license: apache-2.0
task_categories:
- robotics
tags:
- benchmark
- manipulation
- vla
DuoBench Raw Dataset
This is the raw dataset collected for the FR3 Duo Benchmark DuoBench. The format is a raw parquet format that includes additional information such as robot state (real) and sim state (sim) which are filtered out for training. If you are looking for the converted hugging face dataset ready for training, see this dataset.
You can also manually convert this raw dataset to the hugging face format with the following command:
# install rcs: https://github.com/RobotControlStack/robot-control-stack
python -m rcs lerobot-convert <output_path> --dataset-path <path> --repo-id duobench/<task_id> --no-joints --camera head@224x224 --camera left_wrist@224x224 --camera right_wrist@224x224 --video-encoding --video-backend torchcodec --binarize-gripper --no-binarize-gripper-state --clip-joint-limits --n 50
Structure
Each task has a sim and/or real folder with one or more recording sessions (<date>_part-<n>.parquet).
The instruction column holds the task instruction as defined in the DuoBench code.
| Task | sim episodes | real episodes |
|---|---|---|
| ball_maze | 58 | 59 |
| bin_sort | 123 (2 sessions) | 198 (3 sessions) |
| block_balance | 93 | - |
| carry_pot | 61 | - |
| hinge_chest | 63 | 64 |
| join_blocks | 77 | - |
| pour_marbles | 116 | - |
| spring_door | 54 | - |
| transfer_cube | 60 | 53 |
| transfer_gate | 54 | - |
| transfer_reorient | 82 | - |
Episodes include failed demonstrations (see the success column); the conversion command above uses the first --n episodes.
Project page: https://duobench.github.io/ DuoBench Code: https://github.com/RobotControlStack/duobench
Citation
If you find this dataset or DuoBench useful for your work, please consider citing it:
@misc{duobench,
title={{DuoBench}: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World},
author={Tobias J{\"u}lg and Seongjin Bien and Simon Hilber and Yannik Blei and Pierre Krack and Maximilian Li and Sven Parusel and Rudolf Lioutikov and Florian Walter and Wolfram Burgard},
year={2026},
url={https://arxiv.org/abs/2606.11901}
}