RLBench 18 Tasks Dataset

Overview

This repository provides the RLBench dataset for 18 tasks, originally hosted by PerAct in Google Drive. Since downloading large files from Google Drive via terminal can be problematic due to various limits, we have mirrored the dataset on Hugging Face for easier access. To know more about the details of this dataset, please refer to PerAct.

Dataset Structure

The dataset is organized into three splits:

data/
├── train/  # 100 episodes per task
├── val/    # 25 episodes per task
└── test/   # 25 episodes per task

Each split contains zip files corresponding to individual tasks, making it easier to download specific tasks without fetching the entire dataset.

Dataset Details

These pre-generated RLBench demonstrations follow the same splits used in the original PerAct paper. Using these pre-generated demonstrations ensures reproducibility, as the original scene generation process involves randomness.

Train Split: 100 episodes per task
Validation Split: 25 episodes per task
Test Split: 25 episodes per task
Total Size: ~116GB

Acknowledgements

Thanks to the authors of PerAct and RLBench for creating and sharing the original dataset. This mirror aims to make data access more convenient for researchers.

Citation

If you use this dataset, please cite the original PerAct and RLBench papers:

@inproceedings{shridhar2022peract,
  title     = {Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation},
  author    = {Shridhar, Mohit and Manuelli, Lucas and Fox, Dieter},
  booktitle = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
  year      = {2022},
}

@article{james2020rlbench,
  title={Rlbench: The robot learning benchmark \& learning environment},
  author={James, Stephen and Ma, Zicong and Arrojo, David Rovick and Davison, Andrew J},
  journal={IEEE Robotics and Automation Letters},
  volume={5},
  number={2},
  pages={3019--3026},
  year={2020},
  publisher={IEEE}
}