Researchers have introduced HumanCLAW, a novel evaluation framework designed to assess whether vision-language models (VLMs) can act through a physical body. The framework decouples high-level action decision-making from low-level motor execution, allowing researchers to isolate and measure a model's 'action intelligence'—its moment-to-moment choice of what the body should execute next.
In HumanCLAW, a harnessed VLM issues atomic skill commands at each step, which are translated into sub-second chunks of continuous full-body motion with real physical consequences, including gravity and collisions. This setup factors out execution-side disturbances such as balance errors and motor failures, enabling a pure measure of the model's decision-making ability.
Based on this framework, the team built HumanCLAW-Bench, comprising 1,218 long-horizon, egocentric find-navigate-interact episodes across 41 indoor scenes. They tested nine state-of-the-art VLMs and found that none solved the benchmark; the best model achieved only a 16.8% success rate. The bottleneck was not recognizing targets but rather a lack of embodied self-awareness—models lost track of their own body, failing to know where it was, whether it had reached a goal, or whether it had hit an obstacle.
The authors note that a model may know where a sofa is yet still not know how to move its body there and sit down. They call this ability 'Action Intelligence,' and HumanCLAW makes it measurable without tying it to motor control. This work represents a first step toward models that can reason and act through a physical body.