Adaptive Human-Robot Interactions for Multiple Unmanned Aerial Vehicles

Advances in unmanned aircraft systems (UAS) have paved the way for progressively higher levels of intelligence and autonomy, supporting new modes of operation, such as the one-to-many (OTM) concept, where a single human operator is responsible for monitoring and coordinating the tasks of multiple unmanned aerial vehicles (UAVs). This paper presents the development and evaluation of cognitive human-machine interfaces and interactions (CHMI²) supporting adaptive automation in OTM applications. A CHMI² system comprises a network of neurophysiological sensors and machine-learning based models for inferring user cognitive states, as well as the adaptation engine containing a set of transition logics for control/display functions and discrete autonomy levels. Models of the user’s cognitive states are trained on past performance and neurophysiological data during an offline calibration phase, and subsequently used in the online adaptation phase for real-time inference of these cognitive states. To investigate adaptive automation in OTM applications, a scenario involving bushfire detection was developed where a single human operator is responsible for tasking multiple UAV platforms to search for and localize bushfires over a wide area. We present the architecture and design of the UAS simulation environment that was developed, together with various human-machine interface (HMI) formats and functions, to evaluate the CHMI² system’s feasibility through human-in-the-loop (HITL) experiments. The CHMI² module was subsequently integrated into the simulation environment, providing the sensing, inference, and adaptation capabilities needed to realise adaptive automation. HITL experiments were performed to verify the CHMI² module’s functionalities in the offline calibration and online adaptation phases. In particular, results from the online adaptation phase showed that the system was able to support real-time inference and human-machine interface and interaction (HMI²) adaptation. However, the accuracy of the inferred workload was variable across the different participants (with a root mean squared error (RMSE) ranging from 0.2 to 0.6), partly due to the reduced number of neurophysiological features available as real-time inputs and also due to limited training stages in the offline calibration phase. To improve the performance of the system, future work will investigate the use of alternative machine learning techniques, additional neurophysiological input features, and a more extensive training stage.