Researchers led by Arthur Bucker, with Jean Oh among the authors, have built VG-Steer, which corrects a trained robot policy at run time when conditions differ from its training, according to a paper on arXiv. It transfers the motion of earlier successful runs along the distances between pairs of task-relevant objects, which large vision and vision-language models identify.
In a real robot experiment, it more than doubled a frozen policy's success at placing cables one grid step outside its training positions. It also steered a single-task policy towards objects it had never been trained on and almost removed confusion between objects. In simulation it corrected flow-matching policies when object poses and goals shifted.



