Agentic World Action Models
AW-0 turns robot skills into reusable code while predicting future video and motor actions.
The framework couples code, video and action generation in a single architecture, linking multi-step plans to predicted physical outcomes. An automated pipeline labels robot and human video trajectories with programs and collects recurring routines into skill libraries. The authors report better task performance than vision-language-action and world action model baselines on RoboTwin 2.0 and physical two-arm manipulation tasks. They also report gains in language following, semantic reasoning and robustness; the abstract gives no numerical results.
HF Daily Papers' note
The framework couples code, video and action generation in a single architecture, linking multi-step plans to predicted physical outcomes. An automated pipeline labels robot and human video trajectories with programs and collects recurring routines into skill libraries. The authors report better task performance than vision-language-action and world action model baselines on RoboTwin 2.0 and physical two-arm manipulation tasks. They also report gains in language following, semantic reasoning and robustness; the abstract gives no numerical results.
HF Daily Papers' note
score 5