Details regarding the training of pi0-sf？

[yuefanhao]

Hello, excellent work. My question is:
1.When pi0-sf is evaluated on the libero, was it only trained on the libero dataset or did it also use some other publicly datasets? like robotwin?
2.Why wasn't siglip frozen during the training process?
3.I noticed in the paper that LORA training was conducted on Robotwin using pi0_base. Why wasn't it trained in Full-Tuning?
thank you.

[haofuly]

1. We don't provide an official result based on pi0-libero config. But we provided the OpenvlaOFT-libero and pi0-robotwin results. These results are trained only on the target dataset without other public datasets.
2. Just follow the base model training setting.
3. Just for fast training with low computation cost.