AW-0 robot model averages 71.1% across 8 real-robot tasks

A joint UC Berkeley, NVIDIA and University of Pennsylvania team has posted AW-0, an "agentic world action model" that treats code, future video and robot actions as three coupled streams denoised together inside one mixture-of-transformers network.

On eight real-world bimanual tasks spread across two embodiments — the YAM parallel-gripper station and the 22-DoF Tianji–Sharpa dexterous hands — AW-0 averaged 71.1%. The paper reports π₀.₅, a strong vision-language-action baseline, at 43.6%, with DreamZero at 46.0%, WLA at 39.9% and EgoScale at 35.9%. On RoboTwin 2.0's clean setting, AW-0 lands at 80.8% versus 64.8% for the WLA baseline.

The authors describe the central idea as a "unified mixture-of-transformers architecture" in which "a code expert generates parameterized skill routines and control logic while video and action experts generate future visual states and low-level motor commands." Doing the three denoising passes together rather than token-by-token clocks in at 0.23 seconds versus 0.55 for the autoregressive variant, a 2.39× speedup on an RTX 5090.

The skills themselves are built by machine. The paper's annotation pipeline samples an episode per task, has Gemini 3.1 Pro draft subtask-level programs, then uses Claude Opus 4.8 as a coordinator that distills recurring routines into a frozen library of 50 skills — 23 for the YAM robot and 27 for egocentric human video. That library then re-annotates roughly 248,642 pretraining episodes drawn from about 2,000 hours of YAM data and 1,000 hours of human video.

Robustness is uneven. Under the random RoboTwin 2.0 setting the default mode falls to 20.6%, while a hybrid-execution mode that toggles between programmatic skill calls and direct action chunks holds 29.2%. The paper is one more entry in a dense recent run of NVIDIA-anchored robotics work on AI Weekly's robotics tracker, which has logged 166 stories in the last 90 days.