human subject performing movements, interacting with a computer-controlled
haptic interface producing force fields during motion
Impedance Adaptive Controller of the Human Arm
This project is done together with Rieko Osu and Mitsuo
Kawato, Erato and ATR,
Japan, Ted Milner
and Dave Franklin, Simon
Fraser University, Canada, and Yasuharu Koike, Tokyo
Institute of Technology. We infer the neuro-mechanical control of the
human arm by performing experiments using EMG, computer-controlled haptic
and visual interfaces, and by modelling this control using biomechanics,
control theory and artificial neural networks.
Shadmehr&Mussa-Ivaldi [1] have analyzed motor adaptation
when arm movements are repeated in a velocity dependent force field. We
have measured stiffness [2] in these conditions and shown that in this
case stiffness depends mainly on the torque produced by muscles [3], similar
to the postural case [4]. This would mean that endpoint stiffness is just
is a byproduct of muscle mechanics. However, we were recently able to show
a highly significant increase of stiffness in a divergent elastic force
field [5]. This means that humans are able to control the endpoint impedance
to perform stable movements in unstable environment dynamics. This result
was predicted by Hogan in 1985 [6], but previous experiments to confirm
it all failed.
In summary, the human arm controller seems to be composed
of a learning controller with Internal Dynamic Model (IDM) and adaptive
impedance. The learning mechanims seem to be completely different in both
cases: The IDM would be formed automatically, while adaptation of impedance
still requires attention even after learning. We are currently analyzing
the differences between these two kind of learning using EMG. We will design
psychophysiological experiments to find out the features of these two distinct
learning processes, and the role both have for interacting with arbitrary
dynamic environments. While learning of an IDM is probably processed by
supervised learning in the Cerebellum, adaptation of the endpoint stiffness
may be performed by reinforcement learning in the Basal Ganglia and Cerebral
Cortex.
We also develop two kinds of computational models:
-
We are developing a controller with hybrid IDM/Impedance
learning [7] that we will implement on lightweight manipulators.
-
We are using artificial neural networks [8] to identify the
function from the EMG to the torque and the function from the EMG to endpoint
stiffness.
References
[1] R. Shadmehr and F.A. Mussa-Ivaldi (1994) Adaptive
Representation of Dynamics During Learning of a Motor Task, Journal of
Neuroscience 14 (5)
[2] E. Burdet, R. Osu, D. Franklin, T.E. Milner
and M. Kawato (1999) A Method for Measuring Hand Stiffness during Multi-joint
Arm Movements. Subm. to Journal of Biomechanics
[3] E. Burdet, R. Osu, T.E. Milner and M. Kawato
(2000) Measuring stiffness during arm movements in various dynamic
environments, in preparation. Also: Proc. of the 1999 ASME Annual Symposium
on Haptic Interfaces and Virtual Environments for Teleoperator Systems,
Nashville, USA
[4] H. Gomi and R.Osu (1998) Task-Dependent
viscoelasticity of human multijoint arm and Its Spatial Characteristics
for Interaction with Environments, The Journal of Neuroscience 18(21):
8965-8978.
[5] E. Burdet, R. Osu, T.E. Milner and M. Kawato
(2000) Voluntary Adaptation of Muscle Impedance to Unstable Dynamics
(in preparation)
[6] N. Hogan (1985) The Mechanics of Multi-joint
Posture and Movement Control, Biological Cybernetics 52: 315-331
[7] E. Burdet, KP Tee, CM Chew, J Peters and V
Loo (2001) Hybrid IDM/Impedance in Human Movements, Proc. of The First
International Symposium on Measurement, Analysis and Modeling of Human
Functions.
[8] Y. Koike and M. Kawato (1995) Estimation
of dynamic joint torques and trajectory formation from surface electromyography
signals using a neural network model. Biological Cybernetics 73: 291-300
back
to Etienne BURDET homepage
March 2001, by Etienne Burdet