We introduce MA‑SafeDiffuser, a multi‑agent extension of the single‑agent SafeDiffuser that equips diffusion‑based trajectory planners with finite‑time invariance guarantees for joint safety specifications. The method embeds control barrier function (CBF) constraints into the reverse‑diffusion updates, directly constraining sampled trajectories.
Core Techniques
- Joint Safe Set: Constructed as the intersection of per‑agent safe sets and all pairwise barrier sets, defining a global safety domain.
- Constrained Denoising: Derive both centralized and decentralized (communication‑aware) reverse‑diffusion procedures, with provable invariance under mild assumptions.
- Local Traps & Deadlocks: Mitigated by time‑varying specifications and liveness CBFs, preventing local minima and deadlock scenarios.
- Benchmark Suite: A lightweight multi‑agent Maze2D domain is provided to evaluate the trade‑off between safety and planning quality.
Empirical Findings
Compared with unconstrained diffusion baselines, MA‑SafeDiffuser achieves a substantial reduction in safety violations while maintaining comparable planning performance. Full algorithms, proofs, and reference implementations are released.
Review: MA‑SafeDiffuser seamlessly merges safety‑critical control theory with generative diffusion models, delivering a verifiable toolkit for safe multi‑agent planning that balances distributed flexibility with rigorous centralized guarantees.