Embedding Consequence Awareness in Unmanned Aerial Systems with Generative Adversarial Networks

Small unmanned aerial systems (sUAS) are becoming more prevalent, driven by consumer interest and their potential for revolutionizing aspects of commercial applications, such as delivery of urgent goods. The expected ubiquity of such systems raises concerns about their safety, and the ability of such autonomous systems to operate safely in densely populated areas (where their value will be greatest). In this paper, we outline a new framework aiming to add an additional layer of safety to aerial systems operated by a human pilot or autopilot by monitoring the UAVs environment for visual cues, and monitoring the human pilot for signs of distraction. The system will endow a UAS with the ability to reason about its safety, and the consequences of safety failures during its operation. The UAS will furthermore continuously reason about possible safety maneuvers in response to likely failures – in the event of an emergency, the vehicle can then execute its last safe maneuver, thus reducing the systems impending danger. Embedding consequence awareness in sUAS is an obvious appeal to safer and more insurable missions. For pilot skill level awareness, a method utilizing generative adversarial networks, which improves pilot skill level classification accuracy in our experiments, is proposed to compensate limited training data availability.