Vision-language-action (VLA) models have recently made progress in robot manipulation by combining large-scale robot datasets with foundation models. According to the abstract, however, their ability to rapidly adapt to novel situations remains a challenge.

The paper introduces a method for one-shot human video-guided adaptation. This suggests a robot could learn a new task from a single video of a human performing it, rather than requiring many robot demonstrations or extensive fine-tuning.

The abstract does not include experimental details or results, so the effectiveness of the approach is not described in the source. The paper is listed on arXiv under identifier 2610.19746.