Decentralized Multi-Agent Motion Planning in Dynamic Environments

Abstract

In this paper we present a decentralized multi-agent motion planning algorithm for navigation in dynamic environments. Each agent constructs a graph of boundary value problems in the environment considering their own kinodynamic constraints using a learning-based motion planning framework. A game-theoretic approach is then used by each agent to select their individual path through the environment while considering the planned motion of other agents. This path is updated online to ensure collisions are avoided, and to provide a method of counteracting the freezing robot problem. The effectiveness of the algorithm is illustrated in simulations.