AW-0 unifies code planning, video prediction, and robot control in one generative model

AW-0, a collaboration between UC Berkeley, NVIDIA, and the University of Pennsylvania, jointly generates code, future video, and robot actions through a single Mixture-of-Transformers model using joint denoising. Pretrained on roughly 3,000 hours of demonstrations, it distills 50 reusable skills and runs inference in 230 ms on an RTX 5090—40 ms over the base WAM and about 2.4× faster than autoregressive code generation. On RoboTwin 2.0's clean setting it reaches 80.8% success versus 64.8% for the strongest baseline; hybrid code execution lifts randomized-scene success from 20.6% to 29.2%. Across eight real-world tasks, average rubric score is 71.1 versus 46.0.

✨ AI summary of the linked source — not editorial content.

Explore the topic

• Video