Vision-language-action (VLA) models have become the default recipe for robot manipulation: take a vision-language backbone, feed it demonstration data, and train it to emit actions. A new paper from the OmniAI team at Zhejiang University calls out a gap that most of this line of work quietly sidesteps — mainstream VLAs learn a direct mapping from observations to demonstrated actions, and the physical state changes in between (did the object move, did the gripper make contact) are never an explicit learning target. The consequence, the authors argue, is overfitting to observation-action correspondences.
VLA-Δ (read "VLA-Delta") compresses its proposal into one sentence: make the physical state transition itself the supervision signal.
Three supervision channels, one pretraining run
The method constructs three explicit annotations between current and future observations: future-state difference text (a language description of what changed), structured transition slots covering task progress, spatial motion, contact changes and physical constraints, and dense optical flow. To support this, the team built a 484-hour, 24-million-frame pretraining corpus with an automated annotation pipeline. Pretraining runs in three stages: first learning future-state difference text, then initializing slot and optical-flow representations, then joint action prediction. The base is the 2B ZR-0 model, Qwen3-VL serves as the vision-language backbone, and a Flow Matching Action Expert generates continuous actions. The practically important detail: at inference time the model maps observations to actions directly, so all three supervision channels stay on the training side — zero added deployment overhead, in contrast to CoT-style approaches that must generate intermediate text while running.
The numbers: top of both comparison tables, pi0.5 beaten on real robots
On LIBERO, VLA-Δ averages 98.4 (Spatial 98.8, Object 99.8, Goal 98.6, Long 96.4), placing it above openVLA-OFT (97.1), GR00T-N1.7 (97.0), pi0.5 (96.9) and ZR-0 (97.8) in the paper's comparison table. On RoboTwin 2.0's Clean/Random settings it scores 90.06/89.74, ahead of ZR-0 (88.70/87.98), Motus (88.52/87.02) and pi0.5 (82.74/76.76). In real-robot experiments the paper reports a 5.5% advantage over pi0.5. Paper and code: arXiv:2610.21306 and the GitHub repo.
My read: competing on supervision design, not scale
The value here is not the 98.4 itself — the top of the LIBERO board is crowded into the 96-98 band, where fractions of a point prove little. It is the reframing: VLA-Δ pulls the competition axis away from "bigger VLM, more data" back toward "what the model is asked to understand". Physical-state-transition supervision forces the model to learn what its actions change rather than memorizing mappings, and because all of it happens during training, inference collects the benefit for free. There is also a signal for the embodied-AI community: the open-source artifacts are complete — Apache 2.0 license, three downloadable checkpoints (base, LIBERO, RoboTwin), reproduction barriers pushed to the floor.
The cold water is ready too: simulation scores near the ceiling carry limited discrimination, and the real-robot 5.5% is more persuasive but its scale and task coverage deserve a read of the paper itself. Whether VLA-Δ becomes a standard pretraining recipe will depend on whether the annotation pipeline transfers to other people's data.
So the takeaway: as end-to-end observation-to-action mapping hits diminishing returns, the next gains may hide in what the model is asked to understand — supervision-signal design is becoming the new battleground for embodied AI.