Human evaluation of robot-generated spatial descriptions

We describe a system where the semantics of spatial referential expressions have been automatically learned by finding mappings between symbolic natural language descriptions of the environment and non-symbolic representations from the sensory data of a mobile robot used for localisation and map building (SLAM). Although the success of learning can be measured by examining classifier performance on held-out data, this does not in itself guarantee that the descriptions generated will be natural and informative for a human observer. In this paper we describe the results of an evaluation of our embodied robotic system by human observers.