Meta's HumanCLAW strips motor skill, VLMs still fail

Meta built a benchmark called HumanCLAW that strips motor control out of VLM navigation tests, and the best of nine models still scored 16.8%, per the new arXiv paper.

The design is the point. A separate, physically simulated controller handles balance, movement, and collision avoidance, translating a model's high-level commands into sub-second chunks of full-body motion. Whatever's left over after that is pure decision-making, tested across 1,218 long-horizon find-navigate-interact episodes in 41 indoor scenes. All nine state-of-the-art vision-language models tested failed to clear even a fifth of them.

The failure mode isn't the obvious one. The paper's diagnosis: recognizing the target object isn't what trips these models up. They lose track of their own body in space, and can't reliably tell whether they've arrived at a target or just collided with something near it.

The bottleneck moved

That reframes where embodied AI's problem actually sits. It isn't perception, and with motor control factored out, it isn't motor skill either. It's self-localization, a model's sense of where it is and what it's touching. Scaling up the VLM doesn't obviously fix that. Watch whether follow-up work bolts on an explicit spatial-state model instead of a bigger one.