| license: mit | |
| library_name: lerobot | |
| pipeline_tag: robotics | |
| tags: | |
| - lerobot | |
| - eo1 | |
| - vision-language-action | |
| base_model: IPEC-COMMUNITY/EO-1-3B | |
| # EO-1 base for LeRobot | |
| This is a lossless key-layout conversion of | |
| [`IPEC-COMMUNITY/EO-1-3B`](https://huggingface.co/IPEC-COMMUNITY/EO-1-3B) at revision | |
| `fe0a015d9349bdf886d898809213b3574eedf1f9` for LeRobot's native `eo1` policy implementation. | |
| The checkpoint preserves the released EO-1 Qwen vision-language weights, | |
| language-model head, state/action projectors, and flow-matching head. Its defaults | |
| match the release: action chunk size 16, maximum state/action dimension 32, two | |
| action-projector layers, and ten denoising steps. | |
| Use `--policy.path=lerobot/eo1-base` for training or rollout. LeRobot then | |
| applies the dataset or environment's camera, state, and action feature shapes | |
| before instantiating the policy. | |