| --- |
| license: mit |
| --- |
| |
| # OAT RoboMimic Lift (Image) Artifacts |
|
|
| This repository stores intermediate training artifacts for OAT policy training on RoboMimic Lift (image observations). |
|
|
| ## Current checkpoint |
|
|
| - File: `checkpoints/oatpolicy_lift_latest_ep736.ckpt` |
| - Source run: `output/20260523/132544_train_oatpolicy_lift_N200` |
| - Last logged progress before disk-full interruption: `epoch=736`, `global_step=25048` |
|
|
| ## Included files |
|
|
| - `checkpoints/oatpolicy_lift_latest_ep736.ckpt` - policy checkpoint |
| - `logs/lift_train_tmux.log` - raw tmux training log |
| - `plots/lift_training_dashboard.png` - training dashboard (raster) |
| - `plots/lift_training_dashboard.svg` - training dashboard (vector) |
| - `plots/lift_training_dashboard_summary.json` - compact metrics summary |
|
|
| ## Resume training from checkpoint |
|
|
| Use the same config as your original run and pass the checkpoint path: |
|
|
| ```bash |
| uv run scripts/train.py \ |
| --config-name=train_oatpolicy \ |
| task/policy=robomimic/lift \ |
| training.resume=true \ |
| training.checkpoint_path=/workspace/OAT-RoboMimic-Fine-tune/output/20260523/132544_train_oatpolicy_lift_N200/checkpoints/latest.ckpt |
| ``` |
|
|
| If your setup uses a different resume flag/path key, keep the checkpoint file and run with your project's standard resume command. |
|
|
