pi0-LIBERO PyTorch checkpoint
This repository contains an unofficial PyTorch/Safetensors format conversion
of the pi0_libero checkpoint published by Physical Intelligence.
Provenance
- Upstream project: https://github.com/Physical-Intelligence/openpi
- Upstream checkpoint:
gs://openpi-assets/checkpoints/pi0_libero
- Upstream model: pi0 LIBERO
- Local format: PyTorch Safetensors (
model.safetensors)
- Precision: bfloat16
The original JAX checkpoint was converted with OpenPI's
examples/convert_jax_model_to_pytorch.py conversion implementation. The
model was not additionally trained or fine-tuned as part of this conversion.
Modification notice
The checkpoint file format and parameter layout were converted from the
upstream JAX/Orbax representation to a PyTorch Safetensors representation.
This repository is not an official Physical Intelligence or Google release
and does not imply endorsement by either organization.
Files
model.safetensors: converted pi0-LIBERO model weightsconfig.json: model architecture referenceLICENSE_GEMMA.txt: Gemma Terms of Use distributed with OpenPILICENSE_OPENPI: Apache License 2.0 distributed with OpenPINOTICE: required Gemma notice and upstream attribution
Usage
The checkpoint is intended for the PyTorch implementation in OpenPI. For
example, the pi0-WU code in openpi-wu accepts this directory through
--base-pytorch-weight-path.
Users are responsible for reviewing and complying with the Gemma Terms of
Use, the Gemma Prohibited Use Policy, applicable laws, and the upstream
OpenPI license before using or redistributing these files.
