Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

Analysis

Why This Paper Matters

Vision-language-action (VLA) models are a promising paradigm for robot manipulation, but their standard finetuning recipe—behavior cloning on robot demonstrations—has a hidden flaw: it progressively overwrites the rich visual and semantic representations learned during pretraining. This paper, Anchor-Align, identifies and addresses this representation drift, showing that it leads to language-action misalignment that standard benchmarks fail to expose. The work is significant because it challenges the assumption that co-training on web data can prevent this drift, and instead proposes a principled solution that preserves pretrained knowledge while still learning effective actions.

The paper's focus on real-robot validation (xArm7) and large-scale simulation (LIBERO-PRO, LIBERO-Plus, CALVIN) adds credibility. The consistent improvements across two different VLA architectures suggest the method is broadly applicable, not tied to a specific model design. For AI practitioners deploying VLA policies, this work offers a practical, low-overhead finetuning augmentation that could substantially improve robustness and generalization.

Technical Contributions

• Representation Drift Diagnosis: The paper provides empirical evidence that behavior cloning finetuning overwrites pretrained VLM representations, and that standard co-training on web image-text data does not prevent this because language and action losses are applied to separate observations.
• Vision-Language Anchoring: A distillation loss that aligns layer-wise representations of the finetuned model with a frozen copy of the pretrained VLM. This prevents catastrophic forgetting of visual and semantic features without requiring additional web data.
• Language-Action Alignment: Converts each continuous action target into a discrete motion-direction label (e.g., 'move left', 'grasp') and jointly trains language and action prediction on the same robot observation. This forces the model to maintain a shared representation for both modalities.
• Architecture-Agnostic Design: The method is evaluated on two widely used VLA architectures, demonstrating that it can be plugged into existing pipelines without architectural changes.

Results

• Real-Robot xArm7: Success rate improved from 28% to 54% (absolute +26%) on one VLA architecture, and from 37% to 60% (absolute +23%) on another.
• Simulation Benchmarks:
  • LIBERO-PRO (OOD perturbations): Consistent improvements across multiple perturbation types.
  • LIBERO-Plus (perceptual robustness): Improved robustness to visual distractors and lighting changes.
  • CALVIN (long-horizon control): Better task completion rates in long-horizon manipulation sequences.
• The paper does not report absolute numbers for all simulation results, but states 'consistent improvements' across all three benchmarks.

Significance

Anchor-Align offers a simple yet effective solution to a fundamental problem in VLA finetuning: the tension between preserving pretrained representations and learning new action capabilities. By showing that these goals are not fundamentally at odds, the work opens the door to more robust and generalizable robot policies. The method's architecture-agnostic nature and lack of reliance on additional web data make it practical for real-world deployment. Future work could explore extending the anchoring to other modalities (e.g., tactile or depth) and investigating the trade-offs of discrete action labels for fine-grained control.