Neural network position and orientation control of an inverted pendulum on wheels

In this contribution, we develop a feedback controller for a wheeled inverted pendulum in the form of a neural network that is not only stabilizing the unstable system, but also allows the wheeled robot to drive to arbitrary positions within a certain radius and take a desired orientation, without the need to compute a feasible trajectory to the desired position online. While some techniques from the reinforcement learning community can be used to optimize the parameters of a general feedback controller, i.e. policy gradient methods, the method used in this work is an approach related to imitation learning or learning from demonstration. The demonstration data however does not result from e.g. a human demonstrator, but is a set of precomputed optimal trajectories. The neural network is trained to imitate the behavior of those optimal trajectories. We show that a good choice of initial states and a large number of training targets can be used to alleviate a problem of imitation learning, namely deviating from training trajectories, and we demonstrate results in simulation as well as on the physical system.