Title: A Scalable Solution to Lifelong Multi-Agent Path Planning
Date: April 22, 2025
Time: 11:00 AM
Location: GTMI Room 401
Kushal Jignesh Shah
Robotics MS Student
Wallace H. Coulter Department of Biomedical Engineering
Georgia Institute of Technology
Committee:
Dr. Seung-Kyum Choi (Advisor)
Woodruff School of Mechanical Engineering
Georgia Institute of Technology
Dr. Ye Zhao
Woodruff School of Mechanical Engineering
Georgia Institute of Technology
Dr. Roger Jiao
Woodruff School of Mechanical Engineering
Georgia Institute of Technology
Abstract:
Multi-Agent Pickup and Delivery (MAPD) is a fundamental problem in robotics with significant applications in warehouse automation, logistics, and autonomous fleet management. It involves coordinating multiple robots to perform pickup and delivery tasks efficiently in a shared environment, requiring effective task assignment, collision-free path planning, and overall system optimization. Despite progress in MAPD research, existing algorithms often struggle with scalability, adaptability, and real-time performance, especially in dynamic, high-demand settings where robots must frequently respond to unexpected changes. This thesis explores how to overcome these challenges by addressing three key research questions: How can we ensure real-time operation of MAPD systems in dynamic, unpredictable environments while maintaining high-quality task assignment and path planning? What approaches can be employed to scale MAPD algorithms for systems with a large number of agents, ensuring both efficient task completion and minimal computational overhead? And how can MAPD systems effectively handle dynamic obstacles, ensuring that agents can avoid collisions and complete their tasks in environments where obstacles may appear and move unexpectedly? To address these questions, we propose two novel algorithms: Adaptive Task Token Framework (ATTF) and Neural ATTF, that are designed to enable real-time, scalable, and adaptable performance in complex environments. We evaluate these algorithms through extensive simulations and compare their performance against several state-of-the-art MAPD methods, including TP, TPTS, CENTRAL, RMCA, LNS-PBS, and LNS-wPBS. Additionally, a real-world case study conducted in an automotive component manufacturing factory demonstrates the practical value of the proposed framework, highlighting improvements in operational efficiency and guiding decisions on optimal path layouts and fleet sizing for Autonomous Mobile Robots (AMRs).