One-shot learning techniques have recently enabled robots to learn new objects and actions from natural language instructions. We significantly extend this past work to one-shot interaction learning, from both natural language instruction and demonstration, where a robotic learner not only learns the actions appropriate forits role in the interaction, but also the roles of the other interactors. The resulting knowledge can be used immediately such that the robot can assume any role of the learned interaction, to the extent that it can perform the required actions. We demonstrate the operation of the integrated architecture in a handover task in real-time on a robot.
