Peking University and Qiyuan Robot unveil HIL-UMI for arm-free robot post-training
A Peking University team led by Dong Hao, who is also chief scientist at Qiyuan Robot, has introduced HIL-UMI, a human-in-the-loop data collection method for embodied post-training that runs without a robot arm. An operator uses a handheld UMI gripper while a model predicts the next actions in the background; when its predictions diverge from the human demonstration, the system prompts the operator to record corrective data, using an energy score and an advantage estimator. In four real-arm tasks, the team reports an average task-progress score of 94.8 with HIL-UMI versus 58.0 for conventional supervised fine-tuning, and 73.40 ms versus 412.99 ms per effective frame on a desk-cleaning task, a 5.63x collection-efficiency gain.
The report presents HIL-UMI as a way to move corrective data collection for embodied post-training off the robot arm and reduce its cost.
导语：继无本体之后，具身后训练范式有了新进展。
继无本体之后，具身后训练范式有了新进展。
作者丨齐铖湧
编辑丨岑 峰
最近，具身行业讨论“无本体进行后训练的可能性”的热度很高。国内外目前还没有团队做出可行方案。