A Q-learning approach for energy-efficient trajectory design: stochastic event capture using a fixed-wing UAV.

Unmanned aerial vehicles (UAVs) are widely deployed to sense and capture events of interest in dynamic nature or environment. The UAV's size, weight, and power constraints limit its sensing, communication, and computation performance. To tackle such issues, an energy-efficient trajectory design is needed for propulsion energy usage minimization and improving mission quality. A high-quality temporal-spatial trajectory is achieved by exhaustive search and thus at the cost of heavy computation. To this end, we consider discrete the mission time into slots and adopt a Q-learning based approach to cope with dynamic events occurrence and flight state conditions. Simulation results show that the proposed Q-learning approach trains the UAV efficient acceleration actions over time. As a result, our proposed approach significantly improves the event capture energy efficiency of the UAV compared with the benchmark circular trajectory.