GitHub - Perkins729/OmniXtreme (original) (raw)

OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control

Overview

This repository contains the official implementation of OmniXtreme, a unified policy framework for high-dynamic humanoid motion tracking.

We release:

• ✅ Paper and demonstration videos
• ✅ Checkpoints
• ✅ Sim-to-sim evaluation code

The provided pretrained policy under policy/ is trained using the OmniXtreme framework on high-dynamic motions.

Planned Releases

The following components are under consideration for future open-source release:

• Flow matching base policy training and inference code
• Residual post-training and inference code
• C++ real-world deployment code

Installation

1. Create conda environment and set environment variable

git clone https://github.com/Perkins729/OmniXtreme.git
cd OmniXtreme
conda create -n omnixtreme python=3.8
conda activate omnixtreme
conda install -c conda-forge cudnn=8
pip install -r requirements.txt
export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

2. Set motion path

Download the curated subset of our motion data from this link and move it to the policy directory policy/.

3.Run policy