README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - humanoid-locomotion
  - motion-imitation
  - physically-grounded

PHUMA: Physically-Grounded Humanoid Locomotion Dataset

Motion imitation is a promising approach for humanoid locomotion, enabling agents to acquire humanlike behaviors. Existing methods typically rely on high-quality motion capture datasets such as AMASS, but these are scarce and expensive, limiting scalability and diversity. Recent studies attempt to scale data collection by converting large-scale internet videos, exemplified by Humanoid-X. However, they often introduce physical artifacts such as floating, penetration, and foot skating, which hinder stable imitation.

In response, we introduce PHUMA, a Physically-grounded HUMAnoid locomotion dataset that leverages human video at scale, while addressing physical artifacts through careful data curation and physics-constrained retargeting. PHUMA enforces joint limits, ensures ground contact, and eliminates foot skating, producing motions that are both large-scale and physically reliable.

Paper: PHUMA: Physically-Grounded Humanoid Locomotion Dataset Project Page: https://davian-robotics.github.io/PHUMA Code: https://github.com/davian-robotics/PHUMA

Sample Usage

This section provides a quick guide to installing the necessary environment and running examples from the PHUMA data pipeline. For more detailed instructions, please refer to the GitHub repository.

Prerequisites

• Python 3.9
• CUDA 12.4 (recommended)
• Conda package manager

Installation

1. Clone the repository:

   git clone https://github.com/DAVIAN-Robotics/PHUMA.git
   cd PHUMA
   

2. Set up the environment:

   conda create -n phuma python=3.9 -y
   conda activate phuma
   

3. Install dependencies:

   pip install -r requirements.txt
   pip install -e .
   

Dataset Pipeline

1. Physics-Aware Motion Curation

Our physics-aware curation pipeline filters out problematic motions from human motion data to ensure physical plausibility.

Starting Point: We begin with the Humanoid-X collection as described in our paper. For more details, refer to the Humanoid-X repository.

Required SMPL-X Models: Before running the curation pipeline, you need to download the SMPL-X model files:

1. Visit SMPL-X official website
2. Register and download the following files:
   • SMPLX_FEMALE.npz and SMPLX_FEMALE.pkl
   • SMPLX_MALE.npz and SMPLX_MALE.pkl
   • SMPLX_NEUTRAL.npz and SMPLX_NEUTRAL.pkl
3. Place all downloaded files in the asset/human_model/smplx/ directory

Example Usage:

# Set your project directory
PROJECT_DIR="[REPLACE_WITH_YOUR_WORKING_DIRECTORY]/PHUMA"
cd $PROJECT_DIR

# We provide an example clip: data/human_pose/example/kick.npy
human_pose_file="example/kick"

python src/curation/preprocess_smplx.py \
    --project_dir $PROJECT_DIR \
    --human_pose_file $human_pose_file \
    --visualize 0

Output:

• Preprocessed motion chunks: example/kick_chunk_0000.npy and example/kick_chunk_0001.npy under data/human_pose_preprocessed/
• If you set --visualize 1, will also save example/kick_chunk_0000.mp4 and example/kick_chunk_0001.mp4 under data/video/human_pose_preprocessed/

2. Physics-Constrained Motion Retargeting

To address artifacts introduced during the retargeting process, we employ PhySINK, our physics-constrained retargeting method that adapts curated human motion to humanoid robots while enforcing physical plausibility.

Shape Adaptation (One-time Setup):

# Find the SMPL-X shape that best fits a given humanoid robot
# This process only needs to be done once and can be reused for all motion files
python src/retarget/shape_adaptation.py \
    --project_dir $PROJECT_DIR \
    --robot_name g1

Output: Shape parameters saved to asset/humanoid_model/g1/betas.npy

Motion Adaptation:

# Using the curated data from the previous step for Unitree G1 humanoid robot

human_pose_preprocessed_file="example/kick_chunk_0000"

python src/retarget/motion_adaptation.py \
    --project_dir $PROJECT_DIR \
    --robot_name g1 \
    --human_pose_file $human_pose_preprocessed_file \
    --visualize 0

Output:

• Retargeted humanoid motion data: data/humanoid_pose/g1/example/kick_chunk_0000.npy
• If you set --visualize 1, will also save data/video/humanoid_pose/example/kick_chunk_0000.mp4

Motion Tracking and Evaluation

To reproduce our reported quantitative results, use the provided data splits located in data/split/:

• phuma_train.txt
• phuma_test.txt
• unseen_video.txt

LAFAN1 Retargeted Data: Available here.

LocoMuJoCo Retargeted Data: Available here.

For motion tracking and path following tasks, we utilize the codebase from MaskedMimic.

Citation

If you use this dataset or code in your research, please cite our paper:

@article{lee2025phuma,
  title={PHUMA: Physically-Grounded Humanoid Locomotion Dataset},
  author={Kyungmin Lee and Sibeen Kim and Minho Park and Hyunseung Kim and Dongyoon Hwang and Hojoon Lee and Jaegul Choo},
  journal={arXiv preprint arXiv:2510.26236},
  year={2025},
}