---
license: other
license_name: gemma-terms-of-use
license_link: https://ai.google.dev/gemma/terms
pipeline_tag: robotics
tags: [robotics, vla, openpi, manipulation, maniguard, franka]
---

# pi05-real-cab-higher-firsthalf-60-droid-refined-lora

LoRA fine-tune of **pi0.5** (`pi05_droid` warm-start) on
`IDEAS-Lab-Northwestern/real-cab-higher-firsthalf-60-droid-refined` — 60 real
Franka teleop trajectories of the **first half of the cabinet task** (open the
cabinet, keep the door area clear), DROID schema.

- **Task prompt:** "Open the cabinet. Keep the area around the door clear."
- **Warm-start:** `gs://openpi-assets/checkpoints/pi05_droid/params` (clean, fresh)
- **LoRA:** `gemma_2b_lora` (rank 16) + `gemma_300m_lora` (rank 32), batch 4, EMA off.
- **Norm-stats:** reuses `pi05_droid`'s bundled DROID assets (no recompute).

## Checkpoint ladder

| Dir | Cumulative steps | Train loss |
|---|---|---|
| `10000/` | 10k | 0.0046 |
| `20000/` | 20k | 0.0032 |
| `30000/` | 30k | **0.0029** |
| `40000/` | 40k | 0.0038 |
| `50000/` | 50k (saved at step 49999) | 0.0031 |

Train loss bottoms at ~30k; eval-sweep **20k–40k**.

## Refined real setup

Refined real setup (raised-Z cabinet) — refined wrist & main-camera poses and
gripper. Wrist = raw `cam1`, exterior = raw `cam0`. **Collector:** yypeng666.

## Schema (DROID) / inference

`exterior_image_1_left` ← exterior cam0, `wrist_image_left` ← wrist cam1,
`exterior_image_2_left` zero-pad; `joint_position`(7)+`gripper_position`(1) state;
`actions`(8)=joint_velocity(7)+next gripper target(1). Serve with openpi
`pi05_droid_finetune_lora`. `train_state/` omitted (inference-only). Eval on a
real Franka, not the SFT server.

## Paper & Citation

Part of **ManiGuard**: [paper (arXiv:2608.17386)](https://arxiv.org/abs/2608.17386) ·
[code](https://github.com/NU-IDEAS-Lab/ManiGuard) ·
[docs](https://nu-ideas-lab.github.io/ManiGuard/)

```bibtex
@misc{peng2026maniguard,
  title         = {{MANIGUARD}: A Benchmark and Data Suite for Specification-Grounded
                   Safety Evaluation and Improvement of Robotic Manipulation},
  author        = {Peng, Yiyan and Wang, Philip and Zhan, Simon Sinong and Lyu, Yiqi
                   and Ni, Zhenyang and Yan, Jixin and Wong, Fiorelli and Jiao, Ruochen
                   and Yin, Hang and Cao, Xinyu and Shao, Huajie and Li, Manling
                   and Zhang, Ruohan and Zhu, Qi},
  year          = {2026},
  eprint        = {2608.17386},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.17386},
}
```

## License

The fine-tuned weights derive from a Physical Intelligence openpi base model whose
VLM backbone is PaliGemma; use of these weights is therefore subject to the
[Gemma Terms of Use](https://ai.google.dev/gemma/terms) (including the
[Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy)),
which downstream users must pass on. The openpi training code and ManiGuard's own
contributions are Apache-2.0.