HiFi-UMI shows robot policies can deploy from high-fidelity UMI data alone
simple-world-lab · hf · 2026-07-29
HiFi-UMI is a high-fidelity robot-free UMI data system for learning deployable manipulation policies. The paper argues that instead of relying on a small real-robot anchor, improving the fidelity of robot-free data can be enough for direct deployment.
- System design: head-mounted offline stereo-inertial SLAM, native relative pose, microsecond GPIO sync, and two wide-angle cameras per hand.
- Performance: about 3 mm workspace-local end-effector accuracy without external tracking.
- Key result: policies post-trained only on HiFi-UMI data can deploy on real robots and match in-domain teleoperation across three backbones.
- Open source: the authors release HiFi-UMI-2K, 2,000 hours of microsecond-synchronized demonstrations.
Related event: HiFi-UMI Enables Robot Training Without Real-World Teleoperation(2 posts)→
More from Embodied
- Visualizing Distributions in Robot Datasets Using VR — chris_j_paxton · 2026-07-30
- Waymo Resumes Freeway Rides After Suspension; Polymarket Favors Tampa Expansion — Polymarket · 2026-07-30
- HexRunner Robot Hits 33mph Using Pure Mechanical Design Over Sensors — lukas_m_ziegler · 2026-07-30
- Anbernic RG Rotate meets Rabbit R1: a retro AI device mashup — SimonBalmain · 2026-07-30
- Testing Codex Agent to Autonomously Deconstruct Hardware Design and Navigate Supply Chains — mattfreed · 2026-07-30
- OpenAI President Confirms 'Family of Devices' for AI Chatbots in Development — The Verge AI · 2026-07-30