Vehicle automation has the potential to improve driving safety and driver comfort. The Mediator system aims to aid the realization of this potential by mediating between the driver and the automation on who is fittest to drive. Making this trade-off in a timely and safe manner requires both driver and automation fitness to be detected and predicted for the near future. Within this document, a method to quantify automation fitness, including time to automation fitness and time to automation unfitness are defined. In contrast to the driver state, which has known degraded performance markers such as fatigue or distraction as well as methods to quantify them, there are no established methods for assessing automation fitness. This deliverable provides the foundations and concepts that will allow the automation state module to estimate the automation fitness. The ideas and concepts developed in this deliverable will be implemented and validated at a later stage of the MEDIATOR project.
Firstly, “degraded automation performance” is defined by situations in which the driver disengages or overrides the driving automation system (DAS) due to perceived ill-fitting actions, or a situation in which the DAS shuts itself or goes into some fallback due to within-system quality triggers, or a situation where the automation causes a crash with another road user or the infrastructure.
Degraded performance as it manifests in markers (visible system behaviour) is the consequence of internal or external conditions to the DAS, which we refer to as factors. These factors expose functional limitations of the DAS and thereby have an impact on its fitness to drive. The resulting degradation of information quality circulating throughout the interconnected components will introduce uncertainty, at some point affecting the automation fitness.
Understanding the factors leading to degraded automation performance and the resulting effects throughout the various components of the driving automation system (such as perception or decision making) is a key element of the process for estimating how long the automation may remain fit to drive. We have identified two categories of factors relevant for the automation state module as they can be measured and predicted and therefore used as indications for an upcoming degraded performance:
- Factors related to system input such as adverse weather, dense traffic or roadworks, which can be measured and be predicted using driving context information such as weather forecast from an online service,
- Factors related to internal states of the system, which can be measured using information from the driving automation system to compute performance self-assessment indicators.
Taking the (simplified) system architecture of typical automation systems into account, we estimate the current and predicted automation fitness using the correlation between the performance self-assessment indicators, the annotated vehicle behaviour, and the driving context.
To quantify the automation fitness, an automation fitness scale is introduced, which corresponds to the rate of automation system deactivations or overrides (following the definition of degraded performance stated above). The higher on the scale, the less frequent are the system deactivations and so the more fit the automation. Using collected and annotated data for the driving automation system to be assessed, the goal is to correlate both the automation indicators and the driving context with the number of occurrences of system deactivations/overrides/fallback initiations per time unit normalized on the automation fitness scale The final outcome of the method is an estimation of an automation fitness score using both online observations of the automation indicators as well as online observations and predictions of the driving context. The estimation of the automation fitness score is then used to predict the time to automation (un)fitness using cut-off thresholds; for instance the time to automation unfitness (TTAU) would be the shortest time when the estimated automation fitness score becomes lower than a cut-off threshold.
This research led to the main outcome of this deliverable, the functional requirements for the automation state module of the Mediator system. Further refinements to the work presented here will be done as part of the actual development of the automation fitness module.
MEDIATOR has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 814735.