It’s about time! Earlier take-over requests in automated driving enable safer responses to conflicts

Journal article, 2022

Automated driving (AD), which takes full responsibility for the driving task in certain conditions, is currently being developed. An important concern in AD is how to design a take-over request (TOR) that mitigates automation effects (e.g., delayed responses to conflict scenarios) that previous literature from simulator experiments has shown can occur. To address this concern, this study aims to investigate and compare driver responses to TORs and a lead-vehicle cut-out scenario under three conditions: (1) after a period of AD with a TOR issued early (18 s time-to-collision), (2) same as (1) except with a TOR issued late (9 s time-to-collision), and (3) baseline, with adaptive cruise control (ACC). This paper also compares the results to those of a previous study using the same conflict scenario but with near-perfect assisted driving system (SAE Level 2). The lead-vehicle cut-out scenario was encountered on a test track after 30 minutes driving with either ACC or AD. In AD the TOR was issued prior to the conflict object was revealed to the participants when the lead vehicle performed the cut-out (at conflict onset). This TOR strategy differed from previous driving-simulator studies that issued the TOR at conflict onset. The participants had to respond by steering and/or braking to avoid a crash. Our findings show that, independent of TOR timing, the drivers required similar amounts of time to 1) direct their first glance to the human–machine interface, 2) look forward, 3) end their secondary task, 4) put their hands on the steering wheel, and 5) deactivate automation. However, when the TOR was issued early rather than late, they started to brake earlier (even before conflict onset). All participants successfully managed to avoid crashing with the object, independent of the condition. AD with an early TOR resulted in the earliest response, while ACC drivers responded slightly earlier than the drivers in AD with the late TOR. Our findings do not support the findings of severe automation effects in previous driving-simulator studies. One reason for the difference is that when a TOR is issued prior to conflict onset, drivers are given the time needed for their preparatory actions (e.g., placing hands on the wheel, deactivating AD) that is not needed when driving with ACC or in manual driving (baseline), before having to respond to the conflict scenario. Thus, at conflict onset the drivers in AD are as ready to act (hands on wheel, eyes forward) as the drivers in the baseline and can perform an avoidance manoeuvre similar as to the baseline drive. Overall, the present study shows that AD does not need to end up in a highly critical situation if the TOR is issued early enough. In fact, AD with an early TOR may be safer than driving with ACC, because in the former drivers are more likely to brake earlier in preparation for the conflict. Finally, a TOR clearly communicates the need for drivers to resume manual control and handle potential events when AD has been deactivated. In our study, once the drivers had taken control, they clearly understood their responsibilities to respond to the conflict, in contrast to a previous study with a similar, near-perfect assisted driving system.