| --- |
| library_name: lerobot |
| tags: |
| - molmoact2 |
| - robotics |
| - lerobot |
| - vla |
| base_model: allenai/MolmoAct2 |
| --- |
| |
| # molmoact2_block_stack_norm_fix |
|
|
| Fine-tuned [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) (action-expert-only) on SO101 data, with SO101 gripper normalization fix. |
|
|
| | | | |
| |---|---| |
| | **Base model** | [allenai/MolmoAct2](https://huggingface.co/allenai/MolmoAct2) | |
| | **Dataset** | [villekuosmanen/armnetbench_block_stack](https://huggingface.co/datasets/villekuosmanen/armnetbench_block_stack) | |
| | **Task** | `block_stack` | |
| | **Action dim** | 6 (single-arm) | |
| | **Cameras** | `top`, `wrist`, `front` | |
| | **Normalization** | MEAN_STD (action + state + gripper), IDENTITY (visual) | |
| | **Training** | Isambard GH200, batch 64, bf16, gradient checkpointing | |
| | **W&B project** | [molmoact2_block_stack](https://wandb.ai/pravsels/molmoact2_block_stack) | |
| | **W&B run** | [eqx1ihxu](https://wandb.ai/pravsels/molmoact2_block_stack/runs/eqx1ihxu) | |
| |
| ## Checkpoints |
| |
| | Step | Path | |
| |------|------| |
| | 002000 | `checkpoints/002000/pretrained_model/` | |
|
|
| The latest step (`002000`) is also copied to the **repository root** for direct loading. |
|
|
| ## Usage |
|
|
| ```python |
| from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy |
| policy = MolmoAct2Policy.from_pretrained("pravsels/molmoact2_block_stack_norm_fix") |
| ``` |
|
|
