How to make robots coordinate without central control?

Coordinating multiple robots is hard: central planning doesn't scale, but letting each robot plan alone ignores others. This work trains each robot's trajectory generator independently using diffusion models, then steers them toward collision-free solutions by wrapping a multi-agent reinforcement learning value function around the generation process. In a four-robot maze, the approach cuts inter-agent conflicts from 55% to 42% while keeping the system decentralized.