Effect of Automation Instructions and Vehicle Control Algorithms on Eye Behavior in Highly Automated Vehicles

Abstract

Increasingly vehicle automation may convey greater capability than it actually possesses. The emergence of highly capable vehicle automation (e.g., SAE Level 4) and the promise of driverless vehicles in the near future can lead drivers to inappropriately cede responsibility for driving to the vehicle with less capable automation (e.g., SAE Level 2). This inappropriate reliance on automation can compromise safety, and so we investigated how algorithms and instructions might mitigate overreliance. Seventy-two drivers, balanced by gender, between the ages of 25 and 55, participated in this study using a fixed-base driving simulator. Drivers were exposed to one of three vehicle steering algorithms: lane centering, lane keeping, or an adaptive combination. A gaze tracker was used to track eye glance behavior. While automation was engaged, participants were told they could interact with an email sorting task on a tablet positioned near the center stack. Changes in roadway demand-traffic approaching in the adjacent lane-varied across the drive. Instructions indicating the driver was responsible, in combination with the adaptive algorithm, led drivers to be particularly attentive to the road as the traffic approached them. These results also have implications for evaluating more capable automation (SAE Levels 4 and 5), where drivers need not attend to the road: unnecessary attention to roadway demands might indicate lack of trust and acceptance of control algorithms that guide driverless vehicles.

Automation Expectation Mismatch: Incorrect Prediction Despite Eyes on Threat and Hands on Wheel

“…Encouragingly, the visual AR and hands-on-wheel reminders worked, that is, they produced high levels of visual attention and hands on wheel prior to the conflict and during all driving. It is also likely that the AR effect can be enhanced further by affecting vehicle control algorithms, which has been shown to improve eyes on road in highly automated vehicles ( Price, Lee, Dinparastdjadid, Toyoda, & Domeyer, 2017 ).…”

“…Going forward, two main solution alternatives present themselves: either the automation is designed so it does not rely on the driver (as in unsupervised automation), or it is designed so that the driver unmistakably understands that it is only a driving assistance system (a teammate) that needs an active driver who shares control at all times (at least when using highly reliable, near-perfection automation as tested here). Perhaps further research using the shared-control paradigm ( Mulder, Abbink, & Carlson, 2015 ) or adaptive automation ( Parasuraman et al, 2000 ), for example, in the form of adaptive vehicle control or warning/intervention algorithms that respond to driver engagement analysis (e.g., Price, et al, 2017 ), are viable ways forward. Note that if drivers do understand the limitations of the automation, or if drivers do not overtrust the automation, then conflicts can easily be handled, as indicated by the majority (72%) of drivers who easily avoided the conflicts.…”

“…It is also likely that the AR effect can be enhanced further by affecting vehicle control algorithms, which has been shown to improve eyes on road in highly automated vehicles (Price, Lee, Dinparastdjadid, Toyoda, & Domeyer, 2017).…”

Characterizing Driver Trust in Vehicle Control Algorithm Parameters

“…Automated vehicle feedback plays a critical role in aligning driver expectations with the capability of the automation. Our previous research demonstrated that vehicle behavior is an important component in influencing trust and attention in an automated vehicle (Price et al, 2016;Price et al, 2017). The present study used a relatively underused statistical approach in the human factors community with an attempt to expand the scope of the previous trust model.…”

“…The data used in the analysis is a combination of Study 1 outcomes and two follow-up simulator studies (Study 2 and Study 3), each designed to examine a different implementation of a steering algorithm. Study 2 and Study 3 were nearly identical and used the same scenario with some modifications to the objectives (Price et al, 2017). Study 2 explored whether changing steering behaviors (from lane centering to lane keeping) would influence expectations about whether the automation was capable of driving without sustained visual attention and monitoring of the roadway.…”