GridWorld RL Project

A reinforcement learning project that trains a DQN (Deep Q-Network) agent to navigate and solve tasks in a GridWorld environment.

Overview

The agent learns to navigate a grid world with walls and markers, performing actions like moving, turning, picking/placing markers, and finishing tasks. The environment loads different scenarios from JSON files containing grid layouts, wall positions, and start/goal configurations.

Key Components

• main.py - Training loop and DQN implementation
• model.py - Neural network architecture and replay memory
• envs/env_project.py - GridWorld environment implementation
• utils.py - Plotting utilities for training metrics
• datasets/ - Training and validation data (easy, medium difficulty levels)

Actions

The agent can perform 6 actions:

• Move forward
• Turn left/right
• Pick/put markers
• Finish task

Usage

Run training:

python main.py

The script will:

1. Load random task configurations from the dataset
2. Train the DQN agent using experience replay
3. Save the trained model as model.pth
4. Generate training plots (episode duration, loss, rewards)

Requirements

• PyTorch
• Gym
• NumPy
• Matplotlib
• Pygame (for rendering)

Dataset

Contains JSON files with grid configurations including:

• Grid size and walls
• Agent start/goal positions and orientations
• Marker locations (pre/post states)