# pour_pi05_base

PyTorch `pi05` baseline policy fine-tuned on the local UR5 real-robot LeRobot-format pour dataset.

- dataset: `ur5_place_and_pour_nuts_camera_shifts`
- training config name: `pi05_ur5_pour_pytorch_baseline`
- experiment name: `pi05_ur5_pour_pytorch_baseline_tandon_2gpu_b16`
- source Slurm job: `9753605`
- base camera view: `observation.images.context_left_rgb`
- wrist camera view: `observation.images.wrist_right_rgb`
- base model init: `/scratch/yz11445/pi05_base`
- currently included checkpoint steps: `20000`, `25000`, `30000`

This run has completed through `30000` steps.

## Included Files

- `config.json`: base pi05 model config copied from the initialization checkpoint
- `model_architecture_config.json`: baseline architecture settings used by this run
- `training_config_summary.json`: training/data/run summary for this release
- `assets/ur5_place_and_pour_nuts_camera_shifts/norm_stats.json`: normalization statistics
- `checkpoints/<step>/`: hard-linked checkpoint snapshots with `model.safetensors`, `metadata.pt`, and hard-linked assets

## Inference

Serve one of the included checkpoints with:

```bash
uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_ur5_pour_pytorch_baseline \
  --policy.dir=/path/to/pour_pi05_base/checkpoints/30000
```

## Notes

- The policy loader uses the code-defined training config `pi05_ur5_pour_pytorch_baseline` from `src/openpi/training/config.py`.
- This baseline variant has geometric augmentation off and keeps pose/ray/view/cross-view/aux-point modules disabled.
- Release checkpoints intentionally exclude `optimizer.pt`.
- Normalization stats are also present inside each checkpoint asset tree; a top-level hard-linked copy is included for convenience.