Takeover Time

Marc Green

Perception-response time is measured from some event which starts the clock running. In typical research, this might be the appearance of another vehicle, a pedestrian, bicycle, etc. The literature is large, containing investigations of numerous variables on driver PRT.

The newest clock starter in the response time literature is the "take-over request" (TOR) in automated or semi-automated vehicles. It starts the clock for measuring "takeover time" (TOT), the interval required for a "driver"¹ to regain control of his vehicle from an automated system that signals a "critical event." The literature uses a large number of terms and acronyms not encountered in standard PRT research. In addition to TOR and TOT, there are ADS ("automatic driving system"), NDT/NDRT ("non-driving task" or "non-driving related task"), and ODD ("operational design domain"). "Time budget" (which surprisingly is never shortened to TB) is the interval from the TOR until the ADS reaches the limit of its ODD. It is analogous to urgency in TTC, but it typically involves intervals on the order of five to nine seconds, much longer than those employed in analogous PRT research. The literature is too extensive to cover here but several large reviews are available (e.g., Zhang, de Winter, Varotto, Happee, & Martens, 2019; Morales-Alvarez, Sipele, Léberon, Tadjine, & Olaverri-Monreal, 2020; Weaver & DeLucia, 2020). For present purposes, the main interest is whether TOT results add anything new to estimating driver PRTs.

TOT studies generally use clock starters and stoppers that differ fundamentally from typical PRT research. One obvious difference lies in sense modality. When "drivers" (a better term would be "controllers") travel in unautomated vehicles, the clock starters are almost invariably visual, the sight of brake lights or a pedestrian stepping off a curb, etc. TORs are more often auditory (a buzz, tone, or voice) or somatosensory (vibrotactile), since the "driver" may be looking away from the road and the instrument panel. Multimodal TOTs are also possible. Unlike normal visual clock starters where the detection, localization and sometimes identification occur together and invoke a specific response (e.g., a vehicle is approaching from the left so steer to the right), TORs are usually arbitrary signals although some (e.g., a voice command) may supply additional information such as hazard location or suggested response. However, this additional information apparently provides no improvement in the TOR (Weaver & DeLucia, 2020).

The clock stoppers may also differ. Some TOT studies use the conventional responses of braking and steering. Others use ones not commonly seen in the PRT literature, e.g., time to switch gaze, place the hands on the wheel, check the mirrors, or complete a lane change.

To summarize the research very briefly, TOT is longer when the "driver" is engaged in an NDRT (i.e., distracted) and when the NDRT creates higher cognitive load. Of the three types of driver distraction (Chapter 13), visual and manual are more disruptive than purely cognitive because they overlap most with the resources needed to drive. Unsurprisingly, TOT is also longer when expectation is violated. Conversely, take-over research evidence suggests that the auditory and somatosensory clock starters can produce faster response than purely visual ones. This would be expected from basic sensory research since the auditory and somatosensory modalities transduce energy faster and create shorter PRTs than vision, which is a relatively slow sense. The faster response to a somatosensory clock starter has already been shown in PRT research where driver responses to abrupt gusts of crosswind were faster than typical visual PRTs (Wierwille, Casali, & Repa, 1983). Moreover, drivers shifting gaze from near to far must accommodate and verge for the new distance which can take substantial time (e.g., Travis, 1948). Lastly, response is faster when the "driver" is more experienced with the system and when the "time budget" is shorter (greater urgency) (e.g., Eriksson & Stanton, 2017). Some (Radlmayr & Bengler, 2015) argue that longer time budgets produce slower but better quality responses. Humans are subject to a speed-accuracy tradeoff with slower but more accurate responding often leading to a better outcome. Response quality is seldom considered in the PRT literature except for the relative effectiveness of steering and braking avoidance that was discussed above (but see Ayres & Kubose, 2012). In some cases, however, TOT and response quality may not vary inversely. Zeeb, Buchner, & Schrauf (2016) found that NDRTs lowered response quality as measured by a small lateral movement difference but had no effect on a TOT of putting the hands on the wheel.

Some of these effects are small and authors of meta-analyses may disagree about their significance. This reflects a fundamental problem with meta-analyses: it requires the combining of data from studies which vary in methodology and operational definitions and which may have subtle, yet critical, differences that influence the results. A meta-analysis requires the author to gloss over details that may be critical. I mention this problem when discussing PRT computer programs.

Another problem with applying TOT studies to PRT is their relatively long durations. Despite the sensory speed advantage of nonvisual TORs, TOTs range up to four seconds depending on condition. Like PRT the response distributions skew asymmetrically toward longer times. Eriksson & Stanton (2017) found that when variability is considered, a safe TOT time budget extends beyond the oft recommended seven seconds, a speed which would be glacial by PRT standards. The explanation for the slow response speed and the wide range of TOT results is likely due in part to the long time budgets. Events further in the future are both less critical and less certain than those in the near future. It is also like due to the poorer situational awareness of "drivers" in automated vehicles (e.g., Endsley & Kiris, 1995).

TOT studies also suffer from the same methodological limitations as PRT research and a few more. De Winter, Stanton, & Eisma (2021) suggest that TOT studies are generally performed in unrealistic circumstances. They note that the tests are relatively short so drivers do not have time to become fatigued or bored. They also don't have enough time to be lulled into complacency by developing trust in the system. More trusted systems produce poorer takeover performance (Fu, Johns, Hyde, Sibi, Fischer, & Sirkin, 2020). Long term use of an automated system may lead to drivers who skill has atrophied through lack of use or who may never develop driving skill at all. As the level of automation grows, "drivers" will be more willing to perform increasingly engaging NDRTs (e.g., de Winter, Happee, Martens, & Stanton, 2014). No research tests the effects of deeply engaging NDRTs such as reading or writing a complex paper or doing mathematics. Increased automation trust would almost certainly induce more people to "drive" while intoxicated, high, or otherwise impaired, but such "drivers" are never tested. Moreover, the drivers in the TOT studies are subject to the same artificial elements as found in PRT research. The drivers know that they are being tested so they are on high alert and on best behavior especially since a safety driver is in the vehicle. Some PRT studies have attempted to create more ecologically valid surprise conditions, but this is too dangerous to attempt in a TOT study conducted in realistic situations. Lastly, the "drivers" are both self and researcher selected and practiced so representativeness is also an issue.

In sum, the number of TOT studies has exploded over the last few years, but it is doubtful whether this frantic activity has added much to knowledge about driver behavior in unautomated or even automated vehicles. De Winter, Stanton, & Eisma (2021) concluded that TOT research has mostly reinvented the wheel. The studies generally employ different clock starters and often different clock stoppers but the qualitative effects of most variables, such as expectation, distraction, etc., are completely predictable from basic and applied PRT research. The absolute TOTs tend to be much longer than typical PRTs but that is likely because the long time budgets remove the urgency which is a critical variable in driver PRT. Moreover, the task of taking over control of an automated vehicle in response to an arbitrary auditory or vibrotactile signal is fundamentally different than normal driving which involves visually detecting and localizing potential hazards, tasks which normally occur together (e.g., Green, 1992). Lastly, the studies suffer from the same artificialities as other controled research studies.

Endnote

¹Since the automated car is doing most or all the of the driving, the human is clearly no longer the "driver." He is closer to being a passenger but not quite since he may need to take control. A new term is needed. "Supervisor" would be a possibility as it is already often used in other highly automated systems. Unfortunately, "supervisor" has too many syllables and would never catch on. It also implies that the human is keeping a continual eye on the system, something that those in highly automated vehicles are unlikely to do. Perhaps most problematic, it would also be resisted by the impaired driver hysteria lobby. MADD would be mad since "drunken supervisor" doesn't alliterate or have have much pizzazz. "Smashed supervisor?" "Stewed supervisor?" And there is "distracted supervisor." Another alternative is "back up driver," but this would suggest that the automation is not 100% trustworthy.

---

A division of 2057949 Ontario, Inc.
Copyright © 2024 Marc Green, Ph. D.
Home Page
Contact Us