PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models

Technical Analysis "PriorVLA" is a novel framework that preserves pretrained priors and learns to leverage them for effective adaptation. PriorVLA keeps a frozen Prior Expert as a read-only prior source and trains an Adaptation Expert for downstream specialization. Expert Queries capture scene priors from the pretrained VLM and motor priors from the Prior Expert, integrating both into the Adaptation Expert to guide adaptation. Together, PriorVLA updates only 25% of the parameters updated by full fine-tuning. High-level Analogy: Imagine you have a highly skilled, experienced master chef who knows how to cook everything (this is the Pretrained VLA Model). Now, you want a junior chef to learn a specific new recipe for a restaurant (this is the downstream task). Full fine-tuning would be like giving the junior chef the master chef's original, one-of-a-kind cookbook and telling them to just start scribbling changes directly into it to make the new recipe. They might get good at that specific recipe, but they might accidentally ruin or 'forget' the master's other fundamental recipes for diverse dishes (this is prior forgetting leading to poor performance on different, 'out-of-distribution' situations). PriorVLA works differently: The Master's Cookbook is Photocopied and Locked Away (Prior Expert): You make a perfect, unchangeable copy of the master chef's original cookbook. This copy is 'frozen' and can only be read, never written on. It holds all the fundamental cooking…