After Vehicle Automation Fails: Analysis of Driver Steering Behavior after a Sudden Deactivation of Control

Abstract

Vehicles with SAE Level 2 or 3 automation rely on the driver to intervene and resume control when failures occur. In cases which the driver must steer upon regaining control, the initial conditions of the vehicle's state variables can affect the success of the drivers' recovery. Hence, a model to determine the consequences of these initial states could help identify the requirements of shared control to guarantee a smooth recovery after an automation failure. Such a modeling tool should be simple, such as a two-point visual continuous control model of steering. Data to validate such a model were collected from participants driving in the NADS-1 simulator who were placed in a situation similar to an extreme case of automation failure by drifting their vehicle to a target heading angle and lane deviation. This was done while the drivers were distracted with a secondary task that kept their eyes off the road. The maximum lane deviation reached during recovery shows that the initial heading angle and steering wheel angle strongly affected the maximum lane deviation. Moreover, a slightly modified version of the two-point visual control model was used to simulate the drivers' steering profiles. The model was successful at recreating the participants heading angle and lane deviation profiles but failed to replicate the drivers' steering profile. This simple model of steering control could be used to assess the consequences of a vehicle ceding control at various initial conditions, but is not able to reproduce all aspects of steering control.

Predicting takeover response to silent automated vehicle failures

“…The lack of model development is partly due to a lack of empirical work on which to base these models. To the authors' knowledge, there are very few studies that have examined steering responses to automation failures without any alert (exceptions being [10,[12][13][14]), or with a visual-only alert [which effectively becomes a silent failure when the visual icon is not in the driver's current field of view; e.g. [4,15].…”

“…In contrast to previous studies, which examine only a few scenarios (e.g. [4,10,[12][13][14][15]), we systematically examine behaviour across a wide range of failure criticalities in highly controlled takeover conditions. Bayesian hierarchical modelling is employed to closely examine responses to silent failures under both optimal conditions and during increased cognitive load.…”

“…prediction and error accumulation) have been suggested to play a role in manual steering corrections [ 9 ], but as yet have not been employed to examine the steering response to silent failures. Dinparastdjadid et al [ 10 ] showed that a popular model of manual steering control, where drivers generate control outputs based on a weighted combination of angular inputs from a near and a far point [ 11 ], can capture the lane position and orientation profiles of steering recoveries to silent failures (where the vehicle drifted without warning while the driver was looking towards a visual distraction task) but crucially fails to describe how the driver moves the steering wheel. Further development is clearly needed for models to capture the mechanisms underpinning steering behaviour in silent failures [ 3 , 10 ].…”

Designing for the Extremes: Modeling Drivers’ Response Time to Take Back Control From Automation Using Bayesian Quantile Regression

“…The eligibility criteria included having a license without restriction (other than corrective lenses), being able to drive without the aid of special equipment, and having driven at least 2,000 miles per year. More information about the screening process can be found in Dinparastdjadid et al (2018). Each test session ranged between 1.3 and 2.0 hr and participants were compensated for their time.…”

Predicting takeover response to silent automated vehicle failures

“…prediction and error accumulation) have been suggested to play a role in manual steering corrections (Markkula et al, 2018), but as yet have not been employed to examine the steering response to silent failures. DinparastDjadid et al (2018) showed that a popular model of manual steering control, where drivers generate control outputs based on a weighted combination of angular inputs from a near and a far point (Salvucci and Gray, 2004), can capture the lane position and orientation profiles of steering recoveries to silent failures (where the vehicle drifted without warning while the driver was looking towards a visual distraction task) but crucially fails to describe how the driver moves the steering wheel. Further development is clearly needed for models to capture the mechanisms underpinning steering behaviour in silent failures (McDonald et al, 2019;DinparastDjadid et al, 2018).…”