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

Summary

Vision-Language-Action (VLA) models have shown promise for generalizable robot control but struggle when adapting to new domains while preserving prior capabilities.

Abstract Summary

Vision-Language-Action (VLA) models have shown promise for generalizable robot control but struggle when adapting to new domains while preserving prior capabilities. This paper introduces PriorVLA, a method that preserves pre-trained priors during VLA fine-tuning through constrained optimization in ... The method demonstrates significant improvements over existing approaches, providing both theoretical insights and practical benefits for real-world deployment. Comprehensive experiments validate the effectiveness of the proposed approach across diverse scenarios and task settings.

Key Points

Proposes PriorVLA, a novel approach for vla in robotics.
Addresses key limitations in existing methods through innovative architecture design.
Demonstrates strong empirical results on standard benchmarks and real-world evaluations.
Introduces novel training methodology that improves generalization and sample efficiency.
Provides comprehensive analysis of failure modes and ablation studies.

Method

We consider a vision-language-action (VLA) policy that predicts a future action chunk for robot manipulation. Given visual observation $o_{t}$, language instruction $l$, and proprioceptive state $s_{t}$, the policy outputs an action chunk $\mathbf{a}_{t:t+H-1}$, where $H$ denotes the action horizon.

A common pretrained VLA instantiation couples a vision-language model (VLM) with a flow-matching-based action expert (AE) [ 4 , 5 ] .

Method — PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models — Figure 2: PriorVLA architecture. PriorVLA builds on a pretrained VLA and introduces two coupled modules: Dual Action Experts and Expert Queries. Dual Action Experts keep the original AE as a frozen Prior Expert and train an Adaptation Expert for downstream action generation. Expert Queries capture scene and motor priors from pretrained forward paths and integrate them into the Adaptation Expert; the Prior Expert serves only as a read-only prior source, while the Adaptation Expert drives denoising and produces the final action chunk.

Experiments

To comprehensively evaluate PriorVLA, we conduct experiments on RoboTwin 2.0 [ 22 ] , LIBERO [ 25 ] , and eight real-world tasks across two robot embodiments. We organize the evaluation around five questions: (1) Can PriorVLA improve downstream adaptation over full fine-tuning and strong VLA baselines? (2) Does preserving pretrained priors improve OOD generalization?

(3) Are the benefits stronger when downstream data coverage is limited? (4) Do these gains transfer to real-world robots across embodiments and evaluation conditions? (5) Are the g...

Experiments — PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models — Figure 4: Experimental overview. We evaluate PriorVLA on RoboTwin 2.0 ( 22 ) , LIBERO ( 25 ) , and two real-world robot embodiments, covering ID/OOD generalization, data regimes, and component ablations.

Sources and Demonstrations

Figures are reproduced from Guo et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.