One Wrong Action Normalization Ruins Everything: VLA Fine-Tuning Fails to Converge, and 80% Get Stuck at This Step

When VLA fine-tuning fails to converge, 80% of the problems come from a mistake in the action normalization step. The article focuses on the role of action normalization in VLA fine-tuning and explains that an error in this step directly causes training to fail to converge.