Contextual Anomaly Management Interface (CAMI) for Autonomous Systems

To deal with the increasingly complex, dynamic and unpredictable operational environments of the 21st century, unmanned and autonomous system, sensor, and vehicle technologies are being expanded and improved. For example, unmanned air, ground, and maritime systems provide an expanding set of capabilities, such as intelligence, detection, security, targeting, and strike, while reducing the risk to human life. The goal with the employment of these systems is to shift from todays manpower-intensive model of unmanned system control to a future model with fewer users who are supervising autonomous systems (1). However, to achieve this goal, a significant issue that needs to be addressed is determining how users can, and should, supervise these multiple autonomous systems in future environments that are unpredictable, complex, and highly dynamic. A key technology that can help users supervise these autonomous systems is the development and maturation of machine-based anomaly detection, to detect and characterize significant anomalous behaviors that might emerge within an on-going mission and task context. This technology can help users supervise systems by drawing users limited attention to just the most critical, anomalous events and will be a key enabler to reducing the manpower required to manage autonomous systems by monitoring for anomalous events currently performed by humans. The focus of this effort is to make the anomaly detection technology relevant and useful to the future human supervisor. The output of this effort should define, structure, and enable efficient information transactions between users and the anomaly detection technology. Research will be needed to inform the development of system behavior anomaly detection algorithms, including models of system normalcy, deviations from normalcy, and mission context. A central challenge in this domain will be determining what behavior constitutes significant deviations from normal behavior. Deviations in the face of dynamic missions, and operational contexts is difficult to define, and must be relevant to the human user supervising the system. The model must be tailored to needs of user tasks and decisions, and tuned to optimize trust in automation (2, 3, and 4) and avoid the documented pitfalls of automation (5, 6). A user interface layer and an associated business process will be needed to structure and enable interactions between users and anomaly detection algorithms. The proposed research will develop techniques to enable the detection of anomalies in the behavior of command and control systems. The desired solution should be applicable to anomaly detection in a variety of command and control domains, such as multi-echelon military command and control, and the management of multiple autonomous vehicles and systems. For example, for the management of multi-UAV systems, the algorithms will detect anomalies to either make corrections within the UAVs mission scope or alert the operator and provide alternative courses of action. Resultant capabilities are expected to produce cost savings through a reduction in manpower, as Autonomous Warfare evolves from multiple operator vehicles with teams of human controllers, to a single operator managing multiple systems.