Robots Show Us How to Teach Them: Feedback from Robots Shapes Tutoring Behavior during Action Learning (Extract)

Recently, an approach to learning continuous movement skills has been proposed which integrates social cues from the tutor (i.e. prosody, head orientation and gaze direction), though it has not been tested with inexperienced users[14]. We are not aware of related work on imitation of actions on a trajectory-level which incorporates robot feedback into the system.

The presented work challenges the predominant assumption of a unidirectional knowledge transfer based on an extensive user study with an autonomously interacting humanoid robot. We subscribe to a perspective present in research in social human-human interaction emphasizing the process of alignment between mental states, actions' goals[15], and communication[16]. Correspondingly, action learning via interaction has the aim to align the learner's mental states or action goals to those of the tutor in a co-construction, which is not possible through active perception only (active perception refers to strategies involving for instance an autonomous re-positioning of the robot's sensors to increase information gain and improve perception[17]). Subscribing to the interactive view in Human-Robot Interaction (HRI), it is not the user alone who determines what is being demonstrated (Figure 1A) (as it is currently implicitly assumed in robot imitation learning) but the demonstration has to emerge with the feedback of the learner (Figure 1 B).

Feedback behavior is essential since, as we have previously shown in parent-child interactions[3], tutors are highly sensitive to the learner's feedback, which is an important cue to infer the learner's current state of understanding. Parents for instance modify their manual movements with regard to their child's focus of attention. Also robots – similarly to children – can benefit from the input tailored specifically to their perceptional and cognitive capabilities[6]. In current HRI, the interactive view on social cognition and communication has not been tested, because most robots are barely capable of a real interaction. An appropriate setting requires endowing a humanoid robot with autonomous feedback behavior, which is a technically demanding task which we had to solve to conduct the current study (see Methods section). This is opposed to commonly applied Wizard-of-Oz techniques, in which a human operator remotely controls the robot, making them much simpler but unsuitably implying generating human feedback instead of robotic feedback for the robotic system[18].