Project page: https://michaelx.io/parc

Installation

Tested with Ubuntu 22.04

Optional: Install Isaac Gym (https://developer.nvidia.com/isaac-gym) for motion tracking

Install requirements:

conda create -n parc python=3.8.20
conda activate parc
pip install -r requirements.txt

and it should be good to go. If pytorch is not being able to detect CUDA, try reinstalling:

pip install torch==2.2.0+cu118 -f https://download.pytorch.org/whl/torch_stable.html

You can then run scripts such as:

python scripts/run_motionscope.py

Motionscope

Motionscope is my motion/terrain editor, as well as motion generator visualizer, built on top of Polyscope: https://polyscope.run/py/. You can run it by running:

python scripts/run_motionscope.py

Before that, you must edit the config file parc/motionscope/motionscope_config.yaml to load the motion you want, as well as optionally load an MDM model.

Dataset and Models

Download the datasets from the initial iteration and each stage of PARC.

New release: https://huggingface.co/datasets/mxucg/PARC

Two separate PARC experiments (Dec 2024 with 4 iterations, April 2025 with 5 iterations). New small model (~30 mb).

These files are loaded with anim/motion_lib.py and anim/kin_char_model.py. You can view them with scripts/run_motionscope.py, by editing the "motion_filepath" param in parc/motionscope/motionscope_config.yaml

If you only want the data without installing the whole repo, check out the script: scripts/read_motion_data.py You should only need numpy (and maybe pytorch?) to read the data.

(Old release: https://1sfu-my.sharepoint.com/:f:/g/personal/mxa23_sfu_ca/Et16uLMFxoRKouibvBa7LbwBEmX5_iI5a8dZyiMc0wmSTA?e=ihma1b The password is "PARC". The file format is only compatible with v0.1 PARC release.)

User configuration

All configuration files reference data, checkpoints, and generated outputs through a $DATA_DIR placeholder. Set this base directory in user_config.yaml at the repository root:

DATA_DIR: "/absolute/path/to/your/data"

DATA_DIR must be an existing absolute path. The training and pipeline scripts will automatically replace $DATA_DIR in YAML configs with the configured value when they load them.

Motion Tracking

PARC's motion tracking module was written using Isaac Gym, based on an early version of MimicKit. Isaac Gym is deprecated and there are now many great open-source motion tracking repositories. I personally recommend using https://github.com/xbpeng/MimicKit

The recipe for building a blocky terrain style motion tracking environment as shown in PARC is to start from a DeepMimic environment, but to load in a terrain that is composed of all the terrain-motion pairs in your dataset laid out as a grid. The environment should keep track of which motion-terrain each agent is currently learning, book-keeping the position offset for the agent and reference motions.

If you still wish to use Isaac Gym, then install it here: https://developer.nvidia.com/isaac-gym

I recommend using their installation scrip to install it within a conda environment with python 3.8.20 (Other versions may also work, but not tested).

The isaac gym helper yaml should look like this:

name: parc
channels:
  - pytorch
  - conda-forge
  - defaults
dependencies:
  - python=3.8.20
  - pytorch=2.20.0
  - torchvision=0.9.1
  - cudatoolkit=11.1
  - pyyaml>=5.3.1
  - scipy>=1.5.0
  - tensorboard>=2.2.1

Codebase Guide

• PARC Guide

Citation

If you find PARC helpful, please consider citing the references in the citation document.