Cooperative multi‑agent tasks often impose a shared objective or safety constraint. Traditional approaches use a team‑level violation cost to measure constraint breaches, but they cannot identify which agents are responsible. SafeDiSTL introduces a constrained multi‑agent reinforcement learning framework based on Signal Temporal Logic (STL) robustness that enables credit assignment even when constraints are measured jointly.
The key idea is to treat each violation as a running‑maximum cost increment and accumulate these increments into a smooth full‑episode robustness metric. Because robustness is differentiable with respect to the system dynamics, we can compute action sensitivities that indicate how each agent’s action contributes to a cost increment at the same decision time. These sensitivities are used to allocate the increment across agents, taking delayed effects into account, thus providing accurate per‑agent credit.
We evaluate SafeDiSTL on three cooperative scenarios:
- Payload transport with multiple robots
- Connected navigation for a fleet of vehicles
- Networked traffic control at intersections The experiments show that SafeDiSTL improves safe‑success rates by 1.6‑5.1× in transport, raises navigation safety by 11.1%, and increases traffic throughput by 50‑68% compared with constrained baselines, while preserving the performance of unconstrained policies.
These results demonstrate that combining STL robustness with differentiable dynamics for credit allocation yields safer and more efficient policies under shared constraints.
Review