Workshops

Traffic simulation plays an indispensable and pivotal role in the evaluation and enhancement of autonomous driving planning systems. Once deployed on public roads, autonomous vehicles need to interact with human participants who have diverse social preferences. In order to ensure that autonomous vehicles execute maneuvers that are both safe and efficient across various interactive traffic scenarios, it is crucial to test the robustness of autonomous vehicle algorithms by exposing them to various reactive agents in the simulated environment. The primary objective of this workshop is to foster discussion on how to develop diverse and realistic simulation scenarios and reactive agent behavior for efficient autonomous vehicle validation. In particular, we will focus our discussion on synthesizing simulation scenarios with diverse social characteristics. We plan to discuss methods to quantify social characteristics of driving behavior and various algorithms to synthesize diverse reactive agents with and without data support.

Machine learning components, especially neural networks, are critical to enabling the vision of automated and autonomous vehicles, as well as in broader autonomous cyber-physical systems (CPS). However, ensuring the reliability of these components to ensure overall system safety is extremely challenging, as illustrated by accidents in motor vehicles and theoretical limitations such as adversarial perturbations. This tutorial will present an overview of state-of-the-art methods for formal verification of neural networks and their usage within automated vehicles and autonomous CPS. The session will begin with an overview of the theoretical foundations and a survey of state-of-the-art methods. Following these foundations, the Neural Network Verification (NNV) software tool (https://github.com/verivital/nnv) will be described and demonstrated for establishing safety, robustness, and other specifications in neural networks and their closed-loop operation in autonomous systems. Interactive demonstrations will show attendees the capabilities of these tools focusing on NNV, as well as industrial applications of these foundational methods, with use cases described in automated ground and aerial vehicles.