Reproduce “Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through Corrections” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoting Li, Gang Chen, Javier Alonso-Mora, Cosimo Della Santina, Jens Kober
Researchers from Delft University of Technology and University of Stuttgart introduced the Set-Supervised Diffusion Policy (SDP), a framework that integrates set-valued supervision with diffusion policies for robotic action-chunking. SDP effectively utilizes both desired and undesired actions from human-in-the-loop corrections, leading to consistently higher success rates in manipulation tasks and improved robustness to noisy feedback compared to standard diffusion policies and other correction learning methods.
Key insights
- Explicitly leveraging both positive (teacher's) and negative (robot's undesired) action-chunks through set-valued supervision provides a richer and more robust learning signal for diffusion policies.
- Defining a "desired action set" for training, rather than a single target action, enhances policy robustness to noisy human feedback and imperfect corrections by allowing for a region of acceptable behaviors.
- Integrating set-valued supervision with diffusion models using a constrained sampling process effectively extends this learning paradigm to high-dimensional, multi-modal action-chunk spaces.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.