| --- |
| library_name: diffusers |
| pipeline_tag: robotics |
| tags: |
| - robotics |
| - vision-language-action |
| - pi0.5 |
| - diffusers |
| license: other |
| --- |
| |
| # PI0.5 Base — GigaModels PyTorch Conversion |
|
|
| This repository contains a PyTorch Diffusers-format conversion of the |
| Physical Intelligence PI0.5 base checkpoint. The weights were converted from |
| the official OpenPI JAX checkpoint with the following script from |
| [open-gigaai/giga-models](https://github.com/open-gigaai/giga-models): |
|
|
| - [`projects/vla/pi0/scripts/convert_jax_model_to_pytorch.py`](https://github.com/open-gigaai/giga-models/blob/main/projects/vla/pi0/scripts/convert_jax_model_to_pytorch.py) |
| - [GigaModels PI0/PI0.5 documentation](https://github.com/open-gigaai/giga-models/blob/main/projects/vla/pi0/README.md) |
|
|
| No additional fine-tuning was applied to these base weights. The tokenizer |
| files are colocated with the converted policy so consumers can use one Hub |
| repository for both the policy and tokenizer. |
|
|
| ## Loading with GigaModels |
|
|
| ```python |
| from giga_models import PI0Policy |
| |
| policy = PI0Policy.from_pretrained("Miical/pi05-base") |
| ``` |
|
|
| ## Loading with verl-vla |
|
|
| Set the model path directly to the Hub repository: |
|
|
| ```yaml |
| cluster: |
| actor_rollout_ref: |
| model: |
| path: Miical/pi05-base |
| ``` |
|
|
| Normalization statistics are embodiment- and dataset-specific and are not |
| included in this base checkpoint. Configure `model.adapter.norm_stats_path` |
| when training or evaluating with verl-vla. |
|
|
| ## Attribution and license |
|
|
| The conversion implementation is provided by GigaModels under Apache-2.0. |
| Use and redistribution of the converted weights remain subject to the terms |
| applicable to the original PI0.5 checkpoint and its underlying components. |
| Please also cite the original PI0/PI0.5 work as requested by the GigaModels |
| documentation. |
|
|
