Inside AW-0: code, video and action experts jointly denoise in one Mixture-of-Transformers
trevordarrell · x · 2026-10-27
Part 2/8 of the AW-0 thread: structured planning inside the world action model. AW-0 combines three experts—visually grounded skill programs (code), future visual dynamics (video), and continuous robot motion (action). The video expert attends to code; the action expert attends to both. Skill calls, object arguments, and control flow thus shape both what the model predicts and how the robot moves, with all three streams generated together via directional conditioning throughout denoising.
Related event: UC Berkeley and NVIDIA Unveil AW-0, a Unified World Action Model(4 posts)→
More from Embodied
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Bedrock Robotics raises $600M as its fully autonomous excavators do paid work on live job sites — Sethwinterroth · 2026-10-27
- AW-0 hits 80.8% on RoboTwin tasks, beating strongest baseline by 16 points — trevordarrell · 2026-10-27
- AW-0's code stream fixes world model prediction: picks mango where baseline grabs apple — trevordarrell · 2026-10-27
- AW-0 runs at 230ms on a single RTX 5090, 2.4x faster than autoregressive code planning — trevordarrell · 2026-10-27
- AW-0 debuts: one generative world action model plans in code, predicts, and controls robots — trevordarrell · 2026-10-27