Zsolt Kira

In the first paper involving building of models of robot differences, we have shown that confusion matrices, describing mappings between various object properties (such as color and texture), can be learned using instances from each robot in a shared context. These models describe which properties represent similar object properties across different robots, and can subsequently be used to faciliate knowledge sharing.

In the 2009 IROS paper [3] (as well as upcoming papers), we use the models of robot differences to actually transfer object models between robots. This bootstraps the learning of one robot using experiences from another robot. This can be evidenced by higher learning curves when using transfer learning as opposed to the robot learning by itself.

As an initial step towards the overall goal, the first paper published involves proposing to leverage similarity to deal with heterogeneity. Specifically, we show how establishing a physically shared context can be used to learn models of the differences between two robots.  This is similar to the joint attention or gaze following subfields, but in this case the purpose is to ensure a shared context in order to figure out perception differences arising from robot heterogeneity. In the paper, we analyzed the cost and accuracy of several methods for the establishment of the physically shared context with respect to such modeling. We will apply these methods to use information-theoretic measures developed during the candidacy proposal to determine what perceptual differences and similarities exist between two robots.