Distributed reinforcement learning for self-reconfiguring modular robots

In this thesis, we study distributed reinforcement learning in the context of automating the design of decentralized control for groups of cooperating, coupled robots. Specifically, we develop a framework and algorithms for automatically generating distributed controllers for self-reconfiguring modular robots using reinforcement learning. The promise of self-reconfiguring modular robots is that of robustness, adaptability and versatility. Yet most state-of-the-art distributed controllers are laboriously handcrafted and task-specific, due to the inherent complexities of distributed, local-only control. In this thesis, we propose and develop a framework for using reinforcement learning for automatic generation of such controllers. The approach is profitable because reinforcement learning methods search for good behaviors during the lifetime of the learning agent, and are therefore applicable to online adaptation as well as automatic controller design. However, we must overcome the challenges due to the fundamental partial observability inherent in a distributed system such as a selfreconfiguring modular robot. We use a family of policy search methods that we adapt to our distributed problem. The outcome of a local search is always influenced by the search space dimensionality, its starting point, and the amount and quality of available exploration through experience. We undertake a systematic study of the effects that certain robot and task parameters, such as the number of modules, presence of exploration constraints, availability of nearest-neighbor communications, and partial behavioral knowledge from previous experience, have on the speed and reliability of learning through policy search in self-reconfiguring modular robots. In the process, we develop novel algorithmic variations and compact search space representations for learning in our domain, which we test experimentally on a number of tasks. This thesis is an empirical study of reinforcement learning in a simulated latticebased self-reconfiguring modular robot domain. However, our results contribute to the broader understanding of automatic generation of group control and design of distributed reinforcement learning algorithms. Thesis Supervisor: Daniela Rus Title: Professor