GenRobot.AI Turns Egocentric Video Into Whole-Body Manipulation Data With ~3cm Error
CyberRobooo · x · 2026-09-15
GenRobot.AI unveiled a data production system that converts egocentric human video into Whole-Body Manipulation Data for robot policy training. A portable, vision-only multi-view ground-truth system provides supervision, while its DFM model reconstructs temporally coherent whole-body motion unifying body, hands, objects, contact, and task semantics.
Internal evaluation reports 3cm mean whole-body error (≈2cm upper body, 3.5cm lower). The pipeline moves data collection beyond mocap labs into homes, commercial, and industrial spaces. The company also sells DAS-series hardware (hands, fingers, grippers, controllers).
Related event: GenRobot AI turns first-person video into full-body robot data(2 posts)→
More from Embodied
- Pure-Rust visloc-rs adds visual-inertial SLAM, runs 3.46x faster than COLMAP on CPU — rsasaki0109 · 2026-09-15
- PhysBrain 1.5 unifies physical understanding, action, and prediction — DeepCybo · 2026-09-15
- Developer Seeks Peers Working on Drones, Gaussian Splatting and World Models — WesEklund · 2026-09-15
- Sidewalk Robots Already Navigate Real Urban Intersections — CurieuxExplorer · 2026-09-15
- GPT-6 Astra gets a Unitree humanoid into a Tesla Model 3 in MuJoCo simulation — YuXiang_IRVL · 2026-09-15
- Tesla Cybercab Heads to Mainland China for First Public Display — XFreeze · 2026-09-15