Sim-to-Real Policy Transfer in Deep Reinforcement Learning

Deep reinforcement learning holds tremendous potential for robotics applications. However, it requires large amounts of data obtained through the interaction of a learning agent with its environment. Collecting real-world training data for robots poses several challenges, including crash and safety risks, battery limitations, and wear on the driving units. While simulation offers a more accessible way to obtain training data, it often results in performance degradation due to the gap between simulation and reality. In this paper, we propose a method that involves learning a policy in simulation using domain randomization, transferring the learned policy to an actual quadrotor, and utilizing it as a low- level controller. Both simulation and real-world experimental results demonstrate that our approach can learn a robust policy capable of handling situations not encountered during the training process. This method enables the application of policies learned in simulation to real-world environments without the need for additional tuning.