| language: | |
| - en | |
| library_name: lerobot | |
| pipeline_tag: robotics | |
| tags: | |
| - vision-language-action | |
| - lerobot | |
| inference: false | |
| license: gemma | |
| base_model: lerobot/pi05_base | |
| # framepick base checkpoint (pi0.5 + subtask + frame re-expression) | |
| Weights are bit-identical to [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base). | |
| The checkpoint self-describes the **framepick** architecture (LeRobot plugin | |
| `lerobot_policy_framepick`): `config.json` has `type: framepick`, and | |
| `policy_preprocessor.json` carries the subtask tokenizer step plus the framepick | |
| step with a NEUTRAL frame selector. Experiment knobs (which frames, selection | |
| strategy) live in the training config, e.g. | |
| `--policy.frame_selector.assignment='{"arm": "image2"}'`. | |
| Generated by `lerobot_policy_pi05_subtask/scripts/prepare_base_checkpoint.py | |
| --policy-type framepick`. See also `dgrachev/pi05_subtask_base` (same weights, | |
| pi05_subtask architecture) and the `dgrachev/libero_extrinsics` dataset. | |