WEBVTT 00:00:03.517 --> 00:00:10.517 This is the Convergent Science Network podcast. Leading researchers in the domain 00:00:10.517 --> 00:00:16.797 of neuroscience, brain theory and technology are interviewed by Paul Verscher and Tony Prescott. 00:00:20.217 --> 00:00:25.277 Hi, this is Tony Prescott for the Convergent Science Network podcast from the 00:00:25.277 --> 00:00:29.777 Barcelona Summer School on Cognition, Brain and Technology. and I'm talking 00:00:29.777 --> 00:00:34.417 to Andy Filippides from Informatics at the University of Sussex. 00:00:35.137 --> 00:00:43.377 Andy, your group works with ants and other insects looking at their behavior 00:00:43.377 --> 00:00:48.197 in the natural environment and trying to understand the relationship between 00:00:48.197 --> 00:00:51.697 the environment that they live in and the behavior that they have. 00:00:51.697 --> 00:00:57.557 Now, why do you think it's important to study animals in their real environment 00:00:57.557 --> 00:00:58.597 and not just in the laboratory? 00:00:59.097 --> 00:01:06.177 Well, I think you can't do research in the absence of laboratory experiments, 00:01:06.397 --> 00:01:12.957 but you have to test the same behaviors in the real world wherever possible, 00:01:13.117 --> 00:01:15.297 primarily because insects, 00:01:16.777 --> 00:01:22.717 ants, animals, neurons often respond very, very differently when you have natural stimuli. 00:01:23.577 --> 00:01:26.737 And so we need to be able to see what the natural behavior is. 00:01:29.077 --> 00:01:36.137 There are a lot of anecdotal evidence for this, particularly recently with optic 00:01:36.137 --> 00:01:41.337 flow, where people traditionally use very strong optic flow signals and the 00:01:41.337 --> 00:01:43.937 models work in a certain way with these very strong optic flow signals. 00:01:44.157 --> 00:01:49.837 And when they've recently challenged these models with optic flow signals, 00:01:49.877 --> 00:01:54.377 as an animal would perceive moving through the real world, you get very different results. 00:01:55.197 --> 00:01:58.377 So would these be experiments with insects? 00:01:58.797 --> 00:02:03.757 This is experiments. I think they recorded the video from cats moving through 00:02:03.757 --> 00:02:06.817 the undergrowth, and it's just putting them into sort of optic flow-based models. 00:02:07.197 --> 00:02:11.077 So with ants, we typically, in the lab, 00:02:11.077 --> 00:02:14.657 uh we use as blank 00:02:14.657 --> 00:02:19.917 an environment as possible so that the only um visual objects they can use are 00:02:19.917 --> 00:02:27.037 the ones that we've put put in um and this clearly isn't a natural um a natural 00:02:27.037 --> 00:02:34.417 environment so when we're and we typically train into a vertical edge for instance and so um, 00:02:35.931 --> 00:02:40.731 Whilst we can get a lot of information that way, this is not the sort of thing they usually attend to. 00:02:41.071 --> 00:02:45.091 So, for instance, Paul Graham's recent work showing that they used the whole 00:02:45.091 --> 00:02:47.651 panorama wouldn't be... 00:02:47.651 --> 00:02:52.731 That wouldn't work in the lab because you would have to reconstruct an artificial 00:02:52.731 --> 00:02:57.371 panorama that doesn't move in the same way as a real panorama does. 00:02:58.951 --> 00:03:02.391 And primarily it's got a very different distance distribution of objects. 00:03:02.391 --> 00:03:06.771 So this is looking at the visual navigation capability of an ant. Yeah, basically. 00:03:07.071 --> 00:03:11.911 So why would people be particularly interested in the ant as a system in which 00:03:11.911 --> 00:03:13.411 to study navigation behaviour? 00:03:13.871 --> 00:03:16.131 Well, firstly, because they're very good at it. 00:03:16.711 --> 00:03:25.691 Secondly, because they're social insects and social foragers, 00:03:25.831 --> 00:03:28.051 they go out multiple times in the day. 00:03:28.111 --> 00:03:31.151 So practically, they're very good. they go out to get food to give it back to 00:03:31.151 --> 00:03:35.931 the colony um and then go out and forage again so you can train them quickly 00:03:35.931 --> 00:03:40.651 um they learn in one trial they've got lots of interesting behaviors. 00:03:42.371 --> 00:03:48.071 Um the more interesting reason i think is that they're we'd argue they're. 00:03:49.011 --> 00:03:54.851 One of the most complex animals you can study in which you can study them over the course of the 00:03:54.891 --> 00:03:57.631 whole foraging range so it is possible to 00:03:57.631 --> 00:04:00.411 track an ant when it leap from where it needs its 00:04:00.411 --> 00:04:03.611 nest to when it finds food and when it comes back and there 00:04:03.611 --> 00:04:08.471 is some work ongoing in australia to track certain ants um through the course 00:04:08.471 --> 00:04:13.971 of their whole life until one gets bored but at least from so that one can have 00:04:13.971 --> 00:04:18.531 a record of that everything everywhere they've been and all the visual experience 00:04:18.531 --> 00:04:22.531 all the visual um visual input they might have experienced 00:04:22.811 --> 00:04:25.831 so our ants navigation 00:04:25.831 --> 00:04:29.351 capabilities are simpler than mammals safe 00:04:29.351 --> 00:04:32.971 or just specialized well i think 00:04:32.971 --> 00:04:36.151 they're very good at what they do they're different 00:04:36.151 --> 00:04:39.631 because they've got compound eyes so there 00:04:39.631 --> 00:04:42.591 are certain things that uh we would do differently their eyes 00:04:42.591 --> 00:04:46.351 are certainly a lot worse than ours um they don't 00:04:46.351 --> 00:04:49.411 use maps and we do use maps 00:04:49.411 --> 00:04:52.851 i think it's been pretty much categorically shown that they characterically shown 00:04:52.851 --> 00:04:56.551 they don't use cognitive maps um and 00:04:56.551 --> 00:05:04.511 so for instance um if you train an ant to go out to find food and then come 00:05:04.511 --> 00:05:10.831 back it'll come back along the same path um if you then uh place the ant back 00:05:10.831 --> 00:05:12.971 on that path when it's fed. 00:05:14.209 --> 00:05:17.689 Um at the middle of the path it's getting it's got exactly the same visual input 00:05:17.689 --> 00:05:20.549 that it would have got for going out or coming back and it will just return 00:05:20.549 --> 00:05:25.889 home um if you place it where it's empty it will go out to the to the goal so 00:05:25.889 --> 00:05:29.369 this kind of indicates that the memories are kind of insulated from each other 00:05:29.369 --> 00:05:33.249 there's been various quite nice experiments that's shown that that, 00:05:33.909 --> 00:05:38.549 their memories their root memories are kind of insulated from one another by context and they 00:05:38.569 --> 00:05:42.709 can't really share across them so by um not 00:05:42.709 --> 00:05:45.969 having a cognitive map yep they are in some 00:05:45.969 --> 00:05:49.969 sense limited to doing homing behaviors essentially or 00:05:49.969 --> 00:05:53.469 or are they able to learn multiple paths they learn 00:05:53.469 --> 00:05:59.309 multiple paths they learn multiple paths and it's likely if their multiple paths 00:05:59.309 --> 00:06:02.969 are the same well it's possible our model would be that if they multiple paths 00:06:02.969 --> 00:06:10.249 to the same food source then it's maybe stored as one root but certainly it would appear that, 00:06:11.689 --> 00:06:15.469 outbound roots and nest bound roots are insulated from one another and they 00:06:15.469 --> 00:06:19.669 would be primed by the context of being fed and so they could well be primed 00:06:19.669 --> 00:06:24.729 by other things um you know they may well forage at different places at different 00:06:24.729 --> 00:06:29.949 time of day such as bees would um but yeah i think. 00:06:32.809 --> 00:06:38.569 I mean, the other reason that we study them is because they are specialists. 00:06:39.929 --> 00:06:45.389 And when they want to return home having found food, all they care about is getting home. 00:06:46.249 --> 00:06:50.129 And Tom Collett has a very nice phrase where he says that their behavior gives 00:06:50.129 --> 00:06:52.249 a direct readout of their nervous system. 00:06:52.249 --> 00:06:55.809 Them and so without doing 00:06:55.809 --> 00:06:59.269 something invasive and difficult to do in the field you 00:06:59.269 --> 00:07:02.669 can see exactly well you can see what the ant it 00:07:02.669 --> 00:07:07.489 thinks it's doing yeah i think that that is really useful because one of the 00:07:07.489 --> 00:07:12.549 biggest problems when in neuroethology i think of studying an ethology of studying 00:07:12.549 --> 00:07:16.849 animals in the field is that it's very hard to know what the intentions of the 00:07:16.849 --> 00:07:22.629 animal are and therefore to have much insight into what it might be trying to do, 00:07:22.709 --> 00:07:26.769 never mind from what it's actually doing. Yeah, I think that's right. 00:07:28.109 --> 00:07:34.089 That is primarily all they want to do. So you know exactly what its intention is. 00:07:35.389 --> 00:07:38.489 There aren't any distractions. And so, yeah. 00:07:38.889 --> 00:07:44.809 So we know that one of the mechanisms that they use is to do path integration, a kind of step counting. 00:07:44.809 --> 00:07:47.649 And they also have some kind of compass based 00:07:47.649 --> 00:07:50.329 on the sun yeah they've got cells and eyes that are 00:07:50.329 --> 00:07:53.289 sensitive to polarized light um so they have 00:07:53.289 --> 00:08:01.189 and they need both steps and polarized light to do path integration and path 00:08:01.189 --> 00:08:05.109 integration in some senses should be enough to go back to the next it should 00:08:05.109 --> 00:08:10.749 be um but the problem with path integration so in path integration effectively you. 00:08:11.994 --> 00:08:16.674 If you imagine dividing up your whole path into a series of small vectors, 00:08:16.854 --> 00:08:23.394 then you simply sum those vectors and minus the sum vector points you back home. 00:08:23.714 --> 00:08:30.154 Unfortunately, every step you take is going to have some error associated with 00:08:30.154 --> 00:08:34.594 it, particularly if you're being blown by the wind, which the ants often are. 00:08:34.594 --> 00:08:37.854 And so the errors accumulate and so 00:08:37.854 --> 00:08:41.274 when you return home the longer the 00:08:41.274 --> 00:08:44.674 path you've gone the more you're likely to miss the nest by so 00:08:44.674 --> 00:08:47.574 ants actually do very similar things 00:08:47.574 --> 00:08:50.794 to sailors used to do in 00:08:50.794 --> 00:08:54.934 which if a sailor was trying to head back to their port by dead reckoning they 00:08:54.934 --> 00:08:59.954 would always aim to one side of the port so that they knew that when they hit 00:08:59.954 --> 00:09:03.794 the coast they knew which way they had to turn to find it and ants do seem to 00:09:03.794 --> 00:09:07.854 do a similar thing in that they will go one side of the nest so that they know 00:09:07.854 --> 00:09:09.094 which side to bias their search by. 00:09:10.794 --> 00:09:16.254 So path integration is great and is essential for them to find their way home 00:09:16.254 --> 00:09:20.594 the first time, but from the very first trip back, they will learn visual cues, 00:09:21.494 --> 00:09:23.854 because envision the stability is in the world. 00:09:24.594 --> 00:09:28.174 The ants that we study in Australia, Molophorus, Bogotty, 00:09:28.854 --> 00:09:35.594 they don't use pheromones for navigation irrigation because they they burn off 00:09:35.594 --> 00:09:42.934 in the uh in the heat and um the cataglyphus would be very much the same um 00:09:42.934 --> 00:09:47.894 the wood ants we study in sussex in the lab um, 00:09:49.219 --> 00:09:51.899 pheromones the ground is quite unstable because there's quite a lot 00:09:51.899 --> 00:09:55.159 of rain um and so pheromones aren't 00:09:55.159 --> 00:09:58.199 much good um and so again they're primarily 00:09:58.199 --> 00:10:07.059 visual and i and all ants that can use that have vision do use vision and do 00:10:07.059 --> 00:10:12.279 seem to prioritize it over the other elements they might use to get home so 00:10:12.279 --> 00:10:16.939 um ants will have multiple strategies and as you're saying different species, 00:10:17.739 --> 00:10:22.859 some species may use chemical trials but what's quite common in ants is to use 00:10:22.859 --> 00:10:29.219 on top of path integration and compass sense some visual memory of the world 00:10:29.219 --> 00:10:34.559 and that can compensate for errors you might make with your path integration 00:10:34.559 --> 00:10:37.639 so you'd have two mechanisms that are complementary. 00:10:39.819 --> 00:10:43.579 Yes well a complementary and there 00:10:43.579 --> 00:10:46.479 is some evidence that they 00:10:46.479 --> 00:10:56.179 are used they're used together but it looks more like that they they just come 00:10:56.179 --> 00:11:01.039 to rely on one they will use the signals from one and then maybe use the other 00:11:01.039 --> 00:11:03.859 if if the world starts to look different for instance so i'll 00:11:03.939 --> 00:11:07.799 start using vision once they've learned the path they'll use vision and then 00:11:07.799 --> 00:11:11.879 if things start to look wrong maybe they'll turn back to path integration or 00:11:11.879 --> 00:11:14.139 go into a search behavior um, 00:11:15.299 --> 00:11:23.679 it's very difficult to tease those things apart in experiments in the natural world because um, 00:11:25.099 --> 00:11:32.919 you can't really put those two things in direct opposition very easily so the 00:11:32.919 --> 00:11:37.999 ants uh As you said, the ANTAD has a somewhat crude visual system. Yes. 00:11:38.739 --> 00:11:43.039 What is it that we think that they're attending to when they're using visual cues for navigation? 00:11:43.559 --> 00:11:50.239 Okay, so some of our recent experiments say that the skyline, 00:11:50.299 --> 00:11:54.539 Paul Graham and Ken Cheng's work have shown that the skyline, 00:11:54.579 --> 00:11:58.519 which is the shape of trees against the sky, 00:11:59.239 --> 00:12:04.379 is sufficient for them to be able to navigate, gate to be able to recover a direction home. 00:12:06.279 --> 00:12:11.259 And their work also showed that they didn't just use certain key prominent objects. 00:12:11.979 --> 00:12:18.659 So we think that it's likely to be some version of the, 00:12:20.007 --> 00:12:24.547 the shape that things make against the sky, which is very easy for them to pick 00:12:24.547 --> 00:12:27.427 out because they've got UV sensors. 00:12:29.007 --> 00:12:34.407 So they would, in some sense, encode the shape of the horizon as they were leaving 00:12:34.407 --> 00:12:39.067 the nest in order to be able to recognise when they're back there? 00:12:39.327 --> 00:12:43.467 Yeah, that would be the classic snapshot view that you would remember what the 00:12:43.467 --> 00:12:48.187 world looked like from the nest and a very low-dimensional parametrized version 00:12:48.187 --> 00:12:52.547 of the world, whether you remember it as an image, whether you remember it as a height map, 00:12:52.767 --> 00:13:02.047 whether you remember it as in other models as retinotopic positions of significant 00:13:02.047 --> 00:13:05.327 gaps in the world, those things are very difficult to say. 00:13:06.267 --> 00:13:09.547 But you remember that and then when you want to return to your nest, 00:13:09.747 --> 00:13:14.067 you just move to make the world look more like your memory. 00:13:14.067 --> 00:13:19.807 So that's the kind of hill climbing thing that if I move left and it looks a 00:13:19.807 --> 00:13:23.507 bit more like my memory, I can keep moving left or something. 00:13:23.827 --> 00:13:27.327 You just carry on. Yeah, you carry on in the direction you're going while it's getting more similar. 00:13:27.567 --> 00:13:32.507 But I mean, hill climbing strategies are known to get stuck. They do. 00:13:33.227 --> 00:13:38.187 So why doesn't the ant get stuck in local minima for its strategy? 00:13:38.187 --> 00:13:44.987 Primarily because if you're in an area and there aren't any obstacles in the 00:13:44.987 --> 00:13:49.547 way, if the region that you're in doesn't have any obstacles, 00:13:50.047 --> 00:13:53.807 then there won't be any local minima. 00:13:54.147 --> 00:13:57.427 And you can say that from what, from experimental findings? Yes, 00:13:57.467 --> 00:14:04.947 from some work that Jochen Zyl did originally, where he took panoramic images 00:14:04.947 --> 00:14:06.867 is from a series of natural environments. 00:14:07.087 --> 00:14:14.027 And over a cubic meter, there was a clear gradient in this kind of image space. 00:14:14.547 --> 00:14:19.227 And you can kind of do it mathematically as well. In a sense, 00:14:19.327 --> 00:14:20.787 it's kind of the inverse of optic flow. 00:14:22.507 --> 00:14:27.167 That if you haven't got occlusion, then things move in a very predictable way. 00:14:27.667 --> 00:14:31.907 Now the caveat with all that is that for this to work, you have to be lined 00:14:31.907 --> 00:14:37.527 up in the same, you have to be yeah you have to be in the same orientation you 00:14:37.527 --> 00:14:40.307 were for all these images now. 00:14:42.527 --> 00:14:48.707 Most of the visual homing algorithms this is true and this comes from the original 00:14:48.707 --> 00:14:53.567 models were based on models of bees and wasps who do line up very precisely 00:14:53.567 --> 00:14:56.787 in a certain orientation before entering in the nest. 00:14:59.167 --> 00:15:03.787 And for ants, this is somewhat more difficult. So that is a challenge for the models. 00:15:05.287 --> 00:15:09.807 And so we've proposed that visual memories might be used in a slightly different 00:15:09.807 --> 00:15:12.587 way, whereby instead of trying to, 00:15:14.607 --> 00:15:17.027 you try and recover the direction to your. 00:15:19.611 --> 00:15:24.931 Yeah you try and recover the direction try and use the images as a visual compass, 00:15:26.771 --> 00:15:32.991 so with a visual compass what you do is you remember the view from your goal 00:15:32.991 --> 00:15:36.111 when you're in a certain orientation or more properly you remember the view 00:15:36.111 --> 00:15:42.911 when you're pointing at your goal and from nearby positions if you rotate on 00:15:42.911 --> 00:15:45.871 the spot when you're in that same orientation orientation, 00:15:46.071 --> 00:15:52.211 then you will again find a sort of minimum in this image difference landscape space. 00:15:52.911 --> 00:15:56.231 And so you'll be able to recover the orientation of those images. 00:15:57.771 --> 00:16:02.791 Now, that doesn't seem very good for homing. But if you then remember a series 00:16:02.791 --> 00:16:08.651 of images, when you're pointed at your goal from points surrounding the goal, 00:16:08.771 --> 00:16:11.071 then you should be able to get back. 00:16:11.071 --> 00:16:12.331 And we've modelled this and you can 00:16:12.331 --> 00:16:18.111 then get back from anywhere within a reasonable range around that goal. 00:16:20.171 --> 00:16:24.971 So how would you know that the ant was using that strategy? Is there some behavioural 00:16:24.971 --> 00:16:27.351 marker that it's encoded a snapshot? 00:16:28.651 --> 00:16:37.311 Well, we have observed ants visually scanning the world, or they appear to visually scan the world. 00:16:37.311 --> 00:16:42.951 So we've got some high-speed recordings, and this is work in preparation of 00:16:42.951 --> 00:16:46.591 ants, of Malophorus, the Australian desert ant. 00:16:47.091 --> 00:16:53.431 When it's challenged with a new environment, sometimes just sort of naturally, 00:16:54.151 --> 00:16:57.331 it has this kind of saccadic motion where it will run along for a bit, 00:16:57.411 --> 00:17:02.591 then it will stop and it will turn on the spot, seem to pick a direction and 00:17:02.591 --> 00:17:05.491 then head off again in a straight line and then again turn the spot. 00:17:05.491 --> 00:17:09.071 And it does more of these scans when it's 00:17:09.071 --> 00:17:14.891 in an unfamiliar environment and it also seems these scans are directed towards 00:17:14.891 --> 00:17:20.651 the more familiar part of the environment and some really elegant work by Paul 00:17:20.651 --> 00:17:27.511 Graham and Ken Chang and Antoine Wistrach so. 00:17:29.973 --> 00:17:37.693 We at least know that the behaviour is there that would serve to facilitate our model. 00:17:38.253 --> 00:17:44.493 The other thing we also do see in wood ants is they tend to walk in a sinusoidal 00:17:44.493 --> 00:17:52.633 path, which again would enable you to behaviourally scan the world as you went along. 00:17:53.033 --> 00:17:56.253 So as well as having behavioural evidence 00:17:56.253 --> 00:17:59.353 that points towards this possibility of 00:17:59.353 --> 00:18:03.233 them encoding snapshots um and this 00:18:03.233 --> 00:18:06.553 is behavioral evidence from experiments in 00:18:06.553 --> 00:18:10.413 in in natural environments yeah um you 00:18:10.413 --> 00:18:14.893 also are doing computational modeling work to actually demonstrate that those 00:18:14.893 --> 00:18:20.033 hypotheses about the mechanisms could operate to control say a simple robot 00:18:20.033 --> 00:18:26.213 yep and so um what similarities you think there need to be between the robot 00:18:26.213 --> 00:18:28.353 on the ant in order to test this model? 00:18:28.593 --> 00:18:30.853 To test it so that it's a good model of an ant. Yeah, I mean, 00:18:30.853 --> 00:18:32.413 can it be a very crude robot model? 00:18:32.793 --> 00:18:38.953 I think that the real difficulty, and this is going to be the sticking block, 00:18:39.553 --> 00:18:44.053 is to get the eye down to ant level. 00:18:44.893 --> 00:18:53.373 And that's the real problem. I think we could do, I think we have done some tests on indoor robots. 00:18:53.633 --> 00:18:56.213 We've got a gantry robot, but that's quite precise. 00:18:56.713 --> 00:19:00.393 And we've kind of cluttered up the environment the world enough um 00:19:00.393 --> 00:19:03.853 but we don't need we can do it with any sort 00:19:03.853 --> 00:19:07.133 of robot platform and we should be doing this in the next year 00:19:07.133 --> 00:19:11.313 you mean the strong test of actually putting a robot in the australian desert 00:19:11.313 --> 00:19:15.693 where the ants are and showing that it can yeah i think the first test would 00:19:15.693 --> 00:19:20.513 be we would do it in natural environments around sussex i think there's no point 00:19:20.513 --> 00:19:25.593 doing it in the australian desert where it's a bit hot um until we've got 00:19:25.673 --> 00:19:30.013 it working until we can test whether it's working in, in Sussex when it's a 00:19:30.013 --> 00:19:33.133 foot above the ground, say. Um, and. 00:19:35.456 --> 00:19:40.416 Yeah, I think that's our next step. So I think that could be quite a crude robot. 00:19:41.016 --> 00:19:42.356 We don't really care about speed. 00:19:43.476 --> 00:19:51.556 The ant has to deal with an uneven ground, which, because it's small, is very uneven for it. 00:19:52.496 --> 00:19:55.876 And we're not going to be able to recreate that, I think, with any robot we have. 00:19:56.696 --> 00:19:59.976 So there are certain compromises that come into... 00:20:00.636 --> 00:20:03.356 We have to take into consideration in building a robot model of the system. 00:20:03.356 --> 00:20:07.296 But are there, even though it's a crude model, are there already some insights 00:20:07.296 --> 00:20:10.876 that you've got through this modeling approach that you think you may not have 00:20:10.876 --> 00:20:12.776 had from a purely experimental one? 00:20:12.776 --> 00:20:18.836 Um definitely i think the the the thing that we have done i think we've has 00:20:18.836 --> 00:20:26.556 given us a lot of insight is is taking panoramic images from the positions that ants were navigating in, 00:20:27.096 --> 00:20:30.696 and being able to see what their visual input. 00:20:32.396 --> 00:20:35.136 Would have been what the what the input to their eyes 00:20:35.136 --> 00:20:38.176 would have been it doesn't tell us what they're seeing doesn't really 00:20:38.176 --> 00:20:40.936 tell us how the image is then being processed but we 00:20:40.936 --> 00:20:44.696 can tell what the raw visual input is um and um 00:20:44.696 --> 00:20:47.496 this is following again following on from work of yakinzade sort of 00:20:47.496 --> 00:20:50.556 pioneered this work um and it is 00:20:50.556 --> 00:20:55.416 staggering to see um how 00:20:55.416 --> 00:20:59.116 different the world looks most of the world is sky and 00:20:59.116 --> 00:21:03.336 a lot of the world is the ground and there's not much of the world carries any 00:21:03.336 --> 00:21:10.816 kind of a signal for navigation um and so one of the slides we we like to show 00:21:10.816 --> 00:21:14.156 or one of the images we like to show is an image of the environment where as 00:21:14.156 --> 00:21:15.856 a human you instantly pick out the house. 00:21:16.496 --> 00:21:21.036 And when you show it as a panoramic image, the house just kind of disappears into the background. 00:21:21.916 --> 00:21:26.696 And when you blur it down to insect resolution, you really don't see anything. 00:21:28.376 --> 00:21:31.556 So that really has given us a lot of insight and 00:21:31.556 --> 00:21:35.076 has enabled us to do some interesting modelling 00:21:35.076 --> 00:21:38.136 about the scale over which one single visual 00:21:38.136 --> 00:21:40.876 memory would serve to allow you to 00:21:40.876 --> 00:21:45.976 navigate and the insect 00:21:45.976 --> 00:21:52.796 navigation the is a very collaborative area and lots of other people are doing 00:21:52.796 --> 00:21:58.116 this sort of methodology and I think it's really helping the field to see you 00:21:58.116 --> 00:22:03.636 know how different does the world look from this position do I need to I, 00:22:05.181 --> 00:22:09.301 Would the ant need to have a cognitive map to get back from A to B, 00:22:09.381 --> 00:22:12.681 or can it simply navigate with a simple strategy? 00:22:14.181 --> 00:22:22.341 And just getting an idea of what the insect is seeing has been invaluable, I think. 00:22:24.041 --> 00:22:28.981 Do you think that there are useful ideas people can take from this for designing 00:22:28.981 --> 00:22:31.901 artefacts, say micro-robots? 00:22:31.901 --> 00:22:40.741 Well, I think that most people tell me that the SLAM problem is solved and that… 00:22:40.741 --> 00:22:42.041 That's the mapping problem. 00:22:42.181 --> 00:22:48.701 The mapping problem. So the autonomous navigation can be solved by probabilistic 00:22:48.701 --> 00:22:52.841 techniques where essentially you track a large number of features of the world 00:22:52.841 --> 00:22:59.201 and you integrate that with information about your position and your likely movement. 00:22:59.201 --> 00:23:02.741 You integrate probabilistically and you localize 00:23:02.741 --> 00:23:10.801 both yourself and the features of the world in a map and they have um in the 00:23:10.801 --> 00:23:15.661 DARPA grand challenges they navigate very long distances using these methods 00:23:15.661 --> 00:23:21.001 however they take a lot of computation awful lot of computation and so they're quite 00:23:21.081 --> 00:23:28.841 heavy are both with batteries and with power so i think the applications would be for. 00:23:30.281 --> 00:23:36.681 Uavs unmanned air vehicles anywhere where the sort of power and weight considerations 00:23:36.681 --> 00:23:45.861 are important so maybe also space exploration um and also the other thing is other places where 00:23:45.941 --> 00:23:48.201 there is no GPS, a GPS denied environment. 00:23:48.501 --> 00:23:52.501 If you have a GPS, you probably want to use it. But I think in environments 00:23:52.501 --> 00:23:56.421 where you want a low cost solution for whatever reason, 00:23:57.161 --> 00:24:03.141 and you haven't got a global GPS signal, then I think these methods, 00:24:03.541 --> 00:24:06.761 they're surprisingly robust. 00:24:08.683 --> 00:24:14.783 And I think is that almost a lesson from studying insect nervous systems for 00:24:14.783 --> 00:24:16.503 designing technology more generally, 00:24:16.683 --> 00:24:24.823 that what looks like, what is very robust behavior can be generated using a 00:24:24.823 --> 00:24:27.023 relatively simple algorithm. 00:24:27.223 --> 00:24:31.763 It doesn't have to be as sophisticated as what you might imagine at first. 00:24:31.943 --> 00:24:35.263 No, definitely. Definitely. I think that is one of the great lessons. 00:24:36.863 --> 00:24:44.783 With simple eyes, very poor resolution eyes, and with a brain that can't do 00:24:44.783 --> 00:24:50.443 a lot of computation, or can't do very heavy computations, and probably hasn't 00:24:50.443 --> 00:24:51.523 got a very big memory load. 00:24:54.363 --> 00:25:01.703 Answer fantastic navigators. I mean, bees go miles with not much bigger brains. 00:25:03.443 --> 00:25:06.483 And with similar mechanisms or does it get more complex? 00:25:07.623 --> 00:25:14.223 Oh, I would like to, if I was to speculate I'd say I don't see why not I don't 00:25:14.223 --> 00:25:19.123 see why not There is a controversy within the field as to whether bees might use cognitive maps, 00:25:20.143 --> 00:25:25.743 I think the consensus would be that they don't I think they would use similar 00:25:25.743 --> 00:25:31.523 mechanisms I know several people seem to think that a lot of the long distance homing would be, 00:25:32.543 --> 00:25:38.663 based on the dormitory, um, I don't see why they wouldn't use visual information. 00:25:39.783 --> 00:25:45.803 Um, that the horizon, what we've shown when we've shown our algorithms can easily 00:25:45.803 --> 00:25:50.323 work over a hundred meters in an open environment, um, and you can kind of make 00:25:50.323 --> 00:25:51.963 the world open by flying up. 00:25:52.383 --> 00:25:56.183 And the shape of the horizon is a really, really strong queue and a robust one, 00:25:56.263 --> 00:25:58.303 presumably, but it's not going to move very much. 00:25:58.363 --> 00:26:01.963 I don't think, um, particularly over their lifetime. time so I don't see why 00:26:01.963 --> 00:26:05.523 they wouldn't and you kind of only need to match. 00:26:07.270 --> 00:26:12.950 Gross shape and you can get roughly the way back um i think insects have, 00:26:14.610 --> 00:26:17.450 always tend to have multiple strategies which makes 00:26:17.450 --> 00:26:20.350 it to do any one task which makes it complicated 00:26:20.350 --> 00:26:23.410 when you're trying to study them and which is why they're interesting i think 00:26:23.410 --> 00:26:28.250 um but they need to because they need to be robust and they need to get home 00:26:28.250 --> 00:26:34.830 um and so i think that there will always be a combination of strategies so yeah 00:26:34.830 --> 00:26:37.650 i would like to think that bees and wasps. 00:26:37.650 --> 00:26:42.970 We know they use vision certainly for the last part of their to find their nest. 00:26:43.090 --> 00:26:45.550 So I don't see why they wouldn't over a longer distance. 00:26:46.470 --> 00:26:48.790 Thanks very much for talking to us, Andy. Thank you. 00:26:48.400 --> 00:26:53.840 Music. 00:26:53.610 --> 00:26:59.230 The CSN podcast was produced by the Convergent Science Network of Biometrics 00:26:59.230 --> 00:27:06.030 and Biohybrid Systems, a project funded by the European PN7's Research Framework Programme. 00:27:07.130 --> 00:27:12.490 For more interviews, recorded lectures or upcoming conferences in the field 00:27:12.490 --> 00:27:18.730 of biometrics and biohybrid systems, go to csnnetwork.eu. 00:27:19.600 --> 00:27:26.640 Music.