A new arXiv preprint argues that while multimodal reasoning gives robotic foundation models the predictive context needed for reliable action generation, language-based reasoning has a blind spot. It captures task semantics well, but it struggles to express the fine-grained physical details that matter when a robot actually touches, grips, or moves an object.
The paper proposes a solution called "Physical Words" — a vocabulary designed to structure latent reasoning in robot manipulation. Rather than relying solely on natural language, this vocabulary aims to represent physical concepts in a way that bridges the gap between high-level task understanding and low-level motor control.
The abstract is brief, so details about how the vocabulary is constructed or evaluated are not yet available. But the core claim is clear: adding a structured physical layer to latent reasoning could make robotic action generation more precise and dependable.