HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
The paper claims high-fidelity robot-free UMI data can replace the usual real-robot post-training anchor.
HiFi-UMI uses stereo-inertial SLAM, native relative pose capture, microsecond GPIO synchronization, and wide-angle hand cameras to reach 3 mm local end-effector accuracy. Policies post-trained only on its demonstrations deployed directly on a real robot and matched in-domain teleoperation across three tested backbones. The authors also report a 2,000-hour open-source HiFi-UMI-2K dataset, with demonstrations reconstructed and validated through simulation replay. HF Daily Papers' note
HiFi-UMI uses stereo-inertial SLAM, native relative pose capture, microsecond GPIO synchronization, and wide-angle hand cameras to reach 3 mm local end-effector accuracy. Policies post-trained only on its demonstrations deployed directly on a real robot and matched in-domain teleoperation across three tested backbones. The authors also report a 2,000-hour open-source HiFi-UMI-2K dataset, with demonstrations reconstructed and validated through simulation replay. HF Daily Papers' note
score 5