Policy-Based Deep Reinforcement Learning for Visual Servoing Control of Mobile Robots With Visibility Constraints

摘要

In this article, the image-based visual servoing (IBVS) problem for mobile robots with visibility constraints is studied by using a policy-based deep reinforcement learning (DRL) approach. First, the classical IBVS (C-IBVS) method and its feature-loss problem are introduced. Then, a DRL-based IBVS method is presented to solve the feature-loss problem and improve the servo efficiency.Specifically, the formulation of the C-IBVS controller is inherited by the designed controller to ensure the analytical stability, and a policy-based DRL algorithm is proposed to design an adaptive law for tuning the controller gain in the continuous space, which can maintain the feature in the field of the view of the camera as well as improving the servo efficiency. Finally, the effectiveness of the proposed method is demonstrated by various comparative experiments.