Rice researchers predict robot model success without rollouts

A Rice University group has published a training-free way to rank robot foundation models without putting any hardware on the floor. Their metric, called Sensorimotor Alignment or S², correlates with downstream robot task success at Pearson r=+0.831 on real-robot policies and r=+0.717 on vision-language-action models evaluated against SimplerEnv, according to the paper by Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang and Chen Wei.

The pitch is simple. Closed-loop rollouts on real robots are expensive, and the field needs a cheaper way to screen models before paying the fine-tuning or hardware bill. S² is a "training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion." The authors frame it as a robotics-specific extension of the Platonic Representation Hypothesis, which argues that stronger models converge on a shared statistical picture of reality.

The paper reports that visual question answering scores — the standard proxy for VLM capability — correlate at r=−0.154 with VLA success, essentially nothing. S² on the same raw backbones reaches r=+0.681. The authors describe the ablation plainly: "Freezing the encoder reduces S² by 41% and SimplerEnv success rate by 40%." They use that near-matched drop to argue the metric is tracking something the policy cares about, not an artifact.

The accompanying project page lists results across SimplerEnv, RoboTwin, ALOHA and the VLM4VLA framework, with named world action models including Motus, LingBot-VA and Vidar. The abstract itself publishes no per-model table and no confidence intervals on any of the correlation numbers, so the ranking strength at the margins — the part that matters when you are picking between two close candidates — is not visible in the public materials.