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

Predictive Control under Uncertainty for Safe Autonomous Driving: Integrating Data-Driven Forecasts with Control Design

  • Author(s): Carvalho, Ashwin Mark
  • Advisor(s): Borrelli, Francesco
  • et al.
Abstract

Self-driving vehicles have attracted a lot of interest due to their potential to significantly reduce traffic fatalities and transform people's lives. The reducing costs of advanced sensing technologies and the increasing capabilities of embedded computing hardware have enabled the commercialization of highly automated driving features. However, the reliable operation of autonomous vehicles is still a challenge and a major barrier in the large scale acceptance and deployment of the technology.

This dissertation focuses on the challenges of designing safe control strategies for self-driving vehicles due to the presence of uncertainty arising from the non-deterministic forecasts of the driving scene. The overall goal is to unify elements from the fields of vehicle dynamics modeling, machine learning, real-time optimization and control design under uncertainty to enable the safe operation of self-driving vehicles. We propose a systematic framework based on Model Predictive Control (MPC) for the controller design, the effectiveness of which is demonstrated via applications such as lateral stability control, autonomous cruise control and autonomous overtaking on highways. Data collected from our experimental vehicles is used to build predictive models of the vehicle and the environment, and characterize the uncertainty therein. Several approaches for the control design are presented based on a worst-case or probabilistic view of the uncertain forecasts, depending on the application. The proposed control methodologies are validated by experiments performed on prototype passenger vehicles and are executed in real-time on embedded hardware with limited computational power. The experiments show the ability of the proposed framework to handle a variety of driving scenarios including aggressive maneuvers on low-friction surfaces such as snow and navigation in the presence of multiple vehicles.

Main Content
Current View