RL and Domain Randomization Techniques for Learning Push Recovery for Humanoids across different environments

Overview

The project focuses on developing control policies that allow humanoid robots to maintain balance when subjected to external forces (pushes) in varied environments. By using domain randomization during training, the policies become more robust and can generalize to unseen conditions.

Installation

Prerequisites

• Python 3.8+
• Git

Setup

1. Clone the repository

   git clone https://github.com/yourusername/humanoid-push-recovery.git
   cd humanoid-push-recovery

2. Create a virtual environment

   python -m venv pendulum_env

3. Activate the environment

   • Windows:

     pendulum_env\Scripts\activate

   • Linux/macOS:

     source pendulum_env/bin/activate

4. Install dependencies

   pip install gymnasium==0.29.1
   pip install mujoco==3.1.4
   pip install glfw imageio
   pip install stable-baselines3[extra]
   pip install PyOpenGL PyOpenGL_accelerate  # if rendering fails

5. Verify installation

   python test_mujoco_install.py

   This should open a MuJoCo window, confirming successful installation.

Usage

Training

To train a policy with default parameters:

python train_ip.py

The parameters can be changed inside the code.

Evaluation

Visualize the trained policy:

python visualize_ip_video.py