Decentralized Multi-Agent Pursuit Using Deep Reinforcement Learning

Abstract

Pursuit-evasion is the problem of capturing mobile targets with one or more\npursuers. We use deep reinforcement learning for pursuing an omni-directional\ntarget with multiple, homogeneous agents that are subject to unicycle kinematic\nconstraints. We use shared experience to train a policy for a given number of\npursuers that is executed independently by each agent at run-time. The training\nbenefits from curriculum learning, a sweeping-angle ordering to locally\nrepresent neighboring agents and encouraging good formations with reward\nstructure that combines individual and group rewards. Simulated experiments\nwith a reactive evader and up to eight pursuers show that our learning-based\napproach, with non-holonomic agents, performs on par with classical algorithms\nwith omni-directional agents, and outperforms their non-holonomic adaptations.\nThe learned policy is successfully transferred to the real world in a\nproof-of-concept demonstration with three motion-constrained pursuer drones.\n