sourlreapwr/RobotArena-Benchmark

RobotArena Dataset

The RobotArena dataset contains reconstructed robotic manipulation scenes used for evaluating Vision-Language-Action (VLA) policies in simulation.

It is designed to support the RobotArena benchmark, which evaluates robot policies by converting real-world robot videos into physics-consistent simulation environments.

---

Contents

The dataset includes:

• Scene reconstructions used for simulation environments
• RGB image sequences from robot demonstrations
• Depth maps and geometry information
• 3D assets and meshes used for scene generation
• Background and environment assets

---

---

Dataset Size

• Total size: ~4.3 GB
• Number of files: ~15k

---

Intended Use

This dataset is intended for:

• Evaluating robot manipulation policies
• Benchmarking Vision-Language-Action models
• Research on simulation-to-real transfer
• Studying embodied AI systems

---

License

This dataset is released under the CC BY 4.0 License.
See the LICENSE.md file for full license details.

---

Citation

If you use this dataset in academic work, please cite:

@misc{jangir2025robotarenainftyscalablerobot,
  title={RobotArena $\infty$: Scalable Robot Benchmarking via Real-to-Sim Translation}, 
  author={Yash Jangir and Yidi Zhang and Kashu Yamazaki and Chenyu Zhang and Kuan-Hsun Tu and Tsung-Wei Ke and Lei Ke and Yonatan Bisk and Katerina Fragkiadaki},
  year={2025},
  eprint={2510.23571},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2510.23571}, 
}