Optimizing Reinforcement Learning Control Model in Furuta Pendulum and Transferring it to Real-World

Reinforcement learning does not require explicit robot modeling as it learns on its own based on data, but it has temporal and spatial constraints when transferred to real-world environments. In this research, we trained a balancing Furuta pendulum problem, which is difficult to model, in a virtual environment (Unity) and transferred it to the real world. The challenge of the balancing Furuta pendulum problem is to maintain the pendulum's end effector in a vertical position. We resolved the temporal and spatial constraints by performing reinforcement learning in a virtual environment. Furthermore, we designed a novel reward function that enabled faster and more stable problem-solving compared to the two existing reward functions. We validate each reward function by applying it to the soft actor-critic (SAC) and proximal policy optimization (PPO). The experimental result shows that cosine reward function is trained faster and more stable. Finally, SAC algorithm model using a cosine reward function in the virtual environment is an optimized controller. Additionally, we evaluated the robustness of this model by transferring it to the real environment.