HiFi-UMI: Training Deployable Robot Policies Using Only High-Fidelity Data

_akhaliq · x · 2026-07-30

HiFi-UMI introduces a portable, robot-free data collection system designed to solve the cost and fidelity issues in embodied AI data gathering. The system achieves a 3 mm end-effector accuracy and microsecond-level sensor synchronization.

Research shows that post-training using only HiFi-UMI demonstrations matches the performance of real robot teleoperation. Pre-training on 4,000 hours of this data reduces action error by 41% across ten unseen tasks and improves real-robot success rates by 18.1%. The team also open-sourced the HiFi-UMI-2K dataset, containing 2,000 hours and over 482K demonstrations.

Related event: HiFi-UMI Enables Robot Training Without Real-World Teleoperation(2 posts)→

Original post →

More from Embodied

Embodied channel →