Object Detection for a Humanoid Robot Using a Probabilistic Global Workspace

Abstract

We present an architecture for detecting generic objects in unstructured scenes for an embodied visual system. The proposed architecture integrates the contributions of a collection of loosely coupled processes, each supplying a different type of information derived from a robot’s sensors, including vision and kinesthesia. The core of the architecture is a probabilistic global workspace, which is used to incrementally build a representation of the scene, and whose contents are made available to the whole cohort of processes. The loosely coupled nature of the architecture facilitates parallelisation, and makes it easy to incorporate additional processes providing new sources of information. We provide an instantiation of this architecture using five processes on an upper-torso humanoid robot. Preliminary results show that the system can classify the elements of a scene well enough for the robot to be able to detect and touch a variety of movable objects within its reach.