Event

LAW framework unifies language reasoning, action generation, and world modeling

1 reports · 1 sources

Jiajun Cao and other researchers proposed and released the LAW closed-loop VLM-WAM framework, which unifies language reasoning, action generation, and world modeling through semantic subtask progress. The authors report that the framework achieves an average success rate of 94.21% on 50 RoboTwin tasks; after replacing the WAM backbone with OpenWAM-α, the average success rate rises to 95.56%, which the authors say sets a new open-source record on RoboTwin. On three long-horizon real-robot tasks, LAW achieves an average success rate of 65%, compared with 28.33% for π0.5 and 25% for Fast-WAM as reported by the authors. The authors also state that their VLM has only 2B parameters, yet in variant comparisons it outperforms versions using Embodied-R1.5 and GPT-5.5 High.