Speaker(s) : Elmar Rückert, Andrea d'Avella
Event : Neural Control of Movement Conference (NCM 2013)
Place : Puerto Rico, USA
Date and Time : 04/19/2013, 8:20 am
Abstract : One salient feature of human motor skill learning is the ability
similarities across related tasks. In biology a common hypothesis for
shared knowledge are muscle synergies or a coherent activation of a group of
muscles. Studies of human motor behavior have shown that a rich set of
motor skills can be generated by a combination of a small number of such
For motor skill learning on the other hand a popular approach are dynamic
movement primitives. This machine learning approach has many advantages, e.g.
it implements a stable attractor system that facilitates learning and it can
used in high-dimensional continuous spaces. However, it does not allow for
re-using shared knowledge. For each task an individual set of parameters has
We propose a novel movement primitive representation that implements muscle
synergies in the form of parametrized basis functions. For each task a
combination of such muscle synergies modulates a stable dynamical system.
allows for a compact representation of multiple motor skills while preserving
efficient learning in high-dimensional continuous systems. The dynamic
primitive approach can be modeled as special case in our formulation, where
discrete and rhythmic movements can be represented.
We demonstrate in complex humanoid walking experiments
that learning multiple skills
modelling walking patterns with different step heights
is more robust (i.e. good solutions are found reliably)
and more efficient (i.e. with fewer samples) compared to single-task
Furthermore, the proposed movement primitives are also used to learned muscle
excitation patterns for controlling a bio-mechanical model of
a human arm with six muscles.
Partners : - TUG - Institute for Theoretical Computer Science