Refining Compositional Diffusion for Reliable Long-Horizon Planning

Compositional diffusion planning generates long-horizon robot trajectories by stitching together overlapping short-horizon segments via score composition. The problem? When local plan distributions are multimodal, existing methods average incompatible modes, yielding plans that are neither locally feasible nor globally coherent. Researchers Kyowoon Lee, Yunhao Luo, Anh Tong, and Jaesik Choi propose Refining Compositional Diffusion (RCD), a training-free guidance method that steers sampling toward high-density, globally coherent plans. RCD leverages the self-reconstruction error of a pretrained diffusion model as a proxy for the log-density of composed plans, combined with an overlap consistency term that enforces smoothness at segment boundaries. This combined guidance concentrates sampling on high-density plans, effectively mitigating mode-averaging.

RCD was evaluated on challenging long-horizon tasks from OGBench, including locomotion, object manipulation, and pixel-based observations. It consistently outperformed existing compositional methods across all benchmarks — and without requiring any additional training or fine-tuning. This is particularly significant for real-world robotics where retraining is expensive and data is scarce. The method is model-agnostic and can be applied to any pretrained diffusion planner. By fixing a fundamental flaw in compositional planning, RCD opens the door to more reliable, long-horizon autonomy in robotics, from warehouse logistics to household assistants.