PriorVLA: Prior-Preserving Adaptation for Robotics

• PriorVLA is a framework that efficiently adapts vision-language-action models for robotic manipulation by preserving broad pretrained priors while enabling task-specific adaptation.
• The framework integrates learnable scene, motor, and action queries with dual action experts to extract and route prior information for improved generalization in out-of-distribution and few-shot scenarios.
• PriorVLA updates only about 25% of model parameters during training, yielding significant performance gains on simulation benchmarks and real-world robotics with reduced overfitting risks.

PriorVLA is a framework for efficient adaptation of large-scale Vision-Language-Action (VLA) models for robotic manipulation that preserves broad pretrained priors while enabling strong downstream performance. The framework addresses the degradation of generalization commonly observed in conventional full fine-tuning, which shifts model behavior toward narrow, training-distribution-specific patterns. PriorVLA achieves performance and sample efficiency advantages by architectural innovations that separate prior preservation from task adaptation and by integrating learnable interfaces for extracting and routing priors. Significant improvements are demonstrated on standard simulation benchmarks and real-world robotic tasks, particularly under out-of-distribution (OOD) and few-shot learning regimes (Guo et al., 11 May 2026).

1. Architectural Foundations: Dual Action Experts and Policy Decomposition

PriorVLA builds on a base VLA policy composed of a pretrained vision-LLM (VLM) and an action expert (AE) trained via flow-matching to denoise future action sequences. The approach instantiates two parallel branches from the AE:

• Prior Expert (PE): A frozen clone of the pretrained AE, serving exclusively as a read-only provider of internal “motor-prior” features. Its outputs are discarded for control, but its hidden states expose high-level information about generic manipulation intent and initialization.
• Adaptation Expert (AE$_\text{Ada}$): A trainable clone initialized from the same weights as the PE. AE$_\text{Ada}$ is responsible for adapting to and acting within new downstream tasks, predicting denoised action updates that are actually executed.

At each denoising iteration, both experts are run in parallel on the same noisy action chunk input, but only the outputs of AE$_\text{Ada}$ are used to update the control trajectory: $$\tilde{\mathbf a}^{\tau+1} = \mathrm{FM}(\tilde{\mathbf a}^\tau,\;f^\tau_\text{Ada})$$ where $\mathrm{FM}$ denotes the flow-matching update and $f^\tau_\text{Ada}$ is AE$_\text{Ada}$'s denoising output (Guo et al., 11 May 2026).

2. Expert Queries: Mechanisms for Prior Extraction and Routing

To enable the adaptation branch to leverage broad pretrained priors without direct fine-tuning, PriorVLA introduces three groups of learnable query tokens, each employing standard Transformer attention:

• Scene Queries (SQ): Inserted into every VLM layer, these tokens attend to, and are attended by, vision and language tokens. SQs extract “scene-priors” from the frozen VLM, synthesizing a latent summary of visual context relevant to the current scene.
• Motor Queries (MQ): Exclusive to the Prior Expert, MQs attend only to the PE’s noisy action tokens. This design exposes broad motor skills and manipulation strategies encoded in pretraining, without modifying the Prior Expert itself.
• Action Queries (AQ): Appended solely to AE$_\text{Ada}$ layers, these tokens aggregate prior features from SQs, MQs, and AE$_\text{Ada}$’s own internal state, routing comprehensive prior information to the adaptation process.

These queries collectively enable the Adaptation Expert to integrate both scene and motor priors as input for task-specific policy optimization, all while keeping the source branches immutable (Guo et al., 11 May 2026).

3. Training Strategy and Parameter-Efficiency

The optimization objective is the standard flow-matching mean squared error (MSE) loss, applied only to the outputs of AE$_\text{Ada}$. The Prior Expert is completely frozen and not included in any loss computation: $_\text{Ada}$0

Parameter update policy:

• Frozen: All VLM components except the vision encoder, plus the Prior Expert branch.
• Trainable (~25% of the model): Vision encoder, all query modules (SQ, MQ, AQ), and the entire AE$_\text{Ada}$1 branch.

This approach updates only about 25% of the total model parameters compared to standard full fine-tuning (which updates 100%), yielding both computational benefits and reducing overfitting risks (Guo et al., 11 May 2026).

4. Practical Workflow and Algorithmic Steps

The PriorVLA procedure is as follows:

1. Initialize PE and AE$_\text{Ada}$2 from the pretrained AE, freezing PE and rendering AE$_\text{Ada}$3 trainable.
2. Insert SQs into VLM layers, MQs into PE layers, and AQs into AE$_\text{Ada}$4 layers.
3. Freeze all VLM parameters except the vision encoder and all PE parameters.
4. For each minibatch, compute noisy action chunks. At each denoising step:
   • Run VLM+SQs for scene features.
   • Run PE+MQs to extract motor priors.
   • Run AE$_\text{Ada}$5+AQs, integrating scene and motor priors for adapted action prediction.
   • Update actions via flow-matching.
   • Accumulate loss from AE$_\text{Ada}$6 outputs.
5. Backpropagate and update trainable parameters only.
6. Repeat until convergence.

This efficient workflow ensures that large parts of the foundation model remain unchanged, supporting scalability to resource-constrained regimes (Guo et al., 11 May 2026).

5. Experimental Evaluation and Quantitative Results

PriorVLA exhibits statistically significant improvements across diverse settings:

• Simulation (RoboTwin 2.0, LIBERO):
  • On RoboTwin 2.0 (bimanual, 13 tasks): Outperforms $_\text{Ada}$7 by +10 and +11 points on “Easy” (ID) and “Hard” (OOD) modes, respectively (77% vs. 67%, 53% vs. 42%). In few-shot (10 demos) regimes, gains of +12 (Easy) and +11 (Hard) points.
  • On LIBERO (four 10-task suites): Achieves 99.1% average success, exceeding both $_\text{Ada}$8 (96.9%) and OpenVLA-OFT (97.1%).
• Real-World Robotics:
  • Eight tasks (Franka single-arm, AC-One dual-arm): With full data, achieves 81% (ID) and 57% (OOD) success, compared to 69% and 41% ($_\text{Ada}$9). In few-shot setting, 48% (ID) and 32% (OOD), surpassing $_\text{Ada}$0 by 24 and 22 points respectively.

These improvements persist across OOD and low-data conditions, with sign-test consistency at $_\text{Ada}$1 (Guo et al., 11 May 2026).

6. Ablations and Analytical Insights

Ablation experiments clarify the contributions of each architectural component:

• Prior Expert (PE) and MQ: Removing either drops Hard (OOD) task performance (49% → 42%). Replacing with a random or trainable branch also degrades results (max 44%).
• Expert Queries (SQ, MQ, AQ): Elimination of all queries collapses Hard success from 49% to 28%. Single-group removals also incur notable declines, with scene queries exerting the strongest influence.
• Vision Encoder: Freezing it reduces performance but not to the extent of removing prior preservation; visual adaptation is helpful but not a substitute for broad-prior access.

These findings confirm that both the preservation of pretrained priors and the ability to query them explicitly are essential for strong adaptation, especially in OOD regimes (Guo et al., 11 May 2026).

7. Comparative Context and Implications

PriorVLA is positioned among recent advances in prior-guided VLA adaptation and efficient parameter management for generalist robot manipulation. While some alternative methods (e.g., $_\text{Ada}$2VLA (Zhu et al., 9 Mar 2026)) approach the prior problem via causal grounding or discrete latent variation modeling, PriorVLA distinguishes itself by architectural decomposition, explicit query-mediated prior transfer, and parameter-freezing. The demonstrated empirical gains under resource-constrained and distribution-shift conditions substantiate the benefits of its separation-of-concerns approach.

A plausible implication is that, for scaling foundation models to rapidly evolving or sparsely demonstrated robotic tasks, frameworks that explicitly preserve and interface with broad pretraining priors—rather than overwrite them during fine-tuning—will remain critical to generalization and data efficiency.

---

References:

• PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (Guo et al., 11 May 2026)
• $_\text{Ada}$3VLA: Prior-Guided Vision-Language-Action Models via World Knowledge Variation (Zhu et al., 9 Mar 2026)