Memory-enhanced deep reinforcement learning for UAV navigation in 3D environment

It is a long-term challenging task to develop an intelligent agent that is able to navigate in 3D environment using only visual input in an end-to-end manner. In this paper, we introduce a goal-conditioned reinforcement learning framework for vision-based UAV navigation, and then develop a Memory Enhanced DRL agent with dynamic relative goal, extra action penalty and non-sparse reward to tackle the UAV navigation problem. This enables the agent to escape from the objective-obstacle dilemma. By performing experimental evaluations in high-fidelity visual environments simulated by Airsim, we show that our proposed memory-enhanced model can achieve higher success rate with less training steps compared to the DRL agents without memories.