Peking University and Qiyuan Robotics propose HIL-UMI for robot-free embodied post-training

Peking University and Qiyuan Robotics propose HIL-UMI, a method that collects error-correction data for embodied models without a robot arm.

A handheld gripper records human demonstrations, and the model flags segments where its predicted actions diverge from the human's, prompting targeted recording.

The article reports 5.63x higher data collection efficiency than real-robot HG-DAgger and average task progress of 94.8 versus 58.0 for standard SFT under equal frame budgets.