New RL Framework Ensures Safety in Multi-Agent Systems

Multi-agent systems are increasingly deployed in safety-critical applications, demanding coordinated behavior under strict safety constraints. Current methods face a dilemma: learning-based approaches offer strong performance but lack theoretical safety guarantees, while control-theoretic methods ensure safety but often result in overly conservative actions. This paper introduces a hierarchical multi-agent reinforcement learning framework designed to overcome this trade-off. It enforces hard safety constraints at a low level through a constraint manifold, operating under mild assumptions. Simultaneously, it facilitates effective coordination via high-level policy learning. The proposed approach provides theoretical safety guarantees within multi-agent settings and establishes stationary learning dynamics, leading to stable and efficient training. Empirical evaluations demonstrate that this method achieves competitive performance while maintaining nearly perfect safety rates. Furthermore, it exhibits strong generalization capabilities when faced with varying numbers of agents and obstacles.