Learning and generalising semantic knowledge from object scenes

The robot described in this paper learns words that relate to objects and their attributes, and also learns concepts, which may be recursive, that involve relationships between several objects. Once the system is explicitly taught some words by a human teacher it finds new objects that might help to refine its concepts. Once it has found a new object, it tries to generalise its concepts to include the new object and asks the teacher for feedback. The robot learns further properties of objects by interacting with them, by touching them or walking around them to gain a new perspective. The system learns semantic knowledge from spoken interactions, using speech recognition and generation, motion segmentation, feature extraction from images using Ripple Down Rules and generalisation using Inductive Logic Programming.