Interaction between biology and robotic and what we can learn from it

type: 

Invited talk
Speaker: Elmar Rückert - Graz University of Technology - TUG
Event: SLF Seminar Talk
Place: Santa Lucia Foundation - SLF, Italy
Date: 05/09/2012 12:00 pm

 

In AMARSi new compliant robots with dozens of actuators were constructed that
are able to perform more versatile movements. However, controlling such
high-dimensional dynamic systems remains an open problem. Facing this
challenge we are looking for alternative control strategies and movement
representations.

One recently developed movement representation is based on probabilistic
inference in learned graphical models. The parameters specify abstract goals
or features of the motor skill, i.e. the energy state of a system or
via-points. Traditional methods usually parameterize the shape of a movement.
Such methods cannot explain motor variability and it is very likely that a
direct superposition of learned movements will fail. However, the learned
parameters in the proposed formulation have a different interpretation. This
results in new and interesting features. Specifying how important it is to
fulfill a certain goal can control the motor variability of a system. A
superposition of motor skills is more likely to be successful as it is
performed in an abstract cost function space.

Movement primitives are meant to be elementary movement descriptions that
could be sequenced or superimposed in time. However, for most motor tasks,
i.e. different tennis swings individual movement primitives have to be
learned. However, all these tasks share some similarity and it is implausible
that humans memorize hundred thousands of movement primitives. It is more
likely that we exploit these similarities and learn a shared abstract
description. In recent studies we generalize the most widely used dynamic
movement primitives to implement shared abstract knowledge. This idea
originated from a concept called muscle synergies in biology. This novel
movement representation approach allows for learning multiple motor skills
simultaneously. Interestingly, by exploiting the linear relationship among
the previously learned tasks new motor skills can be generated without
learning.

One central question for us is how can humans infer so much about the world
given so little experience. In particular we are interested in model learning
approaches that can operate on multiple motor tasks in continuous spaces. A
potential hypothesis to solve this problem is to learn a meaningful prior
distribution over these multiple tasks that allows for generalizing abstract
knowledge. However, for high-dimensional problems these models usually need a
lot of data samples, which is in reality or in typical real robot experiments
rarely available. We therefore investigated hierarchical Bayesian models with
mixtures of linear drift priors that facilitate fast transfer and adaptation
of learned movement patterns to new tasks. The model can make accurate
predictions already after observing few samples and outperforms standard
approaches.

 

 

 

Recent Publications