Skip to main content
eScholarship
Open Access Publications from the University of California

Intelligent and High-Performance Behavior Design of Autonomous Systems via Learning, Optimization and Control

  • Author(s): Sun, Liting
  • Advisor(s): Tomizuka, Masayoshi
  • et al.
Abstract

Nowadays, great societal demands have rapidly boosted the development of autonomous systems that densely interact with humans in many application domains, from manufacturing to transportation and from workplaces to daily lives. The shift from isolated working environments to human-dominated space requires autonomous systems to be empowered to handle not only environmental uncertainties such as external vibrations but also interaction uncertainties arising from human behavior which is in nature probabilistic, causal but not strictly rational, internally hierarchical and socially compliant.

This dissertation is concerned with the design of intelligent and high-performance behavior of such autonomous systems, leveraging the strength from control, optimization, learning, and cognitive science. The work consists of two parts. In Part I, the problem of high-level hybrid human-machine behavior design is addressed. The goal is to achieve safe, efficient and human-like interaction with people. A framework based on the theory of mind, utility theories and imitation learning is proposed to efficiently represent and learn the complicated behavior of humans. Built upon that, machine behaviors at three different levels - the perceptual level, the reasoning level, and the action level - are designed via imitation learning, optimization, and online adaptation, allowing the system to interpret, reason and behave as human, particularly when a variety of uncertainties exist. Applications to autonomous driving are considered throughout Part I. Part II is concerned with the design of high-performance low-level individual machine behavior in the presence of model uncertainties and external disturbances. Advanced control laws based on adaptation, iterative learning and the internal structures of uncertainties/disturbances are developed to assure that the high-level interactive behaviors can be reliably executed. Applications on robot manipulators and high-precision motion systems are discussed in this part.

Main Content
Current View