UniWAM: Unified World-Action Model
UniWAM ties robot action prediction to a model that also reasons about language and generates future visual states.
The paper says action-only robot training lacks grounding in world dynamics, while video-based world-action models can struggle with semantic reasoning under distribution shifts. UniWAM combines a physical reasoner, world generator, and action predictor in one architecture. The authors also describe a cleaned and annotated mix of human egocentric data, robot data, and VQA supervision, with low-level actions represented in natural language. They report state-of-the-art results across robustness, generalization, instruction following, and long-horizon execution.
HF Daily Papers' note
The paper says action-only robot training lacks grounding in world dynamics, while video-based world-action models can struggle with semantic reasoning under distribution shifts. UniWAM combines a physical reasoner, world generator, and action predictor in one architecture. The authors also describe a cleaned and annotated mix of human egocentric data, robot data, and VQA supervision, with low-level actions represented in natural language. They report state-of-the-art results across robustness, generalization, instruction following, and long-horizon execution.
HF Daily Papers' note
score 4