Forward models applied in visual servoing for a reaching task in the iCub humanoid robot

Abstract

This paper details the application of a forward model to improve a reaching task. The reaching task must be accomplished by a humanoid robot with 53 degrees of freedom (d.o.f.) and a stereo-vision system. We have explored via simulations a new way of constructing and utilizing a forward model that encodes eye–hand relationships. We constructed a forward model using the data obtained from only a single reaching attempt. ANFIS neural networks are used to construct the forward model, but the forward model is updated online with new information that comes from each reaching attempt. Using the obtained forward model, an initial image Jacobian is estimated and is used with a visual servoing controller. Simulation results demonstrate that errors are lower when the initial image Jacobian is derived from the forward model. This paper is one of the few attempts at applying visual servoing in a complete humanoid robot.

Keywords:

• forward models
• ANFIS
• robotics
• visual servoing
• neural networks

Acknowledgment

This work is supported by the ROBOTCUB project (IST-2004-004370), funded by the European Commission through Unit E5 ‘Cognitive SYSTEMS’. Also, we would like to thank the Italian Institute of Technology for its financial support to the Ph.D. courses of the first two authors.