Learning Diffusion Models to Enable Efficient Sampling for Task and Motion Planning on a Panda Robot

Abstract

A search then sample approach to bilevel planning in the context of task and motion planning is one method of effectively solving multi-step robotics problems. In this planning framework, high-level plans of abstract actions are refined into low-level continuous transitions by sampling controller parameters associated with each action. Efficiently sampling these parameters remains a significant challenge, as exhaustive searches often become computational bottlenecks, especially for tasks requiring complex or multimodal parameter distributions. Moreover, relying on samplers hand-designed by humans is both impractical and limiting. To address these challenges, we propose using diffusion models to learn efficient sampling distributions from demonstrations. By avoiding the limitations of hand-specified and naïve sampling methods, our approach enhances planning efficiency and achieves superior performance across diverse tasks that require learning multimodal parameter distributions to solve successfully.