Vision-language-action (VLA) models combine visual perception, language understanding, and robot control, enabling instruction-conditioned manipulation. A new arXiv paper introduces VLRLA, a method for adapting such models at deployment time through in-context reinforcement learning. The abstract begins to describe a limitation of conventional pretrained VLA policies, but the sentence is cut off in the available text.

VLRLA is presented as a deployment-time adaptation approach, meaning the model adjusts its behavior when deployed rather than being retrained. The use of in-context reinforcement learning suggests the model learns from interactions at runtime. The paper is available on arXiv under identifier 2610.19517.