Safe Operation of Connected Vehicles in Unforeseen Environments
Background:
Over the last decade, autonomous vehicles have seen rapid development in their capabilities and efficiency and are expected to handle complex missions with little to no human intervention. Thus, both the public and private sectors have been increasingly deploying autonomous vehicles for various purposes, like autonomous shuttle, truck, and vehicle sharing. Accordingly, the global autonomous vehicle market has been exponentially growing to $54.23 billion in 2019 and is expected to be $556.67 billion by 2026. The last few years witnessed numerous instances when the safe control of autonomous vehicles was challenged by different types of uncertainties. While in some of these instances the cause of failures could be attributed to slippery roads, prediction errors, perception errors, and software failures, it has become obvious that from the perspective of autonomy research one needs to develop a framework that provides rigorous guarantees in the form of certificates for performance, control, planning, perception, and software.
Motivation:
Recent years have seen tremendous efforts in terms of integrating machine learning with robust control methods towards establishing unified frameworks for learning-based planning and control for enhanced autonomy. Given known bounds for uncertainties, such frameworks have shown superior performance in challenging tasks like autonomous racing, nonlinear quadrotor control, and underwater vehicle formation. However, as one considers the highly occupied highways with unforeseen conditions imposed due to the unpredictable weather or other environmental factors, including e.g. traffic congestion, among many, the assumption on known bounds for uncertainties may not be verified. Using data shared over the cloud network(s) and proactively adapting to the new environments can be crucial to improving the safety of all vehicles on the highways. For example, connected vehicles could share local information over a cloud network to effectively avoid collision between them, reduce traffic jams, and be fuel-efficient and robust in operations.
State-of-the-art:
This project seeks to advance the state-of-the-art by designing a proactive/reactive adaptation and learning architecture for connected vehicles, unifying techniques in spatio-temporal data fusion, machine learning (ML), and robust adaptive control (RAC).
Objectives:
This project aims to develop a multi-level adaptive control architecture, where the proactive level leverages data over a cloud network to cope with unforeseen environmental uncertainties and support high-level decision making, while the reactive level uses ML and RAC to compensate for uncertainties. Each component of the framework is designed to produce certificates for performance and control, which can be consolidated into a global certificate verifying the system's safety and performance.
Publications:
- Kim, H., Wan, W., Hovakimyan, N., Sha, L., and Voulgaris, P.G.. (2021) Robust Vehicle Lane Keeping Control with Networked Proactive Adaptation. In 2021 American Control Conference (ACC), pp. 136-141.