From reinforcement learning to classical path planning: motion planning with obstacle avoidance

• Main Author: Ng, Tze Minh
• Other Authors: Yeo Chai Kiat
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Reinforcement learning; Motion planning
• Online Access: https://hdl.handle.net/10356/181149
• Institution: Nanyang Technological University

This project investigates the comparative performance of Reinforcement Learning (RL) and sampling-based motion planning methods in robotics, focusing on obstacle avoidance, illustrated in a 3D and 2D environment respectively with a singular agent and obstacle present. This is broken down into two phases. The approach involves first replicating the results of a chosen research paper on Soft Actor Critic with Prioritised Experience Replay (SACPER) and running it on a simulation software. Then, a comparative analysis of different sampling-based motion planning algorithms is generated. Through this process, insights into how differing scenarios and tasks call for different methods for optimal performance will be uncovered. Phase 1 involving the implementation of SACPER was unable to learn due to a stagnant reward curve, which necessitated the need for increased time and computing resources. Phase 2 investigated how sampling-based methods performed in a 2D environment based on slight changes in the environment. Overall, this project contributes to the understanding of motion planning for robotics, emphasizing the strengths and limitations of learning and sampling-based strategies. Future developments in considering a hybrid approach between learning and sampling-based strategies could be pioneered.