Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment
Anchor-Align improves robot manipulation by augmenting behavior cloning with two objectives: Vision-Language Anchoring to preserve pretrained visual representations and Language-Action Alignment to ensure semantic grounding of actions. This method stabilizes finetuning, prevents catastrophic forgetting, and fixes language-action misalignment by linking continuous actions to discrete motion labels.