Researchers led by Jinxi Liu have built One View, One Action (OVOA-VLA), a vision-language-action model that learns to predict actions in the frame of each of its cameras, not only in the robot's base frame, according to a paper on arXiv. The same movement then gives a separate training signal for every view.
Experts for the base frame and for each view exchange information through two attention graphs, and a consistency term makes their base-frame predictions agree. The model averaged 88.1% success on LIBERO-Plus with full inference and 85.7% using the base expert alone. On three real tasks with shifted cameras and starting positions, the base-only version raised average success from 28.3% for π0.5 to 58.3%.



