PhysWord preprint reports robot reasoning gains using action words

Researchers introduce PhysWord in a preprint reporting robot manipulation success rates of 90.3% on UniVTAC with vision and touch and 99.6% on vision-based LIBERO-Long. They also report that the model outperforms competing methods in real-world tasks using bimanual grippers and dexterous hands. The model predicts simple action verbs and links their internal representations to future observations, giving action generation structured context about physical changes. Predicting these representations across future timesteps captures transitions between execution stages. The framework supports visual representations with optional tactile variants; the two reported benchmark results use different sensory inputs.