Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success
Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
Rice University · NVIDIA
Published October 23, 2026 · on the feed October 29, 2026 · ▲ 56 · arXiv 2610.31670
The abstract
Robot foundation models are typically evaluated through costly closed-loop rollouts, motivating low-cost offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, modalities such as vision and language converge toward a shared statistical model of reality. We ask whether motion trajectories form another such modality: as embodied models become more capable, should their sensory representations converge toward the structure of physical motion? To test this, we propose Sensorimotor Alignment (S^2), a training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion. We find that S^2 correlates positively with downstream robot task success across model families, making it a low-cost offline proxy for prioritizing models before closed-loop evaluation. S^2 can also identify promising vision-language model (VLM) backbones before costly fine-tuning into vision-language-action models (VLAs). Beyond this correlation, S^2 offers representation-level explanations for phenomena observed in VLAs and world action models (WAMs). With S^2, we extend PRH to robotics and make sensorimotor alignment measurable, offering a low-cost, representation-level view of robot foundation models that closed-loop success rates alone cannot reveal.
The Brief
2:01A short podcast summary of the paper. Written and voiced by AI.
Transcript
Sensorimotor Alignment. A Platonic Proxy for Embodied Policy Success. By Yunfei Xie and four others, from Rice University and NVIDIA. Published on the twenty-third of October.
Robot foundation models are usually evaluated through closed-loop rollouts: put the policy on a robot, or in a loop with its environment, and see whether the task gets done. That is costly. It motivates a search for offline proxies, cheap measurements that say something about a model's capability before any rollout happens.
The starting point is the Platonic Representation Hypothesis. It suggests that as representations improve, modalities such as vision and language converge toward a shared statistical model of reality. The authors ask whether motion trajectories form another such modality. As embodied models become more capable, should their sensory representations converge toward the structure of physical motion? Sensorimotor Alignment is the measurement built to answer that. It scores how well a model's sensory representation aligns with the geometry of physical motion.
The metric is training-free and offline. Nothing is fine-tuned to compute it, and no robot needs to run. It is meant to sit before closed-loop evaluation, as a way of deciding which models deserve the expensive test.
The central finding is a positive correlation. Sensorimotor Alignment tracks downstream robot task success across model families, which is what makes it usable as a low-cost proxy for prioritizing models. It can also pick out promising vision-language model backbones before the costly step of fine-tuning them into vision-language-action models.
Beyond correlation, the metric offers explanations at the level of representations for phenomena observed in vision-language-action models and in world action models. With it, the authors extend the Platonic Representation Hypothesis to robotics and make sensorimotor alignment something that can be measured.
A success rate tells you whether a robot finished the job. It says nothing about what the model inside has come to represent. The claim here is that the second question can be asked directly, cheaply, and before any rollout, by checking whether what a model sees has taken on the shape of how things move. That gives a view of robot foundation models that closed-loop success rates alone cannot reveal.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Testing robots is slow and expensive. Is there a cheaper way to tell which model will do well?
You have many candidate robot policies and room to test only a few on real hardware. Each closed-loop trial is slow, noisy and depends on the benchmark you pick.
The road ahead
- 1 · The problem ~7 min Robot models are costly to test for real, and the cheap scores we have measure perception, not whether a model is ready to act.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Score a model by whether episodes it sees as similar also have similar recorded robot motions, with no training and no rollouts.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Encode recorded episodes, compare feature neighbours with motion neighbours, check the score against task success, then use it to rank models.
- 6 · The experiment ~6 min You’ll guess: Does a training-free score of how well a robot model's features align with physical motion track the model's real task success?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.