Intelligent Robotics Group: Machine Learning

Integrative Computation For Autonomous Agents: A Novel Approach Based On The Vertebrate Brain (Reverb)

Professor Mark Lee

Involving the universities of Sheffield, Wales (Aberystwyth), Manchester, Bristol, Oxford Cambridge, Dundee, and B A E Systems Plc.

In many situations it is better to have autonomous agents or robots to "do things that humans don't want to do, in a places they don't want to be, at times they don't want". For example, it would be very useful to have robots that can perform important tasks in dangerous, hostile or inaccessible environments, monitor the environment and the performance of humans for signs of danger, provide assistance to the disabled, aged or infirm, or have artificial agents or virtual characters in computer games that can entertain. However, what is required in each of these situations is for the agent to respond flexibly and adaptively to unpredictable circumstances. To do this successfully it needs to perform many different behaviours, and have the capacity to switch between tasks as circumstances dictate. That is, the agent must combine or integrate all its possible actions and behaviours harmoniously so that it does the right thing at the right time in a rapidly changing world, and changing internal motivations (just like animals and humans). It also needs the ability to learn from its mistakes and successes because it is impossible to pre-programme such a device to deal with all the possibilities it might face. While these requirements are easy to describe, they have proved difficult to achieve in practice, as recently shown by the general failure of contestants in the recent race for autonomous vehicles organised by the American military http://www.darpa.mil/grandchallenge/. This research programme aims to solve this problem of behavioural integration by using our knowledge of brain systems, which appear to solve the multi-tasking problem almost effortlessly. Thus, we will try to 'unravel' the mechanisms the brain uses and build models based on these mechanism that will run on computers, and which can be used to control the behaviour of artificial agents (e.g. robots). In particular, we aim to use information about relevant parts of the brain to build a robot that can do three things: (I) recognise objects using a visual system with a narrow field of view, and use this information to retrieve certain objects (ii) detect and direct attention to sudden events in the wider field of view (iii) detect static features in the wide field , and direct its narrow field visual system to them for closer inspection. All these capabilities will be combined within an overall control system that makes use of a central selection mechanism (just as we believe occurs in the brain). The next stage of our research will again depend on our knowledge of the brain to build into our agent the ability to modify its behaviours on the basis of internal 'motivations', and changes in external context (for example the presence of distracting events). We will also build into the model the capacity to learn from past experience in just the kind of way that we think humans do. After we have constructed our robot we then want to 'stand back' from the particular model we have built and ask the question 'what general features of the model gave it its ability to integrate its behaviours successfully?' By answering this we aim to be able to transfer our work into a wide range of robots designed for many different tasks. To undertake this large and multifaceted project many different forms of expertise will be required. We have therefore assembled a collection of mathematicians, control engineers, computer modellers and neuroscientists who's collaboration on all aspects of the project will be essential. In addition, there will be a significant contribution from an industrial sponsor, BAE systems, who has a strong commercial interest in developing autonomous agents who's behaviour is both flexible and adaptable. Starts: 4 April 2005 Ends: 3 April 2010 £ Value: 1,893,234