Efficient Rare-Event Sampling in Diffusion Policies for Motion Discovery

Dian Yu, Sebastian Sanokowski, Majid Khadiv

Keywords: Diffusion Policies, Rare-Event Sampling, Feynman–Kac Sampling, Motion Discovery, Robotic Manipulation

TL;DR: A complementary Feynman–Kac sampler guides diffusion policies toward repairable rare trajectories, enabling trajectory optimization and dataset expansion to recover missing modes and discover diverse robotic behaviors.

Abstract: Diffusion policies can represent multimodal behavior, but under sparse demonstrations, they often collapse onto dominant modes. We propose an expert-in-the-loop self-bootstrapping framework that biases a diffusion policy toward repairable frontier samples, refines them with trajectory optimization, and adds the repaired trajectories back to the training set. The key component is a complementary Feynman--Kac sampler based on predicted-noise energy, which steers exploration toward underrepresented modes without changing the policy architecture. We test our framework on two settings: In a controlled bimodal task, the full loop recovers a missing valid mode; in an object manipulation task, the framework recovers new behaviors for performing the task.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 13