Bayesian issues in ant navigation

Our brains have to deal with ambiguity and uncertainty, and an increasingly popular explanation of how they do so is based on Bayesian reasoning. In essence, this says we estimate the probability of a certain state of affairs (such as 'I am at home') on the basis of both current sensory inputs ('This looks like my house') and prior expectations ('Given my starting location, and the speed I was travelling, I wouldn't expect to be home yet'). Bayes theorem tells us how we should combine these factors to obtain the best estimate of our current state. But is this form of reasoning universal? An ideal way to investigate this issue is to look at 'simple' animals that have to solve analogous problems. And an effective way to test our understanding of what these animals do is to implement and test our hypotheses in robot models that operate in the same sensory environment. A clear example of an animal solving such problems is found in desert ants, who forage individually and without the use of chemical trails, yet can efficiently relocate their nest or a food source over long distances in barren or complex environments. Recent studies have shown that ants can individually learn and recall specific routes through cluttered environments that force detours and prevent the use of distant landmarks. Ant navigation depends on two main mechanisms: they can keep track of how far they have moved and in which direction from the nest and continuously update a vector that points back home; and they can recognise familiar visual surroundings and use these to determine which way to go. Do they integrate these cues in an optimal fashion? What if one or other cue is more or less variable? Can they use one of these cues to disambiguate the other? We can make the investigation of these issues rigorous and quantitative by drawing on methods developed for robot navigation. We will first determine what ants actually see as they develop new routes, by following ants as they forage, and capturing images from the ant's eye point of view. We will feed this information into algorithms that should be able to learn a map of the area. We can systematically vary the type of information available, its reliability, and the computational methods used to update the map, and compare the performance to ants. Further experiments to see what the ants do when the same variables are manipulated will serve to evaluate the models. Finally, the models will also be tested in the real world by implementing them on a small robot able to navigate in the ant environment.

We want to answer the question: do all animals have Bayesian brains? Ant navigation is an ideal test case, as a complex, cognitive capability displayed by a small brained animal, and as a problem for which there is well developed Bayesian theory in robotics. We will first gather rich data about the sensory experiences of ants navigating in their natural habitat. An ant nest will be moved to a new area and access points restricted. Ants will be followed as they develop new foraging routes. An image database will be collected and used to reconstruct a virtual ant environment for use in modelling. We will then test how different forms of Bayesian mapping algorithms, in particular topological forms of visual Simultaneous Localisation and Mapping (SLAM), perform when provided with the same input as the ant. Due to the high visual similarity of ant environments, this is not a trivial technical problem, and we will be able to use the virtual environment to selectively introduce different forms of uncertainty and noise and make quantitative predictions for ant behaviour. The search densities for ants returning to a nest under path integration, visual guidance or both can be measured, and Bayesian cue integration theory can be used to assess if ants optimally combine the cues, or alternate between them, and if either strategy is altered as the variance of one or other cue is increased. Ants following routes should face the problem of visual aliasing, and we will use the predictions of Bayesian filtering algorithms to evaluate the extent to which they use prior expectation to disambiguate cues. Revised models based on these results will be implemented on a small robot (surveyor SRV-1 black-fin) which will carry a panoramic camera and polarised light compass sensor, and be deployed in the real ant environment to obtain to obtain real-world evaluation.

Although the project outlined here is primarily basic science, there are several ways in which it may have broader societal and economic impact. The methodologies proposed in our work have implications for science policy linked to public views on animal experimentation. We are proposing two ways in which questions about behaviour, cognition and intelligence can be answered, or at least refined, prior to undertaking experiments on vertebrates (including humans), allowing such experiments to be reduced or made more productive. The first is the use of insects as our target system, and their study in ecologically relevant contexts. The task of navigation is broadly similar across a wide variety of animals, and there is likely to be at least some convergence in the underlying algorithms. The second is the use of well-grounded modelling to test hypotheses and produce explicit predictions in advance of experimentation. This allows refinement of the paradigms to carry out only the most informative manipulations. In connection with this issue, we believe important socio-cultural insights can be provoked by the comparison of human behaviour to very different species, promoting deeper ecological appreciation. From another perspective, there are interesting philosophical and ethical considerations raised by mechanistic explanations of behaviour that draw strong parallels between animals and robots. We have a range of connections to public dialogue on these issues as outlined in our attached 'pathway to impact'. Another more practical area in which the research should have some impact is in commercial applications of robotics. Robotics is widely predicted to have an increasingly important economic role. Navigation is one of the fundamental capabilities required for autonomous robots to be employed in areas such as exploration of hazardous environments and disaster areas; yet existing methods are not effective, or not efficient, in unstructured environments. Understanding how the ant solves this problem, and building a robot system to test proposed mechanisms in the same conditions as the ant, could lead directly to new solutions with potential for commercial development. We have a direct route for taking any such developments forward through Informatics Ventures, a scheme funded by Scottish Enterprise and the European regional development fund which provides networking and connections to sources of venture capital for informatics researchers. Finally, we anticipate direct social impact through science communication activities. Our work, connecting biology and robotics, has consistently attracted media interest. It has also been used as the basis of talks and workshops at science festivals, and in events for school children to promote recruitment into science and engineering study. We plan to continue our direct involvement in such activities as outlined in the 'pathway to impact'.