---
license: apache-2.0
library_name: openpi
tags:
  - robotics
  - manipulation
  - pi0.5
  - openpi
  - jax
  - orbax
pipeline_tag: robotics
---

# openpi-pi05-real01c-insert-marker-d1-sobol-r0-ft-iris

A [Pi0.5](https://www.physicalintelligence.company/blog/pi0-5) model fine-tuned using the [OpenPI](https://github.com/Physical-Intelligence/openpi) framework.

## Model Details

| Property | Value |
|---|---|
| OpenPI config | `pi05_sir_droid_finetune` |
| Checkpoint step | 4999 |
| Training data | N/A |
| Precision | bfloat16 |
| Parameter size | ~6.7 GB |
| Source checkpoint | `/iris/u/ankile/openpi-pi05-real01c/checkpoints/pi05_sir_droid_finetune/sir_real01c_sobol_r0_5k_bs32_iris_h200/4999` |
| Hugging Face repo | `ankile/openpi-pi05-real01c-insert-marker-d1-sobol-r0-ft-iris` |
| W&B run | [link](https://wandb.ai/self-improving/real-dagger-mining-01c/runs/dh6au05s) |
| SLURM job ID | `15689480` |

## Usage

### Download and run inference

```bash
# Download checkpoint from HF Hub
huggingface-cli download ankile/openpi-pi05-real01c-insert-marker-d1-sobol-r0-ft-iris --local-dir <local_path>

# Run inference server
cd deps/openpi
uv run python scripts/serve_policy.py pi05_sir_droid_finetune \
  --checkpoint-dir <local_path>
```

### In-process inference (Python)

```python
from openpi.training import config as openpi_config
from openpi.policies import policy_config as openpi_policy_config

train_config = openpi_config.get_config("pi05_sir_droid_finetune")
policy = openpi_policy_config.create_trained_policy(
    train_config, "<local_path>"
)
result = policy.infer(obs_dict)
actions = result["actions"]
```

## Checkpoint Format

[Orbax](https://github.com/google/orbax) format, all parameters in bfloat16.

```
├── _CHECKPOINT_METADATA
├── checkpoint_provenance.json
├── openpi_config.json
├── assets/
│   └── (normalization stats)
└── params/
    └── (orbax checkpoint files)
```

## License

Apache 2.0