WEBVTT 00:00:01.640 --> 00:00:04.040 The following content is provided under a Creative 00:00:04.040 --> 00:00:05.580 Commons license. 00:00:05.580 --> 00:00:07.880 Your support will help MIT OpenCourseWare 00:00:07.880 --> 00:00:12.270 continue to offer high-quality educational resources for free. 00:00:12.270 --> 00:00:14.870 To make a donation or view additional materials 00:00:14.870 --> 00:00:18.830 from hundreds of MIT courses, visit MIT OpenCourseWare 00:00:18.830 --> 00:00:20.000 at ocw.mit.edu. 00:00:22.552 --> 00:00:24.010 TONY PRESCOTT: Great pleasure to be 00:00:24.010 --> 00:00:26.790 in Woods Hole, my first visit here, 00:00:26.790 --> 00:00:29.390 had a wonderful swim in the sea yesterday. 00:00:29.390 --> 00:00:33.790 Sheffield Robotics is across both universities in Sheffield. 00:00:33.790 --> 00:00:36.130 And it's been founded since 2011, 00:00:36.130 --> 00:00:39.490 but we've really been doing robotics since the 1980s. 00:00:39.490 --> 00:00:41.950 And I joined them in 1989. 00:00:41.950 --> 00:00:45.570 And we do pretty much every different kind of robotics, 00:00:45.570 --> 00:00:47.740 but I'm going to talk about biomimetics. 00:00:47.740 --> 00:00:50.650 I also do what you might call cognitive robotics. 00:00:50.650 --> 00:00:53.000 And I collaborate with Giorgio Metta on that, 00:00:53.000 --> 00:00:54.760 but since he's speaking too, I'm going 00:00:54.760 --> 00:00:57.430 to focus on the more animal-like robots 00:00:57.430 --> 00:00:59.680 that we've been developing. 00:00:59.680 --> 00:01:02.890 So this is one of our latest projects. 00:01:02.890 --> 00:01:08.260 This is a small, autonomous, mobile robot, 00:01:08.260 --> 00:01:10.330 which is a commercial platform, which 00:01:10.330 --> 00:01:13.840 will be available in the UK, I think, next January. 00:01:13.840 --> 00:01:16.570 And there will hopefully be a developer program 00:01:16.570 --> 00:01:19.330 for people that are interested in helping 00:01:19.330 --> 00:01:23.270 to develop the intelligence for this robot. 00:01:23.270 --> 00:01:27.310 So this is at a conference we had in Barcelona last month. 00:01:27.310 --> 00:01:31.000 And you can see that it's a robot pet. 00:01:31.000 --> 00:01:34.720 And we've been focusing on giving it 00:01:34.720 --> 00:01:38.650 some effective communication abilities, 00:01:38.650 --> 00:01:40.870 responding particularly to touch. 00:01:40.870 --> 00:01:44.990 You can see that it's orienting to stimuli. 00:01:44.990 --> 00:01:46.135 It has stereo vision. 00:01:46.135 --> 00:01:47.470 It has stereo sound. 00:01:47.470 --> 00:01:51.490 And it can orient to visual stimuli 00:01:51.490 --> 00:01:53.400 and also to auditory stimuli. 00:01:53.400 --> 00:01:55.640 Here we're showing it a picture of itself 00:01:55.640 --> 00:01:57.730 on this magazine cover. 00:01:57.730 --> 00:02:01.550 And the goal is to demonstrate that we can, 00:02:01.550 --> 00:02:05.320 in a commercial robot that will cost less than $1,000, 00:02:05.320 --> 00:02:08.770 considerably less, some of the principles of how 00:02:08.770 --> 00:02:12.140 the brain generates behavior. 00:02:12.140 --> 00:02:14.500 So this robot is called MiRo. 00:02:14.500 --> 00:02:19.360 It is based on some high-level principles abstracted 00:02:19.360 --> 00:02:22.450 from what we know about how mammalian brains control 00:02:22.450 --> 00:02:26.200 behavior, so it's a relatively complex robot, 13 00:02:26.200 --> 00:02:29.440 degrees of freedom, 3 arm processes corresponding 00:02:29.440 --> 00:02:31.360 to different levels, if you like, 00:02:31.360 --> 00:02:35.810 of the neuroaxis, the central nervous system. 00:02:35.810 --> 00:02:41.380 So once you start out with some general ideas, and questions, 00:02:41.380 --> 00:02:44.260 and issues about how we might learn 00:02:44.260 --> 00:02:47.890 from the biology in the brain, how to develop robots, 00:02:47.890 --> 00:02:51.730 and how we might use robots to help us understand the brain. 00:02:51.730 --> 00:02:54.880 And a central question I think that robotics 00:02:54.880 --> 00:02:58.170 can help us answer, and I think that's a core question 00:02:58.170 --> 00:03:01.510 in neuroscience, is what you might call the problem 00:03:01.510 --> 00:03:02.980 of behavioral integration. 00:03:02.980 --> 00:03:05.650 And the neuroscientist Ernest Barrington 00:03:05.650 --> 00:03:07.090 summarized this quite nicely. 00:03:07.090 --> 00:03:10.120 He said, "the phenomenon so characteristic of living 00:03:10.120 --> 00:03:13.660 organisms, so very difficult to analyze, the fact 00:03:13.660 --> 00:03:16.840 that they behave as wholes rather than as the sum 00:03:16.840 --> 00:03:18.700 of their constituent parts. 00:03:18.700 --> 00:03:21.090 Their behavior shows integration, 00:03:21.090 --> 00:03:23.820 a process unifying the actions of an organism 00:03:23.820 --> 00:03:27.140 into patterns that involve the whole individual." 00:03:27.140 --> 00:03:29.500 And this picture of a squirrel over here I think nicely 00:03:29.500 --> 00:03:30.260 demonstrates this. 00:03:30.260 --> 00:03:32.980 So of course this squirrel is leaping from one branch 00:03:32.980 --> 00:03:34.030 to the next. 00:03:34.030 --> 00:03:36.370 And you can see that every part of his body 00:03:36.370 --> 00:03:39.640 is coordinated and organized for this action, 00:03:39.640 --> 00:03:42.430 so you can see that his eyes looking straight ahead, 00:03:42.430 --> 00:03:44.830 his whiskers-- and I'll talk a lot more about whiskers-- 00:03:44.830 --> 00:03:46.250 pointing forward. 00:03:46.250 --> 00:03:49.930 His arms and his feet are all ready 00:03:49.930 --> 00:03:53.140 and they're there ready to catch his fall. 00:03:53.140 --> 00:03:55.840 Even his tail is angled and positioned 00:03:55.840 --> 00:03:57.670 to help him fly through the air. 00:03:57.670 --> 00:04:00.820 So it's the coordination of the different parts of the body, 00:04:00.820 --> 00:04:03.130 and the multiple degrees of freedom of the body, 00:04:03.130 --> 00:04:07.060 and the sensory systems in space and in time, which, I think, 00:04:07.060 --> 00:04:11.170 is a critical problem for biological control 00:04:11.170 --> 00:04:14.200 and also a problem for robots, which we're still struggling 00:04:14.200 --> 00:04:16.290 to address with our robots. 00:04:16.290 --> 00:04:19.209 And I want to give you two very general principles 00:04:19.209 --> 00:04:23.870 for thinking about how brains solve this problem. 00:04:23.870 --> 00:04:27.700 So many of you will have come across Rodney Brooks, 00:04:27.700 --> 00:04:29.710 yes, from MIT. 00:04:29.710 --> 00:04:32.020 And he's famous for, in robotics, 00:04:32.020 --> 00:04:36.340 the notion of layered control, which he called subsumption. 00:04:36.340 --> 00:04:39.370 And I think the ideas that he brought 00:04:39.370 --> 00:04:41.740 into robotics really changed how people thought 00:04:41.740 --> 00:04:44.020 about robots in the 1980s. 00:04:44.020 --> 00:04:49.060 But if we go back to the 1880s, John Hughlings Jackson, 00:04:49.060 --> 00:04:52.950 who was a British neurologist, proposed a similar idea, 00:04:52.950 --> 00:04:56.270 but with respect to the nervous system. 00:04:56.270 --> 00:05:01.030 So in the 1880s, people thought about the higher areas 00:05:01.030 --> 00:05:04.510 of the brain, particularly the cortex, 00:05:04.510 --> 00:05:08.710 as being about higher thought, and reasoning, and language, 00:05:08.710 --> 00:05:11.460 and not so much about perception and action. 00:05:11.460 --> 00:05:14.500 And Hughlings Jackson, I think, was revolutionary in his day 00:05:14.500 --> 00:05:17.770 in saying, that the highest motor senses represent over 00:05:17.770 --> 00:05:20.110 again in more complex combinations what 00:05:20.110 --> 00:05:22.072 the middle motor centers represent. 00:05:22.072 --> 00:05:24.530 In other words, he was saying that, the whole of the brain, 00:05:24.530 --> 00:05:27.470 all the way up, is about coordinating perception 00:05:27.470 --> 00:05:28.490 with action. 00:05:28.490 --> 00:05:31.460 And he described it in many senses as a layered system. 00:05:31.460 --> 00:05:33.080 He talked about how you could take off 00:05:33.080 --> 00:05:37.190 the top layers of the system and the competences of the lower 00:05:37.190 --> 00:05:39.440 layers remained intact, which, of course, is 00:05:39.440 --> 00:05:41.930 very much the idea of Rodney Brooks subsumption 00:05:41.930 --> 00:05:43.580 architecture. 00:05:43.580 --> 00:05:47.660 And some old studies, they did transection in animals 00:05:47.660 --> 00:05:49.805 like cats and rats-- demonstrate this nicely. 00:05:49.805 --> 00:05:53.120 So if you take a cat or a rat, particularly a rat, 00:05:53.120 --> 00:05:56.245 and you remove, in fact, all of the cerebral cortex, 00:05:56.245 --> 00:06:00.530 so if you make a slice here that takes away cortex, 00:06:00.530 --> 00:06:03.110 you get an animal that actually, to all appearances, 00:06:03.110 --> 00:06:04.280 looks fairly normal. 00:06:04.280 --> 00:06:06.660 It does motivated behavioral sequences, 00:06:06.660 --> 00:06:07.737 so it will get hungry. 00:06:07.737 --> 00:06:08.820 And it will look for food. 00:06:08.820 --> 00:06:10.220 And it will eat. 00:06:10.220 --> 00:06:12.200 If there's an appropriate mate nearby, 00:06:12.200 --> 00:06:16.220 it will look to have a sexual relationship. 00:06:16.220 --> 00:06:21.050 And it will fail in some challenges such as learning, 00:06:21.050 --> 00:06:25.000 and perhaps also in dexterous control, but in many ways, 00:06:25.000 --> 00:06:26.910 it will look normal. 00:06:26.910 --> 00:06:31.910 If you slice below the other part of the forebrain, 00:06:31.910 --> 00:06:34.610 the thalamus and the hypothalamus, 00:06:34.610 --> 00:06:37.100 you remove these areas, then you remove this capacity 00:06:37.100 --> 00:06:41.270 for motivated behavior, but you leave intact midbrain systems 00:06:41.270 --> 00:06:43.820 that can still generate individual actions. 00:06:43.820 --> 00:06:46.460 And if you remove parts of the midbrain, 00:06:46.460 --> 00:06:48.710 you leave intact, still, component movements, 00:06:48.710 --> 00:06:51.890 so for example, animals that can run on a treadmill. 00:06:51.890 --> 00:06:57.800 So we are, with our MiRo robot, loosely recapitulating 00:06:57.800 --> 00:07:01.070 this architecture, so we have three processors. 00:07:01.070 --> 00:07:03.720 And the idea with this robot, it's actually a part work, 00:07:03.720 --> 00:07:04.760 so you build it up. 00:07:04.760 --> 00:07:07.790 You get a magazine every week with a new part for the robot. 00:07:07.790 --> 00:07:10.580 And you build, essentially, a spinal robot first. 00:07:10.580 --> 00:07:12.620 And then you add a midbrain processor. 00:07:12.620 --> 00:07:15.500 Eventually you add a cortical processor, 00:07:15.500 --> 00:07:18.470 which gives with it some learning capacities, 00:07:18.470 --> 00:07:22.280 some pattern recognition, some navigation, so that's 00:07:22.280 --> 00:07:24.800 one principle, layered architecture, which 00:07:24.800 --> 00:07:29.360 seems to work both for biology, and perhaps in robotics. 00:07:29.360 --> 00:07:32.600 So second principle, and this goes back 00:07:32.600 --> 00:07:35.270 to another famous neuroscientist, 00:07:35.270 --> 00:07:39.050 this time Wilder Penfield, who is known to many people 00:07:39.050 --> 00:07:42.770 for his discovery of somatotopic maps in the brain, 00:07:42.770 --> 00:07:46.960 that if you stimulate, in the brain, in the area, 00:07:46.960 --> 00:07:49.940 the sensory area, then you find that people 00:07:49.940 --> 00:07:53.240 have experience of tickling on parts of the body. 00:07:53.240 --> 00:07:56.330 And adjacent parts of cortex correspond 00:07:56.330 --> 00:07:58.100 to adjacent parts of the body. 00:07:58.100 --> 00:08:00.830 And he found a similar homunculus in the motor area 00:08:00.830 --> 00:08:03.110 that you stimulate, and you get movement 00:08:03.110 --> 00:08:05.190 in adjacent parts of the body. 00:08:05.190 --> 00:08:08.650 And he also proposed another idea. 00:08:08.650 --> 00:08:13.040 And that was sort of a transencephalic dimension 00:08:13.040 --> 00:08:14.870 to nervous system organization. 00:08:14.870 --> 00:08:17.330 And that's to say that, down the midline 00:08:17.330 --> 00:08:19.330 of the central nervous system, there 00:08:19.330 --> 00:08:22.430 are a group of structures that don't seem to be specifically 00:08:22.430 --> 00:08:26.690 involved in specific aspects of perception and action, 00:08:26.690 --> 00:08:28.940 but seem to be about integration. 00:08:28.940 --> 00:08:32.010 And amongst them, particularly the basal ganglia, 00:08:32.010 --> 00:08:33.919 he noted as being important, and parts 00:08:33.919 --> 00:08:35.809 of the reticular formation. 00:08:35.809 --> 00:08:38.690 So Michael Frank was here talking to you 00:08:38.690 --> 00:08:41.330 about basal ganglia, so I'm not going 00:08:41.330 --> 00:08:45.350 to say much more about this, but this is just to point out, 00:08:45.350 --> 00:08:47.870 in an slice in the rat brain, these 00:08:47.870 --> 00:08:51.320 are the elements of the basal ganglia, particularly, 00:08:51.320 --> 00:08:53.720 the striatum is the input system. 00:08:53.720 --> 00:08:57.210 In the rat, the substantia nigra and part of the globus pallidus 00:08:57.210 --> 00:08:58.850 are the output systems. 00:08:58.850 --> 00:09:03.350 And then you have, in the rat brain, and also in our brains, 00:09:03.350 --> 00:09:05.720 you have massive convergence onto the input 00:09:05.720 --> 00:09:10.100 area, the caudate putamen it is also called, from the cortex 00:09:10.100 --> 00:09:11.610 and from the brain stem. 00:09:11.610 --> 00:09:14.000 So you have signals coming in from all over the brain 00:09:14.000 --> 00:09:16.490 to the striatum, which could be interpreted 00:09:16.490 --> 00:09:18.570 as a request for action. 00:09:18.570 --> 00:09:20.720 And then you have inhibition coming out 00:09:20.720 --> 00:09:23.030 from the output structures of the basal ganglia, 00:09:23.030 --> 00:09:25.370 here I'm showing it for the substantia nigra, 00:09:25.370 --> 00:09:28.040 going back to all of those areas of the brain. 00:09:28.040 --> 00:09:29.810 And this inhibition is tonic. 00:09:29.810 --> 00:09:31.760 And in order to have functional reaction, 00:09:31.760 --> 00:09:33.680 you have to remove the inhibition. 00:09:33.680 --> 00:09:39.770 So this is a system that can give you 00:09:39.770 --> 00:09:41.240 some of that behavioral integration 00:09:41.240 --> 00:09:43.790 that you need, the ability to ensure that you 00:09:43.790 --> 00:09:45.620 do one thing at a time. 00:09:45.620 --> 00:09:47.870 You do that quickly. 00:09:47.870 --> 00:09:50.150 You do that consistently. 00:09:50.150 --> 00:09:52.700 You dedicate all of your resources to the action 00:09:52.700 --> 00:09:54.870 that you want to do. 00:09:54.870 --> 00:09:57.200 Here's a little video of a rat. 00:09:57.200 --> 00:10:00.710 And I'm showing you some integrated behavior over time 00:10:00.710 --> 00:10:01.790 in an intact rat. 00:10:01.790 --> 00:10:05.600 So this is a rat exploring in a large container. 00:10:05.600 --> 00:10:08.089 And rats generally don't like open spaces, 00:10:08.089 --> 00:10:10.130 so when you first put the animal into this space, 00:10:10.130 --> 00:10:12.800 it will tend to stay near the walls. 00:10:12.800 --> 00:10:15.320 And it prefers this corner, which is dark. 00:10:15.320 --> 00:10:18.470 And of course, it's hungry too, so there's a dish of food here. 00:10:18.470 --> 00:10:20.120 And eventually, it gets up the courage 00:10:20.120 --> 00:10:21.710 to go out, collect a piece of food, 00:10:21.710 --> 00:10:26.060 and it will take it back into this dark corner to consume it. 00:10:26.060 --> 00:10:29.080 And one of the first models that we built 00:10:29.080 --> 00:10:32.200 was a model of basal ganglia operating as this, kind of, 00:10:32.200 --> 00:10:34.070 action-selection device. 00:10:34.070 --> 00:10:36.490 And with a simple Kepler robot, this 00:10:36.490 --> 00:10:39.870 is a robot that just really uses infrared sensors and a gripper 00:10:39.870 --> 00:10:40.900 arm. 00:10:40.900 --> 00:10:44.680 And we are using a model of the basal ganglia 00:10:44.680 --> 00:10:48.010 to control decision making about which actions 00:10:48.010 --> 00:10:51.820 to do at which time, and to generate sequencing 00:10:51.820 --> 00:10:53.150 of those actions. 00:10:53.150 --> 00:10:58.120 So as the need to stay close to walls diminishes, 00:10:58.120 --> 00:11:01.780 the robot, like the rat, goes and collects these cylinders. 00:11:01.780 --> 00:11:05.230 And it carries them back into the corners and deposits them. 00:11:05.230 --> 00:11:08.820 So a model of the central brain structures-- and I'm 00:11:08.820 --> 00:11:12.415 happy to discuss in more detail about how that model operates, 00:11:12.415 --> 00:11:14.560 and how similarities to the model 00:11:14.560 --> 00:11:16.270 that Michael will have described to you, 00:11:16.270 --> 00:11:19.000 but that's controlling the behavior switching, if you 00:11:19.000 --> 00:11:21.980 like, in this robot. 00:11:21.980 --> 00:11:27.250 So I spent some time working on this question of how 00:11:27.250 --> 00:11:30.850 central systems in the brain, particularly the basal ganglia, 00:11:30.850 --> 00:11:33.470 are involved in the integration of behavior, 00:11:33.470 --> 00:11:36.550 but I became frustrated with not understanding what 00:11:36.550 --> 00:11:39.740 were the signals coming in to the central brain structures 00:11:39.740 --> 00:11:42.820 and not understanding what effects those brain 00:11:42.820 --> 00:11:47.030 structures were having on the motor system of the animal. 00:11:47.030 --> 00:11:49.000 So I thought that what we needed to do 00:11:49.000 --> 00:11:51.190 was look at complete sensory motor loops. 00:11:51.190 --> 00:11:55.330 We needed to look at sensing and action, and how those interact. 00:11:55.330 --> 00:11:57.640 And in our Psychology Department, 00:11:57.640 --> 00:11:59.260 we have a neuroscience group that 00:11:59.260 --> 00:12:01.930 works mainly with rats, so it was natural for us 00:12:01.930 --> 00:12:03.670 to look at the rat. 00:12:03.670 --> 00:12:07.450 And in the rat, we know that one of the key perception systems 00:12:07.450 --> 00:12:09.200 is the vertebral system. 00:12:09.200 --> 00:12:11.560 So here you see, this is actually a pet rat, 00:12:11.560 --> 00:12:13.480 wandering around on my windowsill 00:12:13.480 --> 00:12:15.040 in my house in Sheffield. 00:12:15.040 --> 00:12:18.400 And the thing to notice is the whiskers here. 00:12:18.400 --> 00:12:20.500 And the whiskers are moving back and forth pretty 00:12:20.500 --> 00:12:24.430 much all the time that the rat is exploring. 00:12:24.430 --> 00:12:29.980 And we understand from nearly 100 years now of research 00:12:29.980 --> 00:12:32.530 that this system is very important for the rat 00:12:32.530 --> 00:12:33.860 to understand the environment. 00:12:33.860 --> 00:12:36.040 In fact, if it's completely dark, 00:12:36.040 --> 00:12:38.620 the rat would move around in much the same way. 00:12:38.620 --> 00:12:40.240 And it would be able to understand 00:12:40.240 --> 00:12:43.030 the world through touch pretty well, even 00:12:43.030 --> 00:12:45.970 in the absence of vision. 00:12:45.970 --> 00:12:48.310 So this is the same video, but now slow down 00:12:48.310 --> 00:12:51.910 10 times, and just to show you these movements 00:12:51.910 --> 00:12:54.400 of the whiskers, and how quite precise they are, 00:12:54.400 --> 00:12:58.540 because the rat isn't just, in a stereotypical way, 00:12:58.540 --> 00:13:00.640 banging its whiskers against the floor. 00:13:00.640 --> 00:13:02.530 It is lightly touching the whiskers 00:13:02.530 --> 00:13:04.690 in places it will get useful information. 00:13:04.690 --> 00:13:07.930 And you can see, when he puts his head over the window sill 00:13:07.930 --> 00:13:10.630 here, the whiskers push forward, as 00:13:10.630 --> 00:13:12.160 if he knows that he's going to have 00:13:12.160 --> 00:13:15.140 to reach further forward if he's going to find anything. 00:13:15.140 --> 00:13:17.590 Here you see him exploring this wooden cup. 00:13:17.590 --> 00:13:20.740 And you can see light touches by the whiskers. 00:13:20.740 --> 00:13:23.440 And you can also see that the movement of the whiskers 00:13:23.440 --> 00:13:26.050 is being modulated by the shape of the surface 00:13:26.050 --> 00:13:28.020 that he's investigating, so there's 00:13:28.020 --> 00:13:30.670 some fairly subtle control happening here. 00:13:30.670 --> 00:13:32.740 And I think it's not too much to say 00:13:32.740 --> 00:13:36.430 that the way in which the rat controls its whiskers 00:13:36.430 --> 00:13:38.200 has almost the same richness as the way 00:13:38.200 --> 00:13:40.030 that we control our fingertips. 00:13:42.560 --> 00:13:44.680 So I'm interested in how this plays out 00:13:44.680 --> 00:13:47.200 in terms of a layered architecture story. 00:13:47.200 --> 00:13:49.880 And of course, many people study this system. 00:13:49.880 --> 00:13:52.630 The beauty of it is, if you're a neuroscientist, 00:13:52.630 --> 00:13:55.040 that you can look in the cortex. 00:13:55.040 --> 00:13:57.310 This is rat cortex here. 00:13:57.310 --> 00:14:02.050 And a huge area of rat cortex is dedicated to somatosensation, 00:14:02.050 --> 00:14:06.010 to touch, of which a large area is dedicated to whiskers. 00:14:06.010 --> 00:14:07.480 In fact, you zoom in, you can find 00:14:07.480 --> 00:14:09.490 this area called barrel cortex. 00:14:09.490 --> 00:14:11.260 And with the right kind of staining, 00:14:11.260 --> 00:14:15.700 you can find groups of cells which preferentially receive 00:14:15.700 --> 00:14:18.070 signals from individual whiskers, 00:14:18.070 --> 00:14:21.430 so for example, you can move one whisker here, 00:14:21.430 --> 00:14:25.560 and you can know exactly where you record in the barrel cortex 00:14:25.560 --> 00:14:27.970 to get a very strong response from that whisker. 00:14:27.970 --> 00:14:30.820 And this means that barrel cortex and the whisker system 00:14:30.820 --> 00:14:35.170 has become one of the prepared, preferred preparations in which 00:14:35.170 --> 00:14:38.890 to study the cortical microcircuit altogether, 00:14:38.890 --> 00:14:41.650 so people study this system to really understand 00:14:41.650 --> 00:14:44.110 how cortex operates. 00:14:44.110 --> 00:14:46.390 Now, if we think about this system 00:14:46.390 --> 00:14:50.110 as a pathway from the whiskers up to barrel cortex, 00:14:50.110 --> 00:14:52.210 we're really only capturing one element 00:14:52.210 --> 00:14:54.940 of what's going on in the vertebral system. 00:14:54.940 --> 00:14:57.570 And that's this pathway here from the vertebrae, 00:14:57.570 --> 00:15:00.520 via the trigeminal complex, goes by the thalamus 00:15:00.520 --> 00:15:02.180 up to sensory cortex. 00:15:02.180 --> 00:15:05.920 And this is probably where 9 out of 10 papers on this system 00:15:05.920 --> 00:15:08.800 are published, but actually, this system 00:15:08.800 --> 00:15:11.420 is only part of a looped architecture, 00:15:11.420 --> 00:15:13.660 or we might say, a layered architecture. 00:15:13.660 --> 00:15:16.210 And at each level of this layered architecture, 00:15:16.210 --> 00:15:20.890 there's a completed loop, so that sensing can affect action, 00:15:20.890 --> 00:15:25.060 so sensing on the vertebrae can affect the movement and control 00:15:25.060 --> 00:15:26.770 of the vertebrae. 00:15:26.770 --> 00:15:28.810 So there's a loop via the brainstem 00:15:28.810 --> 00:15:32.470 here, so that, directly from the trigeminal complex, 00:15:32.470 --> 00:15:35.020 signals come back to the facial nucleus, which 00:15:35.020 --> 00:15:38.290 is where the motor neurons are that move the whiskers. 00:15:38.290 --> 00:15:40.600 There's a loop via the midbrain here 00:15:40.600 --> 00:15:43.390 so that sensory signals ascend very quickly 00:15:43.390 --> 00:15:45.490 to the midbrain superior colliculus. 00:15:45.490 --> 00:15:49.000 And they come back to affect how the whiskers move. 00:15:49.000 --> 00:15:51.970 And then, of course, there's the loop via the cortex 00:15:51.970 --> 00:15:54.750 too, so there's essentially those three loops, at least, 00:15:54.750 --> 00:15:56.920 that we need to think about. 00:15:56.920 --> 00:15:59.350 So since 2003, we've been building 00:15:59.350 --> 00:16:02.800 different whiskered robots, the aim being 00:16:02.800 --> 00:16:05.140 to instantiate our theories about how 00:16:05.140 --> 00:16:08.280 whiskered control works in this layered architecture 00:16:08.280 --> 00:16:11.080 and demonstrate it in a robot platform. 00:16:11.080 --> 00:16:13.540 And often, actually, building a robot platform 00:16:13.540 --> 00:16:15.760 causes us to ask new questions that might not 00:16:15.760 --> 00:16:19.210 be obvious to you just by doing biological experiments 00:16:19.210 --> 00:16:21.952 or even by doing simulation. 00:16:21.952 --> 00:16:23.410 Before I show you some robots, just 00:16:23.410 --> 00:16:28.120 quickly show a little bit more about the rat and its whiskers. 00:16:28.120 --> 00:16:30.230 So we began thinking we could just build robots, 00:16:30.230 --> 00:16:32.260 but we quickly realized that we didn't 00:16:32.260 --> 00:16:34.970 know enough about how rats use their whiskers to do that. 00:16:34.970 --> 00:16:37.874 And that's partly because the experiments that had been done 00:16:37.874 --> 00:16:39.790 haven't been done with the purpose of building 00:16:39.790 --> 00:16:41.260 a whiskered robot. 00:16:41.260 --> 00:16:43.810 So when you try and build a whiskered robot, 00:16:43.810 --> 00:16:46.680 you have to ask questions like, how do the whiskers move? 00:16:46.680 --> 00:16:49.600 And when you look at a video like this filmed from above, 00:16:49.600 --> 00:16:51.580 this is with a high speed camera, 00:16:51.580 --> 00:16:54.130 you think, well, the whiskers are sweeping backward 00:16:54.130 --> 00:16:56.110 and forward, like this. 00:16:56.110 --> 00:17:00.580 But in fact, if you put a mirror just tilted down here 00:17:00.580 --> 00:17:02.146 and you see what happens, then it 00:17:02.146 --> 00:17:03.770 turns out to be a little bit different, 00:17:03.770 --> 00:17:07.450 so you see that the whiskers are going up and down as much 00:17:07.450 --> 00:17:10.010 as they are going backwards and forwards. 00:17:10.010 --> 00:17:12.300 So the whiskers are actually sweeping like this, 00:17:12.300 --> 00:17:16.290 and they're making a series of touches on the surface. 00:17:16.290 --> 00:17:20.560 And if you watch, you can see that the whiskers are 00:17:20.560 --> 00:17:23.800 sort of playing down on the surface, sort of in a sequence, 00:17:23.800 --> 00:17:27.819 quite quickly, so that information might be giving you 00:17:27.819 --> 00:17:31.870 details about the shape of the surfaces in your world. 00:17:31.870 --> 00:17:33.640 So we mainly look at the long whiskers. 00:17:33.640 --> 00:17:35.590 This is a rat that's running up an alley. 00:17:35.590 --> 00:17:38.560 And we put an unexpected object in the alley, which 00:17:38.560 --> 00:17:41.950 could be this aluminum rectangle, 00:17:41.950 --> 00:17:44.350 or it could be this plastic step. 00:17:44.350 --> 00:17:47.050 And what you see is, if the animal encounters something 00:17:47.050 --> 00:17:49.420 unexpected with its long whiskers, then 00:17:49.420 --> 00:17:52.150 it turns very quickly and investigates it. 00:17:52.150 --> 00:17:54.970 So the long whiskers are like the periphery 00:17:54.970 --> 00:17:57.640 of a sensory system that has a fovea. 00:17:57.640 --> 00:18:00.070 And the fovea at the center of that system 00:18:00.070 --> 00:18:02.950 is a set of short whiskers around the mouth, also 00:18:02.950 --> 00:18:06.700 the lips and the nose, so that you 00:18:06.700 --> 00:18:09.670 can sniff and smell the surface that you're investigating. 00:18:09.670 --> 00:18:13.180 So we can zoom in and see that sensory fovea, here you 00:18:13.180 --> 00:18:16.690 see these short whiskers that are being used to investigate 00:18:16.690 --> 00:18:20.219 this plastic puck, and the longer whiskers investigating 00:18:20.219 --> 00:18:21.010 around the outside. 00:18:23.570 --> 00:18:26.710 So we have recapitulated elements 00:18:26.710 --> 00:18:28.890 of this layered architecture in our robot. 00:18:28.890 --> 00:18:31.060 And this is-- these are the loops. 00:18:31.060 --> 00:18:33.260 And this was about five years ago, 00:18:33.260 --> 00:18:35.650 we built a system with a brain stem loop, 00:18:35.650 --> 00:18:37.260 and really, midbrain loop. 00:18:37.260 --> 00:18:39.770 And this is our robot Scratchbot, 00:18:39.770 --> 00:18:43.750 which is the first of the whiskered robots 00:18:43.750 --> 00:18:47.395 that we felt really was capturing whisking in the way 00:18:47.395 --> 00:18:48.380 that the rat does it. 00:18:48.380 --> 00:18:52.780 It's running at about 4 hertz, whereas the real rat is 00:18:52.780 --> 00:18:55.150 whisking from 8 to 12 hertz, but it's 00:18:55.150 --> 00:18:58.780 scaled up to be about four times rat size. 00:18:58.780 --> 00:19:00.640 And what it's doing here is, it's 00:19:00.640 --> 00:19:02.680 using the whiskers to orient to stimuli, 00:19:02.680 --> 00:19:05.380 so this is Martin Pearson from Bristol Robotics Lab. 00:19:05.380 --> 00:19:08.290 He's putting an object in the whisker field. 00:19:08.290 --> 00:19:11.260 And the robot is turning and orienting 00:19:11.260 --> 00:19:12.980 to the touch with the object. 00:19:12.980 --> 00:19:15.520 It's putting its short micro vibrissae, in fact, 00:19:15.520 --> 00:19:18.020 against the object and exploring it. 00:19:18.020 --> 00:19:21.340 Now to do that, to detect a stimulus on the whiskers 00:19:21.340 --> 00:19:24.680 and then turn is not a fantastically hard task. 00:19:24.680 --> 00:19:27.640 The main challenge is to work out 00:19:27.640 --> 00:19:29.650 where the whisker was in its sweep 00:19:29.650 --> 00:19:31.270 when you made contact with an object, 00:19:31.270 --> 00:19:33.322 because the whisker is sweeping back and forth, 00:19:33.322 --> 00:19:34.780 so if you want to know the location 00:19:34.780 --> 00:19:36.370 of the point of contact, you need 00:19:36.370 --> 00:19:41.530 to integrate the position of the whisker in its sweep, what you 00:19:41.530 --> 00:19:44.590 might call a theta signal, and the presence 00:19:44.590 --> 00:19:46.210 of the contact on the whiskers. 00:19:46.210 --> 00:19:49.700 And the coincidence of those two is detected in the brain. 00:19:49.700 --> 00:19:52.210 And we know that there are cells in the barrel cortex 00:19:52.210 --> 00:19:54.520 that respond to that coincidence. 00:19:54.520 --> 00:19:57.340 So we have in our robot a model of the super colliculus, which 00:19:57.340 --> 00:19:59.200 is the location in the brain which we 00:19:59.200 --> 00:20:01.900 think is involved in orienting. 00:20:01.900 --> 00:20:04.480 And in our model of the colliculus, 00:20:04.480 --> 00:20:06.970 we have a head-centered map, which 00:20:06.970 --> 00:20:10.660 looks for this coincidence between a cell encoding 00:20:10.660 --> 00:20:13.370 the position of the whisker in its sweep and a cell 00:20:13.370 --> 00:20:17.710 encoding a contact and makes a turn to orient and explore 00:20:17.710 --> 00:20:20.177 that position. 00:20:20.177 --> 00:20:21.760 And then if we want to actually create 00:20:21.760 --> 00:20:24.860 behavior, which is integrated over time, 00:20:24.860 --> 00:20:28.030 so if the robot was just to orient every time it touched 00:20:28.030 --> 00:20:31.080 something, that wouldn't be very animal-like, 00:20:31.080 --> 00:20:32.960 particularly, you don't want to orient 00:20:32.960 --> 00:20:35.230 every time you touch the ground, so you just 00:20:35.230 --> 00:20:38.140 want to orient when you touch important stimuli. 00:20:38.140 --> 00:20:40.820 So we put into our model the basal ganglia 00:20:40.820 --> 00:20:44.020 so that we can decide whether the contact we've just made 00:20:44.020 --> 00:20:45.850 is something we want to investigate 00:20:45.850 --> 00:20:49.100 or something that doesn't interest us so much. 00:20:49.100 --> 00:20:52.570 So we have a system now with a midbrain 00:20:52.570 --> 00:20:55.630 that does orienting, a basal ganglia that makes decisions 00:20:55.630 --> 00:20:57.160 about sequencing. 00:20:57.160 --> 00:20:59.810 And those two things together give us 00:20:59.810 --> 00:21:03.450 reasonably lifelike behavior in our robot, Scratchbot. 00:21:03.450 --> 00:21:05.800 And that's quite a lot of what we 00:21:05.800 --> 00:21:08.710 have running now on the new robot, MiRo, is this system. 00:21:08.710 --> 00:21:11.180 It's for orienting and exploring. 00:21:11.180 --> 00:21:13.840 And we can use it-- we use it here for tactile orienting, 00:21:13.840 --> 00:21:15.520 but of course, the same system can 00:21:15.520 --> 00:21:17.170 underlie orienting to sounds, if you 00:21:17.170 --> 00:21:21.650 can localize those, and space, and orienting to visual stimuli 00:21:21.650 --> 00:21:23.020 too. 00:21:23.020 --> 00:21:25.990 So it turns out that this isn't a complete solution 00:21:25.990 --> 00:21:29.260 to the problem of orienting for our whiskered robot. 00:21:29.260 --> 00:21:31.900 And that's because sometimes our robot 00:21:31.900 --> 00:21:34.690 would stop as it was moving around 00:21:34.690 --> 00:21:38.024 and just move and investigate a point in space 00:21:38.024 --> 00:21:39.190 where nothing was happening. 00:21:39.190 --> 00:21:41.579 We call that a ghost orient. 00:21:41.579 --> 00:21:43.120 And the problem is that the whiskers, 00:21:43.120 --> 00:21:45.040 because they're moving back and forth, 00:21:45.040 --> 00:21:47.680 they sometimes generate signals in the strain 00:21:47.680 --> 00:21:50.770 gauges that are detecting bending of the whisker. 00:21:50.770 --> 00:21:52.390 And sometimes those signals, just as 00:21:52.390 --> 00:21:55.090 a consequence of the movement and the mass of the whisker, 00:21:55.090 --> 00:21:57.550 are strong enough to be above threshold 00:21:57.550 --> 00:21:59.570 to generate an orient, so you get, 00:21:59.570 --> 00:22:02.350 if you like, these ghost-orienting movements 00:22:02.350 --> 00:22:04.990 towards stimuli that don't exist. 00:22:04.990 --> 00:22:08.590 And we know that rats don't make these kind of ghost orients, 00:22:08.590 --> 00:22:11.920 so something else must be going on in the brain. 00:22:11.920 --> 00:22:15.240 And one part of the brain that might be helping here 00:22:15.240 --> 00:22:18.780 is a region called the cerebellum, which, I'm not sure 00:22:18.780 --> 00:22:21.360 if you've covered that in the summer school, 00:22:21.360 --> 00:22:24.180 but the cerebellum, this large structure 00:22:24.180 --> 00:22:25.980 at the back of the right brain. 00:22:25.980 --> 00:22:29.010 One of its key functions seems to be 00:22:29.010 --> 00:22:32.510 to make predictions about sensory signals, 00:22:32.510 --> 00:22:35.580 and particularly, to be able to predict sensory signals that 00:22:35.580 --> 00:22:37.755 have been caused by your own movement. 00:22:37.755 --> 00:22:39.510 And there's a lovely experiment that's 00:22:39.510 --> 00:22:43.200 been done by Blakemore et al, where they put people 00:22:43.200 --> 00:22:44.760 into a scanner. 00:22:44.760 --> 00:22:49.360 And they investigated how they responded to tickling stimuli. 00:22:49.360 --> 00:22:51.270 So of course, if somebody tickles you, 00:22:51.270 --> 00:22:53.910 that can be quite amusing, but unfortunately, 00:22:53.910 --> 00:22:56.940 if you try to tickle yourself, it's really uninteresting. 00:22:56.940 --> 00:22:59.670 It doesn't work as a stimulus. 00:22:59.670 --> 00:23:04.530 And it's worth thinking about why it is that self-tickling 00:23:04.530 --> 00:23:06.150 is so unrewarding. 00:23:06.150 --> 00:23:08.910 And one of the reasons is that it's just not surprising. 00:23:08.910 --> 00:23:12.160 You know what's going to happen when you tickle yourself, 00:23:12.160 --> 00:23:14.070 whereas if somebody else is doing it, 00:23:14.070 --> 00:23:16.290 it's unexpected and surprising. 00:23:16.290 --> 00:23:20.670 So why is self-tickling unexpected? 00:23:20.670 --> 00:23:22.560 Why is it not surprising? 00:23:22.560 --> 00:23:25.530 It must be because the brain expects and anticipates 00:23:25.530 --> 00:23:27.570 the signal that it's going to get. 00:23:27.570 --> 00:23:30.450 And what Blakemore et al did was to show 00:23:30.450 --> 00:23:32.910 that the cerebellum really lights up 00:23:32.910 --> 00:23:35.220 when you try to tickle yourself, because it's 00:23:35.220 --> 00:23:38.340 estimating and predicting the sensory signal, 00:23:38.340 --> 00:23:40.920 and using that to cancel out, if you like, 00:23:40.920 --> 00:23:44.130 the signal that's coming from your skin. 00:23:44.130 --> 00:23:48.030 The same thing is happening in electric fish, which 00:23:48.030 --> 00:23:50.490 generate this broad electric field which 00:23:50.490 --> 00:23:53.242 they use for catching prey. 00:23:53.242 --> 00:23:55.200 And they need to be able to tell the difference 00:23:55.200 --> 00:23:58.800 between a distortion to the electric field caused by a prey 00:23:58.800 --> 00:24:01.350 animal and a distortion caused by their own movement, 00:24:01.350 --> 00:24:02.370 by swimming. 00:24:02.370 --> 00:24:05.080 And they do that by having a very large cerebellum. 00:24:05.080 --> 00:24:06.780 So we put a model of the cerebellum 00:24:06.780 --> 00:24:08.700 in our whiskered robot. 00:24:08.700 --> 00:24:11.020 And the cerebellum predicts the noise 00:24:11.020 --> 00:24:13.830 you might get due to the movement of the whiskers. 00:24:13.830 --> 00:24:15.930 And it learns online to accurately 00:24:15.930 --> 00:24:18.150 predict what noise signals you might get, 00:24:18.150 --> 00:24:19.920 and to cancel them out, so you get 00:24:19.920 --> 00:24:24.420 a much better signal-to-noise ratio in the robot. 00:24:24.420 --> 00:24:28.460 So we've dealt with whisking. 00:24:28.460 --> 00:24:29.940 And we've dealt with orienting. 00:24:29.940 --> 00:24:33.060 But as you saw with that rat on the windowsill, 00:24:33.060 --> 00:24:35.291 the whisker movements are really precise. 00:24:35.291 --> 00:24:36.540 And they're really controlled. 00:24:36.540 --> 00:24:38.550 And the rat seems to really care about 00:24:38.550 --> 00:24:41.190 how it's moving its whiskers and how it's touching. 00:24:41.190 --> 00:24:43.290 We call this active sensing. 00:24:43.290 --> 00:24:45.360 And if you look at these high-speed videos, 00:24:45.360 --> 00:24:47.220 you can see, for instance, this rat 00:24:47.220 --> 00:24:49.880 when it's exploring this perspex block. 00:24:49.880 --> 00:24:53.354 The whiskers aren't moving in a stereotype symmetric way. 00:24:53.354 --> 00:24:55.770 You can see that here, the whiskers on the right-hand side 00:24:55.770 --> 00:24:57.940 are really reaching round to try and reach 00:24:57.940 --> 00:24:59.770 the other side of the block. 00:24:59.770 --> 00:25:01.740 If you watch this rat here, you see that too. 00:25:01.740 --> 00:25:03.960 You've got asymmetry. 00:25:03.960 --> 00:25:07.260 And you'll see that, even as the rat comes up to the cylinder 00:25:07.260 --> 00:25:10.710 here, the whiskers at the front are pushing forward 00:25:10.710 --> 00:25:14.170 while the ones at the back are hardly moving at all, 00:25:14.170 --> 00:25:16.350 so there's some ability to control even the whiskers 00:25:16.350 --> 00:25:18.130 on one side of the head. 00:25:18.130 --> 00:25:20.064 And when you move your fingers, of course, 00:25:20.064 --> 00:25:22.230 there's some coupling between your finger movements. 00:25:22.230 --> 00:25:24.510 You can't move them entirely independently. 00:25:24.510 --> 00:25:26.760 And each of these whiskers has its own muscle, 00:25:26.760 --> 00:25:29.490 so there's a degree of independence 00:25:29.490 --> 00:25:31.690 in how the whiskers can move. 00:25:31.690 --> 00:25:34.720 And we find that when we record over long intervals. 00:25:34.720 --> 00:25:36.290 So this was a study-- 00:25:36.290 --> 00:25:37.580 [ELECTRONIC NOISE] 00:25:37.580 --> 00:25:42.880 --in which we recorded the whisking muscles using EMG. 00:25:42.880 --> 00:25:46.410 And that's the sound that you can hear as the rat explored. 00:25:46.410 --> 00:25:50.270 And we tracked the rat as he was moving around. 00:25:50.270 --> 00:25:51.920 And we showed that, whenever he came 00:25:51.920 --> 00:25:55.060 close to the edge of the box here, 00:25:55.060 --> 00:25:56.720 the whiskers would become asymmetric. 00:25:56.720 --> 00:25:58.970 And the whiskers that were furthest away from the wall 00:25:58.970 --> 00:26:02.120 would push round to try and touch the sides of the box. 00:26:02.120 --> 00:26:04.490 The whiskers that were close to the wall 00:26:04.490 --> 00:26:06.180 would barely move at all. 00:26:06.180 --> 00:26:10.280 So we want to put that kind of control into our robot. 00:26:10.280 --> 00:26:14.600 So I briefly want to come back to this question of how 00:26:14.600 --> 00:26:17.160 we decompose control. 00:26:17.160 --> 00:26:21.470 So in our original robot that was controlled 00:26:21.470 --> 00:26:24.050 by the basal ganglia, and it's collecting cans, 00:26:24.050 --> 00:26:26.990 we decompose behaviors into the different elements 00:26:26.990 --> 00:26:28.120 of behavior-- 00:26:28.120 --> 00:26:31.520 looking for a can, picking it up, carrying it to the wall, 00:26:31.520 --> 00:26:32.880 these sorts of things. 00:26:32.880 --> 00:26:35.180 And if we look in the ethology literature, 00:26:35.180 --> 00:26:37.400 we find that people have talked about these kinds 00:26:37.400 --> 00:26:38.850 of decompositions. 00:26:38.850 --> 00:26:41.330 There's a very famous paper by Baerends 00:26:41.330 --> 00:26:42.780 about the herring gull. 00:26:42.780 --> 00:26:44.310 And with the herring gull, there's 00:26:44.310 --> 00:26:50.090 this famous experiment where the egg rolls out the nest. 00:26:50.090 --> 00:26:53.360 And the bird will retrieve the egg with its bill 00:26:53.360 --> 00:26:55.220 and push it back into the nest. 00:26:55.220 --> 00:26:59.340 And it will do this same action really reliably and repeatedly. 00:26:59.340 --> 00:27:01.250 And it can do it with eggs of various size. 00:27:01.250 --> 00:27:03.380 It might even do it for a Coke can. 00:27:03.380 --> 00:27:05.960 And if you take the egg away during the movement, 00:27:05.960 --> 00:27:07.970 it will still complete the movement. 00:27:07.970 --> 00:27:11.360 And ethologists have called this a fixed action pattern, 00:27:11.360 --> 00:27:13.100 so it may be that behavior is decomposed 00:27:13.100 --> 00:27:15.650 into action patterns. 00:27:15.650 --> 00:27:17.630 And that's one of the ways, for instance, 00:27:17.630 --> 00:27:20.720 in which Rodney Brooks wants to decompose robot behavior. 00:27:20.720 --> 00:27:23.030 We decompose it into different things 00:27:23.030 --> 00:27:24.605 we might want the robot to do. 00:27:24.605 --> 00:27:26.480 And we can do that with our whiskered robots. 00:27:26.480 --> 00:27:30.020 Here's another one with its behavior decomposed 00:27:30.020 --> 00:27:32.090 into different kinds of, if you like, 00:27:32.090 --> 00:27:36.270 orienting behaviors and fixed action patterns. 00:27:36.270 --> 00:27:37.940 Another way to decompose behavior 00:27:37.940 --> 00:27:40.730 is to think about where your attention is, so 00:27:40.730 --> 00:27:43.070 where you put your attention might decide 00:27:43.070 --> 00:27:44.540 what you're going to do next. 00:27:44.540 --> 00:27:47.180 And for an animal that doesn't have arms, 00:27:47.180 --> 00:27:51.080 and of course most animals except humans and some primates 00:27:51.080 --> 00:27:54.650 don't usually use their forelimbs for much else 00:27:54.650 --> 00:27:56.390 other than locomotion. 00:27:56.390 --> 00:27:58.430 And they primarily are positioning 00:27:58.430 --> 00:28:00.840 their head and their face. 00:28:00.840 --> 00:28:03.600 And their main effector, then, is their mouth. 00:28:03.600 --> 00:28:05.930 So where you position your attention 00:28:05.930 --> 00:28:08.520 could determine what you're going to do next. 00:28:08.520 --> 00:28:10.580 So another way of decomposing control 00:28:10.580 --> 00:28:12.977 is to solve the attention problem first. 00:28:12.977 --> 00:28:14.685 And then once you solve that, the problem 00:28:14.685 --> 00:28:17.180 of what you're going to do is simplified. 00:28:17.180 --> 00:28:19.940 So in this robot, we're controlling it 00:28:19.940 --> 00:28:22.190 by deciding where its attention should go. 00:28:22.190 --> 00:28:26.150 And then the rest of the body kind of follows. 00:28:26.150 --> 00:28:28.910 When humans have special attention, of course, 00:28:28.910 --> 00:28:31.560 we explore that in the visual modality. 00:28:31.560 --> 00:28:34.170 And we look at the saccadic eye movements that people make. 00:28:34.170 --> 00:28:37.940 So in the famous experiment, Albert Yarbus 00:28:37.940 --> 00:28:39.980 had people looking at this picture 00:28:39.980 --> 00:28:41.990 and tracking where their eyes would look. 00:28:41.990 --> 00:28:45.470 And of course, we look at the socially-significant elements 00:28:45.470 --> 00:28:48.080 of the picture, people's faces and so on, 00:28:48.080 --> 00:28:52.750 not just arbitrary points of light, or corners, and so on. 00:28:52.750 --> 00:28:57.170 And we can actually calculate a saliency map for space 00:28:57.170 --> 00:29:00.140 and say what are the important parts of space 00:29:00.140 --> 00:29:02.690 for exploring and attending to. 00:29:02.690 --> 00:29:07.370 And we've taken that idea and transferred it into our model 00:29:07.370 --> 00:29:08.630 for understanding the rat. 00:29:08.630 --> 00:29:11.510 And we thought about tactile saliency maps, so can 00:29:11.510 --> 00:29:13.900 we, with a sense of touch, think about areas 00:29:13.900 --> 00:29:17.000 of the world which are important to explore and understand 00:29:17.000 --> 00:29:18.080 through touch? 00:29:18.080 --> 00:29:21.110 And can we use that to control the movement 00:29:21.110 --> 00:29:24.470 of our robot, or in this case, our simulation? 00:29:24.470 --> 00:29:28.070 So here, we have a form of emergent wall following, which 00:29:28.070 --> 00:29:31.850 is a consequence of the rat's spatial attention being 00:29:31.850 --> 00:29:35.830 driven by contact with vertical objects, which we-- 00:29:35.830 --> 00:29:39.230 we program it so that the vertical surfaces 00:29:39.230 --> 00:29:41.090 are salient and interesting. 00:29:41.090 --> 00:29:43.130 And it has this salient zone. 00:29:43.130 --> 00:29:46.610 And it tries to put its whiskers into the salient zone. 00:29:46.610 --> 00:29:49.700 And then here is a robot instantiating this. 00:29:49.700 --> 00:29:51.380 This is Ben Mitchinson, who's programmed 00:29:51.380 --> 00:29:52.730 many of these robots. 00:29:52.730 --> 00:29:54.800 And so what we're doing now is following 00:29:54.800 --> 00:29:58.820 this biologically-inspired orienting system 00:29:58.820 --> 00:29:59.930 to explore shapes. 00:29:59.930 --> 00:30:03.350 And in this case, he put his own face in front of the robot. 00:30:03.350 --> 00:30:07.580 And you can see the robot making light touches against his face 00:30:07.580 --> 00:30:11.420 and investigating it, looking-- 00:30:11.420 --> 00:30:14.090 making a series of, if you like, exploratory touches, 00:30:14.090 --> 00:30:17.720 somewhat like saccades, somewhat like what 00:30:17.720 --> 00:30:19.670 you might imagine a blind person would 00:30:19.670 --> 00:30:21.740 do if they were investigating your face 00:30:21.740 --> 00:30:23.300 to try and recognize you. 00:30:23.300 --> 00:30:25.310 And Mitra Hartmann from Northwestern 00:30:25.310 --> 00:30:27.200 has shown that you can take signals 00:30:27.200 --> 00:30:30.040 off these kinds of whiskers and reconstruct a face, 00:30:30.040 --> 00:30:33.410 so it should be possible from this 00:30:33.410 --> 00:30:36.960 to build up from the touches, the sequence of touches, 00:30:36.960 --> 00:30:39.080 a lot of rich information about the object that's 00:30:39.080 --> 00:30:40.480 being investigated. 00:30:40.480 --> 00:30:42.202 How much time? 00:30:42.202 --> 00:30:42.910 I need to finish. 00:30:42.910 --> 00:30:44.630 OK, let me just skip through. 00:30:44.630 --> 00:30:47.362 So we've been doing-- working on the cortex. 00:30:47.362 --> 00:30:48.820 We have a number of models of that, 00:30:48.820 --> 00:30:51.580 which I'd like to show you, but I 00:30:51.580 --> 00:30:53.950 want to just finish, just to make contact 00:30:53.950 --> 00:30:56.980 with John's talk, is that we've been doing, 00:30:56.980 --> 00:31:00.640 in our robots, tactile simultaneous localization 00:31:00.640 --> 00:31:01.270 and mapping. 00:31:01.270 --> 00:31:03.520 So this is our whiskered robot. 00:31:03.520 --> 00:31:05.830 And we have various models for this, some of which 00:31:05.830 --> 00:31:07.780 are more hippocampal-like. 00:31:07.780 --> 00:31:10.130 This one, I think, was more of an engineered model. 00:31:10.130 --> 00:31:12.700 But you can see the robot just using touch 00:31:12.700 --> 00:31:15.340 on these artificial whiskers, building up 00:31:15.340 --> 00:31:17.090 a map of its environment. 00:31:17.090 --> 00:31:19.600 These two lines show its dead-reckoning position 00:31:19.600 --> 00:31:21.490 and its calculated position. 00:31:21.490 --> 00:31:24.550 And just using touch, we can build up 00:31:24.550 --> 00:31:28.220 a reasonably accurate map of the world that we're exploring. 00:31:28.220 --> 00:31:31.210 So Giorgio will talk about the iCub. 00:31:31.210 --> 00:31:33.730 And I just wanted to mention that, in the work we're 00:31:33.730 --> 00:31:36.280 doing with Giorgio, we are very much trying 00:31:36.280 --> 00:31:38.360 to understand human cognition. 00:31:38.360 --> 00:31:41.680 I wrote a short article for New Scientist on the possibility 00:31:41.680 --> 00:31:44.640 that robots might one day have selves.