Transferability of Existing Augmentation Recipes to Vision-Language-Action Models

Determine whether image-augmentation recipes developed for conventional convolutional actor-critic policies transfer effectively to modern vision-language-action models with pretrained vision-language backbones and flow-matching action heads.

Background

The paper contrasts established augmentation strategies for small convolutional policies, including methods that augment the value pathway or both actor and critic inputs, with the substantially different architecture of modern vision-language-action models. In these models, augmentation may corrupt representations from a large pretrained vision-LLM, while policy optimization must propagate gradients through a flow-matching action head.

The authors present their work as the first systematic study of augmentation placement, type, and strength for VLA post-training, thereby addressing this previously unresolved transfer question within their experimental scope.

References

Yet this line of work studies small CNN encoders with Gaussian or categorical actors, a regime far from modern vision--language--action models. In VLAs the visual encoder is a large pretrained VLM whose features risk being corrupted, and the actor is a flow-matching head through which gradients must propagate. Whether existing augmentation recipes transfer to this setting remains unknown.

— How (and How Not) to Use Data Augmentation in VLA Post-Training  (2610.05994 - Grooten et al., 5 Oct 2026) in Section 1, Introduction

Our study is bounded in scope: all runs are in simulation, on one benchmark, and with one RL algorithm, so whether the conclusions hold on real hardware, on other task suites, or with off-policy algorithms remains untested.

— How (and How Not) to Use Data Augmentation in VLA Post-Training  (2610.05994 - Grooten et al., 5 Oct 2026) in Section 6, Discussion, paragraph titled “Limitations”