A fuzzy-TD3 hybrid reinforcement learning framework for robust trajectory tracking of the Mitsubishi RV-2AJ robotic arm

Abstract

This paper proposes a novel hybrid control architecture that synergistically integrates a fuzzy logic system with the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to achieve precise, robust trajectory tracking for a 5-degree-of-freedom (5-DOF) robotic manipulator. The design merges the interpretable, rule-based reasoning and rapid transient response of fuzzy logic with the model-free, long-term adaptive optimization capabilities of deep reinforcement learning. Within this framework, a fuzzy supervisor delivers immediate corrective actions using real-time error states, while the TD3 agent concurrently learns an optimal control policy to manage the system’s nonlinear dynamics. Extensive simulation studies on complex trajectories, including N-shaped, helical, and spiral paths, demonstrate the architecture’s superiority. The hybrid fuzzy-TD3 controller reduces tracking error by 27.8–50% compared to a standalone TD3 agent and by 14.8–28.6% compared to a hybrid PID-TD3 baseline. Furthermore, under conditions of parametric uncertainties and internal as well as external disturbances, it maintains performance improvements of 23.5–34.2% over TD3 and 11.0–16.7% over the hybrid PID-TD3, confirming enhanced robustness. Validation through sensitivity analysis, numerical stability verification, and rule activation transparency establishes this method as an effective, adaptive, and explainable solution for advanced robotic control applications.