Reproduce “Generalizability of Graph Neural Networks for Decentralized Unlabeled Motion Planning” on OpenResearch

About this paper

Shreyas Muthusamy, Damian Owerko, Charilaos I. Kanatsoulis, Saurav Agarwal, Alejandro Ribeiro

A Graph Neural Network-based decentralized control policy, trained with a hybrid imitation learning and reinforcement learning approach, enables multi-robot systems to achieve near-zero collisions and high task completion for unlabeled motion planning. This policy demonstrates strong generalizability, controlling up to 500 robots while trained on 100, effectively balancing optimal assignment with robust collision avoidance.

Key insights

• Combining imitation learning for initial policy shaping (from a globally optimal assignment expert) with reinforcement learning for fine-tuning enables balancing global task efficiency and local safety constraints.
• Graph Neural Networks provide a robust framework for decentralized multi-agent control, naturally handling varying numbers of agents and local communication.
• Carefully designed reward functions in RL can effectively incentivize complex behaviors like collision avoidance while maintaining task performance.

Summary by alphaXiv.