metadata
license: mit
tags:
- SMPL-humanoid
- Mocap-tracking
- state-action-pairs
- imitation-learning
size_categories:
- 10K<n<100K
UniPhys Dataset: Offline Dataset for Physics-based Character Control
This dataset is part of the UniPhys, enabling large-scale training of diffusion policies for physics-based humanoid control using SMPL-like characters. The state-action pairs are generated by the PULSE motion tracking policy.
Dataset Overview
amass_state-action-pairs: state-action pairs for motion sequences from AMASS dataset (excluding infeasible motions)babel_state-action-text-pairs: Packaged AMASS motions with BABEL frame-level text annotations.
AMASS state-action pairs
Data Structure
For each sequence, the dataset contains:
| Field | Shape | Description |
|---|---|---|
body_pos |
[T, 24, 3] |
Joint positions in global space |
dof_state |
[T, 69, 2] |
Joint rotations (dim 0) and velocities (dim 1) 69 = 23 joints × 3 DoF each |
root_state |
[T, 13] |
Contains: - Position (0:3) - Quaternion (3:7) - Linear velocity (7:10) - Angular velocity (10:13) |
action |
[T, 69] |
Joint angle targets (23 joints × 3 DoF) |
pulse_z |
[T, 32] |
Latent action space from PULSE policy |
is_succ |
bool |
Tracking success flag (True/False) |
fps |
int |
Frame rate (30 FPS) |
Visualization
To replay the sequence:
python replay_amass_state_action_pairs.py --load_motion_path amass_state-action-pairs/$YOUR_FILE_PATH
BABEL state-action-text pairs
This is the training dataset used in UniPhys.
Visualization
To replay the packaged offline BABEL dataset along with frame-level text annotation:
python replay_babel_state_action_text_pairs.py --load_motion_path babel_state-action-text-pairs/babel_train.pkl
Citation
If using this dataset useful, please cite:
@inproceedings{wu2025uniphys,
title={UniPhys: Unified Planner and Controller with Diffusion for Flexible Physics-Based Character Control},
author={Wu, Yan and Karunratanakul, Korrawe and Luo, Zhengyi and Tang, Siyu},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year={2025}
}
@inproceedings{
luo2024universal,
title={Universal Humanoid Motion Representations for Physics-Based Control},
author={Zhengyi Luo and Jinkun Cao and Josh Merel and Alexander Winkler and Jing Huang and Kris M. Kitani and Weipeng Xu},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=OrOd8PxOO2}
}