Visual-Tactile Registration of Objects in Robotic Assembly Systems

Abstract

The information about the accurate position and orientation of an object relative to a robot plays a fundamental role in robotic manipulation. Registration refers to the process of determining that so-called object pose. Generally, registration is performed manually and must be executed with high precision. For flexible systems, an automated approach, where the robotic system determines poses of objects automatically is desirable. This work proposes a visual-tactile feature-based approach for object registration. Features describe distinctive properties of objects that can be observed by sensors. By using a Bayesian inference approach, the system explicitly models hypotheses and uncertainties of feature poses and infers the hidden world state from measurements. In this work, recursive state estimation is used in order to fuse visual and tactile data. In particular, a sequential Monte Carlo algorithm, namely a particle filter, is applied. Additionally, periodically the belief on the current world state is analysed in order to detect and characterise uncertainties. Based on that, uncertainty lowering strategies are selected, parametrised and automatically executed by the robot. Eventually, the approach is experimentally validated within a robotic work cell. There- fore, hole-shaped features of objects are used. Comparison of solely visual and fused visual-tactile results shows that the addition of tactile measurements and entropy redu- cing strategies increases information and lowers the uncertainty of the believed world state significantly. At the same time better feature poses for object registration are generated.