Reproduce “Diff-DAgger: Uncertainty Estimation with Diffusion Policy for Robotic Manipulation” on OpenResearch

About this paper

Sung-Wook Lee, Xuhui Kang, Yen-Ling Kuo

Researchers at the University of Virginia developed Diff-DAgger, an interactive learning framework that leverages a diffusion policy's intrinsic loss to accurately estimate uncertainty in robotic manipulation tasks. It achieves a 39.0% higher F1 score in task failure prediction and reduces DAgger iteration wall-clock time by 7.8 times compared to baselines.

Key insights

• The intrinsic loss of a diffusion policy can serve as a robust and computationally efficient proxy for uncertainty, effectively identifying out-of-distribution states in complex, multi-modal robotic tasks.
• Existing ensemble-based uncertainty estimation methods are fundamentally ill-suited for multi-modal policies as they conflate multi-modal distributions with high uncertainty, leading to inefficient expert queries.
• An intelligent, model-inherent uncertainty estimation mechanism enables more efficient use of expert demonstrations and significantly reduces the wall-clock time required for interactive learning iterations.

Summary by alphaXiv.