Autonomous Navigation with Improved Hierarchical Neural Network Based on Deep Reinforcement Learning

Compared with traditional navigation strategies in normal environments, the unmanned vehicles in battlefield environments require better navigation strategies. This research formulates the autonomous navigation in battlefield environments as a markov decision process (MDP) and introduces deep deterministic policy gradient (DDPG) to obtain the continuous control signal. Meanwhile, the curriculum learning is employed to increase utilization of samples in this research. Inspired by the biological mechanism, an improved hierarchical neural network is proposed to refine the input information, which plays a better role in coordinating the choice of agent's behavior. Experimental results show that the models we proposed are able to acquire effective navigation strategies without knowing the whole information of environment. At the same time, it is proved that the hierarchical neural network and the curriculum learning are effective for improving efficiency of learning and generalization capability of models.