Object Recognition and Visual Servoing: Two Case Studies of Employing Fuzzy Techniques in Robot Vision

Two Case Studies of Employing Fuzzy Te hniques in Robot Vision Alois Knoll, Jianwei Zhang, Thorsten Graf, and Andr e Wolfram Fa ulty of Te hnology, University of Bielefeld, Germany Summary. The apabilities of observing the world and manipulating obje ts based on visual information are basi requirements in roboti appli ations. Typi ally, for manipulating obje ts in real environments it is ne essary to re ognise and lo ate the substantial obje ts robustly. Due to external in uen es, e.g. partial o lusions of obje ts and illumination hanges, as well as to internal in uen es, e.g. noisy imaging hardware, ina urate measurements and quantization e e ts, the re ognition systems have to ope with in omplete, un ertain and ina urate information. In the rst part of this paper we present the general framework of a robust approa h for re ognising partially o luded obje ts. It ombines the popular onept of using invariant shape des riptions, i.e. des riptions of obje ts whi h remain una e ted by ertain variations of the intrinsi and extrinsi amera parameters, with the exibility and readability of rule-based fuzzy systems by applying invariant obje t shape des riptions in fuzzy if-then lassi ation rules. In the se ond part, we propose a fuzzy ontrol approa h for learning nepositioning of parallel-jaw robot gripper using visual sensor data. The rst omponent of the used model an be viewed as a per eptron network that proje ts highdimensional input data into a low-dimensional eigenspa e. The se ond omponent is a fuzzy ontroller serving as an interpolator whose input spa e is the eigenspa e and whose outputs are the motion parameters. Instead of undergoing umbersome hand-eye alibration pro esses, our system is trained in a supervised learning proedure using systemati al perturbation motion around the optimal grasping pose.