Reproduce “Self-Labeled Preference Transfer for Steerable Robot Manipulation” 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

Kaiwen Hong, Shaoting Peng, Xiangyu Zeng, Pengcheng Wang, Boyuan Tian, Haonan Chen, Yifan Zhu, Yixiao Wang and 1 more

SPT transfers preference-conditioned control to new tasks without preference labels on target demonstrations. A source-trained recognizer assigns preference names to target demonstrations, and the policy adapts using those inferred labels; target demonstrations and a preference vocabulary remain necessary. In RoboTwin, preference following is 83.8% versus 59.4% for Naive FT, while task success is 89.2% versus 88.0%; success is not higher on every axis.

Summary by alphaXiv.

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