Goal recognition using temporal emphasis

The question of what to imitate is pivotal for imitation learning in robotics. When the robot's tutor is a naive user, it is very difficult for the embodied agent to account for the unpredictability of the tutor's behaviour. Preliminary results from a previous study suggested that the phenomenon of temporal emphasis, i.e., that tutors tend to keep the goal state of the demonstrated task stationary longer than the sub-states, can be used to recognise that task. In the present paper, the previous study is expanded and the existence of the phenomenon is investigated further. An improved experimental setup, using the iCub humanoid robot and naive users, was implemented. Analysis of the data showed that the phenomenon was detected in the majority of the cases, with a strongly significant result. In the few cases that the end state was not the one with the longest time span, it was a borderline second. Then, a very simple algorithm using a single binary criterion was used to show that the phenomenon exists and can be detected easily. That leads to the argument that humans may also be able to detect this phenomenon and use it for recognizing, as learners or emphasizing and teaching as tutors, the end goal, at least for tasks with clear and separate sub-goal sequences. A robot that implements this behavior could be able to perform better both as a tutor and as a learner when interacting with naive users.