Skip to main content

​
System Requirements

PARC has been tested on Ubuntu 22.04 with CUDA-enabled GPUs.
Recommended Hardware:
  • NVIDIA GPU with CUDA support (RTX 4090 for real-time kinematic control, RTX 3090+ for training)
  • 16GB+ RAM
  • Ubuntu 22.04 or compatible Linux distribution

​
Installation Steps

1

Clone the Repository

Clone the PARC repository from GitHub:
git clone https://github.com/mshoe/PARC.git
cd PARC
2

Create Conda Environment

Create a new conda environment with Python 3.8.20:
conda create -n parc python=3.8.20
conda activate parc
Python 3.8.20 is the tested version. Other versions may work but are not officially supported.
3

Install Dependencies

Install all required packages from requirements.txt:
pip install -r requirements.txt
This will install:
  • PyTorch 2.2.0 - Deep learning framework
  • Polyscope - 3D visualization for Motionscope
  • Trimesh & embreex - Mesh processing
  • WandB - Experiment tracking (v0.17.4+)
  • NumPy, SciPy, Matplotlib - Scientific computing
  • Gym 0.26.2 - RL environment interface
  • ImageIO - Video/image I/O with FFmpeg support
4

Fix CUDA Detection (if needed)

If PyTorch cannot detect CUDA after installation, reinstall with explicit CUDA version:
pip install torch==2.2.0+cu118 -f https://download.pytorch.org/whl/torch_stable.html
Verify CUDA is available:
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}')"
5

Configure Data Directory

Set up the user configuration file to specify your data directory. Create or edit user_config.yaml at the repository root:
user_config.yaml
DATA_DIR: "/absolute/path/to/your/data"
The DATA_DIR placeholder is referenced throughout configuration files for data, checkpoints, and outputs. Make sure this is an existing absolute path.
6

Verify Installation

Test your installation by running a simple script:
python scripts/run_motionscope.py
You should see the Motionscope viewer launch. If you get configuration errors about missing motions, proceed to the Quick Start guide.

​
Optional: Isaac Gym Setup

Isaac Gym is only required if you want to train motion tracking controllers. The motion diffusion model and Motionscope viewer work without Isaac Gym.
Isaac Gym is NVIDIA’s deprecated physics simulator used for PARC’s motion tracking module. While deprecated, it’s still functional for this use case.

​
Why Isaac Gym?

PARC’s tracking environment was built on an early version of MimicKit using Isaac Gym. For new projects, consider using the recommended open-source alternative: MimicKit.

​
Installation Process

1

Download Isaac Gym

Register and download Isaac Gym from NVIDIA:https://developer.nvidia.com/isaac-gym
2

Create Isaac Gym Environment

Use the Isaac Gym installation script within a conda environment. Create a helper YAML file:
isaac_gym_env.yaml
name: parc
channels:
  - pytorch
  - conda-forge
  - defaults
dependencies:
  - python=3.8.20
  - pytorch=2.2.0
  - torchvision=0.9.1
  - cudatoolkit=11.1
  - pyyaml>=5.3.1
  - scipy>=1.5.0
  - tensorboard>=2.2.1
Then create the environment:
conda env create -f isaac_gym_env.yaml
3

Install Isaac Gym

Follow the Isaac Gym installation instructions from their documentation. Typically:
cd isaacgym/python
pip install -e .
4

Install PARC in Isaac Gym Environment

After Isaac Gym is installed, install PARC’s remaining dependencies:
pip install -r requirements.txt

​
Download Pre-trained Models & Datasets

PARC provides pre-trained models and motion datasets on HuggingFace:

PARC Dataset

Download datasets from initial iteration and each PARC training stage

​
Available Datasets

  • Dec 2024 experiment: 4 PARC iterations
  • April 2025 experiment: 5 PARC iterations
  • Small model: ~30 MB checkpoint files
The datasets include:
  • Motion sequences with terrain data
  • Trained motion diffusion models
  • Tracking controller checkpoints

​
Loading Motion Data

Motion files are loaded using:
  • PARC/anim/motion_lib.py - Motion library utilities
  • PARC/anim/kin_char_model.py - Character model definitions
To view motions, edit the motion_filepath parameter in PARC/motionscope/motionscope_config.yaml.
Standalone Data Reading: If you only want to read motion data without installing PARC, use scripts/read_motion_data.py - it only requires NumPy and optionally PyTorch.

​
Troubleshooting

​
PyTorch CUDA Not Detected

If torch.cuda.is_available() returns False:
  1. Verify NVIDIA driver installation: nvidia-smi
  2. Check CUDA version compatibility with PyTorch 2.2.0
  3. Reinstall PyTorch with explicit CUDA version (see Step 4 above)

​
Import Errors

If you encounter import errors when running scripts: 
# Install PARC as editable package
pip install -e .

​
Polyscope Won’t Launch

Polyscope requires OpenGL. On headless servers:
  • Use X11 forwarding or VNC
  • Consider running Motionscope locally after downloading motion files

​
Old Dataset Format

If using datasets from the old SharePoint release (password: “PARC”), note that the file format is only compatible with PARC v0.1. Use the new HuggingFace datasets for the current version.

​
Next Steps

With PARC installed, you’re ready to:

Quick Start

Run Motionscope to visualize motions and start experimenting

PARC Guide

Learn about the 4-stage PARC training loop and key components

Build docs developers (and LLMs) love

Get started for freeTalk to us