In this work, we provide evidence from a large-scale Tesla Autopilot driving dataset that drivers in this dataset maintain functional vigilance in their use of Autopilot. In this data, drivers use the system extensively and travel 34.8% of miles under Autopilot control. We annotated 18,928 epochs of Au- topilot disengagements for presence of challenging scenarios (termed “tricky situations”) and whether the driver anticipates or responds to these situations in a timely manner. Through this process, 8,729 tricky situations were annotated by mul- tiple individuals until an agreement was reached about the timeliness of the driver response with respect to the critical signal. The resulting annotation categorized the driver as func- tionally vigilant in all cases as determined through subjective annotation of the temporal characteristics of their response. We discuss two possible contributing factors that underlie these results, and propose possibly beneficial expansion of the functional vigilance framework with monitoring systems for both the human and the automation.
