VLA-Δ supervises physical state transitions in vision-language-action models
The paper proposes VLA-Δ, a framework that treats the physical change between current and future states as an explicit learning target, rather than having vision-language-action (VLA) models map observations directly to demonstrated actions — an approach the authors say is prone to overfitting observation-action correspondences. To support this, the team built a large-scale pretraining dataset and derived physical state-transition annotations covering task-relevant semantics and visual motion changes. At inference the model still generates actions directly from current observations with no added overhead; the authors report performance competitive with the state of the art on LIBERO and RoboTwin and better than π0.5 in real-robot experiments, with code released.
If reproducible, this supervision signal could help robot VLA models move beyond surface imitation of demonstrations toward actions grounded in actual physical state change.