Vision-language-action (VLA) models are designed to leverage large-scale pretraining for generalist manipulation. But deployed policies must also support continual learning, so they can pick up new tasks over time without losing what they already know. A new preprint on arXiv introduces SAMBAR, a method aimed at this problem.
According to the abstract, SAMBAR stands for Selective Anchoring via Method of Multipliers. The method is intended to balance knowledge acquisition and retention—that is, learning new tasks while preserving existing capabilities. The abstract does not provide experimental details, so the reported effectiveness of the approach is not yet available.
Because there is only one source, there are no conflicting findings to compare. The paper is listed as a new announcement on arXiv under ID 2609.32108.