Intelligent model structures in visual servoing

• Main Author: Cisneros, Marco Antonio Perez
• Published: University of Manchester 2004
• Subjects: 629.8
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.682243

This thesis focuses on visual servoing (VS) control systems, particularly on image-based visual servoing (IBVS) control structures. In IBVS, the error signal is computed in the image plane and the regulation commands are generated with respect to such error by means of a visual Jacobian. The main design challenge is the high latency of the visual sensor which affects the overall performance and limits the design. The primary objective is to develop a complete framework for simulation and real-time experimentation of VS schemes. One commercial CCD camera is attached to the TQ MA2000 robotic manipulator. The framework has been employed to investigate the use of RL algorithms to increase the performance of the IBVS control structure. The classic RL actor-critic structure has been used to perform on-line adjustment of the gains driving the linear trajectory regulator inside the IBVS control structure. The neural system learns directly from data in the image space and the state of the robot. Two feedforward networks are used, the actor directly modifies the regulator gains whereas the adaptive critic stores and assigns action values. By using the adaptive heuristic critic approach (AHC), the training aims to achieve real-time improvement and adaptation without losing an acceptable regulation of the visual servoing task. A compact model and a flexible framework host the reinforcement learning algorithm in order to enable its inclusion within the IBVS control structure. The approach in this thesis has solved critical neuro-dynamic problems which are derived from the interaction between the imaging model and the robot’s dynamics. The VS toolkit also provides a real-time library to implement and test the IBVS control structure. The libraries have proven effective to construct both the linear IBVS and the RL-supported IBVS system thanks to its layered architecture which facilitates the inclusion of control en› tities of different nature such as the neural networks and the learning framework. Two case studies demonstrate the applicability of the CSC VS toolkit to integrate all the required components and to implement each of the VS experiments in real-time. Performance comparison between the linear IBVS and the RL-supported system are also documented to show the effectiveness of the actor-critic structure.