Peking University and PrimeBot's HIL-UMI collects robot correction data without a robot arm

A team led by Hao Dong at Peking University, with Shanghai robot maker PrimeBot, introduced HIL-UMI, a method for post-training vision-language-action models that gathers correction data using a handheld UMI gripper instead of a physical robot arm. Across four real tasks, average task progress rose from 44.3 to 94.8 under the same frame budget, with collection efficiency 5.63 times higher than HG-DAgger.

Pandaily

Open source

Peking University and PrimeBot's HIL-UMI collects robot correction data without a robot arm

Average task progress score: 94.8 for HIL-UMI vs 58.0 for plain SFT
73.40 ms per useful frame for HIL-UMI vs 412.99 ms for HG-DAgger
Method uses a handheld UMI gripper and an Energy Score to trigger corrections
Authors' own results on arXiv, not independently verified

#PrimeBot#PekingUniversity#QiyuanRobot

PhysisForcing from Peking University and Nvidia Lifts Robot Success From 16% to 24%

AgiBot's Lingxi X2 handled 46,686 patient queries at Peking University People's Hospital

Peking University, Tsinghua and Alibaba Open-Source SparkDiffusion for Up to 265x Faster Wan Video Generation

Science

Peking University and PrimeBot's HIL-UMI collects robot correction data without a robot arm

PhysisForcing from Peking University and Nvidia Lifts Robot Success From 16% to 24%

AgiBot's Lingxi X2 handled 46,686 patient queries at Peking University People's Hospital

Peking University, Tsinghua and Alibaba Open-Source SparkDiffusion for Up to 265x Faster Wan Video Generation

Prime assembly can correct multiple mutations at once, pointing toward universal gene therapies