---
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - LIBERO
  - robotics
  - vision-language-action
  - MiniCPM-RobotManip
configs:
  - config_name: default
    data_files: "*/data/*/*.parquet"
---

# MiniCPM-RobotManip LIBERO

This dataset contains the four LIBERO suites converted to LeRobot v3 format
for the MiniCPM-RobotManip LIBERO full-parameter fine-tuning example in
[starVLA](https://github.com/starVLA/starVLA).

## Dataset summary

| Suite | Episodes | Frames | Videos |
|---|---:|---:|---:|
| LIBERO-10 | 358 | 95,740 | 716 |
| LIBERO-Goal | 405 | 48,131 | 810 |
| LIBERO-Object | 450 | 66,294 | 900 |
| LIBERO-Spatial | 423 | 51,707 | 846 |
| **Total** | **1,636** | **261,872** | **3,272** |

- Format: LeRobot v3
- Frequency: 20 Hz
- Cameras: agent view and wrist view
- Video resolution: 256 × 256
- Action target: `observation.xvla_abs_ee6d`
- Action target dimension: 10

## EE6D target

`observation.xvla_abs_ee6d` contains an absolute single-arm target:

- `[0:3]`: xyz position
- `[3:9]`: rotation6D
- `[9]`: gripper closed (`0` = open, `1` = closed)

The starVLA MiniCPM-RobotManip recipe uses the current row as state and offsets
1–30 as the action chunk. It maps this 10-D vector to the left-arm `[7:17]`
slot of MiniCPM-RobotManip's unified 80-D state/action space; other channels are
masked out of the training loss.

Each suite includes `meta/modality.json` with the mapping expected by starVLA.

### Legacy auxiliary column

The parquet files retain `actino.xvla_abs_ee6d`, a legacy auxiliary field from
the original conversion pipeline. It is intentionally preserved for provenance
and is not consumed by the published starVLA recipe. Use
`observation.xvla_abs_ee6d` for both state and offset action targets.

## Download

```bash
hf download openbmb/MiniCPM-RobotManip-LIBERO \
  --repo-type dataset \
  --local-dir playground/Datasets/LIBERO_EE6D
```

Expected top-level directories:

```text
libero_10_agentview_rot180_wrist_raw_lerobot_v30_xvla_abs6d/
libero_goal_agentview_rot180_wrist_raw_lerobot_v30_xvla_abs6d/
libero_object_agentview_rot180_wrist_raw_lerobot_v30_xvla_abs6d/
libero_spatial_agentview_rot180_wrist_raw_lerobot_v30_xvla_abs6d/
```

## Training with starVLA

```bash
VLM_PATH=openbmb/MiniCPM-RobotManip \
LIBERO_EE6D_ROOT=playground/Datasets/LIBERO_EE6D \
bash examples/modelExtensions/MiniCPM-RobotManip/train_files/run_libero_train.sh
```

The corresponding public base checkpoint is
[`openbmb/MiniCPM-RobotManip`](https://huggingface.co/openbmb/MiniCPM-RobotManip).

## Source and processing

The demonstrations originate from the
[LIBERO benchmark](https://lifelong-robot-learning.github.io/LIBERO/). The four
suites were filtered, converted to LeRobot v3 at 20 Hz, and augmented with
absolute EE6D targets aligned to each parquet row and video frame. Original
observation and action fields are retained.

This release contains simulation demonstrations only. It does not contain
personal or human-subject data.

## License

Released under the Apache License 2.0. See `LICENSE`.

## Citation

Please cite LIBERO and MiniCPM-Robot when using this dataset:

```bibtex
@inproceedings{liu2023libero,
  title={LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author={Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}
```

For the current MiniCPM-Robot citation, see the
[MiniCPM-Robot project](https://github.com/OpenBMB/MiniCPM-Robot).