---
title: AR-WAM-RoboTwin-Dataset
canonical_url: "https://www.modelscope.cn/datasets/Jasonggg15/AR-WAM-RoboTwin-Dataset"
md_url: "https://www.modelscope.cn/datasets/Jasonggg15/AR-WAM-RoboTwin-Dataset.md"
repository: Jasonggg15/AR-WAM-RoboTwin-Dataset
last_updated: 2026-10-08
license: mit
storage_size: "455 GB"
downloads: 409
stars: 0
---

# AR-WAM-RoboTwin-Dataset

> AR-WAM-RoboTwin-Dataset - Jasonggg15 在 ModelScope 开源的数据集。RoboTwin AR-WAM Dataset

Jasonggg15/AR-WAM-RoboTwin-Dataset 是 ModelScope 魔搭社区上的数据集，存储大小 455 GB，采用 mit 许可。

- **Repository**: Jasonggg15/AR-WAM-RoboTwin-Dataset
- **License**: mit
- **Storage size**: 455 GB
- **Downloads**: 409
- **Stars**: 0
- **Last updated**: 2026-10-08

Source: https://www.modelscope.cn/datasets/Jasonggg15/AR-WAM-RoboTwin-Dataset

---

# RoboTwin AR-WAM Dataset

27,500 RoboTwin episodes: 2,500 clean and 25,000 randomized, covering 50 tasks. The dataset contains 7,269,561 frames.

```text
robotwin/
├── clean/
└── random/
schema.json
norm_stats.json
```

Episodes are stored as `<split>/<task>/aloha_agilex/data/episode_*.hdf5` under `robotwin/`.

- `schema.json` defines the fields, dtypes and vocabularies, including 23 contiguous skill IDs and 5 semantic style IDs.
- `norm_stats.json` provides fixed, full-dataset min/max bounds for absolute action14 and measured EE16, combined across clean and random episodes.

EE16 contains left/right position, `(w,x,y,z)` quaternion and measured gripper state. Bounding boxes use normalized `(cx,cy,height,width)`. JPEG images retain the OpenCV decoding convention without a channel swap; semantic masks use the additive role encoding described in `schema.json`.

Additional research data includes wrist/third-view images, camera calibration, joint-state and action components, arm events, per-object bounding boxes, instructions and trigger metadata.

## Acknowledgments

This dataset was newly generated using [RoboTwin 2.0](https://github.com/robotwin-Platform/RoboTwin), with our task-execution adaptations and additional annotations. We sincerely thank the RoboTwin authors for open-sourcing the framework, task environments and supporting resources.