VLA-Δ: Learning Physical State Transition for Vision-Language-Action Models
VLA-Δ adds explicit supervision of physical state changes during training, with no additional inference overhead.
The authors target a weakness in models trained to map observations directly to demonstrated actions: overfitting those pairings. They build a large-scale pretraining dataset with annotations covering task-relevant semantics and visual motion changes. At inference, the model generates actions directly from current observations. The paper reports results competitive with the state of the art on LIBERO and RoboTwin, and better performance than π₀.₅ in real-robot experiments.
HF Daily Papers' note
The authors target a weakness in models trained to map observations directly to demonstrated actions: overfitting those pairings. They build a large-scale pretraining dataset with annotations covering task-relevant semantics and visual motion changes. At inference, the model generates actions directly from current observations. The paper reports results competitive with the state of the art on LIBERO and RoboTwin, and better performance than π₀.₅ in real-robot experiments.
HF Daily Papers' note
score 5