Visual object classification by robots, using on-line, self-supervised learning

摘要

The challenge addressed in this paper is the classification of visual objects by robots. Visual classification is an active field within Computer Vision, with excellent results achieved recently. However, not all of the advances transfer into the study of robots in free environments; two differences stand out. One is that Computer Vision algorithms often rely on batch learning over a large but fixed data set, whereas free robots cannot predict the objects they will encounter, making batch learning inappropriate. The second difference is that Computer Vision algorithms often assume a passive relationship with their input to the world, but robots can actively affect the world around them. The main contributions of the paper are to demonstrate: (i) that an on-line version of a successful batch classifier can be adapted so that objects are treated as topic mixtures rather than single topics; and (ii) that robots can self-supervise their learning of such models by interacting with the environment.