VLA-Δ: teaching robots to anticipate physical changes rather than imitate a movement, and beating π0.5 in real-world conditions

Zhejiang University trains its vision-language-action models on the transition between the current and future states of the scene, at no additional cost at runtime.

Vision-language-action (VLA) models pilot robots based on knowledge acquired during image-to-text pre-training. But most directly associate an observation with the demonstrated action, without ever being guided on what needs to change physically in the scene. As a result, according to the ZJU-OmniAI team at Zhejiang University, they tend to memorize observation-action correspondences rather than understand the task.

Status difference as a target

VLA-Δ makes this physical transition, the gap between the present state and the future state, an explicit learning objective. The researchers built up a large pre-training dataset annotated with these transitions, covering both task direction and visible movements. Once trained, the model generates its actions directly from the current observation, without any additional computational steps to inference.

The system is up to state-of-the-art level on the simulated LIBERO and RoboTwin test benches, and exceeds π0.5, the Physical Intelligence model, in experiments on real robots. The code is published on GitHub.