Adaptive Learning of Escape Strategy for an Autonomous Robot
Yusuke Kon, Sadayoshi Mikami · The proceedings of the JSME annual meeting · 2004
A serious problem for autonomous robots working in out-door environments is its possibility to be stuck by unknown and unobservable obstacles. The main causes are due to slip of wheels by slippery surface, inclination of terrain, and some small obstacles. This research-is to realise a rapid escape from stuck situations, by acquiring a sequence of effective actions of tire drive. The strategy of finding a sequence from trial and error is just the same structure as the reinforcement learning methods. We apply the reinforcement learning method that can rapidly find and converge into a good sequence. From computer simulations, we clarify that the on-policy method with eligibility trace is better to be used for this application. The method is now under way to be implemented into a robotic mowing machine.