ABSTRACT Diffusion‐based motion planning has shown strong capability in modelling complex, multimodal trajectory distributions through guided sampling. However, existing diffusion‐based planners primarily focus on single‐robot settings and do not readily extend to multirobot scenarios due to the rapidly growing trajectory dimensionality and the challenge of handling inter‐robot collisions. We propose a diffusion‐based framework for multirobot motion planning that explicitly accounts for inter‐robot interactions during trajectory generation. Instead of jointly modelling all robot trajectories in a single high‐dimensional diffusion process, our method reuses a pretrained single‐robot Motion Planning Diffusion (MPD) model and employs an alternating trajectory generation scheme guided by inter‐robot collision costs. This results in a Gibbs‐style inference procedure for coordinated multirobot planning without additional multirobot training. To resolve residual conflicts caused by temporal coupling, we further introduce a temporal coordination module based on time–space search that searches for a collision‐free execution schedule while preserving the geometric structure of the generated trajectories. Experimental results in simulated multirobot manipulation environments show that the proposed approach achieves the best average success rates among the tested methods under fixed time budgets.
Zhang et al. (Tue,) studied this question.