Reflex: Humanoid robots catch thrown boxes without falling using one RGB-D camera
DJiafei · x · 2026-11-07
Researchers from UW, USC, Stanford, NUS, UNC Chapel Hill and AI2 present Reflex (Visually Grounded Reactive Humanoid Control), enabling a Unitree G1 to catch boxes thrown by people using only an onboard RGB-D camera — without falling.
Instead of catching with its hands, the robot repositions its feet and torso, forms a support with forearms and chest, absorbs the momentum, and keeps hold.
Training runs in three stages in Isaac Lab:
- Train a state-based teacher with PPO plus an AMP-style reward from human motion references
- Train with PPO while inferring box dynamics from delayed, incomplete observations (position, velocity, orientation, size, observation age) compressed into a latent that conditions the controller
- Freeze the controller and learn downstream skills
More from Embodied
- AI smart blackboard shown in Shanghai turns handwriting and equations into interactive 3D models — AnnaCiaunica · 2026-11-07
- Do as I Do: Human video as robot data — 'most robot data will soon be made without robots' — carlosdponx · 2026-11-07
- Yandex to build 1,300+ delivery robots a month, targeting 20K fleet by 2028 — teortaxesTex · 2026-11-07
- AI agents design and operate robot-powered labs, Nature news reports — AnnaCiaunica · 2026-11-07
- Robot party in San Francisco features a bot actually giving a human a tattoo — Scobleizer · 2026-11-07
- Dev controls coding agent via Rokid smart glasses to ship Black Friday campaign — yihui_indie · 2026-11-07