Amazon's Whole-Body Dexterity hits 90% success teaching Unitree G1 to haul boxes like humans
CyberRobooo · x · 2026-10-19
Amazon researchers with UC San Diego and University of Michigan developed Whole-Body Dexterity (WBD): grasp strategies generated in simulation, trained with RL, and distilled onto a real Unitree G1. With coarse teleoperation guidance the robot tucks boxes under one arm and carries them shelf-to-table. Across 50 real-robot trials WBD hit 90% success vs 18% for direct retargeting and 4% for a generalist whole-body controller.
More from Embodied
- Developer torn over foldable iPhone, fears unadapted apps for years — emax · 2026-10-19
- rabbit r1 OS3 promises natively rendered 60FPS apps as vibe coding takes center stage — mervenoyann · 2026-10-19
- British Airways: Filming Crew With Meta Smart Glasses Could Get Your Tickets Cancelled — bookofjoe · 2026-10-19
- Canton Fair diary: service robotics hall outshines industrial automation for robot AI — DominiqueCAPaul · 2026-10-19
- Walmart's billion-dollar warehouse robots can't handle real-world work, WSJ reports — ZehanWang · 2026-10-18
- Families Say Tesla FSD Keeps 80s and 90s Drivers Independent — elonmusk · 2026-10-18