Distributed Learning for Controlling Modular Robots

What: Recently there has been an important research effort into modular, distributed robotics and in particular, self-reconfiguring robotics [2, 5, 8]. Issues with designing controllers for such systems range from constructing motor control primitives to ensuring cooperation between modules. For simpler tasks, such as locomotion in one direction, hand design is easy. However, as modular robots are tried in more complex domains or at more intricate tasks, designing effective and efficient controllers becomes a problem. We propose to have such robots learn their behaviors instead. We research the strategies and algorithms for learning controllers for self-reconfigurable robotic systems, in which each element has some computational and motor power, focusing in particular on applying reinforcement learning techniques. More generally, we are interested in developing distributed reinforcement learning algorithms that can be used to control distributed teams of robots and sensors.