Developing a reinforcement learning-based actor–critic neural network controller for general form of serial chain robots with torque constraints

• Published in: Mechanism and machine theory Vol. 199; p. 105676
• Main Authors: Bahrami, Vahid, Tale Masouleh, Mehdi, Kalhor, Ahmad
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2024
• Subjects: Actor–critic neural network; Desired steady-state torque; Regulator torque; Reinforcement learning; Desired steady-state torque; Actor–critic neural network; Reinforcement learning; Regulator torque
• ISSN: 0094-114X; 1873-3999
• DOI: 10.1016/j.mechmachtheory.2024.105676

This paper uses reinforcement learning techniques to introduce a novel approach to the controller design for a wide range of serial chain robots and cable-driven serial chain robots. The method addresses the regulation problem by designing a regulator and steady-state torques without requiring knowledge of the Coriolis vector, a complex part of the robot’s dynamic model. The approach relies on Denavit–Hartenberg parameters to determine regulator and desired steady-state torques, using an actor–critic neural network for adaptation and a multi-layer perceptron neural network for desired steady-state torque. Stability is ensured through Lyapunov functions, with only the mass matrix and gravity vector required. The proposed control scheme offers adaptability, robustness, scalability, and optimization features compared to traditional controllers. Simulations on a 4-degrees-of-freedom cable-driven serial chain robot validate the proposed control method’s effectiveness. A comparative benchmark is performed, demonstrating that the reinforcement learning controller outperforms other controllers in terms of regulation error, reducing it by at least 17.14% compared to deep deterministic policy gradient and robust integral of the sign of the error methods. •Regulation control system for Serial Chain Robots (SCRs).•Exploring RL as a computationally intensive alternative to traditional optimal control.•Reduced reliance on robot dynamics knowledge.•Limiting computational torque.•Neural Network-based adaptation laws and stability guaranteeing.