l2aggle/teleop_demo_data

TeleOp Demo Data for Humanoid Whole-Body Manipulation

Dataset Summary

This dataset contains 10 minutes of real robot teleoperation data used for post-training in humanoid whole-body manipulation tasks. It includes 4 different humanoid whole-body manipulation tasks and serves as a reference for data format and collection methodology.

Dataset Size: 282.5 MB

Related Resources

• Research Paper: TrajBooster Paper
• Project Page: https://jiachengliu3.github.io/TrajBooster/
• GitHub Repository: https://github.com/jiachengliu3/OpenTrajBooster

⚠️ Important Notice

This dataset is intended for reference purposes only!

If you want to reproduce the results from our project, we strongly recommend NOT using this dataset directly for post-training. Instead, you should:

1. Collect your own teleoperation data using your Unitree G1 robot
2. Use your own collected data for fine-tuning to achieve meaningful reproduction and improvement results

This recommendation is crucial because:

• Hardware variations: Each robot has subtle differences in motor performance, including variations in motor torque characteristics, response latency, gear backlash, and calibration offsets, which can significantly affect motion execution
• Camera specifications: Camera models, resolution settings, field of view, and color calibration may vary between robots
• Mounting and positioning: Camera installation brackets, mounting angles, and positioning relative to the robot body may differ
• Sensor calibration: IMU calibration, joint encoder offsets, and other sensor parameters are robot-specific
• Using our dataset for post-training may not achieve optimal performance on your specific robot setup due to these hardware-specific characteristics

Dataset Description

This dataset contains teleoperation demonstrations for 4 humanoid whole-body manipulation tasks. The data format and structure can serve as a valuable reference for researchers working on similar projects.

Tasks Included

The dataset includes demonstrations for 4 different humanoid whole-body manipulation tasks (specific task descriptions would be added based on your actual tasks).

Data Format

The dataset follows the format described in the TrajBooster paper and is compatible with the OpenTrajBooster codebase.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("l2aggle/teleop_demo_data")

# Use as reference for data format
print(dataset)

Citation

If you use this dataset as a reference or build upon the TrajBooster methodology, please cite:

@article{liu2025trajbooster,
  title={TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning},
  author={Liu, Jiacheng and Ding, Pengxiang and Zhou, Qihang and Wu, Yuxuan and Huang, Da and Peng, Zimian and Xiao, Wei and Zhang, Weinan and Yang, Lixin and Lu, Cewu and Wand, Donglin},
  journal={arXiv preprint arXiv:2509.11839},
  year={2025}
}

License

This dataset is released under the Apache 2.0 License. See the LICENSE file for more details.

Contact

For questions about this dataset or the TrajBooster project, please refer to the GitHub repository or the project page.