RL Platformer

A Gymnasium + Pygame 2D platformer environment, trained end-to-end with Deep Q-Network (DQN) using only pixel observations.

The trained agent can reliably complete the entire level without failing.

Features

• Play the level yourself (game.py)
• Watch the trained agent play (test.py)
• Race head-to-head against the AI (compete.py - human in red, agent in green)

Reward Function

Condition	Reward	Explanation
Every step	+0.1 × Δdistance - 0.1	Δdistance = reduction in Euclidean distance to goal since last step. Strong forward progress signal + small living penalty
No forward movement (stuck)	extra -0.01	Penalizes being blocked / zero x-velocity
Jumping while already in the air	-0.1	Discourages jump-spamming
Falling off the bottom of the world	-20	Death penalty → episode terminates
Reaching the goal (green block)	+100	Success → episode terminates

Files

File	Description
game_env.py	Core Gymnasium environment with physics, camera, collision and the complete level
compete.py	Human vs Agent race mode (first to the goal wins)
game.py	Simple human-playable version
test_agent.py	Watch the trained agent play
final_model.pth	Trained DQN weights (included in the repo)
train.py	Training script for the RL Agent

Observation & Action Space

• Observation: 84×84 grayscale frames, 4-frame stacked
• Actions (Discrete(2)):
  • 0 → no action
  • 1 → jump (up arrow key)

Installation

# Clone or download the repo
git clone https://github.com/Its-a-Me-2345/RL_Platformer.git
cd RL_Platformer

# Recommended: create a virtual environment
python -m venv venv
source venv/bin/activate    # Linux/Mac
# or
venv\Scripts\activate       # Windows

# Install dependencies
pip install -r requirements.txt