Edinburgh Napier University

  Social robotics aim to equip robots with the ability to exhibit socially intelligent behaviour while interacting in a face-to-face context with human partners. An important aspect of face-to-face social interaction includes the efficient recognition of their surroundings, the environment and the objects within it, so as to be able to discuss, describe and provide instructions to assist continuous collaboration between the speaker and the listener. Although humans can efficiently learn from their interlocutors to perceptually ground word meanings of visual objects from just a single example, teaching robots to ground word meanings remains a very challenging, expensive and resource-intensive task. In this paper, we present a novel framework for robot concept acquisition on the fly, by combining few-shot learning with active learning. In this framework, a robot learns new concepts through collaboratively performing tasks with humans. We compared different learning strategies in a task-based evaluation with human participants, and we found that active learning significantly outperforms a non-active learning alternative, and is more preferable by the participants while increasing their trust in the social robot's capabilities.