How iCub Learns to Imitate Use of a Tool Quickly by Recycling the Past Knowledge Learnt During Drawing

Abstract

We show that the iCub humanoid can generalize motor skills by representing movements as shape primitives. After learning to draw shapes, the robot rapidly adapts to tool use tasks by recombining existing shape modules rather than learning from scratch.

Methodology

Motion trajectories are encoded via a passive motion paradigm into basis shapes. Tool trajectories are decomposed onto these bases. A mapping network aligns tool affordances to known shape modules, enabling quick motor command synthesis without full re-training. We show that the iCub humanoid can generalize motor skills by representing movements as shape primitives. After learning to draw shapes, the robot rapidly adapts to tool use tasks by recombining existing shape modules rather than learning from scratch.

Motion trajectories are encoded via a passive motion paradigm into basis shapes. Tool trajectories are decomposed onto these bases. A mapping network aligns tool affordances to known shape modules, enabling quick motor command synthesis without full re-training. We show that the iCub humanoid can generalize motor skills by representing movements as shape primitives. After learning to draw shapes, the robot rapidly adapts to tool use tasks by recombining existing shape modules rather than learning from scratch.