Reproduce “CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents” 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
Yijia Fan, Xilin He, Yilei Jiang, Dongming Wu, Jiaxu Wang, Xiangyu Yue
CaPtain trains a small open vision-language model to control robots by writing programs over low-level perception and motion tools. It uses supervised fine-tuning followed by multi-turn reinforcement learning, with a simulator critic that uses hidden physical state during training. On 240 held-out LIBERO episodes, host success rises from 55.4% to 73.8%; this result does not establish uniform transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.