Invited Talks

Mario Trapp

Towards Managing Perception Uncertainty for Automated Driving

Automated driving requires the fundamental ability to safely navigate through diverse surroundings. Also, when enforcing supervision e.g. through rule-sets, a dependable perception of the environment is inevitable, as it forms the basis for the safety-critical decisions and monitoring. However, with present perception capabilities remains a high degree of uncertainty. This stems from the open world scenarios with unforeseen variations as well as through perception sensors and algorithms, which do not inherently provide sufficient reliability. For achieving performant and safe driving with high utility, we propose the adaptive management of an autonomous vehicle’s safety. Our approach includes raising the self-awareness of a vehicle by considering uncertainty in the perception explicitly and deriving the actual given risks, instead of anticipated worst cases. By this, we can continuously adapt the automated driving system under derived operational conditions to efficient but safe states.

Julia Nitsch

Robust automotive perception in urban environments

Within urban environments, we need to recognize a variety of road users reliably to ensure a safe and anticipatory driving style for autonomous vehicles. Therefore, data-driven approaches are applied for road user recognition, due to their variety of appearances it is nearly impossible to model them individually. Machine Learning (ML) approaches, in particular, neural networks (NNs), are such data-driven approaches that achieve state-of-the recognition performance. Within this setting, NNs often apply the softmax normalization to generate confidences for classification results. However, it is known that softmax normalization achieves unreasonable high confides to out-of-distribution (OOD) objects. These objects have not been present during training but might appear during inference. Within this talk, we will highlight the possibilities ML approaches provide in urban automotive perception. Furthermore, we will also cover the disadvantages of these algorithms and possible solutions to these disadvantages.

Francesca Favaro

Waymo Readiness Evaluation for Fully Autonomous Ride-Hailing

Waymo launched the first commercial fully autonomous ride-hailing service open to the public in Phoenix in October 2020. At that time, Waymo also started an open discussion to share with its customers and the general community the safety methodologies that lay the ground for the determination of when the Waymo Driver(TM) can safely be deployed in its operational design domain. This talk will walk through an overview of such methodologies, while diving more specifically on a few examples from Waymo’s performance data. Furthermore, it will explore Waymo’s engagement in the continued discussions that shape the AV safety conversation.

Kevin Gay

Principles and Applications of Aurora’s Safaty Case Framework

Aurora has adopted a safety case-based approach because we believe that it is the most logical and efficient manner to show and explain how Aurora determines that our self-driving vehicles are acceptably safe to operate on public roads. This talk will provide context and insight about our process and intentions for applying our Safety Case Framework, including how we tailor specific safety cases based on the product and operations. In addition, this presentation will also cover the role that Safety Performance Indicators play in helping to establish the validity of our specific safety cases as well as throughout our development and testing processes.

Daniel Asljung

On Automated Vehicle Collision Risk Estimation using Threat Metrics in Subset Simulation

This paper presents a method for accelerated evaluation of an automated driving function using the subset simulation method. The focus of the paper is to investigate how the evaluation is affected by the choice of metric that is used to steer the subset simulation towards failure. It is found that all investigated metrics provide results relatively close to the reference, but the metrics relating to a state where collision is deemed to be unavoidable proved a little better.