Researchers led by Yefim Nudelman have built Spatially Conditioned Trajectory Diffusion (SCTD), which generates robot paths that avoid obstacles by reading the map throughout the generation process, according to a paper on arXiv. One model trained on single-robot demonstrations needs no obstacle-cost guidance while it samples.
A conflict-based search coordinates the paths of several robots by replanning under constraints. On 15,000 held-out tasks it succeeded more often than a guided diffusion baseline, especially on tasks that need long detours, and produced smoother paths. The combined system did better than other diffusion methods with teams of up to 15 robots.



