Gaze stabilization of a humanoid robot based on virtual linkage

Abstract

Gaze stabilization is a fundamental function for humanoid robots. Stabilizing the image being perceived facilitates the processing and thus the interpretation of visual data. In parallel, fixation should also guarantee that the visual target remains centered in the image. Several approaches exist to address the problem of gaze stabilization: closed-loop algorithms processing the visual data or inferring head movements from kinematic measurements, and feed-forward algorithms anticipating head movements from the lower-body commands. In this contribution, we develop a feed-forward controller addressing both image stabilization and target fixation into a unified framework. The addition of a virtual linkage between the robot eye and the visual target offers to elegantly rephrase the gaze control problem as the classical control of a redundant serial robot manipulator. Furthermore, a novel method to estimate the self-induced optical flow based on the robot kinematics - extended with this virtual linkage - is developed. It is then possible to solve the redundancy (i.e. guaranteeing target fixation) through a minimization of the optical flow (i.e. achieving image stabilization). This method is validated in simulation with a model of the head of the ARMAR IV humanoid. It is shown that the proposed controller offers to accurately estimate and minimize the optical flow, while keeping the visual target exactly in the center of the image.