Robots show us how to teach them: feedback from robots shapes tutoring behavior during action learning

Abstract

Robot learning by imitation requires the detection of a tutor's action demonstration and its relevant parts. Current approaches implicitly assume a unidirectional transfer of knowledge from tutor to learner. The presented work challenges this predominant assumption based on an extensive user study with an autonomously interacting robot. We show that by providing feedback, a robot learner influences the human tutor's movement demonstrations in the process of action learning. We argue that the robot's feedback strongly shapes how tutors signal what is relevant to an action and thus advocate a paradigm shift in robot action learning research toward truly interactive systems learning in and benefiting from interaction.

Figure 1. Action learning concept graphics.

(A) Unidirectional concept of current imitation learning approaches: The tutor demonstrates the action (white oval) according to his/her knowledge (upper hatched oval). The learner passively observes the action demonstration and learns the action. (B) Interactionist concept of learning: The tutor demonstrates the action (upper white oval) corresponding to his/her knowledge (upper hatched oval) emphasizing what is relevant to the action accordingly. The learner's level of understanding or knowledge of the action (lower hatched oval) is communicated by his/her feedback (lower white oval). This feedback directly influences the tutor's action demonstration. The tutor monitors the learner's feedback, builds hypotheses about the learner's understanding, and reacts by changing his/her demonstration accordingly as will be shown in this contribution.

Figure 2. Experimental setting.

Human tutor is sitting across from the robot learner at a table. Green marks on the table indicate the starting points for both the tutor's and the robot's demonstrations. Note that the individual in this Photograph ( Figure 2 ) has given written informed consent (as outlined in PLOS consent form) to publication of her photograph.

Figure 3. Items, task instructions, and example trajectories.

Figure 4. Technical setup.

For sensing the subject's movements, a Vicon system with 8 IR cameras (blue) was used. Additionally the object's position was tracked using a Polhemus Liberty magnetic-field-based tracking system (dark grey). This information was fed into the “robot control system” for generating appropriate robot behavior. For evaluating the study, additional data was recorded by 2 RGB cameras (green) directly synchronized with Vicon, 2 RGB cameras in the robot's head, 2 high-definition cameras (red) and an additional simple hand camera (light red) for showing interesting scenes during the interview after the study.

Figure 5. Definitions of measures of tutor behavior.

A visual depiction for the measure roundness can be found in Figure S4 of the Supporting Information.

Figure 6. Mean values for movement measures as a function of action knowledge condition.

Paired-sample t tests (df = 53) revealed significant differences between conditions for the presented movement measures, (A), velocity (t = −16.11, p<0.001), (B), pace (t = −6.81, p<0.001), (C), total length of motion pauses relative to action length (t = 2.47, p = 0.017), and (D), range (t = −5.1, p<0.001). Additionally, the measures action length, acceleration, and roundness revealed significance. Error bars represent standard errors.

Figure 7. Influence of robot feedback on tutors' action demonstrations.

(A and B), Mean values for number of repetitions as a function of turn-based feedback condition and action knowledge. A two-way repeated measures ANOVA (df = 40) revealed a significant interaction effect (Λ = 0.71, F = 16.23, p<0.001) between conditions, (A) a main effect for turn-based feedback (Λ = 0.22, F = 140.93, p<0.001), and a main effect for action knowledge (Λ = 0.71, F = 16.45, p<0.001). (B) A Scheffé post hoc comparison indicated that the turn-based feedback effect was greater in the manner-crucial action knowledge condition than in the goal-crucial condition. Error bars represent standard errors. (C and D) Mean values for movement measures as a function of online feedback. A one-way between subjects ANOVA (df = 53) revealed significant differences between gaze conditions for the presented movement measures, (C) action length (F = 4.18, p = 0.021) and (D) velocity (F = 7.302, p = 0.002). Scheffé tests used to make post hoc comparisons between conditions revealed that subjects in the social gaze condition demonstrated significantly slower and longer than in the static gaze condition (velocity: p = 0.002, action length: p = 0.049). The comparison between the other groups did not reveal any significant results. Additionally, significance was found for the measure acceleration. Error bars represent standard errors.