I would like to know the normalization parameters of the base model.

#7
by yiming0720 - opened

Hi,

I'm trying to deploy pi05_base outside of the LeRobot runtime
(e.g. ONNX / TensorRT deployment).

The released checkpoint contains:

  • model.safetensors
  • config.json
  • policy_preprocessor.json
  • policy_postprocessor.json

However, I could not find the normalization statistics
(mean/std or q01/q99) used for STATE and ACTION features.

Could you clarify:

  1. Are the normalization statistics included somewhere in the released checkpoint?

  2. If not, where can they be obtained?

  3. Were the original OpenPI normalization parameters removed during the conversion to LeRobot?

  4. If the original normalization statistics are not available, would it be reasonable to use the normalization parameters from the released finetuned checkpoints (e.g. pi05_libero_finetuned_quantiles_v044 ) for deployment validation and ONNX/TensorRT testing?

I'm trying to reproduce inference independently from the LeRobot runtime.

Thanks!

Sign up or log in to comment