A Method for Acquiring Body Movement Verbs for a Humanoid Robot through Physical Interaction with Humans

In this paper, we propose a language grounding method that relates verbs which imply body manipulation to motor angle patterns in a humanoid robot to make robots more entertaining. In established methods, verbs are represented by statistical models based on trajectories or motor patterns of a trajector. In our method we use a novel representation model that has six features including both trajector-reference point relationships and the trajector’s trajectory. By using this model, some verbs which do not depend on a trajectory, e.g. ”move the right hand close to the left hand.” are represented more adequately. In our language grounding method a humanoid robot generates abstract verb meanings independent of context. As input it uses sets of a user input textual command and a motor pattern. The motor pattern is taught using direct physical feedback resembling playing with child. We implemented the algorithm in a humanoid robot and conducted a verb acquisition experiment. As a result, four problematic verbs, ”place-on”, ”move-close-to”, ”move-away-from”, and ”touch-with” were acquired correctly.