Deep Reinforcement Learning-Driven Automatic Controller Design for Swarm Robotics

Abstract

The growing reliance on robots in modern society, especially in sectors involving laborious or hazardous tasks, has accentuated the limitations of single-robot systems in dynamic and complex environments. Swarm robotics emerges as a promising alternative, leveraging the collective capability and decentralised nature akin to social insects to address these challenges, thus paving the way for robust, adaptable, and resilient systems capable of handling real-world complexities autonomously. Consequently, the urgent need to design adept and adaptive controllers for swarm robotic systems has spurred rapid advancements in the domain of automatic controller design. This thesis focuses on the pivotal role of automatic controller designs in advancing swarm robotic systems. Despite substantial advancements, the current approaches face a vital trade-off between performance efficacy and pragmatic design processes. This research emphasises the development and enhancement of Deep Reinforcement Learning (DRL)-driven automatic controller designs, which encapsulate the potential merits of both on-line and off-line processes, aiming to bridge the reality gap encountered in the transition from simulated to real-world environments. To address the centralisation issues in the traditional Multi-Agent Reinforcement Learning (MARL) frameworks, which contradict the decentralised essence of swarm robotic systems, this thesis proposes a Federated Learning (FL)-based DRL training strategy. The main objective is to foster a decentralised approach in DRL-driven automatic controller design, potentially motivating the evolution of more proficient and adaptive swarm robotic systems. Structured into three primary aims with corresponding objectives, the thesis endeavors to scrutinise the impacts of realistic factors on swarm robotic systems, advocate for DRL-driven automatic controller designs, and propose a novel FL-based DRL strategy to alleviate the centralisation issues inherent in conventional MARL approaches. Through a methodical investigation and critical analysis, this thesis aspires to encourage a new direction in swarm robotics research, steering closer to the realisation of systems proficiently navigating complex real-world scenarios.

Date of Award	31 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Barry Lennox (Supervisor) & Farshad Arvin (Supervisor)