Motion planning (also known as path planning) is a fundamental problem in the field of robotics and autonomous systems, where the objective is to find a collision-free path for an agent from a starting position to a goal state. Despite the importance of motion planning, comparing the performance of various algorithms under the same environment has been rarely explored. Furthermore, the lack of sufficient evaluation metrics in reinforcement learning (RL) studies can hinder the understanding of each algorithm's performance. This thesis investigates the problem of finding the optimal path in 3D environments using both sampling-based and RL algorithms. The study evaluates the performance of six algorithms, including Rapidly-exploring Random Trees (RRT), RRT*, Q-learning, Deep Q-Network (DQN), Trust Region Policy Optimization (TRPO), and Proximal Policy Optimization (PPO), while considering the impact of different features in complex 3D spaces. Simulation results indicate that RRT* outperforms other algorithms in completing a specific path planning task in a 3D grid map. The significance of this study lies in providing a comprehensive comparison of different path planning algorithms under the same environment and evaluating them using various metrics. This evaluation can serve as a useful guide for selecting an appropriate algorithm to solve specific motion planning problems.
