AW-0 couples code planning with video and action prediction
Original title: Agentic World Action Models
The paper reports better RoboTwin 2.0 and bimanual task results than VLA and WAM baselines.
The paper introduces AW-0, an agentic world action model that jointly generates code, future video, and robot actions in a mixture-of-transformers. A code expert writes parameterized skill routines and control logic, while video and action experts predict visual states and motor commands. An automated pipeline labels robot and human trajectories with aligned programs and builds reusable skill libraries. Evaluations on RoboTwin 2.0 and physical bimanual tasks report gains over VLA and WAM baselines in task performance, language following, semantic reasoning, and robustness.
Unconfirmed: Performance gains are the authors' reported evaluations, not independently verified.
- Code structures multi-step prediction and stores skills as reusable programs.
- Curation labels robot and human video trajectories with aligned programs.
- Tests cover RoboTwin 2.0 and physical bimanual manipulation tasks.
- Reported gains include language following, semantic reasoning, and robustness.
From the original article
1 min read at huggingface.co
Recent world action models provide visual foresight and future prediction for robot manipulation, but lack explicit mechanisms for compositional skill reuse. Agentic robotics combines code-based planning and reasoning with reusable skills, but often lacks predictive grounding in physical dynamics.
By Zhuoyang Liu, Dantong Niu, Botong Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.