Reproduce “Decentralized Multi-Agent Pursuit using Deep Reinforcement Learning” on OpenResearch

About this paper

Cristino de Souza Jr, Rhys Newbury, Akansel Cosgun, Pedro Castillo, Boris Vidolov, Dana Kulic

This research introduces a Deep Reinforcement Learning framework for decentralized multi-agent pursuit, enabling non-holonomic robots to cooperatively capture an evader using only local observations. The method successfully transfers learned policies from simulation to a team of real drones, demonstrating emergent cooperative strategies and improved performance compared to classical and other DRL baselines across various challenging scenarios.

Key insights

• Deep Reinforcement Learning, when combined with appropriate state representations and reward shaping, can enable non-holonomic multi-robot teams to learn complex, cooperative pursuit strategies from local observations.
• Curriculum learning significantly accelerates training convergence and yields more robust policies in complex multi-agent DRL tasks with sparse rewards, fostering the development of sophisticated behaviors.
• Careful simulation design, including realistic kinematic constraints and local observation models, allows for direct and successful sim-to-real transfer of DRL policies to physical robotic platforms.

Summary by alphaXiv.