Skip to main content

​
Available Datasets

PARC provides comprehensive motion and terrain datasets for physics-based character animation research.

PARC Dataset on HuggingFace

Official PARC motion and terrain dataset repository with multiple experimental iterations

​
Dataset Contents

The HuggingFace repository includes:
  • December 2024 Release: 4 iterations of PARC training data
  • April 2025 Release: 5 iterations of PARC training data
  • Compact Model: Small model (~30 MB) for efficient deployment

​
Motion Library Structure

The motion data files are structured to work with PARC’s motion library system:
  • Motion files are loaded via anim/motion_lib.py and anim/kin_char_model.py
  • Data includes:
    • Root positions and rotations
    • Joint rotations
    • Body contact information
    • Frame rate (FPS) data

​
Viewing Motion Data

You can visualize the motion files using Motionscope: 
python scripts/run_motionscope.py
Configure the motion file to view by editing the motion_filepath parameter in: 
parc/motionscope/motionscope_config.yaml

​
Downloading and Using the Data

​
Option 1: Full Installation

Follow the installation guide to set up the complete PARC environment.

​
Option 2: Data Only (Standalone)

If you only need to read the motion data without installing the entire repository, use the standalone script: 
python scripts/read_motion_data.py
Minimal Requirements: Only numpy (and possibly PyTorch) needed to read the data.

​
Reading Motion Data Programmatically

Here’s how to access motion data in your own scripts: 
import parc.util.file_io as file_io

MOTION_FILE = "data/motion_terrains/sfu.pkl"

ms_file_data = file_io.load_ms_file(MOTION_FILE)

# Access motion data components
root_positions = ms_file_data.motion_data.root_pos
root_rotations = ms_file_data.motion_data.root_rot
joint_rotations = ms_file_data.motion_data.joint_rot
body_contacts = ms_file_data.motion_data.body_contacts
fps = ms_file_data.motion_data.fps

​
Legacy Dataset

The old release (v0.1) is available on OneDrive with password “PARC”, but uses a file format only compatible with v0.1 of PARC. We recommend using the new HuggingFace dataset for all new projects.

​
Data Directory Configuration

All configuration files reference data through a $DATA_DIR placeholder. Set this in user_config.yaml at the repository root: 
DATA_DIR: "/absolute/path/to/your/data"
The training and pipeline scripts will automatically replace $DATA_DIR with your configured path.

Build docs developers (and LLMs) love

Get started for freeTalk to us