Benchmarking Virtual Reinforcement Learning Algorithms to Balance a Real Inverted Pendulum

We benchmark common reinforcement learning algorithms on a modified version of OpenAI Gym’s Cartpole: a virtual environment containing a simulation of an inverted pendulum. While Policy Gradient, Actor-Critic, and Proximal Policy Optimization are all able to balance the pendulum, only Policy Gradient and Actor-Critic are able to quickly and consistently learn to balance the pendulum in a simulation. By transferring the trained models to the real world, all of the algorithms are able to satisfactorily balance a real inverted pendulum. On the real pendulum, Actor-Critic is best able to adequately reject disturbances among the algorithms tested.