Training autonomous navigation agent in the natural environment usually requires expensive investments in terms of the cost of the machine, the environment, and the manual time consumption caused by repeated experiments to obtain a large number of training data. In this study, a deep reinforcement learning method for training visual navigation intelligence agent in virtual scenarios is proposed. The virtual scenarios which simulate the natural environment, are constructed with Unity3D engine. The intelligence agent gradually learns the spatial position relationship through many iterations of training, which is finally used in every step of the action decision.

Keywords : Autonomous Navigation, Deep Reinforcement Learning, Unity3D