A Prototypical Expert-Driven Approach Towards Capability-Based Monitoring of Automated Driving Systems

TL;DR

The paper tackles runtime capability monitoring for automated driving by introducing an expert-driven framework that links capability graphs to architectural views and parameterizes Bayesian Networks with fuzzy rules. It derives a DAG-based model of system element qualities and uses Mamdani fuzzy logic to generate CPTs, enabling online inference of capability quality from runtime measurements. A UNICARagil-based example demonstrates processing of both binary error signals and continuous performance indicators, showing how degraded capabilities influence decision making and maneuver admissibility. The work highlights the potential and limitations of expert-driven modeling and outlines directions for formalization, data-driven parameterization, and broader applicability. The resulting framework offers a template for building self-aware automated driving systems with runtime capability assessment and safer behavior selection.

Abstract

Supervising the safe operation of automated vehicles is a key requirement in order to unleash their full potential in future transportation systems. In particular, previous publications have argued that SAE Level 4 vehicles should be aware of their capabilities at runtime to make appropriate behavioral decisions. In this paper, we present a framework that enables the implementation of an online capability monitor. We derive a graphical system model that captures the relationships between the quality of system elements across different architectural views. In an expert-driven approach, we parameterize Bayesian Networks based on this structure using Fuzzy Logic. Using the online monitor, we infer the quality of the system's capabilities based on technical measurements acquired at runtime.

Figures (6)

• Figure 1: Architectural viewpoints, adapted from bagschik_systems_2018. Arrows indicate correspondences between viewpoints.
• Figure 2: Functional architecture extract, adapted from nolte_model_2017ulbrich_towards_2017.
• Figure 3: A network displaying the interdependencies between the quality of elements across different architectural views for the longitudinal stop and follow speed maneuver. The latter requires additional capabilities, i.e., nodes and edges indicated by dashed lines.
• Figure 4: Membership functions for the horizontal standard deviation of position as a performance measure for the state estimation function.
• Figure 5: Example scenario -- adapted from stolte_towards_2020. Not true to scale.