Long-horizon robot manipulation requires more than predicting the next action; it also needs language-level reasoning about subtask goals and progress. The LAW framework addresses this by coupling a vision-language model (VLM) with a world-action model (WAM) in a closed loop. The VLM plans remaining subtasks, scores and fuses action proposals from the WAM, and asynchronously checks whether the current step has succeeded before moving on.

Inside the WAM, a Mixture-of-Transformer Experts architecture handles joint video-action modeling, while a dual-stream hierarchical visual memory keeps temporal context for both models. The authors also introduce subtask-local value learning, which ties action evaluation to semantic objectives and preserves supervision from successful prefixes of failed trajectories.

Across 50 RoboTwin tasks, LAW achieved 94.21% average success. Replacing the WAM backbone with OpenWAM-alpha improved this to 95.56%, a new open-source state of the art on that benchmark. The 2B-parameter VLM outperformed variants using Embodied-R1.5 and GPT-5.5 High. On three long-horizon real-robot tasks, LAW reached 65% average success, compared with 28.33% for pi_0.5 and 25% for Fast-WAM.