| license: gemma | |
| library_name: lerobot | |
| pipeline_tag: robotics | |
| tags: | |
| - robotics | |
| - lerobot | |
| - pi052 | |
| - vision-language-action | |
| base_model: lerobot/pi05_base | |
| # PI0.5 v2 initialization checkpoint | |
| `lerobot/pi052_base` is the native LeRobot initialization checkpoint for the | |
| PI0.5 v2 (`pi052`) policy. | |
| It contains the same pretrained tensor values as | |
| [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base), converted once | |
| to PI052's native LeRobot state-dict layout and paired with a `pi052` | |
| configuration and processor pipeline. It can therefore be loaded through | |
| LeRobot's standard checkpoint loader without runtime key conversion. | |
| This checkpoint has **not** received additional PI052 hierarchical-language or | |
| FAST-objective training. It is the starting point for PI052 fine-tuning, not a | |
| trained hierarchical policy. | |
| ## Load | |
| ```python | |
| from lerobot.policies.pi052.modeling_pi052 import PI052Policy | |
| policy = PI052Policy.from_pretrained("lerobot/pi052_base") | |
| ``` | |
| ## Fine-tune | |
| ```bash | |
| lerobot-train \ | |
| --dataset.repo_id=${HF_USER}/my_language_annotated_dataset \ | |
| --policy.type=pi052 \ | |
| --policy.pretrained_path=lerobot/pi052_base \ | |
| --policy.recipe_path=recipes/subtask_mem.yaml \ | |
| --policy.dtype=bfloat16 \ | |
| --policy.device=cuda | |
| ``` | |
| PI0.5 is derived from PaliGemma and is distributed under the Gemma license. | |