Simulation meets real-world : deep reinforcement learning on inverted pendulum system

Abstract

In this thesis, we investigate the differences between the idealized gymnasium cartpole environment and a real cartpole with the aim to train robust agents in the simulation, that perform well in realworld tasks. In this work, we not only consider the classical upright task, but also the so-called swingup. Models for friction and force are implemented and their effectiveness is evaluated on the real cartpole. The robustness of an agent with regards to changing parameters of the cartpole is also examined and possible solutions presented.