Half-Cheetah-RL-Project

A reinforcement learning implementation for training a Half-Cheetah agent in the MuJoCo simulation environment.

Overview

This project implements reinforcement learning algorithms to train a Half-Cheetah agent to run forward as efficiently as possible. The Half-Cheetah environment is part of the MuJoCo environments in Gymnasium (formerly OpenAI Gym) and features a 2D cheetah model with multiple joints that must learn to coordinate movement to maximize forward speed.

🐆 Trained Half-Cheetah in Action!

🎥 Watch Demo

Environment Description

The Half-Cheetah is a 2D robot consisting of:

• 9 links and 8 joints (including two paws)
• 6 degrees of freedom in the continuous action space (values between -1 and +1)
• 17-dimensional observation space (joint positions and velocities)
• Reward based on forward distance traveled (We have experimented with Custom Reward functions- Available in Codes/Reward Comparison

Installation

# Clone the repository
git clone https://github.com/yashphalle/Half-Cheetah-RL-Project.git
cd Half-Cheetah-RL-Project

# Set up a virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Implemented Algorithms

This repository implements and compares Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO).

How to Run

1. Training

• Go to Codes directory and find the appropriate notebook of DDPG or PPO
• It can take several hours to run depending upon your system configuration
• Ensure to use CUDA if you have GPU support

2. Run Trained Model

• Go to Trained model directory
• Run test.ipynb with our trained model

PPO Vs DDPG?

Reward Modifications Results: