Lya-DRL-SMC: a Lyapunov-stability-constrained deep reinforcement learning enhanced sliding mode control method for remotely operated vehicles

Shenao Yan, Zini Wang, Hongwen Yu

Remotely operated vehicles (ROVs) operating in complex marine environments are subject to multimodal disturbances, such as wave forces, ocean currents, and model uncertainties, which pose significant challenges to the robustness and stability of the control system. This article proposes a Lyapunov-stability-constrained Deep Reinforcement Learning enhanced Sliding Mode Control (Lya-DRL-SMC) framework. This framework dynamically optimizes the sliding surface parameters of the SMC via a Lyapunov-constrained Deep Reinforcement Learning (Lya-DRL) approach, achieving an optimal balance among robustness, tracking accuracy, and energy consumption. Simultaneously, a frequency-decoupled multimodal Extended State Observer (ESO) is introduced to accurately estimate and compensate for the system’s lumped disturbances. The finite-time stability of the closed-loop system is rigorously proven. Comparative simulation results against conventional SMC (CSMC) and Active Disturbance Rejection Control (ADRC) demonstrate the superior performance of the proposed Lya-DRL-SMC in terms of trajectory tracking accuracy (MAE), energy consumption, and chattering suppression.