Meta's HumanCLAW: top VLMs solve 16.8% of embodied sit tasks

Nine of the top vision-language models were dropped into a simulated body and told to find a couch and sit on it. The best of them, Gemini-3.1, managed the sit only 16.8% of the time. That is the headline from HumanCLAW, a new benchmark from Meta, Nanyang Technological University, the University of Washington, Brown and Northwestern that spans 1,218 find-navigate-interact episodes across 41 indoor scenes.

The design of the test is what makes the number sting. The physics side is handled for the model, since motor commands like walk, turn and sit are executed by a motion generator that hits its target with near-perfect fidelity. So the benchmark is not measuring whether a humanoid can balance. It is measuring whether the VLM knows what its own body is doing. Gemini-3.1 finds targets 64.9% of the time and reaches them 42.4% of the time, but the drop from finding to sitting is where the wall shows up.

The authors' error analysis makes the failure mode concrete. In navigation, about 34% of misses come down to body awareness: the agent has already arrived but doesn't know it, or it has jammed into geometry and keeps issuing forward commands. In interaction, 58% of the sit failures are the agent sitting into thin air with no surface beneath its pelvis. The paper describes the models as behaving like ghosts, fluent about the world but oblivious to the self.

Why this matters if you are not building a humanoid: the industry's current bet is that a frontier VLM plus a good motion layer gets you most of the way to an embodied agent. This paper argues the reasoning layer, not the actuator, is where the ceiling sits right now, and that recognition being effectively solved has been masking that. Take the specifics as a snapshot of a July 2026 result on a single benchmark, not a settled law. The paper only tests frozen models, only tests a sit interaction, and gives the VLM no proprioceptive input.

The upside is that the gap is legible. Body-aware post-training, richer contact signals and mid-level reasoning routines all have room to move the 16.8% number without waiting on new motion data or bigger models.