Publication

Using Kinematically Complex Robots for Case Studies in Embodied Cognition

Yohannes Kassahun, Mark Edgington, José de Gea Fernández, Elsa Andrea Kirchner, Dirk Spenneberg, Frank Kirchner

In: Proceedings of the 9th International Conference on Climbing and Walking Robots (CLAWAR 2006). International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines (CLAWAR) Brussels Belgium Page 258 2006.

Abstract

We present two case studies in embodied cognition which use kinematically complex robots for spatial cognition and concept forming. The first case study involves substrate classification on the basis of pri- marily proprioceptive data. During walking over vari- ous substrates a legged robot generates certain substrate specific sensory motor patterns. The acquired data is used for training a growing self-organizing neural net- work, which is connected with a standard output layer representing different substrates. The second case study is concerned with a recognition system which learns to recognize objects based on multimodal sensorimotor co- ordination. The sensorimotor coordination is generated through interaction with the environment. The system uses a learning architecture which is composed of reac- tive and deliberative layers. The reactive layer consists of a database of behaviors that are modulated to produce a desired behavior. We have implemented in the learning architecture an object manipulation behavior inspired by the concept that infants learn about their environment through manipulation [1]. While manipulating objects, the agent records both proprioceptive data and extero- ceptive data. Both of these types of data are combined and statistically analyzed in order to extract important parameters that distinctively describe the object being manipulated. This data is then clustered using the stan- dard k-means algorithm and the resulting clusters are labeled. The labeling is used to train a radial basis func- tion network for classifying the clusters. It has been found that the trained neural network is able to clas- sify objects even when only partial sensory data is avail- able to the system. Our preliminary results in both case studies demonstrate that kinematically complex robots are suitable for learning about their environment from experience and provide a new useful class of propriocep- tive information in contrast to wheeled systems.

clawa2006-em.pdf (pdf, 661 KB)

German Research Center for Artificial Intelligence
Deutsches Forschungszentrum für Künstliche Intelligenz