molmoact2_mpc_50 / README.md
jstm's picture
Add model card with per-env demonstration counts
11d8041 verified
|
Raw History Blame Contribute Delete
875 Bytes
---
library_name: lerobot
license: apache-2.0
tags:
- robotics
- molmoact2
- maniskill
- mpc
pipeline_tag: robotics
---
# MolmoAct2 MPC N=50
MolmoAct2 (`allenai/MolmoAct2-LIBERO`) finetuned on motion-planned ManiSkill MPC tabletop demos with absolute end-effector pose control (`pd_ee_pose`).
- **Dataset:** [`jstm/mpc_lerobot_pd_ee_pose_50`](https://huggingface.co/datasets/jstm/mpc_lerobot_pd_ee_pose_50)
- **Checkpoint:** training step 10 000
- **Control mode:** `pd_ee_pose`
- **Cameras:** `camera_center`, `camera_left`, `camera_wrist` (378×378)
## Demonstrations
| Env | Demonstrations |
| --- | ---: |
| PickCube-v2-wrist | 48 |
| PushCube-v2 | 48 |
| LiftPegUpright-v2 | 48 |
| PullCubeTool-v2 | 48 |
| **Total** | **192** |
Target collection size was N=50 successes per env; 48 successful demos were retained per env after motion planning / replay.