Human inspiration for autonomous vehicle tactics

Abstract

Tactical control is needed in environments characterized by uncertainty and continuous, dynamic change. Given the likelihood of time constraints and high risks associated with poor tactical choices, current autonomous vehicles do not possess the decision making abilities to successfully perform in these environments. However, human experts frequently operate in these domains where they are forced to make quick, reactive decisions based on incomplete information. We propose, then, that the first step in augmenting autonomous vehicles (AVs) with improved tactical control capabilities is to learn, encode, and apply tactics exhibited by human experts. To test the method, five human subjects were given the task of performing an armed reconnaissance mission in a simulation environment over multiple cases with varying terrain and probability of enemy contact. By scoring the performance in each case, the best actions and decisions were filtered out and analyzed in depth to understand the strategies and tactics behind them. Human cognitive models and decision making theories were utilized to determine the cognitive processes underneath the decisions as displayed by the human subjects' think aloud reports and surveys.

(cont.) A baseline autonomous vehicle controller was designed independent of the human-in-the-loop experiments that could also perform the reconnaissance mission. After capturing the human tactics and encoding them into statechart form, a revised AV displayed a superior ability to engage enemy contacts uncovered during the reconnaissance when compared to the baseline AV. A final framework is presented that outlines how to learn and apply human-inspired tactics in future settings.