HIL-UMI: Robots Learn From Handheld Demos Without Touching Real Hardware
jiqizhixin · x · 2026-10-19
Qiyuan Robotics' Dong Hao team presents HIL-UMI (ArXiv 2609.20659), which collects training experience on a handheld device—no real robot execution needed during data collection.
The system systematically compares what the human demonstrator does against what the current policy would do, flagging states where the gap exceeds a threshold and prompting the operator to keep recording from that state until the step completes. Validated with π0.5 on a Franka arm across four tasks.
More from Embodied
- DIY scanner produces Gaussian Splats trained with LichtFeld Studio — janusch_patas · 2026-10-19
- Tesla rolls out first batch of Optimus humanoid robots at Giga Texas — emmanuelvivier · 2026-10-19
- Singapore targets 2028 rollout of humanoid robots for public safety duties — emmanuelvivier · 2026-10-19
- Sam Altman confirms OpenAI is building a humanoid robot, first demo eyed for 2027 — emmanuelvivier · 2026-10-19
- Google closes the loop: agents now run real biology lab robots and found a cancer-binding molecule — aigclink · 2026-10-19
- KUKA CEO Christoph Schell: Physical AI won't replace deterministic automation — the two coexist — lukas_m_ziegler · 2026-10-19