Human motion prediction for indoor navigation of mobile robots

• Main Author: Chue, Jie Ying
• Other Authors: Soong Boon Hee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/139571
• Institution: Nanyang Technological University

With the help of autonomous robots across different industries such as medical and manufacturing, processes could be done remotely and minimizes manual labour. With the global pandemic COVID-19 in mind, the hospitality industry has seen a growth in the need for autonomous robots to aid in the work processes, healthcare workers and increase efficiency to better serve the patients. Nonetheless, safe navigation is still deemed as a difficult task despite the wide variety of algorithms available. Hence, safe and efficient navigation of a mobile robot through crowds would be the key focus in this report. This report presents two different methodologies which is a classic search-based approach and a deep reinforcement learning (DRL) approach. Firstly, the classic search-based control algorithm with the use of a laser sensor (LIDAR) would be described, tested and analyzed in hospital setting with Gazebo generated actors in Gazebo. However, as the complexity of the problem scenario increases, the mobile robot could not navigate safely and seamlessly with respect to human’s social norms. Therefore, a second approach with the use of DRL algorithms is investigated for the safe navigation in the indoor crowd simulation. Moreover, a Menge crowd framework is used instead of Gazebo actors for crowd generation as it allows the focus on the development of the project. Analysis of existing DRL algorithms such as Q-learning and SARSA provided a basis of the creation for a modified SARSA algorithm that allows collision avoidance and safe navigation of the mobile robot. Therefore, the use of the DRL algorithm provides more efficient, effective and safer navigation as compared to the classic search-based method. Further improvements such as implementation using an omni-directional robot and integration of force torque sensor have proposed for future developments.