LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence
LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence
This research presents LAW, a closed-loop framework that combines vision-language models (VLM) with world-action models (WAM) to enable agentic physical intelligence. The system uses semantic subtask progress to plan, evaluate, and execute actions across long-horizon tasks. LAW achieves 94.21% average success on 50 RoboTwin tasks and 65% on real-robot tasks, outperforming existing methods like $pi_{0.5}$ and Fast-WAM.
The paper introduces LAW, a novel framework that integrates language reasoning, action generation, and world modeling for physical intelligence tasks. It demonstrates strong performance on both simulated and real-world robotic tasks, outperforming existing methods like $pi_{0.5}$ and Fast-WAM.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
95.56%
View original evidence
-stream hierarchical visual memory maintains temporal context for both models, while subtask-local value learning aligns action evaluation with semantic objectives and preserves supervision from successful prefixes of failed trajectories. LAW achieves 94.21% average success across 50 RoboTwin tasks. Replacing the WAM backbone with OpenWAM-$alpha$ further raises the average success rate to 95.56%, establishing a new o
Open source S4tasks
50
View original evidence
-stream hierarchical visual memory maintains temporal context for both models, while subtask-local value learning aligns action evaluation with semantic objectives and preserves supervision from successful prefixes of failed trajectories. LAW achieves 94.21% average success across 50 RoboTwin tasks. Replacing the WAM backbone with OpenWAM-$alpha$ further raises the average success rate to 95.56%, establishing a new o
Open source S4tasks
3
View original evidence
pen-source state of the art on RoboTwin. Despite having only 2B parameters, our VLM also outperforms variants using Embodied-R1.5 and GPT-5.5 High. On three long-horizon real-robot tasks, LAW achieves 65% average success, compared with 28.33% for $pi_{0.5}$ and 25% for Fast-WAM.
Open source S5
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
-stream hierarchical visual memory maintains temporal context for both models, while subtask-local value learning aligns action evaluation with semantic objectives and preserves supervision from successful prefixes of failed trajectories. LAW achieves 94.21% average success across 50 RoboTwin tasks. Replacing the WAM backbone with OpenWAM-$alpha$ further raises the average success rate to 95.56%, establishing a new o
Open source S4
pen-source state of the art on RoboTwin. Despite having only 2B parameters, our VLM also outperforms variants using Embodied-R1.5 and GPT-5.5 High. On three long-horizon real-robot tasks, LAW achieves 65% average success, compared with 28.33% for $pi_{0.5}$ and 25% for Fast-WAM.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org