Vision Real-Time Simulation Training Platform for Quadrotor

Considering that agent is difficult to train in the real-world environment, a real-time simulation training platform for quadrotor is proposed. This article emphasizes the issue of unmanned aerial vehicle simulation training platforms based on deep reinforcement learning (DRL). Previous training platforms either could only handle low dimensional privileged information or needed strong computing power to handle high-dimensional information such as vision. This situation makes it difficult to design algorithms that utilize onboard sensors. For this reason, a real-time simulation platform for quadrotor is designed, which has been successfully used for common tasks such as quadrotor obstacle avoidance and landing. Finally, its effectiveness is verified through simulation.