Extracting general task structures to accelerate the learning of new tasks

Teaching a robot new tasks through kinesthetic demonstrations can be a long and complicated process. For example, a human has to demonstrate a new 'pick and place' task each time the object or the target location has changed. However, obtaining the abstract representation of such task can significantly reduce the learning time as the human only has to teach the necessary parameters required for the successful execution, e.g. the location of an object. In this work, we present a framework which allows to extract general task structures which together with the obtained knowledge can improve and accelerate the teaching of new tasks. Additionally, our framework exploits the semantic similarities between task parameters in order to infer the possible structure of unknown tasks. Our proposed method utilises symbolic representations of tasks combined with an ontology which makes it applicable to different environments in various domains. We analysed our framework in an orange sorting scenario and a cleaning scenario to demonstrate that it allows reducing the time required for teaching from 136.3 to 53 seconds (61.12%) and from 48.7 to 21 seconds (56.87%) respectively compared to learning only by kinesthetic demonstrations.