pi0-base PyTorch checkpoint

This repository contains an unofficial PyTorch/Safetensors format conversion of the pi0_base checkpoint published by Physical Intelligence.

Provenance

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-base model weights
  • config.json: model architecture reference
  • LICENSE_GEMMA.txt: Gemma Terms of Use distributed with OpenPI
  • LICENSE_OPENPI: Apache License 2.0 distributed with OpenPI
  • NOTICE: 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.

Downloads last month
37
Safetensors
Model size
4B params
Tensor type
BF16
·
Video Preview
loading