Spatially Conditioned Trajectory Diffusion
Abstract
Multi-Robot Motion Planning (MRMP) requires trajectories that avoid environmental obstacles and collisions between robots in a shared workspace. Diffusion-based planners often address obstacle avoidance through sampling-time guidance, while some conditioned approaches aggregate map geometry into a global representation. We present Spatially Conditioned Trajectory Diffusion (SCTD), a map-conditioned trajectory generator for multi-robot planning. SCTD retains multiple position-aware occupancy tokens accessed through cross-attention throughout a temporal U-Net, with hard endpoint inpainting enforcing start and goal states. One model trained on single-robot demonstrations across procedural environments generates obstacle-aware trajectories without sampling-time obstacle-cost guidance. High-level Conflict-Based Search (CBS) coordinates these trajectories through constrained replanning, retaining inter-agent guidance during sampling and reusing the model across workspace layouts and team sizes. Experiments assess single-robot success and trajectory quality, generalization to unfamiliar layouts with added obstacles, and coordinated multi-robot planning. On 15,000 held-out tasks, SCTD achieves higher success than the evaluated guided diffusion baseline, with the largest gains on hard-detour tasks, while trajectory-quality comparisons show smoother trajectories. Conditioning ablations show higher success with spatial tokens than with a global map embedding, and map-replacement tests demonstrate dependence on the supplied geometry. The resulting multi-robot system achieves higher team success than the compared diffusion baselines, with experiments covering teams of up to 15 robots.
- Comment
- Project page: https://spatial-trajectory-diffusion.github.io/