:24:53.940 We know that something's happening, 622 00:24:53.960 --> 00:24:54.680 and we should look there. 623 00:24:56.660 --> 00:24:59.760 He, the robot, communicated his internal state to me, 624 00:24:59.800 --> 00:25:01.740 the human, by nodding and smiling slightly, 625 00:25:01.900 --> 00:25:02.980 giving me a little bit of feedback 626 00:25:02.980 --> 00:25:04.640 that he was understanding what was happening. 627 00:25:04.740 --> 00:25:06.440 If this had been a really complicated thing 628 00:25:06.440 --> 00:25:09.120 with 16 screws and 83 wires, 629 00:25:09.420 --> 00:25:10.940 he may have gotten a bit lost, 630 00:25:11.000 --> 00:25:13.040 and he may have looked puzzled at some point, 631 00:25:13.120 --> 00:25:14.600 and that would have given me a social cue 632 00:25:15.120 --> 00:25:17.160 to sit down or go over one point a little bit more. 633 00:25:18.880 --> 00:25:21.120 He recognized the socially communicated reinforcement 634 00:25:21.120 --> 00:25:22.280 when I said, good robot, 635 00:25:22.820 --> 00:25:24.280 or when I nodded to him. 636 00:25:24.620 --> 00:25:27.820 And there was a regulation of social interaction 637 00:25:27.820 --> 00:25:29.460 between the two of us going on. 638 00:25:29.600 --> 00:25:32.260 So we tried to put all those things in Kismet. 639 00:25:33.980 --> 00:25:36.160 This is its overall architecture, 640 00:25:36.360 --> 00:25:38.920 and I'm not going to go into any detail at all about this, 641 00:25:39.000 --> 00:25:40.780 but the important thing is to understand 642 00:25:40.780 --> 00:25:42.340 that there is no central control. 643 00:25:42.480 --> 00:25:43.260 There's no homunculus. 644 00:25:43.740 --> 00:25:46.900 There's just little pieces out there doing different things. 645 00:25:47.000 --> 00:25:48.020 There's an emotional system, 646 00:25:48.140 --> 00:25:49.560 there's a visual attention system, 647 00:25:49.740 --> 00:25:50.520 there's a motor system. 648 00:25:50.900 --> 00:25:52.820 All of these things are running in parallel 649 00:25:52.820 --> 00:25:53.900 on different computers, 650 00:25:54.060 --> 00:25:55.660 sending messages to each other, 651 00:25:55.740 --> 00:25:57.580 but there's no single computer in control, 652 00:25:57.760 --> 00:25:59.060 there's no single place in control. 653 00:25:59.140 --> 00:26:00.220 It's totally distributed. 654 00:26:00.520 --> 00:26:03.040 I'm going to show you some pieces of this 655 00:26:03.040 --> 00:26:04.100 and how they work, 656 00:26:04.200 --> 00:26:05.580 and then show you an example of them, 657 00:26:05.600 --> 00:26:06.540 a demonstration of them, 658 00:26:06.560 --> 00:26:07.880 all put together at the end. 659 00:26:08.640 --> 00:26:09.620 The first thing, 660 00:26:09.740 --> 00:26:11.500 and perhaps the most important thing 661 00:26:11.500 --> 00:26:13.160 for making the behavior of the robot, 662 00:26:13.160 --> 00:26:13.400 coherent, 663 00:26:13.900 --> 00:26:15.400 is its visual attention system. 664 00:26:16.780 --> 00:26:19.140 It looks, at all times, 665 00:26:19.160 --> 00:26:19.900 through its eyes, 666 00:26:20.040 --> 00:26:22.080 it looks for three different things. 667 00:26:22.180 --> 00:26:23.560 It looks for things with skin tone, 668 00:26:23.760 --> 00:26:25.780 and human skin tones across all races 669 00:26:25.780 --> 00:26:26.760 are basically the same, 670 00:26:26.800 --> 00:26:28.000 with just a little pigment change. 671 00:26:28.280 --> 00:26:29.140 So we look in a, 672 00:26:29.220 --> 00:26:31.920 we map the colors into a thing called HSV space, 673 00:26:32.100 --> 00:26:34.180 and we can pick out human skin tone very well. 674 00:26:34.640 --> 00:26:36.460 So as it sees something in the world, 675 00:26:36.500 --> 00:26:38.180 you can see there there's a hand and a face 676 00:26:38.180 --> 00:26:41.000 that it's picked out as being possibly human skin tone. 677 00:26:41.640 --> 00:26:42.520 And someone, 678 00:26:42.640 --> 00:26:43.140 the person, 679 00:26:43.160 --> 00:26:43.700 is moving, 680 00:26:43.840 --> 00:26:44.520 on the other hand, 681 00:26:44.540 --> 00:26:45.940 a brightly colored toy around, 682 00:26:46.120 --> 00:26:47.420 looks for saturated colors, 683 00:26:47.920 --> 00:26:49.020 and it looks for motion, 684 00:26:49.220 --> 00:26:51.360 so that block that's being moved around 685 00:26:51.360 --> 00:26:53.360 on the left hand of the person really shows up, 686 00:26:53.540 --> 00:26:56.100 takes those three images 687 00:26:56.100 --> 00:26:57.720 of things that might be of interest, 688 00:26:59.360 --> 00:27:00.460 adds them up together, 689 00:27:01.280 --> 00:27:03.640 and has another thing which we call habituation, 690 00:27:03.760 --> 00:27:05.640 which over time decays, 691 00:27:05.680 --> 00:27:07.360 which stops it getting too interested in one thing, 692 00:27:07.480 --> 00:27:10.040 and then it selects the most interesting thing to look at, 693 00:27:10.160 --> 00:27:11.620 based on these weights, 694 00:27:11.740 --> 00:27:12.360 W there, 695 00:27:12.360 --> 00:27:16.380 and saccades its eyes towards that interesting thing. 696 00:27:16.460 --> 00:27:17.940 Just as when you see something interesting, 697 00:27:18.540 --> 00:27:20.640 if you're sitting there in a room 698 00:27:20.640 --> 00:27:21.640 and a person walks in, 699 00:27:21.680 --> 00:27:22.040 there's motion, 700 00:27:22.180 --> 00:27:22.680 you look up, 701 00:27:22.700 --> 00:27:23.520 you saccade towards them, 702 00:27:23.560 --> 00:27:24.120 you see their face, 703 00:27:24.200 --> 00:27:24.920 you look at their face, 704 00:27:25.080 --> 00:27:26.040 figure out what's going on, 705 00:27:26.240 --> 00:27:27.940 this system autonomously does that. 706 00:27:28.400 --> 00:27:29.460 At the same time, 707 00:27:29.460 --> 00:27:31.600 other behaviors that are going on inside the robot 708 00:27:31.600 --> 00:27:33.400 might set these weights. 709 00:27:33.680 --> 00:27:35.640 So if the robot is feeling lonely, 710 00:27:35.740 --> 00:27:37.800 and I'll talk about what loneliness means in a minute, 711 00:27:38.440 --> 00:27:41.540 it might put the weight up higher for skin tone, 712 00:27:41.540 --> 00:27:43.920 and be much more interested in looking at things with skin tone. 713 00:27:44.320 --> 00:27:46.060 If it's seen lots of faces recently, 714 00:27:46.140 --> 00:27:47.680 and it's pretty satiated about that, 715 00:27:47.840 --> 00:27:49.480 but maybe it's bored and wants to play, 716 00:27:49.680 --> 00:27:51.200 it might start looking for color, 717 00:27:51.280 --> 00:27:53.500 because its toys are all highly saturated, 718 00:27:54.040 --> 00:27:55.480 have highly saturated colors, 719 00:27:55.540 --> 00:27:59.120 like the Fisher-Price toys that you all see for your kids. 720 00:28:00.020 --> 00:28:01.620 So that visual attention mechanism 721 00:28:02.820 --> 00:28:04.860 then controls how the eyes move, 722 00:28:05.540 --> 00:28:06.400 and here we see, 723 00:28:07.080 --> 00:28:08.660 did I press the button a second time? 724 00:28:10.660 --> 00:28:11.520 Here we see this, 725 00:28:11.540 --> 00:28:13.480 this is the robot looking around for a toy at this point, 726 00:28:13.600 --> 00:28:16.300 and you, as someone who's never seen this robot before, 727 00:28:16.480 --> 00:28:18.540 can tell from where its eyes are pointing, 728 00:28:18.700 --> 00:28:19.700 which way it's looking, 729 00:28:19.860 --> 00:28:22.280 and you know that it wasn't seeing the toy, 730 00:28:22.380 --> 00:28:24.700 but now its eyes are looking at the toy. 731 00:28:25.440 --> 00:28:27.900 It changes its expression a little, 732 00:28:27.980 --> 00:28:31.180 you really are pretty sure that it's seeing the toy, 733 00:28:31.300 --> 00:28:31.820 it's a bit happier, 734 00:28:31.920 --> 00:28:33.720 but it will still pay attention to other things. 735 00:28:33.920 --> 00:28:35.620 Here comes Cynthia in from the right, 736 00:28:35.740 --> 00:28:36.380 it sees her, 737 00:28:36.540 --> 00:28:38.360 and tracks her and follows her. 738 00:28:38.800 --> 00:28:40.520 So, two things going on here. 739 00:28:40.520 --> 00:28:44.180 One is it has very natural human-like visual behaviors, 740 00:28:44.420 --> 00:28:46.700 and secondly, because it's got eyeballs 741 00:28:47.360 --> 00:28:48.740 that look like eyes, 742 00:28:48.900 --> 00:28:51.160 and you can tell which way they're looking, 743 00:28:51.360 --> 00:28:53.120 you, as an observer, 744 00:28:53.260 --> 00:28:55.560 can figure out what the robot is looking at. 745 00:28:55.660 --> 00:28:56.700 There's no doubt in your mind. 746 00:28:57.420 --> 00:28:58.740 In the movie 2001, 747 00:28:59.200 --> 00:29:00.760 Hal just had that red eye, 748 00:29:00.860 --> 00:29:02.500 you couldn't ever tell where it was looking, 749 00:29:02.580 --> 00:29:03.260 it was really spooky. 750 00:29:03.620 --> 00:29:06.140 Here you can understand what the robot is doing, 751 00:29:06.680 --> 00:29:07.340 you think. 752 00:29:09.060 --> 00:29:09.540 Um, 753 00:29:10.200 --> 00:29:11.460 oh, Tim, 754 00:29:12.500 --> 00:29:14.860 now, I'm always reticent to do this, 755 00:29:15.380 --> 00:29:16.580 but, you know, if I start, 756 00:29:17.020 --> 00:29:18.160 see, he pulled back, 757 00:29:18.320 --> 00:29:20.580 if I start doing stuff right in his face, 758 00:29:22.260 --> 00:29:23.860 he reacts in some way. 759 00:29:23.940 --> 00:29:24.580 You can sit down again. 760 00:29:28.060 --> 00:29:31.720 We use social cues to regulate other people. 761 00:29:31.840 --> 00:29:33.440 If things are moving around too fast, 762 00:29:33.620 --> 00:29:36.120 we'll draw back if things are getting too close to our face, 763 00:29:36.260 --> 00:29:37.920 and that controls other people. 764 00:29:38.100 --> 00:29:39.520 So we put some behaviors in there, 765 00:29:39.540 --> 00:29:40.500 things like that, into Kismet, 766 00:29:40.600 --> 00:29:43.500 and it controls what other people are doing, 767 00:29:43.680 --> 00:29:44.700 and the other people, 768 00:29:44.740 --> 00:29:45.640 the people don't, 769 00:29:45.640 --> 00:29:46.240 other people, 770 00:29:46.360 --> 00:29:49.840 the people don't have to really know anything about the robot. 771 00:29:50.200 --> 00:29:52.900 So we've had a whole lot of naive subjects, 772 00:29:53.000 --> 00:29:53.640 which we take, 773 00:29:53.700 --> 00:29:55.260 we know nothing about the robot, 774 00:29:55.380 --> 00:29:56.360 and we test them with the robot, 775 00:29:56.600 --> 00:29:57.980 and here's a couple of examples 776 00:29:57.980 --> 00:29:59.880 of what we call social amplification, 777 00:30:00.800 --> 00:30:02.400 where they get too close, 778 00:30:02.500 --> 00:30:04.260 this is taken from many hours of videotape, 779 00:30:04.440 --> 00:30:06.460 they happen to get too close to the robot, 780 00:30:06.920 --> 00:30:07.920 it draws back. 781 00:30:08.060 --> 00:30:09.260 Am I a little too close to you? 782 00:30:09.260 --> 00:30:10.560 I can stand back. 783 00:30:10.720 --> 00:30:12.720 And he realizes he's too close and draws back. 784 00:30:12.820 --> 00:30:13.420 Here's a kid, 785 00:30:14.280 --> 00:30:15.860 who really draws back, 786 00:30:16.020 --> 00:30:17.340 but he's only 13, 787 00:30:17.480 --> 00:30:19.100 so he wants to push the boundary now, 788 00:30:19.240 --> 00:30:20.920 and he goes in really close, 789 00:30:21.060 --> 00:30:23.160 but he still can't help himself but react 790 00:30:23.160 --> 00:30:24.320 when it pulls back, 791 00:30:24.520 --> 00:30:26.160 even though he's doing this experiment. 792 00:30:26.700 --> 00:30:28.220 So here we've got the robot 793 00:30:28.880 --> 00:30:29.320 triggering 794 00:30:31.680 --> 00:30:33.840 subconscious behaviors of people. 795 00:30:34.140 --> 00:30:36.120 So we've got the people reacting to the robot, 796 00:30:36.180 --> 00:30:38.560 we've made the right behaviors of the robot 797 00:30:38.560 --> 00:30:40.400 to trigger the social behaviors 798 00:30:40.400 --> 00:30:42.340 that people normally engage in, 799 00:30:42.380 --> 00:30:43.300 cross-culturally. 800 00:30:44.920 --> 00:30:46.300 You've noticed, probably, 801 00:30:46.440 --> 00:30:48.320 that Kismet has facial expressions, 802 00:30:48.560 --> 00:30:50.540 and there's a lot of literature about 803 00:30:50.540 --> 00:30:52.480 the production of facial expressions in humans, 804 00:30:52.660 --> 00:30:55.760 this is one particular thing from the literature, 805 00:30:57.120 --> 00:30:59.320 hypothesizing a two-dimensional underlying space 806 00:30:59.320 --> 00:31:01.160 of pleasure versus displeasure, 807 00:31:01.240 --> 00:31:02.860 and arousal versus sleepiness, 808 00:31:03.060 --> 00:31:05.460 and a continuous control of facial expression 809 00:31:05.460 --> 00:31:07.060 on the right there. 810 00:31:07.060 --> 00:31:08.640 So we took that, 811 00:31:08.640 --> 00:31:11.540 and we tried to emulate those facial expressions 812 00:31:11.540 --> 00:31:12.340 with Kismet. 813 00:31:13.380 --> 00:31:15.360 We then asked naive subjects 814 00:31:15.360 --> 00:31:17.780 to rate Kismet's facial expression, 815 00:31:17.980 --> 00:31:20.000 and we also got some of the standard things 816 00:31:20.000 --> 00:31:20.520 from the literature, 817 00:31:20.620 --> 00:31:22.160 and got them to rate human facial expressions, 818 00:31:22.300 --> 00:31:22.900 compared them, 819 00:31:22.980 --> 00:31:25.720 tweaked up our facial expressions for Kismet, 820 00:31:26.540 --> 00:31:28.940 even though it doesn't have all the facial muscles 821 00:31:28.940 --> 00:31:30.200 that humans have, 822 00:31:30.340 --> 00:31:32.440 until we got pretty good correlations 823 00:31:32.440 --> 00:31:35.280 between what people thought about Kismet's facial expression 824 00:31:35.980 --> 00:31:37.040 and what they thought about Kismet's facial expression 825 00:31:37.060 --> 00:31:38.240 and what they thought about the human facial expressions. 826 00:31:38.920 --> 00:31:41.340 In fact, we built it on top of a three-dimensional 827 00:31:41.340 --> 00:31:42.680 emotional space, 828 00:31:42.960 --> 00:31:44.680 so that at any point in time, 829 00:31:44.840 --> 00:31:48.760 Kismet has its emotions represented by a point 830 00:31:48.760 --> 00:31:50.220 in this three-dimensional space. 831 00:31:50.600 --> 00:31:52.780 And different actions, 832 00:31:53.060 --> 00:31:55.000 or different things in the world, 833 00:31:55.040 --> 00:31:57.100 can push its emotional state around 834 00:31:57.100 --> 00:31:57.960 in different directions. 835 00:31:58.300 --> 00:31:59.980 So the same behavior from a human, 836 00:32:00.100 --> 00:32:01.100 when it's happy, 837 00:32:01.260 --> 00:32:03.180 will have a different influence on it 838 00:32:03.180 --> 00:32:05.380 from when Kismet is sad. 839 00:32:07.060 --> 00:32:07.780 So we have these three-dimensional axes, 840 00:32:08.040 --> 00:32:10.540 some prototypical facial expressions 841 00:32:10.540 --> 00:32:11.500 out of the extremes, 842 00:32:11.700 --> 00:32:12.960 and a couple places in the middle, 843 00:32:13.020 --> 00:32:14.880 like disgust and anger and fear, etc. 844 00:32:15.520 --> 00:32:18.640 And then we interpolate over that three-dimensional space, 845 00:32:18.820 --> 00:32:19.620 so no matter where, 846 00:32:19.740 --> 00:32:22.820 whatever Kismet's current emotion is 847 00:32:22.820 --> 00:32:24.480 as a point in that three-dimensional space, 848 00:32:24.640 --> 00:32:27.320 will be the facial expression it tends to give. 849 00:32:27.520 --> 00:32:30.620 We also make it generate different prosody in its voice. 850 00:32:31.100 --> 00:32:33.320 So here it's going to say a string of phonemes, 851 00:32:33.340 --> 00:32:34.300 which it doesn't understand, 852 00:32:36.060 --> 00:32:37.040 in different emotions, 853 00:32:37.060 --> 00:32:37.540 or different emotional states. 854 00:32:40.780 --> 00:32:41.920 It's one emotional state. 855 00:32:42.020 --> 00:32:42.980 Here it is at a different point. 856 00:32:42.980 --> 00:32:44.880 Do you really think so? 857 00:32:46.320 --> 00:32:48.560 Do you really think so? 858 00:32:50.120 --> 00:32:51.700 Do you really think so? 859 00:32:54.020 --> 00:32:55.960 Do you really think so? 860 00:32:57.920 --> 00:33:00.220 Do you really think so? 861 00:33:02.000 --> 00:33:04.220 Do you really think so? 862 00:33:05.320 --> 00:33:10.000 So Kismet is showing facial expressions, 863 00:33:10.100 --> 00:33:12.960 and it's showing prosody in its voice. 864 00:33:12.960 --> 00:33:16.160 But humans, of course, show prosody in their voice. 865 00:33:16.200 --> 00:33:18.720 We want the robot to understand the prosody in our voices. 866 00:33:20.140 --> 00:33:21.800 The general case is very difficult, 867 00:33:21.840 --> 00:33:23.620 so we did a much simpler case. 868 00:33:26.380 --> 00:33:27.920 Mothers, cross-culturally, 869 00:33:28.080 --> 00:33:31.940 tend to show the same basic contours in their voices 870 00:33:31.940 --> 00:33:35.220 when expressing approval, prohibition, attention, 871 00:33:35.860 --> 00:33:37.700 and comfort to their babies. 872 00:33:37.980 --> 00:33:39.600 And there's a characteristic pattern 873 00:33:40.160 --> 00:33:42.640 in the pitch of their voices for that. 874 00:33:43.720 --> 00:33:45.660 Kismet looks somewhat baby-like. 875 00:33:45.700 --> 00:33:46.360 It has big eyes. 876 00:33:46.440 --> 00:33:47.840 We tried to make it look baby-like. 877 00:33:47.900 --> 00:33:50.040 When we put human subjects in front of Kismet 878 00:33:50.040 --> 00:33:52.920 and asked them to express approval, attention, etc., 879 00:33:52.920 --> 00:33:54.860 and looked at the pitch in their voices, 880 00:33:54.860 --> 00:33:56.760 we saw similar sorts of contours 881 00:33:56.760 --> 00:34:00.520 as mothers and parents in general 882 00:34:00.520 --> 00:34:03.780 do for their babies, independent of language. 883 00:34:04.360 --> 00:34:06.420 So we then built a classifier for that, 884 00:34:06.520 --> 00:34:10.080 which looked at the vocal patterns 885 00:34:10.080 --> 00:34:12.040 in human voices that Kismet heard, 886 00:34:12.040 --> 00:34:15.980 and tried to classify what of those basic emotions 887 00:34:15.980 --> 00:34:17.380 was coming out of the voice. 888 00:34:17.540 --> 00:34:19.480 By the way, we had to have different classifiers 889 00:34:19.480 --> 00:34:20.780 for women and for men, 890 00:34:20.860 --> 00:34:22.540 because the vocal tracks are so different 891 00:34:22.540 --> 00:34:25.700 that we couldn't get one classifier to work across the two. 892 00:34:26.300 --> 00:34:29.560 So here you're going to see a bunch of female subjects, 893 00:34:29.720 --> 00:34:32.000 four of them are naive subjects in this case, 894 00:34:32.120 --> 00:34:34.500 being asked to variously praise, 895 00:34:34.560 --> 00:34:37.640 get attention of the robot, prohibit, and soothe the robot, 896 00:34:37.740 --> 00:34:38.860 and we use multiple languages. 897 00:34:39.100 --> 00:34:41.740 In each case, the robot starts off in the same situation, 898 00:34:42.040 --> 00:34:44.000 and then it's going to respond differently. 899 00:34:44.360 --> 00:34:45.620 The first case, praise, 900 00:34:46.420 --> 00:34:49.100 there's not much of a loop closing. 901 00:34:49.260 --> 00:34:50.900 The second case, attention, 902 00:34:51.260 --> 00:34:54.120 we asked the human subjects to get the robot's attention 903 00:34:54.120 --> 00:34:56.060 and then indicate when they had its attention. 904 00:34:56.440 --> 00:34:59.120 So here's a few different people giving praise to the robot, 905 00:34:59.240 --> 00:35:01.320 and you'll see the robot understanding the praise. 906 00:35:01.540 --> 00:35:03.320 You're a very good child. 907 00:35:03.640 --> 00:35:05.800 You do a very good job. 908 00:35:06.280 --> 00:35:08.560 Of course. 909 00:35:09.480 --> 00:35:10.520 Nice robot. 910 00:35:10.520 --> 00:35:12.620 You're such a cute little robot. 911 00:35:13.400 --> 00:35:14.400 You're so cute. 912 00:35:15.620 --> 00:35:17.140 Good job, Kismet. 913 00:35:19.700 --> 00:35:21.160 Very good, Kismet. 914 00:35:23.920 --> 00:35:25.200 Look at my smile. 915 00:35:29.180 --> 00:35:30.480 Oh, that was cute. 916 00:35:30.700 --> 00:35:32.040 Now it's not understanding the words, 917 00:35:32.100 --> 00:35:33.380 just the prosody in the voice. 918 00:35:33.480 --> 00:35:34.700 Here they're asked to get its attention 919 00:35:34.700 --> 00:35:36.560 and indicate when they have its attention. 920 00:35:37.700 --> 00:35:38.180 Kismet? 921 00:35:38.800 --> 00:35:39.620 Hey, Kismet. 922 00:35:41.980 --> 00:35:43.300 Ah, there we go. 923 00:35:44.640 --> 00:35:45.680 Vite, vite. 924 00:35:46.860 --> 00:35:47.380 Regarde. 925 00:35:47.700 --> 00:35:48.880 Ah, ici. 926 00:35:49.700 --> 00:35:51.680 So she understands she has its attention. 927 00:35:51.680 --> 00:35:52.300 Kismet, do you like the toy? 928 00:35:53.300 --> 00:35:53.820 Oh. 929 00:35:55.900 --> 00:35:57.240 Now here, this first case, 930 00:35:57.320 --> 00:35:59.700 this woman pushes the robot into an emotional corner. 931 00:35:59.780 --> 00:36:00.600 She keeps pushing. 932 00:36:01.400 --> 00:36:01.920 No. 933 00:36:02.940 --> 00:36:03.460 No. 934 00:36:03.920 --> 00:36:05.360 You're not to do that. 935 00:36:05.500 --> 00:36:06.020 No. 936 00:36:07.560 --> 00:36:08.300 Uh-uh. 937 00:36:08.300 --> 00:36:11.420 That's not appropriate. 938 00:36:11.720 --> 00:36:12.280 No. 939 00:36:14.200 --> 00:36:14.760 No. 940 00:36:19.060 --> 00:36:20.060 Ya, du. 941 00:36:20.240 --> 00:36:22.580 Now watch this woman's reaction here. 942 00:36:22.620 --> 00:36:23.260 This is very common. 943 00:36:27.500 --> 00:36:29.760 The people start imitating the robot. 944 00:36:30.660 --> 00:36:32.300 Where did you put your body? 945 00:36:35.060 --> 00:36:36.520 And she hangs her head 946 00:36:36.520 --> 00:36:38.140 to imitate the robot hanging the head. 947 00:36:38.140 --> 00:36:39.600 I'm going to skip the soothing. 948 00:36:40.800 --> 00:36:42.420 So, now we've got the robot 949 00:36:42.420 --> 00:36:44.660 starting to understand the person and their emotions, 950 00:36:44.860 --> 00:36:46.680 the people understanding the robot and their emotions. 951 00:36:47.380 --> 00:36:48.380 We then put in 952 00:36:48.840 --> 00:36:50.560 turn-taking behaviors, which 953 00:36:50.560 --> 00:36:52.480 we'd observed Cynthia driving 954 00:36:52.480 --> 00:36:54.520 before, but now we really put in turn-taking, 955 00:36:54.660 --> 00:36:56.640 vocal turn-taking. In vocal turn-taking, 956 00:36:56.780 --> 00:36:58.720 you take the floor, you speak, 957 00:36:58.920 --> 00:37:00.700 then you relinquish the floor through some 958 00:37:00.700 --> 00:37:02.260 eye motion, head motion, 959 00:37:02.460 --> 00:37:04.340 then it's the other person's turn to speak, 960 00:37:04.440 --> 00:37:05.540 and it goes back and forth. 961 00:37:06.080 --> 00:37:07.520 We had naive subjects. 962 00:37:08.820 --> 00:37:10.580 This videotapes from about a year ago. 963 00:37:10.800 --> 00:37:12.540 At that point, we had about 70 naive 964 00:37:12.540 --> 00:37:14.480 subjects. We've had over 300 naive 965 00:37:14.480 --> 00:37:16.500 subjects now come in and interact with the 966 00:37:16.500 --> 00:37:18.100 robot. We videotape what goes on. 967 00:37:18.200 --> 00:37:20.480 We then blind score what 968 00:37:20.480 --> 00:37:22.200 happens and look at 969 00:37:22.200 --> 00:37:24.580 transitions, etc. But I'm just going to show you 970 00:37:24.580 --> 00:37:26.620 two excerpts from one particular 971 00:37:26.620 --> 00:37:28.640 person. Now, the trick with 972 00:37:28.640 --> 00:37:30.620 this conversation is that 973 00:37:31.360 --> 00:37:32.760 the robot doesn't understand 974 00:37:32.760 --> 00:37:34.560 what the people are saying. It only 975 00:37:34.560 --> 00:37:35.360 understands the 976 00:37:35.360 --> 00:37:35.520 the 977 00:37:35.540 --> 00:37:37.060 emotion in their voice. It doesn't understand 978 00:37:37.060 --> 00:37:38.300 the words at all. 979 00:37:39.320 --> 00:37:41.040 And the robot is just babbling 980 00:37:41.040 --> 00:37:43.220 random English phonemes so the 981 00:37:43.220 --> 00:37:44.560 people don't understand the person. 982 00:37:45.060 --> 00:37:47.140 Okay? So this is a conversation with 983 00:37:47.140 --> 00:37:47.660 no meaning. 984 00:37:49.460 --> 00:37:51.100 And, you know, some of us 985 00:37:51.100 --> 00:37:53.100 in bars late at night have had 986 00:37:53.100 --> 00:37:54.840 similar sorts of conversations. 987 00:37:57.900 --> 00:37:58.380 Some 988 00:37:59.020 --> 00:38:01.280 naive subjects came in. We sit them down 989 00:38:01.280 --> 00:38:03.180 and the only instruction we said to them was 990 00:38:03.180 --> 00:38:04.860 talk to the robot. They knew nothing else. 991 00:38:04.980 --> 00:38:05.520 They were like, oh, I'm not going to talk to you. 992 00:38:05.520 --> 00:38:07.720 And the robot would start babbling. Look at them. 993 00:38:07.780 --> 00:38:09.440 And they'd sort of get the idea they were supposed to talk, 994 00:38:09.520 --> 00:38:10.580 but they didn't know what to say. 995 00:38:11.120 --> 00:38:12.120 And they'd say, hello. 996 00:38:13.140 --> 00:38:14.380 The robot would go, blah, blah, blah. 997 00:38:15.860 --> 00:38:17.520 And the person, the people didn't know. 998 00:38:17.640 --> 00:38:19.440 But other people, so we measure 999 00:38:19.440 --> 00:38:21.540 that turn-taking and we conclude that in our data sets 1000 00:38:21.540 --> 00:38:23.280 and that will all be coming out in Cynthia's book. 1001 00:38:24.420 --> 00:38:25.540 Other people really 1002 00:38:25.540 --> 00:38:27.340 got into it. And this one particular person 1003 00:38:27.340 --> 00:38:29.160 spoke for 25 minutes to the robot. 1004 00:38:29.260 --> 00:38:31.160 Just chatted to it. About nothing. 1005 00:38:33.280 --> 00:38:34.940 Right near the start here, you'll see 1006 00:38:34.940 --> 00:38:37.060 him, he wants to tell the robot 1007 00:38:37.060 --> 00:38:39.060 about his watch. And he says, I want you to take 1008 00:38:39.060 --> 00:38:41.020 a look at my watch. And look 1009 00:38:41.020 --> 00:38:42.940 at where the robot's eyes go. The robot 1010 00:38:42.940 --> 00:38:45.000 looks at the watch. Not because the robot understood 1011 00:38:45.000 --> 00:38:46.980 what he said, because he made this motion 1012 00:38:46.980 --> 00:38:49.060 cue and the robot's visual 1013 00:38:49.060 --> 00:38:50.800 attention system was drawn to the motion. 1014 00:38:50.960 --> 00:38:52.920 And so there's redundant channels going 1015 00:38:52.920 --> 00:38:55.040 on and enough of them work together 1016 00:38:55.040 --> 00:38:56.620 to make the right behavior happen. 1017 00:38:58.440 --> 00:38:59.260 No show this? 1018 00:39:00.500 --> 00:39:00.900 Why? 1019 00:39:00.960 --> 00:39:01.400 I don't know. 1020 00:39:02.160 --> 00:39:03.260 I want to show you something. 1021 00:39:04.940 --> 00:39:06.380 This is a watch that my 1022 00:39:07.920 --> 00:39:09.020 girlfriend gave me. 1023 00:39:09.360 --> 00:39:09.960 How the hell? 1024 00:39:10.100 --> 00:39:12.120 Yeah, look, it's got a little blue light in it, too. 1025 00:39:12.280 --> 00:39:14.020 I almost lost it this week. 1026 00:39:20.240 --> 00:39:21.140 How are you? 1027 00:39:21.500 --> 00:39:22.540 There was a collision. 1028 00:39:23.620 --> 00:39:25.660 It's not perfect, back and forth. 1029 00:39:29.060 --> 00:39:30.180 How is this? 1030 00:39:32.320 --> 00:39:33.120 What is this? 1031 00:39:33.800 --> 00:39:34.240 Yeah. 1032 00:39:36.100 --> 00:39:37.240 It's looking at his eyes. 1033 00:39:37.460 --> 00:39:38.680 I'm going to try closing my eyes. 1034 00:39:39.580 --> 00:39:40.920 Making eye contact with him. 1035 00:39:41.480 --> 00:39:43.980 I'm going to try opening my eyes. 1036 00:39:45.380 --> 00:39:46.460 I'm opening my teeth. 1037 00:39:48.900 --> 00:39:49.920 Yeah, I think so. 1038 00:39:50.280 --> 00:39:51.060 Am I too close? 1039 00:39:51.760 --> 00:39:53.560 Then he realized he got too close. 1040 00:39:54.300 --> 00:39:55.980 So we've got the fundamentals of, 1041 00:39:55.980 --> 00:39:58.180 we think, of dialogue 1042 00:39:58.180 --> 00:40:00.160 between the robot and humans 1043 00:40:00.160 --> 00:40:01.040 without any content. 1044 00:40:02.760 --> 00:40:03.900 But we realized that 1045 00:40:03.900 --> 00:40:04.860 there was a lot more. 1046 00:40:04.940 --> 00:40:05.820 There was a lot more that we needed to do 1047 00:40:06.720 --> 00:40:08.460 to have real interactions. 1048 00:40:08.620 --> 00:40:10.140 And some of these things that you saw 1049 00:40:10.140 --> 00:40:12.500 was going on between robot Tim and me. 1050 00:40:13.840 --> 00:40:15.660 That the robot 1051 00:40:16.800 --> 00:40:17.700 needs to understand 1052 00:40:17.700 --> 00:40:18.700 what it is the human knows 1053 00:40:18.700 --> 00:40:19.420 about the world 1054 00:40:21.160 --> 00:40:23.020 and what the human knows about it. 1055 00:40:23.500 --> 00:40:24.580 And it's got to understand 1056 00:40:24.580 --> 00:40:25.960 what the human is doing and why. 1057 00:40:26.340 --> 00:40:28.960 And so we've been building a theory of mind 1058 00:40:28.960 --> 00:40:30.800 for our first robot, 1059 00:40:30.940 --> 00:40:32.340 COG, and I'm not going to go into 1060 00:40:32.340 --> 00:40:33.460 great detail on this, 1061 00:40:34.140 --> 00:40:36.060 except to show you that it's based 1062 00:40:36.060 --> 00:40:37.500 on that same visual attention system. 1063 00:40:37.900 --> 00:40:40.720 It looks at everything in the world 1064 00:40:40.720 --> 00:40:42.620 and sees how they move 1065 00:40:42.620 --> 00:40:45.200 and classifies and gets the trajectory. 1066 00:40:45.340 --> 00:40:46.940 So you can see here as everything moves around 1067 00:40:46.940 --> 00:40:48.100 it tracks different things. 1068 00:40:48.360 --> 00:40:50.060 As he throws that ball up in the air 1069 00:40:50.060 --> 00:40:50.840 he tracks that. 1070 00:40:53.620 --> 00:40:55.280 Then he's going to walk off into the distance 1071 00:40:55.280 --> 00:40:57.300 and it's going to track his hands and head. 1072 00:40:58.280 --> 00:41:00.860 And then after getting that trajectory formation 1073 00:41:00.860 --> 00:41:02.940 it goes up to the theory of a body. 1074 00:41:03.460 --> 00:41:06.240 And tries to classify motions in the world 1075 00:41:06.240 --> 00:41:07.700 as being animate or inanimate. 1076 00:41:07.820 --> 00:41:09.100 The robot, after all, 1077 00:41:09.800 --> 00:41:11.320 shouldn't try and socially interact 1078 00:41:11.320 --> 00:41:13.400 with a tree that's waving in the breeze, 1079 00:41:13.980 --> 00:41:16.120 but perhaps it should try and socially interact 1080 00:41:16.120 --> 00:41:17.340 with human Tim. 1081 00:41:17.980 --> 00:41:20.140 So it needs to classify things 1082 00:41:20.140 --> 00:41:21.420 as animate or inanimate. 1083 00:41:21.480 --> 00:41:22.580 And we're pretty good at doing that, 1084 00:41:22.660 --> 00:41:23.940 as are most animals. 1085 00:41:24.140 --> 00:41:26.120 What's the difference between animate and inanimate things? 1086 00:41:26.220 --> 00:41:29.460 We don't get too confused by trees blowing in the wind, 1087 00:41:29.540 --> 00:41:31.460 by water rippling along in streams, etc. 1088 00:41:32.100 --> 00:41:32.580 So, 1089 00:41:33.460 --> 00:41:35.100 Brian Scazzolatti built a system 1090 00:41:35.100 --> 00:41:36.760 which knows about some naive physics 1091 00:41:36.760 --> 00:41:38.540 and knows what sorts of things happen, 1092 00:41:38.600 --> 00:41:40.720 the quality of motions that can happen 1093 00:41:40.720 --> 00:41:42.880 and has these different experts 1094 00:41:42.880 --> 00:41:44.120 which come up and together 1095 00:41:44.120 --> 00:41:45.200 come up with a classification. 1096 00:41:45.820 --> 00:41:46.940 So in this case 1097 00:41:47.720 --> 00:41:50.040 he's swinging his right arm around 1098 00:41:50.040 --> 00:41:52.740 and it classifies that as something animate. 1099 00:41:53.120 --> 00:41:54.620 It's got the quality of motion 1100 00:41:54.620 --> 00:41:56.320 of an animate object. 1101 00:41:56.800 --> 00:41:58.300 Here he pushes the chair 1102 00:41:58.300 --> 00:41:59.840 which just rolls across the floor 1103 00:41:59.840 --> 00:42:01.900 and it classifies that as inanimate 1104 00:42:01.900 --> 00:42:03.180 because it's sort of, 1105 00:42:03.180 --> 00:42:05.940 it has smooth deceleration, etc. 1106 00:42:06.520 --> 00:42:08.760 Here he's got a chair attached to a stick 1107 00:42:08.760 --> 00:42:09.680 and he's pushing it. 1108 00:42:09.700 --> 00:42:12.820 He was actually trying to make life easier for himself 1109 00:42:12.820 --> 00:42:15.920 and it kept classifying the chair moving as animate 1110 00:42:15.920 --> 00:42:17.380 which indeed it was animate 1111 00:42:17.380 --> 00:42:19.660 because he was the one who was providing the motive force 1112 00:42:19.660 --> 00:42:21.840 and it successfully classifies that as animate. 1113 00:42:22.600 --> 00:42:24.340 And so he built a system 1114 00:42:24.340 --> 00:42:26.620 which works pretty well with figuring out 1115 00:42:27.440 --> 00:42:29.260 whether things are animate or inanimate, 1116 00:42:29.340 --> 00:42:30.840 what needs to be interacted with. 1117 00:42:33.180 --> 00:42:35.260 If we figure out something is animate 1118 00:42:35.260 --> 00:42:37.000 we want to know where their eyes are looking, 1119 00:42:37.120 --> 00:42:38.460 whether they're paying attention to us 1120 00:42:38.460 --> 00:42:40.060 or whether they're paying attention to something else. 1121 00:42:40.200 --> 00:42:42.840 So the next thing is to find faces and eyes 1122 00:42:42.840 --> 00:42:45.060 or find eyes in animate objects. 1123 00:42:45.220 --> 00:42:47.740 And this is Cog's view of its environment 1124 00:42:48.400 --> 00:42:49.940 and I don't know if you can see 1125 00:42:49.940 --> 00:42:51.460 right in the middle there is Brian 1126 00:42:51.460 --> 00:42:53.120 sitting at a computer terminal 1127 00:42:53.120 --> 00:42:56.020 and you probably can't tell 1128 00:42:56.020 --> 00:42:57.640 which way his eyes are looking right now 1129 00:42:57.640 --> 00:42:59.300 but Cog is able to do it 1130 00:42:59.300 --> 00:43:01.740 by foveating on the face, 1131 00:43:02.560 --> 00:43:03.800 extracting the face, 1132 00:43:03.940 --> 00:43:05.580 finding the eyes and the head 1133 00:43:05.580 --> 00:43:08.160 and then getting a fairly crude estimate 1134 00:43:08.800 --> 00:43:10.020 by human standards 1135 00:43:10.020 --> 00:43:11.200 to within about six degrees 1136 00:43:11.200 --> 00:43:13.320 of where he's looking. 1137 00:43:14.060 --> 00:43:15.400 We're very, very good 1138 00:43:15.400 --> 00:43:17.100 at knowing where someone's eyes are looking 1139 00:43:17.100 --> 00:43:18.460 because evolutionarily 1140 00:43:18.460 --> 00:43:20.980 if somebody or something was looking at us 1141 00:43:20.980 --> 00:43:22.900 they were likely to be wanting to eat us 1142 00:43:22.900 --> 00:43:24.820 and so it was a good thing to know 1143 00:43:25.300 --> 00:43:26.460 where their eyes are looking. 1144 00:43:26.680 --> 00:43:28.380 And Cog knows snakes can tell 1145 00:43:28.380 --> 00:43:30.200 whether a person is looking at them or not. 1146 00:43:30.520 --> 00:43:31.620 It's all through the animal. 1147 00:43:31.740 --> 00:43:31.940 It's all through the animal kingdom 1148 00:43:31.940 --> 00:43:34.600 across species knowing about eyes. 1149 00:43:34.840 --> 00:43:37.160 So here is, let me find the faces. 1150 00:43:38.020 --> 00:43:40.500 Here you see the wide angle view 1151 00:43:40.500 --> 00:43:41.900 in the left of those three images. 1152 00:43:42.440 --> 00:43:45.060 Then on the right is the foveal view 1153 00:43:45.060 --> 00:43:47.200 and in the middle is the software zoom 1154 00:43:47.200 --> 00:43:48.620 and you can maybe just see 1155 00:43:48.620 --> 00:43:50.740 the red dots being attached to his eyes. 1156 00:43:50.860 --> 00:43:52.220 Maybe the contrast isn't quite good enough 1157 00:43:52.220 --> 00:43:53.680 and the green dot being attached 1158 00:43:54.400 --> 00:43:56.760 to his face as he moves around. 1159 00:43:57.080 --> 00:43:58.700 It's figuring out where his eyes are 1160 00:43:58.700 --> 00:43:59.720 and figuring out where he's looking. 1161 00:44:00.440 --> 00:44:01.460 I'm going to skip over, 1162 00:44:01.740 --> 00:44:03.420 a couple of examples here 1163 00:44:03.420 --> 00:44:04.200 in the interest of time. 1164 00:44:04.280 --> 00:44:05.180 I'll just say one thing. 1165 00:44:06.180 --> 00:44:07.440 This particular example, 1166 00:44:08.480 --> 00:44:11.000 the robot is imitating his left side 1167 00:44:11.000 --> 00:44:11.700 if he looks left, 1168 00:44:11.780 --> 00:44:12.880 his right side if he looks right. 1169 00:44:12.980 --> 00:44:14.180 That mirror is just there 1170 00:44:14.180 --> 00:44:15.620 so that we can see what's happening. 1171 00:44:17.180 --> 00:44:19.160 But when we put the mirror in front of Cog 1172 00:44:19.160 --> 00:44:21.660 it thinks it sees itself in the mirror 1173 00:44:21.660 --> 00:44:23.220 and thinks itself is animate 1174 00:44:23.220 --> 00:44:24.720 and it starts imitating itself. 1175 00:44:25.060 --> 00:44:27.000 But it doesn't imitate itself very well 1176 00:44:27.000 --> 00:44:29.080 and so it sort of drifts over time 1177 00:44:29.080 --> 00:44:30.800 what the imitation is. 1178 00:44:30.860 --> 00:44:31.720 I'm going to skip over this. 1179 00:44:31.740 --> 00:44:32.480 Object imitation. 1180 00:44:33.120 --> 00:44:34.200 I'm going to skip over that one. 1181 00:44:34.400 --> 00:44:35.340 And now I want to show you 1182 00:44:35.340 --> 00:44:36.540 just where we are today, 1183 00:44:36.720 --> 00:44:37.940 the start of shared attention. 1184 00:44:38.980 --> 00:44:42.200 Here's Christy getting the robot's attention. 1185 00:44:43.220 --> 00:44:44.700 Notice Cog does not have 1186 00:44:44.700 --> 00:44:46.640 a human-like face as Kismet. 1187 00:44:46.760 --> 00:44:48.080 We haven't been concentrating on that. 1188 00:44:48.160 --> 00:44:50.160 She looks over at the block. 1189 00:44:50.300 --> 00:44:51.660 It figures out where she's looking 1190 00:44:51.660 --> 00:44:53.620 and it's curious and looks there too. 1191 00:44:54.300 --> 00:44:56.180 So this is the start. 1192 00:44:56.380 --> 00:44:59.000 It's not just something happening between them 1193 00:44:59.000 --> 00:44:59.520 but it's something, 1194 00:44:59.660 --> 00:45:00.720 a reference to something, 1195 00:45:00.720 --> 00:45:02.320 a third object in the world. 1196 00:45:02.460 --> 00:45:04.040 Now she reaches for that object 1197 00:45:04.040 --> 00:45:07.720 and Cog wants to reach for the same thing. 1198 00:45:07.980 --> 00:45:10.060 So we're starting to get the basis of shared attention. 1199 00:45:10.360 --> 00:45:13.440 Shared attention is something that 1200 00:45:13.440 --> 00:45:16.060 autistic children never really get. 1201 00:45:16.260 --> 00:45:17.960 Down syndrome children do get. 1202 00:45:18.180 --> 00:45:20.080 The great apes only have partially 1203 00:45:20.080 --> 00:45:23.500 and so it's a fairly sophisticated set of modules. 1204 00:45:24.100 --> 00:45:27.960 All this is built in terms of the same 1205 00:45:27.960 --> 00:45:30.560 sorts of little computational processes 1206 00:45:30.720 --> 00:45:32.020 that we built for the insects. 1207 00:45:32.200 --> 00:45:36.440 And out of that comes human-like behavior. 1208 00:45:36.740 --> 00:45:38.980 It's an open question how far that can be pushed. 1209 00:45:39.280 --> 00:45:41.320 I like to think that we'll be able to push it all the way 1210 00:45:41.320 --> 00:45:43.320 to full human level performance. 1211 00:45:44.000 --> 00:45:45.480 But maybe I'm wrong. 1212 00:45:46.020 --> 00:45:48.360 But there are two big questions that people ask 1213 00:45:48.360 --> 00:45:49.260 when they see this work. 1214 00:45:49.840 --> 00:45:53.080 Well if you really do make these robots truly intelligent, 1215 00:45:53.500 --> 00:45:54.680 will we accept them? 1216 00:45:54.900 --> 00:45:56.540 And I'll get back to that question in a minute. 1217 00:45:56.660 --> 00:45:58.740 And the other question is will they want to take over? 1218 00:45:59.160 --> 00:46:00.700 And tomorrow morning at 10 a.m. 1219 00:46:00.700 --> 00:46:03.080 Margaret Wertheim and I are going to be talking about 1220 00:46:03.080 --> 00:46:04.180 will they want to take over. 1221 00:46:04.280 --> 00:46:05.700 So I'm going to skip that question today. 1222 00:46:05.800 --> 00:46:09.060 My basic answer is not if we don't want them to. 1223 00:46:10.120 --> 00:46:11.760 And then in the afternoon session on 1224 00:46:11.760 --> 00:46:12.720 you ain't seen nothing yet. 1225 00:46:14.500 --> 00:46:16.340 Not if we don't want them to. 1226 00:46:16.420 --> 00:46:18.500 But there won't be any us for them to take over from 1227 00:46:18.500 --> 00:46:20.640 because we're going to change ourselves into robots anyway. 1228 00:46:20.900 --> 00:46:22.480 But that's tomorrow afternoon. 1229 00:46:23.340 --> 00:46:25.160 So the will we accept them question. 1230 00:46:25.260 --> 00:46:27.120 And this is a common question that you see in Hollywood. 1231 00:46:27.180 --> 00:46:30.220 Here's a bunch of movies that have addressed this question. 1232 00:46:30.340 --> 00:46:30.680 C3. 1233 00:46:30.700 --> 00:46:30.780 P. 1234 00:46:30.840 --> 00:46:31.020 O. 1235 00:46:31.060 --> 00:46:32.680 always thought he wasn't getting enough respect. 1236 00:46:33.120 --> 00:46:34.980 Data always wanted to be human. 1237 00:46:35.020 --> 00:46:36.840 And he wanted to get that emotion chip. 1238 00:46:37.200 --> 00:46:39.460 And then in Bicentennial Man down below, 1239 00:46:40.760 --> 00:46:44.440 Robin Williams wanted to be declared a person 1240 00:46:44.440 --> 00:46:45.580 by the United Nations. 1241 00:46:45.840 --> 00:46:49.540 And there's a new movie out from Spielberg called A.I. 1242 00:46:49.620 --> 00:46:51.480 which has somewhat the same theme. 1243 00:46:51.540 --> 00:46:53.180 It's really a retell of Pinocchio. 1244 00:46:58.160 --> 00:46:59.000 Sherry Turkle, 1245 00:46:59.000 --> 00:47:01.300 who has long been a critic of A.I. 1246 00:47:01.320 --> 00:47:04.360 when she came to my lab a few years ago, 1247 00:47:04.440 --> 00:47:07.540 and she's now doing experiments in my lab with the robots actually, 1248 00:47:08.960 --> 00:47:11.100 was rather surprised by Cog, 1249 00:47:11.240 --> 00:47:15.600 by Cog making eye contact with her. 1250 00:47:15.860 --> 00:47:18.200 And she wasn't surprised, as she says down below, 1251 00:47:18.360 --> 00:47:20.300 by anything Cog was really capable of. 1252 00:47:20.340 --> 00:47:23.000 She was surprised by her own reaction to Cog. 1253 00:47:23.020 --> 00:47:26.300 That she found herself wanting to catch its attention. 1254 00:47:26.300 --> 00:47:31.320 That she had acted as though in the presence of another being. 1255 00:47:31.760 --> 00:47:35.320 And we've certainly seen that with people we've brought into the lab. 1256 00:47:35.420 --> 00:47:37.440 They act as though they're in the presence of another being. 1257 00:47:37.580 --> 00:47:41.980 And the students who build the robot find themselves continually acting 1258 00:47:41.980 --> 00:47:43.820 as though they're in the presence of another being. 1259 00:47:43.900 --> 00:47:46.640 When Brian Scazzolatti is working on coding, 1260 00:47:48.120 --> 00:47:50.320 and Cog is sitting there looking around at things, 1261 00:47:50.500 --> 00:47:53.700 he eventually has to switch Cog off while he's working on something else. 1262 00:47:53.820 --> 00:47:56.020 Because when Cog looks at something, 1263 00:47:56.020 --> 00:47:58.540 he finds himself impelled to look at it too. 1264 00:47:58.660 --> 00:48:01.800 Even though he knows, so it triggers this stuff in him, 1265 00:48:01.840 --> 00:48:02.360 and he switches it off. 1266 00:48:02.720 --> 00:48:05.740 Cynthia, who built Kismet, will be interacting, 1267 00:48:05.900 --> 00:48:07.480 will be sitting, working elsewhere in the room, 1268 00:48:07.560 --> 00:48:09.880 Kismet's looking around, and suddenly another student 1269 00:48:09.880 --> 00:48:12.380 somewhere else in the lab, connected via telnet, 1270 00:48:12.840 --> 00:48:15.400 or over X-windows to one of the modules, 1271 00:48:15.960 --> 00:48:18.940 reboots the system, and when Kismet gets rebooted, 1272 00:48:19.060 --> 00:48:21.920 it first goes through this system where it goes to all the extremes, 1273 00:48:22.140 --> 00:48:25.360 in its head position, like this, 1274 00:48:25.360 --> 00:48:28.200 and then it goes to something much more inanimately than I just did, 1275 00:48:28.320 --> 00:48:32.580 and she finds herself completely taken every time. 1276 00:48:32.640 --> 00:48:34.900 What's happened? This being has suddenly turned into a robot. 1277 00:48:35.020 --> 00:48:36.860 She knows that it's a robot, she built the thing, 1278 00:48:37.020 --> 00:48:39.080 but it gets into her subconscious, 1279 00:48:39.460 --> 00:48:41.700 and she can't stop herself from thinking that. 1280 00:48:42.520 --> 00:48:46.540 Sherry Turkle, likewise, found herself treating the robot as a being. 1281 00:48:46.720 --> 00:48:51.620 Will we treat the robots as they are beings for more than a few seconds, though? 1282 00:48:51.660 --> 00:48:55.240 Will we ultimately want to say that, 1283 00:48:56.100 --> 00:48:57.880 Kismet has real emotions? 1284 00:48:58.840 --> 00:49:00.580 Does Kismet have real emotions? 1285 00:49:00.700 --> 00:49:02.800 It's just a computer program, after all. 1286 00:49:02.840 --> 00:49:04.700 It happens to be running on 24 PCs, 1287 00:49:04.740 --> 00:49:07.060 but it's just a computer program. 1288 00:49:07.280 --> 00:49:08.320 Could that be real emotions? 1289 00:49:09.160 --> 00:49:10.600 We have real emotions. 1290 00:49:10.900 --> 00:49:12.560 A robot can't have real emotions. 1291 00:49:13.500 --> 00:49:14.680 You know, a simple question. 1292 00:49:14.940 --> 00:49:16.020 Can a robot be afraid? 1293 00:49:16.520 --> 00:49:19.600 You know, I think most artificial intelligence researchers right now 1294 00:49:20.240 --> 00:49:23.180 would be willing to say that programs can reason about facts, 1295 00:49:23.380 --> 00:49:24.360 can make decisions, 1296 00:49:24.360 --> 00:49:25.620 can have goals, 1297 00:49:25.860 --> 00:49:27.960 but I think most AI researchers, 1298 00:49:29.220 --> 00:49:32.700 while they'd be willing to say that robots can act as if they're afraid, 1299 00:49:32.880 --> 00:49:33.980 can seem to be afraid, 1300 00:49:34.060 --> 00:49:35.260 and simulate having fear, 1301 00:49:35.320 --> 00:49:36.680 or simulate having emotions, 1302 00:49:37.240 --> 00:49:40.080 they're not willing to say that a robot can be viscerally afraid 1303 00:49:40.080 --> 00:49:41.740 in the same way people are. 1304 00:49:43.960 --> 00:49:44.940 But why not? 1305 00:49:45.680 --> 00:49:48.380 Is there something different in kind, ultimately, 1306 00:49:48.920 --> 00:49:51.120 between these robots that are machines, 1307 00:49:51.360 --> 00:49:52.820 and us humans? 1308 00:49:52.820 --> 00:49:54.800 My hypothesis is there's not. 1309 00:49:55.460 --> 00:49:58.660 And I think that a lot of our reactions, 1310 00:49:58.760 --> 00:49:59.820 like, no, they're just machines, 1311 00:50:00.900 --> 00:50:01.760 they're just machines, 1312 00:50:02.100 --> 00:50:05.280 is the same sort of thing that we've gone through many times 1313 00:50:05.280 --> 00:50:06.420 over the last few hundred years. 1314 00:50:07.140 --> 00:50:09.600 Humans have always thought of themselves as special. 1315 00:50:10.260 --> 00:50:11.460 Five hundred years ago, 1316 00:50:12.680 --> 00:50:15.900 the Catholic Church had a big argument 1317 00:50:15.900 --> 00:50:18.320 about whether Earth was the center of the universe, 1318 00:50:18.540 --> 00:50:20.840 and humans ultimately had to give up 1319 00:50:20.840 --> 00:50:22.800 the notion that the Earth was the center of the universe. 1320 00:50:22.820 --> 00:50:24.800 We weren't so special, after all. 1321 00:50:27.480 --> 00:50:29.420 With Darwin, in the 19th century, 1322 00:50:29.580 --> 00:50:31.480 humans and animals had common ancestors. 1323 00:50:31.700 --> 00:50:35.320 We weren't different from animals, after all. 1324 00:50:35.380 --> 00:50:36.240 And so on over time, 1325 00:50:36.400 --> 00:50:39.080 we've had to give up more and more degrees of specialness. 1326 00:50:39.160 --> 00:50:40.080 Once we discovered DNA, 1327 00:50:40.400 --> 00:50:41.820 we found we had a lot of genes 1328 00:50:42.520 --> 00:50:44.020 pretty similar to yeast genes. 1329 00:50:45.700 --> 00:50:48.440 More recently, we found that human flesh and body plans 1330 00:50:48.440 --> 00:50:51.560 can be subject to technological manipulation. 1331 00:50:51.560 --> 00:50:53.660 You can mess with the homeobox genes, 1332 00:50:53.920 --> 00:50:56.200 which are pretty much the same as the fruit fly homeobox genes, 1333 00:50:56.340 --> 00:50:57.060 to do that. 1334 00:50:57.560 --> 00:50:59.500 And very recently, we've had a couple of shocks. 1335 00:51:00.800 --> 00:51:03.460 You know, people get really upset 1336 00:51:03.460 --> 00:51:05.860 that maybe we didn't have our own creation event here on Earth. 1337 00:51:05.940 --> 00:51:08.280 Maybe life came from Mars via a meteorite. 1338 00:51:08.300 --> 00:51:10.840 Well, gee, we thought we, you know, 1339 00:51:11.280 --> 00:51:15.080 started here from random chemical interactions here on Earth, 1340 00:51:15.140 --> 00:51:16.860 but now they're telling us it came from another planet. 1341 00:51:16.980 --> 00:51:17.720 That's not so great. 1342 00:51:18.560 --> 00:51:19.740 And even more recently, 1343 00:51:19.880 --> 00:51:21.540 it turns out we maybe only have a life 1344 00:51:21.560 --> 00:51:22.900 of lousy 35,000 genes. 1345 00:51:23.240 --> 00:51:24.740 Potatoes can do much better than that. 1346 00:51:24.920 --> 00:51:29.840 But, you know, this sort of gets at our sense of specialness. 1347 00:51:29.860 --> 00:51:34.500 So we're not that complex as we may have thought. 1348 00:51:34.920 --> 00:51:36.840 And I think the same sorts of things, 1349 00:51:37.460 --> 00:51:39.880 and maybe we can talk about this in the questions a little, 1350 00:51:41.460 --> 00:51:45.020 whether we're ever willing to say that robots can have real emotions. 1351 00:51:45.400 --> 00:51:50.520 When Garry Kasparov, the world chess champion, 1352 00:51:50.520 --> 00:51:53.540 was beaten by Deep Blue about three years ago, 1353 00:51:54.580 --> 00:51:56.920 after years and years of people saying 1354 00:51:56.920 --> 00:51:59.540 the chess champion will never be beaten by a computer, 1355 00:51:59.700 --> 00:52:00.960 he was finally beaten by a computer, 1356 00:52:01.100 --> 00:52:04.440 and he said, well, at least it didn't enjoy beating me. 1357 00:52:04.640 --> 00:52:06.760 So that was his retreat into specialness. 1358 00:52:06.920 --> 00:52:09.940 Well, it can beat me, but it didn't, you know, no pleasure. 1359 00:52:10.720 --> 00:52:13.340 So, you know, we don't want to give up that specialness. 1360 00:52:13.420 --> 00:52:14.580 That's what's left for us now. 1361 00:52:14.780 --> 00:52:15.980 We are emotional beings. 1362 00:52:16.360 --> 00:52:18.680 Everything else, machines and other stuff, can do. 1363 00:52:18.680 --> 00:52:23.320 And I think that's going to take a while for us to overcome 1364 00:52:23.320 --> 00:52:28.820 what I see as a very similar, in kind, sort of retreat from specialness, 1365 00:52:28.960 --> 00:52:32.720 which I think is part of what's going on in the other hall here 1366 00:52:32.720 --> 00:52:35.200 with some of the arguments with the church still 1367 00:52:35.200 --> 00:52:38.380 and how we view ourselves. 1368 00:52:38.560 --> 00:52:42.500 And tomorrow night we'll be talking about science and religion 1369 00:52:42.500 --> 00:52:43.380 and those aspects. 1370 00:52:45.040 --> 00:52:46.560 So with that, I think I'll close, 1371 00:52:46.620 --> 00:52:48.340 and we'll have time for a few questions. 1372 00:52:48.680 --> 00:52:51.720 Which Tim will handle and direct, I guess. 1373 00:52:51.740 --> 00:52:52.140 Thank you. 1374 00:53:02.860 --> 00:53:05.300 Thank you very much, Rodney, on behalf of us all. 1375 00:53:05.340 --> 00:53:07.740 I think that was just stunning, just fantastic. 1376 00:53:08.040 --> 00:53:10.880 We've got time for probably two or three very quick questions. 1377 00:53:10.920 --> 00:53:13.000 So if anyone has any burning issues, yes. 1378 00:53:13.120 --> 00:53:16.300 There's a microphone right next to you there you can stand up and speak into 1379 00:53:16.300 --> 00:53:17.060 so we can all hear. 1380 00:53:17.320 --> 00:53:18.040 That would be great. 1381 00:53:20.220 --> 00:53:20.960 Thank you. 1382 00:53:21.040 --> 00:53:21.520 Yeah. 1383 00:53:22.520 --> 00:53:25.880 This, would you then say that the sum of what we are, 1384 00:53:26.040 --> 00:53:27.580 or, can you hear me? 1385 00:53:27.840 --> 00:53:28.360 Yes, I can. 1386 00:53:28.700 --> 00:53:29.320 Just about. 1387 00:53:29.420 --> 00:53:34.780 The sum of what we are, then, is the sum of all these primitive reactions, 1388 00:53:35.320 --> 00:53:38.880 and is this driving you to remodel what you think of as mind? 1389 00:53:39.240 --> 00:53:42.780 Can you say anything about this, this concept of mind, 1390 00:53:42.920 --> 00:53:46.340 which is also getting a lot of attention recently? 1391 00:53:47.040 --> 00:53:47.520 Yeah. 1392 00:53:47.520 --> 00:53:50.840 Well, my hypothesis, and this is a hypothesis, 1393 00:53:51.000 --> 00:53:54.420 my hypothesis is that we are machines. 1394 00:53:54.540 --> 00:53:55.740 That is all we are. 1395 00:53:55.800 --> 00:53:56.740 We're biochemical machines. 1396 00:53:56.840 --> 00:54:00.320 And that's essentially the unsaid hypothesis of molecular biology, 1397 00:54:00.520 --> 00:54:05.500 that we are machines where the biomolecules interact according to laws. 1398 00:54:05.700 --> 00:54:09.000 So out of that emerges what we call mind, 1399 00:54:09.120 --> 00:54:14.280 happens to be an emergent property of our brain there. 1400 00:54:14.460 --> 00:54:17.500 So my position is that it does not exist. 1401 00:54:17.500 --> 00:54:20.080 It doesn't necessarily need to be built out of the same matter, 1402 00:54:20.220 --> 00:54:21.540 the same biomolecules. 1403 00:54:21.740 --> 00:54:24.680 If we get the right abstractions, 1404 00:54:25.560 --> 00:54:30.700 then maybe we can build mind-like things out of simple mechanical devices, 1405 00:54:31.240 --> 00:54:33.120 electromechanical devices, silicon and steel, 1406 00:54:33.320 --> 00:54:36.360 out of which emerges mind in the same way. 1407 00:54:36.980 --> 00:54:38.600 The analogy would be with flight. 1408 00:54:38.760 --> 00:54:40.220 We see birds, how they fly. 1409 00:54:40.420 --> 00:54:42.140 We understand the laws of aerodynamics. 1410 00:54:42.260 --> 00:54:45.840 We build wings with the same aerodynamic laws. 1411 00:54:46.000 --> 00:54:46.920 They operate a little differently, 1412 00:54:46.920 --> 00:54:48.840 but we get flight out of that. 1413 00:54:48.860 --> 00:54:50.340 The planes are not simulating flying. 1414 00:54:50.500 --> 00:54:52.960 They are flying in a somewhat different manner. 1415 00:54:53.420 --> 00:54:57.280 I think we can ultimately get mind out of these sorts of machines 1416 00:54:57.280 --> 00:54:59.120 from lots of little interactions. 1417 00:55:01.160 --> 00:55:05.840 And philosophy often has gone into this regress of the homunculus, 1418 00:55:05.920 --> 00:55:08.020 but it never gets to the nub. 1419 00:55:08.060 --> 00:55:10.500 And I think it never gets to the nub because there is no nub. 1420 00:55:12.340 --> 00:55:13.500 It's a hypothesis though. 1421 00:55:13.860 --> 00:55:14.480 Yes, please. 1422 00:55:14.720 --> 00:55:16.120 Would you care to use a microphone? 1423 00:55:16.220 --> 00:55:16.900 At the back. 1424 00:55:16.920 --> 00:55:18.020 Or you can shout. 1425 00:55:18.140 --> 00:55:19.160 You've got two choices. 1426 00:55:21.140 --> 00:55:21.660 Loudly. 1427 00:55:22.040 --> 00:55:23.060 My name is Rebecca, 1428 00:55:23.320 --> 00:55:25.780 and perhaps you could make a connection for me. 1429 00:55:25.800 --> 00:55:29.440 At the beginning, you spoke about NASA and intuition, 1430 00:55:35.060 --> 00:55:37.240 intuitively moving the robot. 1431 00:55:37.500 --> 00:55:37.800 Yes. 1432 00:55:37.980 --> 00:55:42.540 And making a connection with touch, please. 1433 00:55:43.340 --> 00:55:43.860 Okay. 1434 00:55:43.920 --> 00:55:45.060 So, at the beginning, 1435 00:55:45.100 --> 00:55:46.900 I talked about the NASA humanoid robot, 1436 00:55:46.920 --> 00:55:49.580 which is a robot that is not scheduled to fly any time soon, 1437 00:55:49.700 --> 00:55:50.920 but they're doing research and development. 1438 00:55:51.160 --> 00:55:53.680 The idea would be that it's outside the space station. 1439 00:55:53.860 --> 00:55:57.080 The astronauts are inside, 1440 00:55:57.240 --> 00:56:01.180 and they're teleoperating and controlling the robot outside. 1441 00:56:01.840 --> 00:56:04.280 It turns out that force feedback and touch, 1442 00:56:04.480 --> 00:56:09.160 not necessarily touch in the same way of being able to feel things 1443 00:56:09.160 --> 00:56:10.260 in different parts of the skin, 1444 00:56:10.360 --> 00:56:11.340 but force feedback, 1445 00:56:11.540 --> 00:56:13.880 feeling that something is pushing back on you, 1446 00:56:13.940 --> 00:56:15.020 that gross force, 1447 00:56:15.140 --> 00:56:16.440 turns out to be very important. 1448 00:56:16.440 --> 00:56:19.060 It's very important for us to control machines remotely. 1449 00:56:21.840 --> 00:56:22.880 It seems, 1450 00:56:23.260 --> 00:56:26.120 and I'd point you to books by Ramachandran, 1451 00:56:27.020 --> 00:56:28.700 V.K. Ramachandran on this, 1452 00:56:28.840 --> 00:56:31.340 it seems that that sense of force feedback 1453 00:56:31.340 --> 00:56:34.380 is important for our own understanding of our own bodily self. 1454 00:56:34.680 --> 00:56:37.120 And so there seems to be something fairly fundamental there. 1455 00:56:37.260 --> 00:56:39.900 But I must say that in all the robots I showed you today, 1456 00:56:40.080 --> 00:56:41.720 we haven't addressed that question. 1457 00:56:41.880 --> 00:56:43.580 Not because we think it's unimportant, 1458 00:56:43.700 --> 00:56:46.420 just we haven't gone there yet. 1459 00:56:46.440 --> 00:56:48.360 And I think that is rather important. 1460 00:56:48.740 --> 00:56:50.700 I think we've got time for one final question. 1461 00:56:51.600 --> 00:56:53.560 Perhaps you were the first one to put up your hand. 1462 00:56:53.740 --> 00:56:56.720 Do you think a robot will ever be able to experience pain? 1463 00:56:57.520 --> 00:56:59.320 The question for those of you who may not have heard, 1464 00:56:59.420 --> 00:57:01.820 do I think robots will ever be able to experience pain, 1465 00:57:01.920 --> 00:57:03.700 I think that's exactly the same question 1466 00:57:03.700 --> 00:57:07.400 as do we think robots will ever be able to have true emotions, 1467 00:57:07.600 --> 00:57:08.720 visceral emotions. 1468 00:57:10.680 --> 00:57:15.680 And I think an issue here is, 1469 00:57:15.680 --> 00:57:18.820 when we understand mechanistically how the robot works, 1470 00:57:19.040 --> 00:57:23.540 we're unwilling to say that's the real stuff, 1471 00:57:24.020 --> 00:57:25.100 that that's really pain. 1472 00:57:25.180 --> 00:57:27.620 No, that's just this electrical connection down here, 1473 00:57:27.840 --> 00:57:29.860 and then this register adds a number here, 1474 00:57:29.920 --> 00:57:31.680 and it compares it with this threshold over here. 1475 00:57:31.820 --> 00:57:32.720 That's not pain. 1476 00:57:32.880 --> 00:57:35.060 Pain is the stuff I feel. 1477 00:57:35.600 --> 00:57:36.360 I'm special. 1478 00:57:36.900 --> 00:57:39.000 I'm not just a bunch of machines and registers. 1479 00:57:39.380 --> 00:57:41.040 So I think ultimately it will, 1480 00:57:41.180 --> 00:57:42.840 but it's going to take a different view of ourselves, 1481 00:57:42.960 --> 00:57:44.660 and I think that different view of ourselves 1482 00:57:44.660 --> 00:57:47.000 is coming along from molecular biology 1483 00:57:47.000 --> 00:57:48.880 as we understand more and more of the mechanisms, 1484 00:57:48.980 --> 00:57:51.220 as we understand them, the mechanisms of pain. 1485 00:57:51.660 --> 00:57:54.580 Our intuition of what's going on inside us 1486 00:57:54.580 --> 00:57:57.120 is completely wrong, 1487 00:57:57.260 --> 00:57:59.180 as we see by those split brain experiments. 1488 00:57:59.380 --> 00:58:01.560 We just make stuff up. We rationalize. 1489 00:58:01.600 --> 00:58:03.820 As we understand the mechanisms of ourselves, 1490 00:58:04.120 --> 00:58:07.380 our views as humans of ourselves is going to change, 1491 00:58:07.500 --> 00:58:10.580 and ultimately I think we will attribute real pain 1492 00:58:10.580 --> 00:58:12.660 to certain robots. 1493 00:58:12.720 --> 00:58:14.400 Not maybe to the robots I've shown you today, 1494 00:58:14.660 --> 00:58:16.520 maybe they'll always be too simple, 1495 00:58:16.540 --> 00:58:18.200 but more complex robots in the future. 1496 00:58:18.920 --> 00:58:21.660 I think it's probably about time to wrap up now, 1497 00:58:21.840 --> 00:58:24.120 but I'd just like to add one observation, Rodney, 1498 00:58:24.160 --> 00:58:24.860 that fascinated me. 1499 00:58:25.040 --> 00:58:27.520 Whether we'll end up developing a psychopathic, 1500 00:58:27.520 --> 00:58:30.220 hell-like robot or not, 1501 00:58:30.280 --> 00:58:32.460 I think is going to be very much determined 1502 00:58:32.460 --> 00:58:35.220 by the evolutionary pressures we place these machines under. 1503 00:58:35.660 --> 00:58:37.580 And a good analog might be the way 1504 00:58:37.580 --> 00:58:39.620 that we've used human slaves in the past. 1505 00:58:40.040 --> 00:58:42.300 Maybe Spartacus does have something to teach us 1506 00:58:42.300 --> 00:58:43.220 in the long term. 1507 00:58:43.540 --> 00:58:45.560 Thank you very, very much, Rodney, for that, 1508 00:58:45.620 --> 00:58:46.800 and thank you all for being here today.