Abstract:

In urban battlefield environments, unmanned reconnaissance vehicles help command centers better understand the situation in target areas, enhance decision-making accuracy, and reduce the threat of military operations. At present, unmanned reconnaissance vehicles mostly use Ackermann steering geometry. The path planned by the traditional algorithms does not conform to the kinematic model of the unmanned reconnaissance vehicle. Thus, the combination of bicycle motion model and deep Q-network are proposed to generate the motion trajectory of unmanned reconnaissance vehicles in an end-to-end manner. In order to solve the problems of slow learning speed and poor generalizing of deep Q-network, a deep Q-network based on experience classification according to the training characteristics of neural network and a state space with certain generalization ability are proposed. The simulation experiment results show that compared with the traditional path planning algorithms, the path planned by proposed algorithm is more in line with the movement trajectory of the unmanned reconnaissance vehicle, and which improve the learning efficiency and generalization ability of the unmanned reconnaissance vehicle.

Key words: deep reinforcement learning, unmanned reconnaissance vehicle, path planning, deep Q-network