Reproduce “BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation” 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
Antong Zhang, Han Qi, Heng Yang
BEACON learns which simulated demonstration windows are useful for a target task, rather than weighting all source examples equally. It scores source windows by encoder distance to nearby target examples, then alternates weight updates with diffusion-policy training. With 500 simulated and 40 real demonstrations, its mean scores across three tasks were 0.79 in-distribution and 0.56 on OOD initial states, versus UOT’s 0.62 and 0.36, over 20 rollouts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.