How people talk when teaching a robot

Abstract

We examine affective vocalizations provided by human teachers to robotic learners. In unscripted one-on-one interactions, participants provided vocal input to a robotic dinosaur as the robot selected toy buildings to knock down. We find that (1) people vary their vocal input depending on the learner's performance history, (2) people do not wait until a robotic learner completes an action before they provide input and (3) people naïvely and spontaneously use intensely affective vocalizations. Our findings suggest modifications may be needed to traditional machine learning models to better fit observed human tendencies. Our observations of human behavior contradict the popular assumptions made by machine learning algorithms (in particular, reinforcement learning) that the reward function is stationary and pathindependent for social learning interactions.We also propose an interaction taxonomy that describes three phases of a human-teacher's vocalizations: direction, spoken before an action is taken; guidance, spoken as the learner communicates an intended action; and feedback, spoken in response to a completed action.

How Do You Like Me in This: User Embodiment Preferences for Companion Agents

“…This evidence supports our hypothesis that the embodiment influences the user's mental model of the communicative capabilities of the agent. Moreover, as in [10], people tailored the way they communicate with the agent to account for performance of the agent so far. This is just the case for the r-agent group.…”

“…It is therefore important to gain insight into how naive users naturally try to communicate and provide input in different tasks. As in [10], this study recruited subjects that had not had much contact with robots or virtual characters. Apart from helping with the design of agent communication behaviours, this may help to predict what communication behaviours agents in different embodiments are likely to meet.…”

Robot feedback shapes the tutor’s presentation

“…In the future, such analyses and ideas need to be considered in conjunction with turn-based feedback (Vollmer et al 2010), and should be systematically integrated with sophisticated learning algorithms (Kim et al 2009) where the connection between the robot's feedback and actual progress realized as internal changes in the system can be investigated.…”

A formal framework for combining natural instruction and demonstration for end-user programming

“…Some approaches propose the use of situated instruction, where PbI is grounded in an example state [12]. Some recent work on combining PbD and PbI has focused on reinforcement learning frameworks [17,21,13]. However, there are no general approaches for incorporating broader forms of instruction and at the same time only require one or very few examples.…”