A common assumption across many industries is that inserting advanced autonomy can often replace humans for low-level tasks, with cost reduction benefits. However, humans are often only partially replaced and moved into a supervisory capacity with reduced training. It is not clear how this shift from human to automation control and subsequent training reduction influences human performance, errors, and a tendency toward automation bias. To this end, a study was conducted to determine whether adding autonomy and skipping skill-based training could influence performance in a supervisory control task. In the human-in-the-loop experiment, operators performed unmanned aerial vehicle (UAV) search tasks with varying degrees of autonomy and training. At the lowest level of autonomy, operators searched images and, at the highest level, an automated target recognition algorithm presented its best estimate of a possible target, occasionally incorrectly. Results were mixed, with search time not affected by skill-based training. However, novices with skill-based training and automated target search misclassified more targets, suggesting a propensity toward automation bias. More experienced operators had significantly fewer misclassifications when the autonomy erred. A descriptive machine learning model in the form of a hidden Markov model also provided new insights for improved training protocols and interventional technologies.
The Impact of Requirement Splitting on the Efficiency of Supervisory Control Synthesis