| license: apache-2.0 | |
| tags: | |
| - robotics | |
| - vla | |
| - pi0 | |
| - droid | |
| # MolmoBot-Pi0 DROID — absolute joint-pose full fine-tune | |
| Full fine-tune of MolmoBot-Pi0 (PaliGemma-decoder VLA) on the filtered DROID | |
| mixture (`shrg7/droid-equad-stage-1`, LeRobot v3.0), with **absolute joint-pose** | |
| action targets (arm[t] = q[t+1], 7 arm joints + gripper). | |
| - Base: MolmoBot-Pi0-DROID | |
| - Config: `molmobot_pi0_lerobot_droid_absjoint` | |
| - Checkpoint: step 5000 (flat pi0 state dict) | |
| - Vision tower frozen; decoder + action expert + projections trained. | |
| `assets/` holds the normalization stats (`norm_stats`) and train config the eval | |
| policy loads. `metadata.pt` is the trainer metadata. Load `model.safetensors` | |
| into a plain `PI0Pytorch`. | |