Reinforcement learning for distributed cooperative control

Abstract

Distributed robotic systems, such as teams of vehicles, robotic swarms, or modular robots - where each individual agent has to engage in decision-making under uncertainty and coordinate their activity - present a challenge in controller design. Also, they need to be robust and adaptable to changing environments. Both these challenges can be addressed by the application of reinforcement learning.

In this talk, we use a class of policy search algorithms to learn distributed controllers, in particular for self-reconfiguring modular robots. Our experiments in simulation demonstrate the challenges of the problem, and we present novel algorithmic extensions and compact search space representations to address some of these challenges.