zuzuzzy/NS-VLA-Dataset

NS-VLA Dataset: Primitive-Annotated Robotic Manipulation Data

Dataset Description

This dataset provides primitive annotations for robotic manipulation demonstrations used to train and evaluate the NS-VLA framework. Each demonstration trajectory is segmented and labeled with structured manipulation primitives.

Primitive Vocabulary

Primitive	Description	Frequency
pick	Grasp a target object	44.4%
place_in	Place object inside a container	25.0%
place_on	Place object on a surface	16.7%
close	Close an appliance (e.g., microwave)	5.6%
place_rel	Place relative to another object	—
turn_on	Activate an appliance (e.g., stove)	—
open	Open an appliance door	—

Dataset Structure

NS-VLA-Dataset/
├── libero/
│   ├── spatial/          # LIBERO-Spatial task annotations
│   ├── object/           # LIBERO-Object task annotations
│   ├── goal/             # LIBERO-Goal task annotations
│   └── long/             # LIBERO-Long task annotations
├── calvin/
│   └── ABC_D/            # CALVIN ABC→D split annotations
└── metadata.json         # Dataset statistics and splits

Annotation Format

Each annotation file is a JSON with the following structure:

{
  "task_id": "libero_spatial_01",
  "instruction": "put the white mug on the left plate",
  "primitives": [
    {"op": "pick", "args": {"object": "white_mug"}, "start": 0, "end": 45},
    {"op": "place_on", "args": {"object": "white_mug", "support": "left_plate"}, "start": 46, "end": 102}
  ],
  "total_steps": 102
}

Usage

⚠️ Note: Dataset files will be released upon paper acceptance. Please check back soon.

from datasets import load_dataset

dataset = load_dataset("Zuzuzzy/NS-VLA-Dataset")

Associated Benchmarks

• LIBERO — Language-conditioned robotic manipulation
• LIBERO-Plus — Robustness evaluation with perturbations
• CALVIN — Long-horizon language-conditioned manipulation

Citation

@article{zhu2026nsvla,
  title={NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models},
  author={Zhu, Ziyue and Wu, Shangyang and Zhao, Shuai and Zhao, Zhiqiu and Li, Shengjie and Wang, Yi and Li, Fang and Luo, Haoran},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2026}
}

License

This dataset is released under the Apache 2.0 License.