Reflex teaches humanoid robots to catch thrown boxes from egocentric vision, 65% success on real Unitree G1
UW · hf · 2026-10-12
Researchers from University of Washington, Stanford, USC, NUS and Allen Institute for AI present Reflex, a three-stage framework for visually grounded reactive whole-body control on humanoids, using catching thrown boxes as the representative task.
- Stage-wise training: RL first learns state-based whole-body catching; the policy then infers box dynamics from delayed, incomplete observations; finally an egocentric RGB-D encoder recovers this compact dynamics representation from vision while the whole-body controller stays frozen.
- Simulation: 85.7% catch success from egocentric RGB-D vs 87.9% with privileged box observations. Temporal visual history is critical — removing it drops success by 35.6 and 39.8 points on centered and offset throws.
- Real robot: On a Unitree G1, Reflex catches human-thrown boxes with 65% and 40% success at 2.5 m and 5 m, reaching 76% and 59% of human catchers' performance at the same distances.
More from Embodied
- Atomic Machines exits stealth with Matter Compiler, an AI-native manufacturing system — mikeflache · 2026-10-12
- Meta debuts 'I Let You In,' a Talk to Me universe VR horror series for Quest — ZeroStateReflex · 2026-10-12
- RealHorse app quantifies horse lameness from smartphone video, dubbed 'Horse Tinder' — RexDouglass · 2026-10-12
- Yeah Hand: open-source humanoid robotic hand kits start at just $217 — CyberRobooo · 2026-10-12
- Fly-by-Code: coding agents program a tiltrotor drone with an arm, sim-to-real — furongh · 2026-10-12
- StreamRender-H3 renders interactive game worlds in real time with a 2-step diffusion model at 24 FPS — songhan_mit · 2026-10-12