Event Actions
SD4AS: Safe Driving for Autonomous Systems
Abstract:
Autonomous Driving Systems (ADSs) are becoming increasingly widespread, with companies deploying it for various tasks such as taxi services, delivery, personal transportation, and even agricultural applications like autonomous tractors for farming. As these systems become integral to daily life, ensuring their safe operation is crucial. There are numerous difficulties for achieving ADS safety; this work focuses on two key problems that are critical for reasoning about ADS conformance with safe driving properties. The first problem is specifying and checking the satisfaction of safe driving properties over raw sensor inputs like images. This requires addressing four critical challenges: finding a suitable abstraction to represent raw sensor inputs, enabling the definition of safe driving properties over this abstraction, ensuring that these properties can be checked both during ADS validation and in real-time deployment, and addressing the limitations of current abstraction generators, which often lack the accuracy needed to represent real-world driving scenarios effectively.
The second problem lies in effectively incorporating safe driving properties early into the system development process to ensure conformance to safety rules from the ground up. This requires addressing three critical challenges: ensuring training datasets adequately cover the conditions needed for ADS to learn safe driving behaviors, penalizing the ADS during training for any unsafe actions, and managing the complexities of ADS architectures to enforce these safety principles across the entire system. To overcome these problems, I propose SD4AS, a comprehensive framework that leverages scene graphs as a structured representation of raw sensor inputs to enable the evaluation of safe driving properties, and their integration early into the ADS development lifecycle. Preliminary results demonstrate the potential of SD4AS to successfully monitor a wide range of safe driving properties and reduce the rate of ADS property violations. This work aims to advance the development of safer ADSs that can reliably operate in complex, real-world environments.
Committee:
- Matthew Dwyer, Committee Chair (CS/SEAS/UVA)
- Sebastian Elbaum, Advisor (CS/SEAS/UVA)
- Ferdinando Fioretto (CS/SEAS/UVA)
- Thomas Fletcher (ECE, CS/SEAS/UVA)
- Corina Pasareanu (CyLab, Carnegie Mellon University)