| license: apache-2.0 | |
| tags: | |
| - robotics | |
| - vla | |
| - molmoact2 | |
| - droid | |
| # MolmoAct2-DROID — absolute joint-pose full fine-tune | |
| Full fine-tune of `allenai/MolmoAct2-DROID` on the filtered DROID mixture | |
| (`shrg7/droid-equad-stage-1-jointabs`, LeRobot v3.0) with **absolute joint-pose** | |
| action targets (arm[t] = q[t+1], 7 arm joints + gripper). | |
| - Trainer: lerobot-native `MolmoAct2Policy` (continuous flow action expert) | |
| - `control_mode="absolute joint pose"`, `action_mode=continuous`, chunk 15 | |
| - Norm stats derived from this dataset (avoids the franka_droid clip issue) | |
| - Checkpoint: step 16000 (training in progress; this is an intermediate checkpoint) | |
| Standard lerobot MolmoAct2 checkpoint dir: `model.safetensors` + `config.json` + | |
| pre/post-processor pipelines (`policy_preprocessor*`, `policy_postprocessor*`) + | |
| `train_config.json`. | |