SMC-Based, Stability-Guided Deep Reinforcement Learning for Control of Uncertain Nonlinear Dynamic Systems
Mahya Ramezani, M. Amin Alandihallaj, Andreas Makoto Hein, Holger Voos · IEEE Access · 2025
This paper presents a novel control framework that integrates the stability insights of traditional control with reinforcement learning (RL) to tackle uncertain nonlinear dynamic systems. By embedding an explicit Lyapunov-decrease term in the RL reward and shaping the reward with sliding-mode control concepts, the policy learns stability-aware and robustness-promoting behavior, optimizing tracking accuracy and energy efficiency while promoting practical stability under uncertainty. The method is applied to the underactuated satellite attitude control problem using magnetorquers, where inherent under-actuation and external uncertainties present challenges. Simulations, complemented by hardware-in-the-loop experiments, demonstrate that the proposed RL controller achieves better performance and robustness compared to traditional control strategies. These results underscore the potential of combining robust control theory with data-driven learning techniques to enhance the safety and efficacy of control systems in aerospace and robotic applications.