Constrained Predictive Learning Control for Omnidirectional Wheeled Mobile Robots via Reducing Iteration Horizon

This paper proposes a predictive learning controller for an omnidirectional wheeled mobile robot (OWMR) with the target of achieving high-precision tracking of the desired trajectory within repetitive tasks. A prediction mechanism, based on the concept of “learning from the future,” is presented to accelerate learning convergence. Moreover, a constrained mechanism is leveraged to guarantee learning safety by introducing constraints on the input difference. The stability of OWMR under the proposed predictive learning controller is realized by formulating an optimization problem that incorporates terminal costs and constraints, together with its performance across iterations being evaluated through an analysis method based on composite energy functions. Additionally, a reducing prediction horizon strategy is adopted to mitigate computational burden, which both enhances the suitability and ensures the initial feasibility of the predictive learning controller for real-world applications. Subsequently, our predictive learning controller is implemented on the OWMR, together with several practical discussions for its deployment. Comparative simulations and real-world experiments are conducted to validate the practicality and effectiveness, highlighting its advantages over existing approaches. Note to Practitioners—Precise trajectory tracking control is a challenging but essential task, particularly for mobile robots. This paper proposes a learning-based controller that enhances performance through practice in repetitive tasks. However, traditional iterative learning controller often requires numerous repetitive operations and does not guarantee dynamic safety during each trial. To address these issues, a novel constrained predictive learning controller is introduced, which not only learns from historical experiences but also predicts future actions while accounting for input constraints. Furthermore, a new prediction strategy is incorporated to reduce computational burden. The effectiveness of proposed predictive learning controller is demonstrated through real-world experiments on the omnidirectional mobile robot.