Inquiry regarding DexVLA pretrained model evaluation and training recipe
First of all, thank you for sharing your pretrained model. Since I lack the resources to train DexVLA models from scratch, I am very interested in utilizing your model for inference and potential fine-tuning.
However, I am facing an issue where the model outputs NaN values when I try to run it using the official DexVLA evaluation code. I have verified that the statistics in the .pkl file (mean/std) appear normal, but the runtime outputs remain NaN.
So, I have a few questions:
1. Training Recipe: Did you use the exact same training recipe as the official one? If there were any modifications, could you please share them?
2. Evaluation Environment: How did you evaluate this model? Is there a specific benchmark or a modified evaluation script I should use to avoid the NaN issue?
3. Input Preprocessing: Are there any specific preprocessing steps or environment settings I might be missing that could lead to these numerical instabilities?
Thank you for your time and for contributing to the community.)