ImDy: Human Inverse Dynamics from Imitated Observations (ICLR 2025)

A large-scale synthesized human dynamics dataset.

Dataset Details

Dataset Description

• Curated by: Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu

Dataset Sources

• Project: https://foruck.github.io/ImDy
• Repository: https://github.com/Foruck/ImDy
• Paper: https://arxiv.org/pdf/2410.17610

Uses

1. Clone the dataset and the GitHub repo.
2. Run python convert.py; python generate_cand.py to convert the raw data into a different format with per-sample pickle files including axis-angle format SMPL parameters, joints, and markers. The torques stored are acquired by summing two consecutive torques in the simulation.
3. Follow the instructions in the GitHub repo for further usage.

Dataset Structure

- utils
- data
 |- raw_test
   |- grf.pkl
   |- pos.pkl
   |- rot.pkl
   |- torque.pkl
   |- weight.pkl
 |- raw_train
   |- ...
- convert.py

Dataset Creation

Curation Rationale

The dataset was created given the scaling issue of real-world human dynamics data: Either intrusive measurement or unstable optimization based on high-precision but motion and external force capture. Both limited the diversity and scalability of human dynamics data.

Source Data

The source data were acquired from AMASS. We rolled out a modified PHC motion tracking policy and recorded the related information to create the dataset.

Citation

BibTeX:

@inproceedings{
liu2025imdy,
title={ImDy: Human Inverse Dynamics from Imitated Observations},
author={Xinpeng Liu and Junxuan Liang and Zili Lin and Haowen Hou and Yong-Lu Li and Cewu Lu},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=br8YB7KMug}
}

APA:

Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, & Cewu Lu (2025). ImDy: Human Inverse Dynamics from Imitated Observations. In The Thirteenth International Conference on Learning Representations.

Dataset Card Authors

Xinpeng Liu

Dataset Card Contact

[email protected]