Contract-based shielding enables safe multi-agent RL without centralized control

A new paper from researchers Omar Adalat, Edwin Hamel-De le Court, and Francesco Belardinelli tackles the challenge of safe coordination in multi-agent reinforcement learning (MARL). When agents act independently under decentralised execution, their individual actions may conflict with global safety constraints. Traditional decentralized shields enforce safety by restricting each agent's permissions, but this often discards team-optimal behaviors that are only safe through coordination. The authors propose contract-based compositional shielding: each agent selects from a library of local LTL_safe obligations, and the entire tuple of obligations must jointly imply a global safety specification. A non-stationary multi-armed bandit chooses the optimal tuple during learning, maximizing team reward without violating safety. The approach is evaluated across 6 environments and 15 algorithmic variants, showing that it recovers team-optimal safe behavior without requiring centralised runtime oversight.

The key innovation is that agents rely on each other's local obligations as assumptions, certified simultaneously to allow projection into local action masks. This means the system can guarantee safety end-to-end while still enabling coordinated, high-reward strategies that purely factorised permissions would block. The method is particularly relevant for autonomous driving fleets, warehouse robots, or drone swarms where global safety cannot be enforced by any single agent alone. By shifting from reactive shielding to contract-based ex-ante coordination, the paper opens a practical path toward safe, scalable multi-agent systems.