A Simulation Platform for MARL Training and Evaluation in Swarm Confrontation

In swarm confrontation, robots must swiftly formulate strategies in transient environments, a challenge well-suited for multi-agent reinforcement learning (MARL). However, existing platforms suffer from the lack of comprehensive confrontation scenario modeling and scalable frameworks, hindering MARL's widespread applications. We introduce a novel platform for training, simulating, and evaluating MARL algorithms in swarm confrontation tasks. It constructs a holistic simulation framework by integrating robot, environment, and rule models for complex confrontation scenarios. Equipped with a decentralized task allocator and path planner for each robot, the platform enables scalable cooperation across dynamic environments. Extensive experiments demonstrate that our platform simulates confrontations involving up to twenty agents per side, providing empirical guidance for algorithm selection in various settings.