Download

Please send an e-mail to limingzhe@stu.xmu.edu.cn, including contact details (title, full name, organization, and country) and the purpose for downloading the dataset. By sending the e-mail you accept the following License.

Note: Due to the open source license restrictions of the SMPL-X, we cannot directly provide .pkl and .npz files containing the body shape parameters directly. If you have questions about missing files, please contact the first author.

Data structure

The detailed directory structure and file formats will be updated here upon the official release of the dataset.

Dataset Explanation

  1. For the convenience of training and evaluation in the Motion Turing Test benchmark, we organize the data into Dataset_Human and Dataset_Humanoid. In order to be compatible with standard motion processing pipelines, we provide two types of files: .npz and .pkl. The .npz files contain the standard SMPL-X motion parameters (pose, translation) which are mainly used for visualization and kinematic analysis. The .pkl files contain the pre-processed spatial-temporal graph data, which are mainly used for training and evaluating graph-based models like our proposed PTR-Net.

  2. The original raw data in this dataset comes from various sources, including real-world humanoid robot SDK outputs (with different degrees of freedom), simulated environment logs, and in-the-wild human videos. If you process these raw formats directly, it will introduce severe domain gaps and appearance biases. But don’t worry, the data in the .npz and .pkl files mentioned above have solved this problem. We utilized GVHMR to unify all motions into the standardized SMPL-X representation, ensuring that the evaluation strictly focuses on motion kinematics rather than visual appearance.

  3. The human-likeness annotations for each motion sequence contain the rigorous 0-5 Likert scale ratings from 30 annotators, including the mean score and variance for each of the 1,000 motion clips (500 human, 500 robot).

  4. Finally, regarding visualization, we provide the standardized rendering script to recreate the exact visual conditions used in our Motion Turing Test. You can directly load the .npz sequences into the provided lightweight rendering environment to generate the appearance-free gray avatars as shown in the paper.