Safe Offline Multi-Agent Reinforcement Learning
Combining diffusion models and control barrier functions for safe offline multi-agent decision making.
This project investigates safe offline multi-agent reinforcement learning using generative trajectory models.
We combine a return-conditioned multi-agent diffusion model with individual Control Barrier Functions (CBFs) to improve safety during trajectory generation while preserving coordination among agents.
The work was accepted to the 23rd IFAC World Congress 2026.