Adapting a frozen flow-matching VLA by learning only the noise it starts fromExpert
Noise Space Learning for VLA Adaptation
On four real-robot tasks, a noise actor trained for about 1.5 minutes per task lifted a frozen pi0.5 policy's average success from 37.5% to 71.3%, and to 83.8% when a human added corrections. The preprint from Microsoft Research, UTS, Nanjing and Tsinghua leaves the big model untouched and only learns where its random starting noise should come from, which makes adaptation cheap and keeps the pretrained skills intact.