PGTT: Phase-Guided Terrain Traversal for Perceptive Legged Locomotion

A perceptive reinforcement learning locomotion framework developed for the Unitree GO2 in simulation (MuJoco MJX) and deployed using unitree sdk2py, on real hardware.

Terrain Generation

Terrain are produced using Wave Function Collapse in MuJoCo.

Training Pipeline

Real World Experiment

The resulting policy deployed in the real world. We use Point-Lio for odometry and Gridmap to extract the desired heightmap.

pgtt_real.mp4

Installation

Install the required dependencies:

pip install mujoco==3.3.0
pip install mujoco_mjx==3.3.0
pip install brax==0.12.1
pip install jax==0.5.0
pip install playground

Training

To train the policy run the bash file as:

bash training/training.sh

Additionally, you could modify the hyperparameters from inside this file or the training difficulty.

Instructions

For more detailed instructions, please refer to the README file in each folder.