An Exploratory Study of Driver Response to Reduced System Confidence Notifications in Automated Driving

An exploratory study in a moving-base simulator with 34 volunteers examined the effects of automated driving system reduced-confidence notifications on driver response. Reduced confidence in this context is defined as a system performance level that is lower than optimal, yet not so low as to disable the system. The threshold for such a level is essentially a policy decision based on perceived costs and benefits of keeping the automation engaged versus disengaging it. So, a reduced-confidence notification was designed to simply indicate that the driver should be especially vigilant to the increased potential to have to take manual control of the vehicle. An actual manual takeover request from the automated system would require the driver to take manual control of the vehicle. To investigate this, a 32-min highway driving scenario was constructed for this exploratory investigation that included 16 potential loss-of-lane tracking segments within this time frame. Three factors at two levels (low/high) each were crossed and presented in a between-subjects experimental design. One factor was system confidence/competence match, which refers to the ratio of the number of actual automation system failures over the number of reduced-confidence notifications. A second factor was the consequence of automation failure, i.e., drifting toward a busy traffic lane or drifting toward a wide shoulder. The third factor was detectability, or being able to see versus not see why the automated driving system began to drift out-of-lane. Data analyses indicated that automated driving with reduced-confidence notifications was judged to be both useful and acceptable regardless of the manipulated factors. However, the lowest incidence of preemption (i.e., manual takeovers prior to the start of a lane drift) was associated with low automation competence/confidence match and low detectability, regardless of consequences. It was also associated with responding later by some drivers. On the other hand, when the reduced-confidence notifications were perfectly matched with the need to manually takeover, there were fewer instances of delayed driver responses. Results are discussed in terms of "cry wolf" effects.