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

2026-05-11

AI Summary

This work addresses the issue of catastrophic forgetting in existing vision-language-action (VLA) models during downstream fine-tuning, which compromises their generalization by overwriting pretrained priors. To mitigate this, the authors propose PriorVLA, a novel framework that decouples prior preservation from task adaptation: the pretrained model is frozen as a read-only source of prior knowledge, while a trainable Adaptation Expert is introduced to selectively integrate scene and action priors via an expert query mechanism. Requiring only 25% of the parameters updated in full fine-tuning, PriorVLA substantially outperforms both full fine-tuning and state-of-the-art baselines across RoboTwin 2.0, LIBERO, and real-robot tasks, achieving a remarkable 99.1% average success rate on LIBERO. On real-world benchmarks, it attains 81%/57% in-distribution/out-of-distribution success rates under standard settings and 48%/32% with only 10 examples—significantly surpassing pi0.5.

Abstract

Large-scale pretraining has made Vision-Language-Action (VLA) models promising foundations for generalist robot manipulation, yet adapting them to downstream tasks remains necessary. However, the common practice of full fine-tuning treats pretraining as initialization and can shift broad priors toward narrow training-distribution patterns. We propose PriorVLA, 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. Across RoboTwin 2.0, LIBERO, and real-world tasks, PriorVLA achieves stronger overall performance than full fine-tuning and state-of-the-art VLA baselines, with the largest gains under out-of-distribution (OOD) and few-shot settings. PriorVLA improves over pi0.5 by 11 points on RoboTwin 2.0-Hard and achieves 99.1% average success on LIBERO. Across eight real-world tasks and two embodiments, PriorVLA reaches 81% in-distribution (ID) and 57% OOD success with standard data. With only 10 demonstrations per task, PriorVLA reaches 48% ID and 32% OOD success, surpassing pi0.5 by 24 and 22 points, respectively.

Problem

• Vision-Language-Action Models
• Prior Preservation
• Downstream Adaptation
• Out-of-Distribution Generalization
• Robot Manipulation

Innovation

• Prior Preservation
• Vision-Language-Action Models
• Parameter-Efficient Adaptation
• Expert Queries
• Out-of-Distribution Generalization