Adaptive PID Control of Robotic Manipulators Using Machine Learning-Based Tuning Methods

Abstract

Robotic manipulators are widely used in industrial automation, precision manufacturing, medical robotics, and intelligent service systems, where accurate trajectory tracking and stable motion control are essential. However, the nonlinear dynamics of robotic manipulators, including joint coupling, friction, payload variation, parameter uncertainty, and external disturbances, limit the effectiveness of conventional fixed-gain PID controllers. This study proposes an adaptive PID control framework for robotic manipulators using machine learning-based tuning methods to dynamically optimize proportional, integral, and derivative gains under changing operating conditions. The proposed approach integrates a supervised machine learning model with the PID control loop to estimate suitable gain values based on real-time control indicators such as position error, velocity error, error rate, torque response, overshoot, and settling behavior. The data used in this study were generated from a MATLAB/Simulink-based simulation environment of a two-link robotic manipulator following multiple reference trajectories under different payload conditions, friction levels, and disturbance scenarios. The dataset includes joint position responses, reference trajectories, tracking errors, angular velocities, control torques, and corresponding optimized PID gain values obtained through offline tuning and performance-based optimization. The generated dataset was divided into training, validation, and testing sets to develop and evaluate the machine learning-based gain tuning model. Simulation results show that the proposed ML-tuned adaptive PID controller significantly improves trajectory tracking performance compared with a conventional fixed-gain PID controller. The proposed controller reduced root mean square tracking error by 34.8%, decreased settling time by 28.6%, and minimized overshoot by 31.4%. In addition, torque fluctuation was reduced by 22.7%, indicating smoother actuator response and improved control stability. Under payload variation and external disturbance conditions, the adaptive controller maintained stable tracking behavior, while the fixed-gain PID controller showed higher deviation and slower recovery. The findings demonstrate that machine learning-based PID tuning can enhance the robustness, adaptability, and tracking accuracy of robotic manipulators. This study provides a practical simulation-based control framework that can be extended toward real-time robotic control, reinforcement learning-based tuning, and experimental validation on physical robotic platforms.