Bi-manual Skills from Kinesthetic Teaching

type: 

Invited talk

Representation and Generalization of Bi-manual Skills from Kinesthetic Teaching

By Dr. Felix Reinhart (Bielefeld University, CoR-Lab)

Event: DGR-Tage 2012 (German Society of Robotics)

Date and Place: 6 September 2012, Technische Universität Berlin

Abstract

The talk presents a modular architecture for bi-manual skill
acquisition from kinesthetic teaching. In most previous work, the observed
data is directly given or transformed into a suitable task space, e.g. tool tip or
hand coordinates. In this task space, a compact representation of the move-
ment is learned. For bi-manual and even more complex movements involving
specific body coordination constraints, the talk argues that the represen-
tation of a skill requires more specifications and amounts to at least four
aspects: the task space movement, the expansion into bi-manual or other
body coordinates respecting particular features of the task, a modulation
and transformation binding the abstract primitive to a particular part of the
robot’s workspace for generalization, and a redundancy resolution scheme
which again respects particular task constraints. This talk shows that gen-
eralization of a learned skill strongly depends on the way how the different
partial representations are embedded. By means of comparative evaluation
of different architectural configurations (qualitatively and quantitatively), a
specific modulation of skill execution with respect to the embedding in the
robots workspace is identified to achieve optimal generalization performance
from few training samples. This representation scheme yields to very robust
teaching and execution of bi-manual skills for the humanoid robot iCub.

Recent Publications