Distributed Reinforcement Learning Based Optimal Controller for Mobile Robot Formation

This paper addresses a problem of attaining desired geometric formation for a group of homogeneous robots using distributed reinforcement learning. The challenges for learning by experience requires huge time and data samples. In multi-agent system (MAS), individual learning becomes more complex as it has to cooperate with its neighboring agent. In this work, a group of homogeneous robots models a single controller while performing a task in a decentralized manner. The framework uses an actor-critic architecture for local learning and its update law is identified using Lyapunov stability analysis. However, a global single controller is achieved by using average consensus protocol. Simulation as well as the experimental results have been given to demonstrate the proposed algorithm.