Reproduce “Learning to Be Different: Component Specialization for Compositional Diffusion Policies” 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

Alec Laprevotte, Haonan Chen, Xiaoshen Han, Yilun Du

The paper introduces a training-only penalty that discourages overlap between component predictions in compositional diffusion policies, addressing redundant factors despite a correct combined prediction. It leaves inference unchanged but adds pairwise training computation. On RoboCasa, mean success rises from 17.8% to 23.2%; on RLBench MT8 it reaches 55.3%, below FDP's 56.1%, showing specialization does not always improve task performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.