Vision-language-action (VLA) models are often described as a step toward generalist robots, but their training data is scattered across different platforms and environments. That fragmentation makes it hard for a single model to consolidate skills learned in one context and apply them in another. A new arXiv paper proposes NestedVLA, a framework designed to address this by learning to consolidate and generate skills.

While the abstract does not detail the internal mechanics, the framing suggests a hierarchical or nested structure that lets the model build on existing skills rather than relearning them from scratch. The authors position this as a way to move beyond the current practice of training on isolated datasets.

Because the abstract is truncated, the full evaluation and method are not available from this source alone. The contribution, as stated, is a system that targets the fragmentation problem directly.