Generalized Path Planning for Collaborative UAVs using Reinforcement and Imitation Learning

Cellular-connected Unmanned Aerial Vehicles (UAVs) need consistent cellular network connectivity to effectively accomplish their designated missions. However, when navigating through regions with partial coverage, such as rural areas, the task of planning the flight paths for these UAV missions becomes notably intricate. Algorithms designed to solve this issue require significant computational resources, making them infeasible for active deployment where an algorithm must run in real time using small compute power. Furthermore, these algorithms exponentially scale in run-time with respect to the number of UAVs being considered. To tackle this problem, we model the parameter space as a discrete grid-world, enable collaboration between drones, and gather supervised data from nonlinear programming and unsupervised data from a simulated version of the environment with associated rewards. We then train a Deep Neural Network (DNN) on this data and approximate optimal results by combining imitation and reinforcement learning methods. This DNN can successfully be deployed at fast speeds using relatively small computational power and can generalize to unseen maps where drone collaboration can be used to reduce mission time. By using the results of a network trained on supervised data as a guiding hand during training, our reinforcement learning approach achieves results better than either method in isolation.