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.
More from Embodied
- FoveaVLA replaces wrist cameras with learned foveation, hitting 70.8% success — Haoyu_Xiong_ · 2026-10-16
- Musk teases two-way Grok x Optimus integration: Grok in the bot, bot in Grok — beffjezos · 2026-10-16
- SimVTA uses simulator-rendered video to post-train driving world action models with RL — OpenDriveLab · 2026-10-16
- Robotics data startups compared: Scale AI tops $1B revenue as 9 LLM vs robotics data suppliers benchmarked — FinanceYF5 · 2026-10-16
- Kroger cuts 3,000 corporate jobs (37%) for AI; humanoid robots next in the 2030s — csuwildcat · 2026-10-16
- Tesla dares its humanoid robot to do a kickflip in playful demo — yunta_tsai · 2026-10-16