The overall objective of our project is to design, implement, and evaluate a learning-based navigation policy that can guide an autonomous agent through 2D obstacle environments.
We will build gridworld environments to train and compare several RL agents (Q-learning, DQN, and PPO).
We evaluate them based on reward functions and analyze how reward shaping and hyperparameter choices influence stability, sample efficiency, and learned behavior.
A summary of the project can be found here: final_report.pdf
All code for training and testing are in the test_envs folder. Each file trains, tests, and evaluates models for the respective environment.
- 3x3 trained on DQN (more details in report)
