Vision-language-action (VLA) models have shown promise for general-purpose robotic manipulation, but they still fall short when asked to handle unseen tasks that require transferring relevant experience across different objects. That gap is the target of a new paper on arXiv, which proposes a method called VLA-Pro.

The paper frames the problem as a need for cross-task procedural memory transfer: instead of treating each manipulation task in isolation, a robot should be able to draw on procedural knowledge gained from prior tasks and apply it to novel situations. VLA-Pro is presented as a way to enable that transfer, though the abstract does not provide full implementation details or experimental results.

As the source is a single arXiv announcement with a truncated abstract, the article cannot confirm how VLA-Pro performs in practice or how it compares with other approaches. The contribution, as described, is a conceptual and methodological step toward improving generalization in VLA models by making procedural memory reusable across tasks.