HiFi-UMI: Robots Learn Skills from Human Videos with Zero Real-World Practice

jiqizhixin · x · 2026-08-14

HiFi-UMI is a new data capture system using high-precision handheld cameras instead of expensive teleoperation, recording human demonstrations with 3mm accuracy. Policies trained solely on this data deploy directly to real robots, matching or beating teleoperation baselines (within 2.5% success). Pre-training on 4,000 hours cuts action error on unseen tasks by 41%, achieving 85% success on precision insertion. The team open-sources HiFi-UMI-2K, a 2,000-hour dataset with microsecond-level sync and 200° views.

Original post →

More from Embodied

Embodied channel →