Vision-language-action (VLA) models transfer semantic knowledge from vision-language pretraining to robot control, achieving strong results on various benchmarks. The new preprint, VLA-Δ, notes that most existing VLAs directly map observations to actions, without explicitly modeling how the physical world changes in response to those actions.
The authors propose learning physical state transitions as an intermediate step. Instead of only outputting a motor command, the model would predict the next state given the current state and action. According to the abstract, this shift from direct mapping to transition learning is intended to improve performance across benchmarks.
Because only the abstract is available in the source, the specific architecture, training data, and experimental details are not covered here. This is a single preprint, so there are no conflicting accounts to compare; the summary reflects only what the abstract states.