Question about data transform for Pi-0 fine-tuning
Dear authors, thank you for your wonderful work and for releasing the real-world dataset. I'm working on following your work and want to do fine-tuning on Pi-0 model with your dataset. I have a question about whether you use data transforms on the actions. On one hand, Pi-0 model is pretrained on normalized actions, so I think adding a similar normalization transform on the UAV-Flow dataset will help. On the other hand, the states in UAV-Flow all start from (0, 0, 0), and adding normalization will influence this. Therefore, I want to know how you process this problem, and I am looking forward to your helpful insights!
Thank you for your interest in our work! We normalized the actions in the UAV-Flow dataset using the 1st and 99th percentile values to scale them to the range [-1, 1] during training. This ensures consistency with the Pi-0 model’s pretraining.