Learning-Based Trajectory Tracking and Balance Control for Bicycle Robots With a Pendulum: A Gaussian Process Approach

Abstract

This article presents a learning-based control framework for a class of underactuated bicycle robots with an active pendulum attached to the rear body as a balancer. In contrast with the extent methods that require the exact dynamic modeling, the proposed solution deals with the uncertainties, including the impacts of changing road environment, unmodeled dynamics, and external disturbances by virtue of Gaussian process and disturbance cancellation on control device. The tracking controller using traction and steering is presented based on Lyapunov redesign. The self-balancing control of the roll-pendulum underactuated subsystem is considered. This subsystem is first converted into a pure-feedback form, and a novel backstepping controller based on dynamic inversion technique is developed to maintain the roll and pendulum angles in the balance equilibrium manifold. By using an event-triggered update in the learning process, efficient prediction can be achieved, and the tracking error and the roll angle are limited to smaller bounds that are explicitly given. The simulation verifies that adjustable performance can be offered for tracking by tuning the control gains and the hyperparameters of the Gaussian process predictor. Finally, experiments are conducted on a real bicycle platform to illustrate the effectiveness of the control framework.