Reproduce “VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling” on OpenResearch

About this paper

Weiqi Li, Quande Zhang, Ruifeng Zhai, Liang Lin, Guangrun Wang

Li et al. demonstrated that Vision-Language-Action (VLA) model brittleness under visual perturbations primarily stems from visual representation misalignment rather than limits in high-level reasoning. Their one-shot adaptation methods, Feature Token Modulation and Feature Linear Adaptation, restore robustness, with FLA achieving 94.8% success on diverse visual conditions with a 99x reduction in trainable parameters compared to previous fine-tuning approaches.

Key insights

• VLA model performance degradation under visual perturbations is primarily due to misaligned visual representations in the Spatial Modeling component, not an inherent limitation of the Physical Modeling (high-level policy).
• Targeted, lightweight adaptation of visual representations can efficiently reactivate latent robustness in pre-trained VLA models without extensive retraining or architectural changes.
• A one-shot adaptation protocol, utilizing a single human demonstration per task, significantly reduces the data and computational resources required for domain adaptation.

Summary by alphaXiv.