Our stance on cognition coincides with the emergent systems approach: cognition is the process whereby an autonomous system becomes viable and effective in its environment. From the implementation point of view, we see cognition as the ability to predict the future unfolding of events initially dealing with the immediate (closer to direct sensory perception) and gradually developing to span a longer time frame (where sensory perception alone cannot get). Cognition cannot be hand-coded, it is necessarily the product of a process of embodied development. Thus, studying and modeling learning and development is a crucial aspect of our research agenda.
|Confidence map when learning motor skills [see reference 1]. Map learning is achieved by an ANN trained using backpropagation.||Random-feature approximation of RBF kernel in passive-aggressive learning . Abscissa are the number of samples; ordinates the average computation time. Note as constant time per sample can be obtained. Here we learn the dynamics of the robot arm for point to point movements.|
|The iCub learning to reach for visually identified targets. Machine learning is used to learn the kinematics and Jacobians online through various babbling strategies.||Quality of the random feature approximation of RBF. Abscissa represent the number of components of the approximation; ordinates are the MSE. The training set is from the robot dynamics learning task.|
1 R. Saegusa, G. Metta, G. Sandini, and S. Sakka. Active motor babbling for sensory-motor learning. In 2008 IEEE International Conference on Robotics and Biomimetics (ROBIO2008), 14th-17th December 2008, pp. 794–799.
D. Vernon, G. Metta, G. Sandini. A Survey of Cognition and Cognitive Architectures: Implications for the Autonomous Development of Mental Capabilities in Computational Systems. In IEEE Transactions on Evolutionary Computation, special issue on AMD. Vol. 11, No. 2, April 2007.
L. Berthouze and G. Metta. Epigenetic robotics: modelling cognitive development in robotic systems. In Cognitive Systems Research. Volume 6 Issue 3. September 2005.
L. Natale, F. Nori, G. Sandini. G. Metta. Autonomous learning of 3D reaching in a humanoid robot. In proceedings of IROS 2007. USA. November, 2007.
A. Gijsberts, G. Metta, L. Rothkrantz. Evolutionary Optimization of Least Squares Support Vector Machines. In Special issue on Data Mining in the Annals of Information Systems. Springer. March 2008.