[论文] Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success

论文概要

研究领域: CV
作者: Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei
发布时间: 2026-10-22
arXiv: 2610.31670

中文摘要

机器人基础模型通常需要通过成本高昂的闭环 rollout 来评估，这催生了对低成本离线代理指标的需求。柏拉图表示假说（PRH）认为，随着表示能力的提升，视觉和语言等不同模态会收敛于对现实的共享统计模型。我们提出一个问题：运动轨迹是否构成另一种这样的模态——随着具身模型能力的增强，其感知表示是否应当趋近于物理运动的结构？为此我们提出 Sensorimotor Alignment（S²），一个免训练的离线指标，用于衡量模型的感知表示与物理运动几何结构的对齐程度。实验发现，S² 与下游机器人任务成功率呈正相关，可作为一种低成本的离线代理指标，在闭环评估之前对模型进行优先排序。S² 还能在将视觉语言模型（VLM）微调为视觉语言动作模型（VLA）之前，就识别出有潜力的 VLM 骨干网络。

原文摘要

Robotic foundation models are typically evaluated through costly closed-loop rollouts, which motivates cheap offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, different 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 examine this, we propose Sensorimotor Alignment (S²), a training-free offline metric that measures how well a model's sensory representation aligns with the geometry of physical motion. We find that S² correlates positively with downstream robot task success across emb...

自动采集于 2026-10-24

#论文 #arXiv #CV #小凯