ONLINE REINFORCEMENT LEARNING FOR ADAPTIVE GAIN TUNING IN IMAGE-BASED VISUAL SERVOING

This thesis presents the development and implementation of a novel image based visual servoing (IBVS) controller for robotic manipulators that combines an optimized Extreme Learning Machine algorithm with an online Reinforcement Learning strategy. The classical IBVS approach is first reviewed, highlighting its primary challenges, namely the accurate estimation of the image interaction matrix and the singularities inherent in its pseudoinverse solution. To overcome these limitations, we introduce an IBVS framework that uses an Extreme Learning Machine optimized by Particle Swarm Optimization to approximate the pseudoinverse of the image interaction matrix, thereby reducing the impact of camera calibration errors. Building upon this foundation, a Continuous Action Q-Learning algorithm is employed to adjust the visual servo gain adaptively in a continuous action space. This integration significantly enhances both convergence speed and overall system stability. Finally, the proposed controller is validated through comparative simulation experiments on a 6-DOF robot manipulator. The results demonstrate that the CAQL

This thesis presents the development and implementation of a novel image based visual servoing (IBVS) controller for robotic manipulators that combines an optimized Extreme Learning Machine algorithm with an online Reinforcement Learning strategy. The classical IBVS approach is first reviewed, highlighting its primary challenges, namely the accurate estimation of the image interaction matrix and the singularities inherent in its pseudoinverse solution. To overcome these limitations, we introduce an IBVS framework that uses an Extreme Learning Machine optimized by Particle Swarm Optimization to approximate the pseudoinverse of the image interaction matrix, thereby reducing the impact of camera calibration errors. Building upon this foundation, a Continuous Action Q-Learning algorithm is employed to adjust the visual servo gain adaptively in a continuous action space. This integration significantly enhances both convergence speed and overall system stability. Finally, the proposed controller is validated through comparative simulation experiments on a 6-DOF robot manipulator. The results demonstrate that the CAQL based approach outperforms the conventional Q-Learning algorithm in terms of tracking accuracy and overall system robustness.