Predicting takeover response to silent automated vehicle failures

It remains a huge challenge to create Automated Vehicles (AVs) that are able to respond safely in all possible circumstances. Silent failures will occur when an AV fails to keep within the safety envelope and does not detect this failure or alert the human driver. To ensure AV safety, it is crucial to have a better understanding of human capabilities responding to silent failures. A highly controlled experiment was conducted to test drivers detecting and steering in response to a range of lane keeping failures of automation, using Time-to-Lane-Crossing (TLC) as the primary performance metric. Bayesian hierarchical modelling was used to construct predictive models that showed drivers responded more slowly (and less consistently) during less critical failures (for each 1 s increase in TLC at failure there was a 0.36 s increase in TLC at takeover). A manipulation that increased cognitive load impaired driver performance further (TLC at takeover decreased by 0.1 s and variability increased by 10\%). Steering response magnitudes scaled according to TLC at takeover, but increased cognitive load dampened steering. Whilst these results demonstrate increased risk caused by additional cognitive load, the magnitude of the effect was fairly small compared to the within and between participant variability. Modelling this variability allowed simulations of hypothetical silent failures to be run based on different road conditions (varied curvature, width and speed) and various delays in response times. This modelling suggests that a high proportion of silent failures would result in unsafe transitions of control from AV to a human driver.