Zhejiang's VLA-Δ Supervises Vision-Language-Action Models on Physical State Transitions, Beats π₀.₅ on Real Robots

Zhejiang University's ZJU-OmniAI group introduces VLA-Δ, a framework that supervises vision-language-action models with the physical change between current and future states rather than just the next action. The researchers release a large pretraining dataset annotated with task-relevant semantics and motion deltas and report competitive LIBERO and RoboTwin numbers plus a win over π₀.₅ in real-robot experiments — with no additional inference overhead at deployment.