Human Demo Data Matches Robot Teleoperation in Embodied AI Training

CyberRobooo · x · 2026-08-06

Beijing Simple AI open-sourced the HiFi-UMI-2K dataset, featuring 2,000 hours of high-fidelity bimanual manipulation data. Built from over 20,000 hours of recording, it contains more than 4.32M processed episodes across 480+ collection scenes, with a 98% valid trajectory rate.

The most surprising finding is that policies trained exclusively on this human demonstration data matched the performance of those trained on real-robot teleoperation data, providing strong evidence for lowering the data collection costs in embodied AI.

Related event: Simple AI Open-Sources 2000-Hour High-Fidelity Robot Dataset(3 posts)→

Original post →

More from Embodied

Embodied channel →