ETH Zurich's ADAPT maps text directly to humanoid joint actions with ~2ms diffusion denoising
breadli428 · x · 2026-11-05
ADAPT, a CoRL paper from ETH Zurich, uses a single diffusion policy to map text commands directly to joint actions on a Unitree G1, switching skills mid-motion. A browser demo and paper are live.
Motivation: Pipelines like NVIDIA Kimodo generate kinematic motion from text, then track it with a separate controller—but the generator doesn't know what the robot can track, so references become infeasible on fast command switches.
How it works:
- Data: human motions (AMASS) with frame-level text labels (BABEL) are replayed on G1 in simulation by an RL motion tracker; successful rollouts become state-action-text training data.
- Model: a small causal transformer jointly denoises the next 15 frames of actions and robot states, conditioned on 5 past frames and a CLIP text embedding—switching commands just swaps the embedding. Two denoising steps take 2 ms, enabling closed-loop control.
More from Embodied
- DIY robot arm build log: J0 joint ~90% done, held together by two M4 bolts — _Stocko_ · 2026-11-06
- Agility Robotics Becomes First Dedicated Humanoid Robot Company to Go Public in the US — davidpattersonx · 2026-11-06
- VFlash decodes robot actions in one shot: 5.1x faster inference, 61% lower 3D trajectory error — DrJimFan · 2026-11-06
- AI smartwatch detects hypertension with 66% sensitivity, 90% specificity, Nature study finds — EricTopol · 2026-11-06
- Apple shows off a polished humanoid robot research project page — carlosdponx · 2026-11-06
- NVIDIA runs four hands-on physical AI training labs at SIGGRAPH Asia 2026 — tomaszbednarz · 2026-11-06