Found first: a primary source the press has not covered yet.
Researchers from Rice University and NVIDIA introduce S² (Sensorimotor Alignment), a training-free offline metric for ranking robot foundation models before hardware deployment. In a paper submitted October 22, 2026, they report Pearson r=+0.831 (p=0.041) between S² scores and real-robot task success, and r=+0.717 (95% CI: [+0.436, +0.912]) for vision-language-action models evaluated on SimplerEnv.
What the source says
Yunfei Xie, Howard H. Qian, and Kaiyu Hang (Rice University), Shiyi Lan (NVIDIA), and Chen Wei extend the Platonic Representation Hypothesis to embodied AI. The hypothesis holds that as models improve, their representations across modalities converge toward a shared statistical model of reality. The authors test whether physical motion trajectories form such a modality, and compute S² as the geometric alignment between a model's sensory features and physical motion geometry, requiring no labeled data or fine-tuning. Tested across raw VLMs (r=+0.681, p=0.031), world action models (r=+0.718, p=0.054), and VLAs on SimplerEnv (r=+0.717, 95% CI: [+0.436, +0.912]), the metric correlates positively with downstream task success in each family.
Why it matters
Selecting a VLM backbone for robotics currently requires fine-tuning each candidate into a VLA and running closed-loop hardware trials. S² computes before any of that, giving teams a representation-level score to narrow the field before fine-tuning begins. The paper shows VQA scores yield r=−0.154 on the same VLM ranking task where S² reaches r=+0.681. Beyond model ranking, S² provides representation-level explanations for performance patterns in VLAs and world action models that closed-loop success rates alone cannot reveal.