Navigation Agent for a 2D Maze Environment Using Classical and Neural Network Techniques

This project involves creating a navigation agent that learns to navigate a 2D maze using a combination of classical supervised learning and advanced neural network methods. Initially, the agent's behavior was trained by imitating expert demonstrations using traditional supervised learning techniques. The project was then extended to incorporate neural networks, increasing the complexity of the environment and enhancing the agent's navigation capabilities.

Key Features:
- Data Collection and Preprocessing: Collected RGB image observations representing the agent's perspective in the maze, and corresponding ground truth locations. Cleaned and prepared data for training, addressing missing values and standardizing formats.
- Model Development with Classical Techniques: Implemented a Position Regressor class for predicting the agent's location based on RGB images. Developed two more classes for determining the agent's next action using both low-dimensional data and visual observations. Used classification algorithms from scikit-learn to handle the discrete action space (up, left, or right).
- Model Development with Neural Networks: Trained Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs) to predict the optimal actions based on the agent's position and RGB images of the maze environment.
- Randomized Environment: Enhanced the complexity by randomizing agent and goal positions and introducing new obstacle maps.
- Behavioral Cloning: Trained the models to imitate expert actions using supervised learning techniques. The models learned to map observations to actions, ensuring accurate and reliable navigation.
- Model Evaluation: Assessed model performance using accuracy metrics and consistency in reaching the goal, and robustness in complex environments. Evaluated the agent's ability to minimize the distance to the goal over multiple runs, ensuring robust and reliable navigation behavior.