Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du, Dingkang Liang, Xiang Bai (H-EmbodVis) · In-Context Robot Learning

World In-Context Learning for Agentic Robot Manipulation

The paper proposes World In-Context Learning (WICL), where a frozen model learns a task from one demonstration by rehearsing it in a simulator and storing the feedback in its context rather than its weights. The authors build RoboScript, which reaches 83.4% mean success across 50 RoboTwin 2.0 tasks with one demonstration per task and no gradient updates, reported as over 20 points above the strongest trained action policy. It also transfers to a physical dual-arm robot, and the code is available on GitHub.