---
library_name: lerobot
tags:
- molmoact2
- robotics
- lerobot
- vla
- bimanual
base_model: allenai/MolmoAct2-SO100_101
---

# molmoact2_transfer_cube_20k

Fine-tuned [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) (action-expert-only) for `transfer_cube` on SO101 data.

| | |
|---|---|
| **Policy** | MolmoAct2 (`policy.type=molmoact2`) |
| **Init checkpoint** | [allenai/MolmoAct2-SO100_101](https://huggingface.co/allenai/MolmoAct2-SO100_101) |
| **Dataset** | [villekuosmanen/armnetbench_transfer_cube](https://huggingface.co/datasets/villekuosmanen/armnetbench_transfer_cube) |
| **Task** | `transfer_cube` |
| **Action dim** | 12 (bimanual) |
| **Cameras** | `top`, `left_wrist`, `right_wrist` |
| **Training** | 20k steps, batch 32 (8/GPU × 4 DDP), 4× RTX 5090 on Vast.ai |
| **Prior HF repo** | [pravsels/molmoact2_transfer_cube](https://huggingface.co/pravsels/molmoact2_transfer_cube) |
| **W&B project** | [molmoact2_transfer_cube](https://wandb.ai/pravsels/molmoact2_transfer_cube) |
| **W&B run** | [lrmsb00d](https://wandb.ai/pravsels/molmoact2_transfer_cube/runs/lrmsb00d) |

## Checkpoints

The final checkpoint (local step `020000`, 20k total steps) lives at the **repository root** for direct loading.

## Usage

```python
from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy
policy = MolmoAct2Policy.from_pretrained("pravsels/molmoact2_transfer_cube_20k")
```