HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
HiFi-UMI claims robot-free demonstration data can be good enough to deploy manipulation policies without a real-robot post-training anchor.
The system reports 3 mm workspace-local end-effector accuracy using portable head-mounted stereo-inertial SLAM, native relative pose, microsecond GPIO synchronization, and wide-angle hand cameras. Policies trained only on HiFi-UMI demonstrations matched in-domain teleoperation across three tested model backbones, with the strongest reaching 85% on a precision insertion task. The authors also say pre-training on 4,000 hours from the corpus cut action error on ten unseen tasks by 41% and improved StarVLA-QwenPI real-robot success by 18.1 percentage points. They are open-sourcing HiFi-UMI-2K, a 2,000-hour synchronized demonstration dataset. ArXiv · AI/CL/LG's note
The system reports 3 mm workspace-local end-effector accuracy using portable head-mounted stereo-inertial SLAM, native relative pose, microsecond GPIO synchronization, and wide-angle hand cameras. Policies trained only on HiFi-UMI demonstrations matched in-domain teleoperation across three tested model backbones, with the strongest reaching 85% on a precision insertion task. The authors also say pre-training on 4,000 hours from the corpus cut action error on ten unseen tasks by 41% and improved StarVLA-QwenPI real-robot success by 18.1 percentage points. They are open-sourcing HiFi-UMI-2K, a 2,000-hour synchronized demonstration dataset. ArXiv · AI/CL/LG's note
score 5