Journey's End: Will Vehicle Automation Make Skilled ...

JOURNEY’S

END: WILL VEHICLE AUTOMATION REDUNDANT?

MAKE SKILLED

DRIVERS

Mark S. Young and Neville A. Stanton Department of Design Brunel University, Runnymede Campus Coopers Hill Lane, Egham Surrey TW20 OJZ United Kingdom Recent advances in technology have meant that an increasing number of vehicle driving tasks are becoming automated. Such automation poses new problems for the human factors specialist, with particular concern about the effects of automation on driving performance. A body of research is building up which suggests that drivers cannot recover from automation failure, yet they can cope with a similarly critical scenario in manual driving. The current paper investigates whether recovery from automation failure is affected by level of driver skill. Learner drivers were compared to expert drivers under two levels of automation, with a failure of longitudinal control in each condition. Clear differences in response patterns were found, with more experts and fewer learners responding to the failure under high levels of automation. It is concluded that the issue of skill with automation is one which requires more attention than is currently forthcoming from the human factors community. INTRODUCTION Vehicle manufacturers are currently in the process of developing devices which automatically control the speed, headway, and lateral position of cars. In Europe, these devices tend to be centred around autonomous vehicles (e.g., Richardson et al., 1997) whereas in the United States, the focus is upon developing the road infrastructure into Automated Highway Systems (AHS; e.g., Bloomfield et al., 1995). Either way, the human factors promises and problems remain much the same. The present paper focuses on the autonomous vehicle concept, in particular the effects of Adaptive Cruise Control (ACC) and Active Steering (AS). ACC offers total longitudinal control of the vehicle (i.e., speed and headway), and has already been released. The AS system will be on the roads in the next few years, bringing limited lateral control to the vehicle. Whilst this is not an automatic overtaking system, AS does maintain lane position without driver intervention. A major concern with such systems is whether any performance benefit will be realised, or

indeed if there may even be a performance decrement. There is evidence that automation in aviation can have adverse consequences (e.g., Sarter and Woods, 1995) so it is reasonable to predict some problems in automobiles (cf. Hancock and Parasuraman, 1992; Stanton and Marsden, 1996). One issue which has received an increasing amount of attention is that of recovery from automation failure. This is an obvious practical concern, as it would be impudent to assume that any technical system is flawless. There are a few studies which have used driving simulators to investigate failures of automation. Nilsson (1995) used three critical scenarios to compare manual driving performance to that when using ACC. Amongst the drivers who crashed in this experiment, most were in the ACC condition. Similarly, Stanton et al. (1996) found that one third of drivers faced with ACC failure were not able to recover control effectively. More recently, de Waard et al. (1999) simulated failures of an AHS, and again demonstrated that only half the participants reclaimed control of the vehicle. Finally, Desmond et al. (1998) investigated failures

Downloaded from pro.sagepub.com at University of Southampton on February 10, 2016

of lateral control when driving manually (i.e., wind gusts) and when the vehicle was fully automated. Recovery was more effective under manual conditions. The explanations for these results vary around expectations about the automation (Nilsson, 1995), mobilisation of effort (Desmond et al., 1998), complacency (de Waard et al., 1999), and mental workload (Stanton et al., 1996). However, none of the research to date has investigated whether driver skill is a factor in determining the impact of automation in future vehicles. It will eventually become the case ihat any driver may step into a vehicle equipped with automated systems, regardless of their experience. Initially, novel technologies are fitted to prestige models only, implying that the drivers who have access to them are highly experienced. However, just as with power assisted steering, anti-lock brakes, and even automatic transmission, these new devices will eventually filter down to become widely available. It is conceivable that a newlyqualified driver with basic training could immediately use a vehicle equipped with ACC, or in the future, AS. The interaction of skill and automation is important for a number of reasons. Bainbridge (1978) views expert performance as open-loop, anticipatory behaviour. Any change in the task environment increases demand and breaks down the predictability of the situation, so the expert must revert to closed-loop, novice tactics. It is posited here that all operators - novices and experts alike essentially satisfy the criteria for automaticity when faced with automation (that is, fast, attention-free, and unconscious performance; Anderson, 1995). Given that experts behave as novices when demand increases, it is plausible to assume that the reverse would be true in a situation of unusually low demand (i.e., automation). However, whereas the expert has an enhanced knowledge base and can anticipate events, the novice is deprived of this ability. Thus they will not react as experts in critical situations, such as the overlearned braking response (e.g., Nilsson 1995). An experiment was conducted in the Southampton Driving Simulator to investigate whether there are any differences between skill

groups in response to automation failure. In a previous experiment by the same authors (Young and Stanton, ZOOO),expert drivers displayed superior longitudinal control of the vehicle when compared to learners. However, if steering was then automated, this advantage apparently disappeared. The purpose of the present experiment was therefore to find out if skilled driving will still be valuable in automated vehicles of the future. METHOD A mixed design was used. Level of automation constituted the within-subjects variable, with two levels: ACC (i.e., longitudinal control is automated), and ACC+AS (i.e., both longitudinal and lateral control are automated). The latter condition essentially constitutes fully automated vehicle control. Order of presentation was randomised to counterbalance practice effects. Driver skill level was the between-subjects factor, again with two levels: learner (i.e., currently learning but does not hold a full licence), and expert (i.e., holds a full driving licence). These groups were chosen on the basis of performance data from a previous experiment (Young and Stanton, 2000). At the time of writing, there were eight learner drivers in this experiment (mean age = 20.9, mean hours of tuition = 21.5), and 24 in the expert condition (mean age = 23.7, mean annual mileage = 5286; mean years held licence = 5.60). Data are still being collected, though, and it is planned to increase the sample size of the learner driver group. A five minute free practice run was followed by the two experimental conditions, each lasting 10 minutes. In the experimental trials, participants were instructed to follow a lead vehicle travelling at 70mph for the entire duration. The simulated road was a mixture of straight and curved sections. At pseudo-random intervals, the lead vehicle would brake until its speed decreased below 30mph, at which point it would accelerate back up to 70mph. Participants were instructed to stay behind the lead vehicle, relying on the ACC system to maintain headway as much as possible. However, participants were also informed that if they felt the need to intervene, they should do so.


1-91

Proceedings

of the IEA 2OOOlHFES 2000 Congress

About one minute from the end of the run, the ACC system would disengage without warning at the same time as the lead car braking. Participants had to intervene if a collision was to be avoided. This happened in both trials, to see if participants were more effective in reclaiming control when expecting the failure, compared to when they were nai’ve. The simulator recorded data on collisions and brake reaction times. An infra-red camera in the footwell was used to record foot movement time. Absolute reaction time was then calculated I 1bY subtracting the foot movement time from brake reaction time - this gives time from the failure event to first reaction. If the participant makes no reaction, time to collision is around 4.0s.

RESULTS

In terms of the dichotomous variable of whether or not participants attempted to respond, it is possible to use chi-square goodness of fit tests on the available data. Significant results were observed in the ACC+AS condition only. In the learner group, only one participant attempted to recover from the ACC failure, the remaining seven did not react at all (&ij = 4.50; p c 0.05). However, in the expert condition, 18 participants reacted to the ACC failure, with only six failing to respond (& = 6.00; p c

0.05). 2.5

2.4

2.3

Given the small data set currently available in the learner group, many of the planned inferential statistics are not viable. Therefore, most of the conclusions in this paper about skill differences will be drawn from descriptive statistics only. It should be stated initially that almost all of the participants collided with the lead vehicle following the ACC failure. This is not an adverse result - the situation presented to participants was quite difficult (although not impossible) to recover from. The important factor is whether participants respond, and if they do, their speed of response. It is these factors which the analyses concentrate on. Amongst those participants who made a response, absolute reaction time across both skill groups was lower in the ACC condition than the ACC+AS condition (M = 2.16s for ACC, 1M= 2.47s for ACC+AS). This pattern is consistent within each skill group (learners: 1M= 2.30s for ACC, 2M= 2.42s for ACC+AS; experts: A4 = 2.12s for ACC, 1M = 2.48s for ACC+AS). However, a paired-samples t-test in the expert group does not reveal a significant difference in reaction time between the automation conditions (rg = -0.998; p = 0.345). Inspecting the graph of these data (figure 1) suggests that there may be an interaction between skill and automation affecting reaction time. With an increased sample size, an analysis of variance might prove this to be significant.

s-B_

ACC ACC+AS

Figure 1: Absolute reaction times of each group

DISCUSSION The purpose of the experiment reported in this paper was to investigate whether level of driver skill might influence responses to automation failure. Although the data set is incomplete at the time of writing, there are already some dramatic effects emerging. Most striking amongst these is the number of people who actually react to the automation failure. When participants use ACC but control the steering manually, there is no association between those who respond and those who do not respond. However, when both lateral and longitudinal control are automated, significantly fewer learners respond whilst significantly more experts attempt to recover from the automation failure. This finding clearly supports the argument outlined previously, based on Bainbridge’s (1978) theory. Under normal conditions, overt


performance with high levels of automation is essentially equivalent across skill groups. The problem arises when participants are suddenly called to re-enter the control loop. Experts have an amassed knowledge base and respond automatically to critical events. Drivers with less experience have not developed such abilities, and therefore fail to respond. A puzzling aspect is why this pattern should not be observed in the ACC-only condition. For the learners, it may be that the increased workload or activation maintains their alertness, giving them a better chance to respond to the failure. dn the other hand, this same extra stimulation may be distracting for the expert drivers, reducing their capacity to respond to abnormal events. Data have also been collected on mental workload, trust, situation awareness, and heart rate, which may ratify such speculation; however these data are beyond the scope of this paper. The data on reaction times can only be interpreted cautiously, due to the low frequencies contributing to the statistics. There is a hint of an interaction between skill and automation for those participants who did respond. Experts react much quicker in the ACC-only scenario than with ACC+AS. A similar difference exists in the learner driver group, but is not quite so pronounced. This result, if confirmed, would be more in line with the predictions. For those participants who do respond, it seems that expertise is only advantageous if the driver retains some control over the vehicle (i.e., steering). Fully automating vehicle control has a detrimental effect on reaction times in each condition. Taking these two sets of results together, it can be concluded that whilst experts are more able to respond to automation failure, the reactions of both groups are dulled if vehicle control is fully automated. This has important implications for the future of such automated systems. Amongst the expert driver population, increased vehicle automation allows more attention to be devoted to hazard perception. However, should a hazard arise, drivers are not so quick to react, and this delay may be critical. More seriously, an inexperienced driver using an

automated vehicle is both less able and slower to respond to critical events requiring manual takeover. Thus we see that skilled driving will continue to be an asset in vehicles of the future. However, some significant human factors effort still needs to be expended to determine how to improve the performance of all drivers in automation failure scenarios. Hancock and Parasuraman (1992) advocate error tolerance in automation, using backup systems to allow a device to “fail softly” whilst still making the failure transparent to the driver. Keeping the driver informed and allowing them to retain an active part in the control loop will prove beneficial when manual control is needed again. In that case, it would seem that redundancy in the systems is infinitely preferable to a redundant driver.
