VLA-Δ trains robot policies on state deltas, beats π₀.₅

The claim is that supervising a robot policy on the physical change between the current frame and a future one, rather than on the next demonstrated action, lifts benchmark scores and beats Physical Intelligence's π₀.₅ model on real robots at no inference cost.

That is the pitch of VLA-Δ, posted to arXiv on October 15 by Qi Lu, Jiarun Li and six collaborators releasing code under the ZJU-OmniAI handle. "VLA-Δ achieves performance competitive with the state of the art on LIBERO and RoboTwin, and outperforms π₀.₅ in real-robot experiments," the paper reports.

The authors' diagnosis of the field is where the paper opens. Most VLA models, they argue, "directly map observations to demonstrated actions without explicit supervision of intermediate physical state changes, making them prone to overfitting observation-action correspondences." Their fix is to train on deltas: "physical changes between current and future states as explicit learning targets," learned from a new large-scale pretraining dataset annotated for "task-relevant semantics and visual motion changes."

At deployment the trick is free. "The model generates actions directly from current observations without additional inference overhead," the abstract says.

The abstract reports no per-benchmark numbers for LIBERO, RoboTwin, or the real-robot comparison, so the size of the π₀.₅ gap stays inside the full paper. It joins Sber's GreenWAM, flagged in our tracker two days earlier, on a short list of recent models claiming wins over π₀.₅.