Raymond's Homepage
| In the grand tradition of those undertaking Ph.D. research, my current exact topic is, well, not quite so exact ... In general terms, I'm working on ways in which robot systems may be endowed with the ability to learn to accomplish tasks in a real world, continuous, robot-centric environment. My application domain is that of the RoboCup Rescue Robot League and my goal is to be able to tell the robot to "move forward 10m and stop" and have it do that whilst traversing unstructured terrain, such as random stepfields. The trick however is to do so using behaviours that have not been hand coded - we intend to use reinforcement learning! To make the task somewhat interesting, the robot we've chosen to apply this to is the Redback robot (see below and http://redback.web.cse.unsw.edu.au/ ) which has 4 degrees of freedom - the tracks on each side and the flippers at either end. Controlling this robot is a highly non-trivial exercise, even for a human, and can become quite difficult as it is often not possible to traverse a variety of obstacles without doing particular actions. Clearly plain reinforcement learning (and even any real variant thereof) will have a very difficult time with such a task so to make life easier, I'm looking at ways of incorporating background knowledge in order to reduce the search space. Currently I'm looking at behavioural cloning and the use of fast, low fidelity environment simulators in order to define a subspace of the search space that the reinforcement learner in the real world can then have a shot at. We've already had some success in getting a cloned behaviour working for our CASTER robot that allows it to traverse some stepfields so it's a start! Challenges that I'll probably need to address in the process of developing such a system include:
|
