The Greenhouse Policy Learned to See the Obstacle Separately

ObstaDiff decomposes a cluttered manipulation scene into target, obstacle and background representations before a diffusion policy plans toward a target-centered bottleneck pose. The authors evaluated six methods over 61 greenhouse trials each, for 366 real-robot executions. ObstaDiff reported 75.41% average task success and an 8.20% average obstacle-collision rate, outperforming the selected imitation-learning baselines. The result suggests that explicit obstacle structure can improve generalization beyond clean training backgrounds, but it is limited to the reported agricultural tasks and test setup.