Effects of Automation Reliability on Driver Vigilance

Abstract

Partially automated vehicles rely on drivers to monitor for automation errors and environmental hazards to which the automation may not respond. However, driver vigilance declines over time, leading to potential safety risks associated with detection failure. This study measured driver trust, workload, stress, and ability to detect automation errors. Automation reliability was manipulated (high vs. low). Findings revealed no significant differences in performance, workload, or trust between high and low reliability conditions. The lack of reliability differences for performance, workload, and trust suggests more severe reliability variations may be needed to produce observable effects. However, the low reliability condition experienced heightened distress and worry emphasizing the effect of reliability on emotional state. All drivers became less responsive as a function of time (i.e., conservative bias shift). Drivers also reported decreased engagement and increased distress. This underscores the need for automation improvements that enhance engagement without increasing negative emotion states.

Partially automated vehicles (PAVs) are capable of both longitudinal (e.g., changes in speed) and lateral (e.g., changes in lane position) vehicle motion, but these vehicles are limited in their ability to detect and/or respond appropriately to various objects and events encountered on the roadway (SAE International, 2021). This leaves the driver responsible for monitoring automated vehicle behavior and the roadway for potential hazards. However, previous research has shown that a driver’s ability to perform this monitoring role declines over time, meaning that drivers of PAVs are less likely to detect and respond to a hazardous event as the drive continues (Greenlee et al., 2018). Prior research only evaluated drivers’ ability to monitor for roadway hazards, such as vehicles stopped in unsafe positions at intersections. Research has yet to examine drivers’ ability to monitor for hazardous automation behavior. Thus, one aim of the current study was to investigate drivers’ ability to monitor for hazardous automation behavior during a prolonged drive.

The current study was also designed to evaluate the effects of automation reliability on drivers’ hazard detection performance, workload, stress, and trust in vehicle automation. Automation reliability is a key predictor of automation monitoring behavior and trust in automation (Lee & See, 2004). Because high reliability automated systems make few errors, users tend to have a high degree of trust in the automation. High trust increases the likelihood of complacency and overreliance, which makes attentive monitoring and detection of automation failures less likely when automation reliability is relatively high (Muir & Moray, 1996). Because active monitoring is effortful and stressful (Warm et al., 2008), higher reliability automation may result in lower task induced workload and stress compared to less reliable automation.

The current study employed a driving simulator to examine drivers’ ability to detect hazardous automation behaviors, trust in automation, workload, and stress. Drivers were tasked with monitoring either a high reliability PAV (one that rarely executed unsafe behaviors) or a low reliability PAV (one that executed unsafe behaviors more often). We predicted that, compared to the low reliability PAV, the high reliability PAV would result in poorer detection performance, higher levels of trust, and less workload and stress.

Twenty-nine participants were included in each reliability condition. The design of the study was a 2 (reliability: high vs. low) × 4 (period of watch). The study included a high reliability condition (2.5% error rate, 97.5% reliability) and a low reliability condition (10% error rate, 90% reliability). The 40-minute vigilance task was split into four 10-minute periods to evaluate performance changes over time.

The task required participants to monitor a simulated PAV on a roadway which consisted of two right lanes traveling in the same direction. Occasionally, the automation would make unsafe maneuvers, such as deviating from the lane or changing lanes using the shoulder of the roadway. The frequency of these events was dependent on reliability condition and equaled the error rates mentioned previously. When the simulated vehicle executed one of these unsafe maneuvers, the driver was to respond by pressing a button on the steering wheel. Participants were told not to respond when the vehicle was continuing safely (i.e., lane keeping or changing lanes safely). Surveys assessed trust, workload, and stress.

Overall, detection performance was unaffected by reliability level. Drivers in both conditions committed fewer false alarms and became less likely to respond over time. Self-reported trust was unaffected by reliability level. Workload ratings suggested that monitoring for hazardous automation behavior was demanding, but those demands were unaffected by reliability level. Reliability condition did affect task-induced stress. Drivers in the high reliability condition reported less severe stress responses than those in the low reliability condition; distress and worry increased to a greater degree when reliability was low. For both conditions engagement declined from pre- to post-task similarly.

In sum, the current results demonstrate that monitoring automated vehicle behavior is a stressful and demanding assignment. This finding expands upon conclusions from previous research arguing that monitoring for roadway hazards is demanding and stressful for PAV drivers (Greenlee et al., 2018). Current results also indicate that monitoring relatively low reliability automated vehicle systems is especially stressful. More research will be needed to determine the cause of this effect. Designers of PAVs should be aware of the stress associated with automation monitoring.