Zero-Shot Sim-To-Real Transfer of Robust and Generic Quadrotor Controller by Deep Reinforcement Learning

The goal of this paper is to develop a controller that can be trained in a simulation environment and seamlessly applied to different types of real-world quadrotors without requiring any additional adaptation or fine-tuning. First, a training environment framework for a generic quadrotor based on the high-fidelity dynamics model is designed. The input for the training environment consists of angular velocity and thrust. Next, the policy network and the detailed policy learning procedure are presented. The training process includes investigating and mitigating differences in dynamics, sensor noise, and environmental conditions between the simulation and real-world quadrotor systems. Efforts are also made to increase the continuity of the action output from the policy during training. The efficiency of the proposed approach is demonstrated through a series of real-world experiments. The trained controller exhibits remarkable robustness and versatility across different quadrotor models, successfully completing flight tasks in real-world scenarios without requiring additional training or modifications. These results highlight the potential of deep reinforcement learning for achieving zero-shot sim-to-real transfer in the domain of quadrotor control.