Event

PhysWord: Structuring Robot Manipulation Reasoning with Physical Tokens

1 reports · 1 sources

Jiaming Liu et al. propose PhysWord, a reason-then-act foundation model for robot manipulation that uses atomic action words as semantic anchors to align continuous physical dynamics representations with language. The model predicts subsequent atomic action words and aligns their contextual hidden states with implicit goals derived from future observations, constructing a temporally structured PhysWord space with visual and optional tactile variants. The paper reports that the model achieves a 90.3% success rate on the visual-tactile UniVTAC and 99.6% on the vision-only LIBERO-Long, and outperforms comparison methods on real-world tasks using dual grippers and dexterous hands; the authors state that its structured implicit reasoning framework can generalize to unseen scenarios and support long-horizon manipulation.