Skill refinement through cerebellar learning and human haptic feedback: An iCub learning to paint experiment

Abstract

This article presents a model of the control of hand movements learned from human imprecise feedback. The setup consists of an iCub humanoid robot able to paint with a pencil on an interactive display. It learns to avoid an abstract boundary, which is drawn on the table and that the robot can perceive. Before learning, the robot does not know that it has to paint only inside the boundary. This is instead learned from human feedback provided whenever the pencil goes outside of the shape. We use a neurocomputational model of the cerebellum, which learns to anticipate the human feedback from the perception of the previously meaningless shape boundaries. We show how this biologically-grounded adaptive mechanism, which is plausible in terms of human infant development, allows the learning of a precise painting behavior, using the perception of the shape boundaries as a predictive signal of aversive stimuli. This mechanism can be generalized to any kind of task where some aversive feedback can be considered correlated with available sensory cues. Consequently, the model allows a human teaching to a robot any kind of complex task, as long as the human knows how to provide consistent feedback and the robot has available the sufficient sensory cues. This becomes specially relevant for educating a robot on social behaviour or bringing it to a socially rich environment.