Composite deep learning control for autonomous bicycles by using deep deterministic policy gradient

Abstract

In this paper, we investigate the problem of balance and tracking controller design for an autonomous bicycle subject to unmodeled dynamics, unknown parameters, and unmeasured states. A composite deep learning based control strategy is proposed, comprising of active disturbance rejection control (ADRC) and Deep Deterministic Policy Gradient (DDPG). Different from most conventional approaches that fail to consider path information and depend critically on exact dynamics, the proposed control scheme uses the DDPG algorithm to learn a virtual control action and then employ ADRC to handle uncertainties and stabilize the bicycle. Extensive simulations are conducted to assess the performance of the composite learning method. The results indicate that the bicycle controlled by our method can follow along a predetermined trajectory while maintaining balance.