Navigation in a complex dynamic environment is one of the most attractive tasks. Although most of such algorithms can achieve navigation tasks effectively, they ignore the necessity of the mission planning in the process of navigation. Given the situation, a novel end-to-end two-stage deep reinforcement learning architecture for a time-sequence navigation and action-decision in a dynamic environment with randomly rapidly
moving obstacles is proposed in this paper. During the first-stage
training, a network with spatial and temporal information is designed to process the navigation task while a conventional recurrent full-connected network is adopted to resolve the action-decision task. During the second-stage training, the two networks are integrated and trained online with dynamic entropy to obtain a stable policy for dynamic missions. Simulations demonstrate that the navigation and the action-decision in
different environments can be completed effectively under our architecture.