metadata
library_name: lerobot
tags:
- molmoact2
- robotics
- lerobot
- vla
base_model: allenai/MolmoAct2
molmoact2_tool_insert_quantile_norm_fix_25k
Fine-tuned MolmoAct2 (action-expert-only) for tool_insert on SO101 data.
| Policy | MolmoAct2 (policy.type=molmoact2) |
| Init checkpoint | allenai/MolmoAct2 |
| Dataset | villekuosmanen/armnetbench_tool_insert |
| Task | tool_insert |
| Action dim | 6 (single-arm) |
| Cameras | top, wrist, front |
| Training | 25k steps, QUANTILES norm, freeze, batch 32 global, Isambard GH200 |
| Prior HF repo | pravsels/molmoact2_tool_insert |
| W&B project | molmoact2_tool_insert_quantile_norm_fix_25k |
| W&B run | oiew4qmt |
Checkpoints
The checkpoint (local step 025000, 25k training steps) lives at the repository root for direct loading.
Verification
| Checkpoint step | 025000 |
| Source path | checkpoints/025000/pretrained_model/ |
| model.safetensors | 10,884,573,720 bytes, sha256 3535baf761b841961728d93b6f5ab6895b9057cf04c3846989040baa2e517782 |
| policy_preprocessor.json | 2,495 bytes, sha256 e7c8b8293cb0265a01f83278033272efcb21b4c7bfb031cfbd683ed74ee7b139 |
| policy_postprocessor.json | 757 bytes, sha256 6dbed1e1ec69e8c50f3a04c1f144a54231e3ef508f15fd7896ead43ea645b033 |
| train_config.json | 8,257 bytes, sha256 346d840c7769cee2cf86ec30041a738c02c95aea24b4e983903b9de172b0904c |
Verify after download:
sha256sum model.safetensors
# expected: 3535baf761b841961728d93b6f5ab6895b9057cf04c3846989040baa2e517782
Usage
from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy
policy = MolmoAct2Policy.from_pretrained("pravsels/molmoact2_tool_insert_quantile_norm_fix_25k")