Fangyu Wu

EECS Department, University of California, Berkeley

Technical Report No. UCB/

December 1, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/Hold/cb3db8aa10453b0cd58f130816fdd8e1.pdf

This dissertation addresses the complexities of planning and control in Multi-Agent Systems (MAS), focusing on optimizing interactions within three configurations: 1) sequentially chained MASs, 2) partially meshed MASs, and 3) centrally controlled MASs. As autonomous technologies increasingly permeate sectors ranging from automotive to industrial applications, the need for robust multi-agent systems capable of efficient coordination among themselves and with human operators becomes paramount. This research provides a comprehensive analysis and development of methodologies that enhance the reliability, efficiency, and safety of these systems through advanced planning and control strategies in the context of connected and automated driving.

The primary objective of this study is to refine the planning and control frameworks used in MAS by leveraging recent advancements in vehicle automation, optimization, and machine learning. By integrating these fields, the dissertation introduces novel solutions that serve real-world applications such as automated car-following, decentralized collision avoidance, and scalable multi-agent planning. The work addresses critical challenges in these systems, including stop-and-go waves, collision avoidance in understructured environments, scalable network-agnostic multi-agent planning, the efficient, stable integration of learning into control frameworks, and hardware and experimental design for MASs.

The main contributions of this research include the development of novel algorithms that reduce error propagation in sequentially chained MAS configurations, both with and without nonlocal measurements, decentralized conflict resolution methods for partially meshed MASs, and scalable scheduling techniques for centrally controlled MASs. These methodologies are validated through a series of simulations and physical experiments, demonstrating their effectiveness and practical applicability.

Moreover, this dissertation explores the integration of machine learning with traditional control frameworks and reflects on the hardware and experimental design utilized in this study. First, we present the memory-augmented model predictive control method, in which empirical data enhance the amortized running time performance of model predictive control models without compromising stability. Next, we discuss the hardware and experimental design employed in this dissertation to document the lessons learned from empirical hands-on applications.

In summary, this dissertation establishes a theoretical and experimental foundation for the next generation of MAS research in the context of connected and automated driving, aiming at the effective and safe integration of these systems into the fabric of daily and industrial applications.

Advisors: Alexandre Bayen

---

BibTeX citation:

@phdthesis{Wu:31422,
    Author= {Wu, Fangyu},
    Editor= {Bayen, Alexandre},
    Title= {On Planning and Control for Multi-Agent Robotic Systems},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Number= {UCB/},
    Abstract= {This dissertation addresses the complexities of planning and control in Multi-Agent Systems (MAS), focusing on optimizing interactions within three configurations: 1) sequentially chained MASs, 2) partially meshed MASs, and 3) centrally controlled MASs. As autonomous technologies increasingly permeate sectors ranging from automotive to industrial applications, the need for robust multi-agent systems capable of efficient coordination among themselves and with human operators becomes paramount. This research provides a comprehensive analysis and development of methodologies that enhance the reliability, efficiency, and safety of these systems through advanced planning and control strategies in the context of connected and automated driving.

The primary objective of this study is to refine the planning and control frameworks used in MAS by leveraging recent advancements in vehicle automation, optimization, and machine learning. By integrating these fields, the dissertation introduces novel solutions that serve real-world applications such as automated car-following, decentralized collision avoidance, and scalable multi-agent planning. The work addresses critical challenges in these systems, including stop-and-go waves, collision avoidance in understructured environments, scalable network-agnostic multi-agent planning, the efficient, stable integration of learning into control frameworks, and hardware and experimental design for MASs.

The main contributions of this research include the development of novel algorithms that reduce error propagation in sequentially chained MAS configurations, both with and without nonlocal measurements, decentralized conflict resolution methods for partially meshed MASs, and scalable scheduling techniques for centrally controlled MASs. These methodologies are validated through a series of simulations and physical experiments, demonstrating their effectiveness and practical applicability.

Moreover, this dissertation explores the integration of machine learning with traditional control frameworks and reflects on the hardware and experimental design utilized in this study. First, we present the memory-augmented model predictive control method, in which empirical data enhance the amortized running time performance of model predictive control models without compromising stability. Next, we discuss the hardware and experimental design employed in this dissertation to document the lessons learned from empirical hands-on applications.

In summary, this dissertation establishes a theoretical and experimental foundation for the next generation of MAS research in the context of connected and automated driving, aiming at the effective and safe integration of these systems into the fabric of daily and industrial applications.},
}

EndNote citation:

%0 Thesis
%A Wu, Fangyu 
%E Bayen, Alexandre 
%T On Planning and Control for Multi-Agent Robotic Systems
%I EECS Department, University of California, Berkeley
%D 2024
%8 December 1
%@ UCB/
%F Wu:31422