Human motor control

Gregor Schöner
Movement generation in animals
movement generation adapted to and
directed at a sensed environment is the core
of animal experience… and a key
evolutionary factor
=> animal are amazing autonomous
movement machines..
=> the brain is strongly organized around
movement generation… (the basis of a
tradition of thought called “embodied
cognition”)
 Human movement

humans are particularly skilled at movement

directed at objects
manipulation, compliant acting on objects

humans are particularly flexible, versatile in

their movement generation
while some other animals excel at particular specialized
motor acts
A landscape of human movement
looking: eye and head movements (gaze)
orienting the body in space, upright stance
legged locomotion
navigation
steering
reach, grasp, manipulate
sequences of motor acts
speech articulatory movement
Qualities of human movement

involuntary (reflexive)
automatic/habitual (requires little attention)
voluntary/intentional
Qualities of human movement

whole body movements in space

movements of hand/arm or other
extremities while anchored in space
Qualities of human movement

rhythmic
discrete (in time)
 Textbooks

David Rosenbaum: Human motor control,

Academic Press, 2009 (2nd edition)
Richard A Schmidt, Timothy D Lee: Motor
Control and Learning, Human Kinematics,
2011 (5th edition)
James Tresilian: Sensorimotor Control &
Learning. Palgrave McMillan, 2012
 What is entailed in generating an
object-oriented movement?
scene and object perception
movement preparation
movement
movement initiation and preparation
termination
movement timing and
coordination
motor control
degree of freedom problem control
timing
=> spans perception, cognition
and control
 What is entailed in generating an
object-oriented movement?

tightly interconnected movement

processes preparation

which this is why

movement is so hard to
study
critical to understand
integration control
timing
 isual s stem

A neural frontal brain

motor
goals
isual corte

visual searc attention

architecture parietal corte

of reaching
sequence s atial re resentation
generation eed
ac
premotor corte
basal ganglia motor corte cerebellum

movement initiation movement re aration

termination

movement timing

ac
cou ling
neuronal d namics o
decou ling
descending activation

spinal cor

re e

sensori motor perip er

muscle d namics
iomec anics ro rioce tion

[adapted from: Martin, Scholz, Schöner.

Neural Computation 21, 1371–1414 (2009] mo ement
 isual s stem
frontal brain isual corte
motor
goals visual searc attention

parietal corte

sequence s atial re resentation

generation eed
ac
premotor corte
basal ganglia motor corte cerebellum

the perception and movement initiation

termination
movement re aration

cognition on which
object-oriented action is movement timing

ac
based…. topic of my neuronal d namics o
cou ling

course in the WS
decou ling
descending activation

spinal cor

re e

sensori motor perip er

muscle d namics
iomec anics ro rioce tion

mo ement
 lower activation level. peaks are bui
Scene perception
The detection decision of DFs translates
this input into a macroscopic decision, whether
input from t
slice input fro
a visual region is interesting enough for further of the boost
inspection (i.e., it contains a scene object) or can solutions and
be ignored as background. Here, the retinal space scene space
field (see Figure  9.1a) encodes these extracted positions cha
neural fields… dynamic field theory
foreground blobs of objects in a neural way,
containing multiple peaks at the same time in a
the new loca
information
multi-item regime, which represent interesting tity. A  secon

u
y x

FIGURE 9.7: Scanning. The retinal level contains a multipeak retinal space f

each salient object location in the image seen on the left. Its activation feeds
on the right. The leftmost object is selected, the two other objects are supp
[Zibner,
inhibitory input from theFaubel: In DFTfields
scene space-feature Primer
and the(2016)]
right one throug
 Movement preparation
coordinate transform into initial position of hand

activation

0
movement
amplitude
mo
v nt
am emen e m e
movement plit
ude t mov on
direction irec ti
d

[Erlhagen Schöner, Psych Rev 2002]

 Sequence generation

every action is
represented as a stable
vement to be executed, and d = x −x
target is the
real
activation sate in an
maining distance.
To summarize, a single timed movement consists of
“intentional
ee separate field”
behaviors: the postural, movement, and
date behavior. In order to function properly, these be-
iors must be activated and deactivated in the correct
that predicts its
uence: the initial position must be memorized before
rting to move and the movement has to suppress the
“condition of
tural behavior. The necessity of organizing behaviors
ime becomes even more apparent when building entire
satisfaction”
hitectures based on discrete behaviors.
The framework for behavioral organization is based on
T, which we now briefly review.
instabilities drive the
Dynamic Field Theory
transition
Dynamic Field from
Theory (DFT) [16] is one
a neural variant Fig. 2. Elementary behavior (EB) in Dynamic Field Theory.
the attractor dynamics approach. We use it here as Each EB consists of two parts: the intention represents the desired
intention
integrating to another
framework between the low level sensory-
change of the EB in the world, while the condition of satisfaction
(CoS) represents the sensory signal expected for the successful
tor streams of the robot and the higher level cognitive completion of the EB.
ctions of the model, for instance its perceptual repre-
tations and its organization of behaviors.
[Sandamirskaya,
Within DFT, dynamic neural fields Zibner,
(DNFs)Schneegans,
are used to Schöner:
Fig. 2). New Ideas
While the inintention
Psychology (2013)]
node simply determines
 isual s stem
frontal brain isual corte
motor
goals visual searc attention

parietal corte

sequence s atial re resentation

generation eed
ac
premotor corte
basal ganglia motor corte cerebellum

movement initiation movement re aration

termination

timing and coordination: movement timing

Lecture 7/Exercise 6 ac
cou ling
neuronal d namics o
decou ling
descending activation

spinal cor

re e

sensori motor perip er

muscle d namics
iomec anics ro rioce tion

mo ement
 isual s stem
frontal brain isual corte
motor
goals visual searc attention

parietal corte

sequence s atial re resentation

generation eed
ac
premotor corte
basal ganglia motor corte cerebellum

movement initiation movement re aration

termination

degree of freedom
problem: Lecture 6/ movement timing

ac
Exercise 5 neuronal d namics o
cou ling

decou ling
descending activation

spinal cor

re e

sensori motor perip er

muscle d namics
iomec anics ro rioce tion

mo ement
 isual s stem
frontal brain isual corte
motor
goals visual searc attention

parietal corte

sequence s atial re resentation

generation eed
ac
premotor corte
basal ganglia motor corte cerebellum

movement initiation movement re aration

termination

human motor control:

how forces are movement timing

ac
generated and regulated neuronal d namics o
cou ling

decou ling
descending activation

spinal cor

re e

sensori motor perip er

muscle d namics
iomec anics ro rioce tion

mo ement
 Human motor control

consider a single DoF, the elbow angle..

in a fixed posture
 Posture is controlled

the elbow does not behave like a

passive mechanical system with a free
joint at the elbow: J ✓¨ = 0
where J is inertial moment of
forearm (if upper arm is held fixed)
Instead, the elbow resists, when
pushed => there is active control=
stabilization of the joint

=>experiment
 Posture is controlled

human effectors are not very

stiff…. unlike robotic actuators
stiffness expressed in
Eigenfrequency => time scale ~
of the same order of magnitude
as movement time
=> human movement is highly
compliant…
The problem of human motor control

=> leads to major problems in

human motor control: how to
make a soft spring move fast to
precisely reach a target and
softly stop there…
 The “mass spring” model

a simplified macroscopic force applied

description
of the mechanics of the muscles

and the reflex control of the muscles

the invariant characteristic

 The mass-spring model

elastic component: proportional to

position
viscous component: resistance depends
on joint velocity

J ✓¨ = k(✓ ) µ✓˙

active torques generated by the muscle

 Agonist-antagonist action

muscles only pull, so the force applied

invariant characteristic comes
antagonist
from pairs of muscle groups
one lambda per muscle
agonist
co-contraction varies stiffness
 Nrm/rad. The other three terms in the stiffness matrix follo

Stiffness
roughly the same form but show a less pronounced decrea
in the middle of the movement. The elbow term, Ree increas
from Ç5 Nrm/rad at movement start to 20–25 Nrm/r
during movement, and the two double-joint terms, Rse a
Res , increase from Ç2 Nrm/rad at movement start to Ç
10 Nrm/rad during movement. Ree , Res , and Rse all decrea

the stiffness, k, can be

measured from
perturbations
the viscosity “mu” is
more difficult to
determine
J ✓¨ = k(✓ ) µ✓˙
 Muscle dynamics
n et al. increasing level of
muscle activation

passive
muscle

[Buchanan et al. 2014]

 Muscle dynamics
force generated depends on speed of
lengthening / shortening
less force for shortening
B. Force properties
more for stretching

1 1
shortening stretching

0 0

[Song 2017]
 Muscle dynamics
Hill type models
A. Muscle-tendon unit

PE
SE
CE

BE

[Song 2017]
Neural basis of invariant characteristic:
stretch reflex

alpha-
gamma
reflex loop
generates
the stretch
reflex

[Kandel, Schartz, Jessell, Fig. 37-11]

[Kandel, Schartz, Jessell, Fig. 37-11]
 spinal cord: reflex loops
the stretch reflex acts as a negative feedback loop

37-12

[Kandel, Schartz, Jessell, Fig. 31-12]

spinal cord: coordination

Ia inhibitory interneuron
mediates reciprocal
innervation in stretch
reflex, leading to
automatic relaxation of
antagonist on activation
of agonist

[Kandel, Schartz, Jessell, Fig. 38-2]

 The name equilibrium-point hypothesis emphasizes the idea that movements of ph
ical objects may be viewed as transitions between equilibrium states (equilibri

Reflex model
points). An equilibrium state, by definition, is a state in which the resultant force act
on the object is zero. For example, a muscle acting against an external load may be
an equilibrium state when muscle force is equal in magnitude and directed against
load force. So, if a muscle is in equilibrium with external forces, its state may be ch
acterized
254 with two variables, length and force. These two variables
Biomechanics and Motorform
Controlthe equi

rium point for the muscle on the lengtheforce plane (Figure 12.3).
There is considerable confusion due to the existence of two versions of the
hypothesis addressed as the a-model and the l-model. The original Feldman’s hypo
monotonic esis (the l-model) assumes that the neural control of a muscle can be adequat
described as setting only one variable, threshold of the tonic stretch reflex (
relationship force- Figure 12.3 illustrates the control of a muscle according to the l-model. Setting a va
of l defines a range of muscle length values (those larger than l) associated w
length muscle activation—to the right of l in Figure 12.3. If muscle length is smaller th
l, no activation is seen. The muscle shows larger activation levels and larger for
for larger deviations of its length from l. The dependence of muscle force on
reflex threshold can difference between its length and l is addressed as the muscle invariant characteri
Figure 12.5 (A) A scheme of three inputs into an alpha-motoneuron (A e afferent; C1 and
be varied by C2 e central). (B) An illustration of the three methods of control of a neuron. C1 and C2 define
the effective distance to the threshold (DTh). The afferent input (A) is assumed to define

descending
the ramp-like change in the membrane potential (V). AP show the moments of action
potential generation.

activation signals action potential). In this case, the frequency of firing of the neuron will not depend on
the afferent input. In contrast, if the central inputs depolarize the neuronal membrane
below the threshold, activity of the neuron will depend on the A input. In other words,
the neuron would behave as an adaptive element, changing its activation as a function
of the state of peripheral structures, from which it receives the afferent input.
Figure 12.5(B) illustrates the membrane potential and its changes under the action
of the three inputs shown in Figure 12.5(A). For simplicity, assume that the afferent
input leads to a ramp-like change in the membrane potential (dashed ramp lines).
[Latash, Zatsiorsky, 2016] One of the central inputs (C1) induces a steady shift of the membrane potential
(V) toward depolarization. The other input (C2) can change the threshold value itself.
Clearly the frequency of firing of the neuron will be a function of all three inputs: C1
Figure
and C212.3 Setting
define a valuedistance
the effective of l defines a range
from the of muscle
steady-state levellength values (those
of depolarization to thelarger than
Movement entails change of posture

the threshold lengths of the muscles must be shifted

during movement so that after the movement, the
postural state exists around a new combination of
muscle lengths (<=> joint configuration)

force
equilibrium
point

joint angle,
Movement entails change of posture

many models account for movement in terms of muscle

activation/desired torques….
=> the shift of the EP is the single most overlooked fact
in control models of movement generation

force
equilibrium
point

joint angle,
 Does the “motor command”
specify force/torque?
Not necessarily ..
because the same descendent neural command
generates different levels of force depending on the
initial length of the muscle
force
equilibrium
point

joint angle,
 Virtual trajectory

Shifting the threshold lengths is necessary, but is it also

sufficient?
first answer: yes… simple ramp-like trajectories of the
combined threshold lengths of the antagonistic muscles
(“r” command ~ virtual trajectory) may model movement
force
equilibrium
point

joint angle,
 Pilon, Feldman, 2006

actual
first50 A yes… simple
answer: trajectory
ramp-like trajectories of the “r”
command (“virtual trajectories”) shift the equilibrium point
smoothly C
in time… .
Elbow angle (°)

θ
e.g.0Pilon, Feldman, 2006
2
R Rc = 0.99, p < 0.01
ErrRMS = 1.7881 °
–50 θ
virtual
trajectory
–100 d = 30 ms κ = 0.006
 Shifting the equilibrium point is
necessary, but is it also sufficient?

such simple ramp-like trajectories of the “r”

command (“virtual trajectories”) may be
sufficient when movements
are sufficiently slow

interaction torques/mechanical conditions unchallenging

but is this generally true?

(answer: no)
 Limit case: velocity dependent force field
after adapting to a velocity dependent force field the
hand reproduces the “natural” path, but must generate
compensatory forces on the way
The Journal

[Shadmehr, Mussa-Ivaldi, 1994]

The Journal of Neuroscience, May 1994, 74(5) 3213

3210 Shadmehr and Mussa-lvaldi * Representation of a Learned Motor Task

1x
left workspace right workspace

150 1

100
100
0.5
5c F
50
5
.8
0 B0 0
s
A
z
-50m
-50 = -0.5
-100
-100
Figure2. Configurations of a model two-joint arm, representing typical
kinematics of the human arm, at two workspace locations where reach- A
-150
-’ i
:.
-.
=. Figu
ing movements were performed. Typical shoulder and elbow angles at -a jecto
-150 workspaces were 15” and 100” at right and 60” and 145” at
these two -0.5 0 0.5 1 l
‘*
50 100 1 -150 -100 -50 0 50 100 150 : the
left, using coordinates-100defined -50
in Figure 01. f
%
plott
Hand x-velocity (m/s)
Displacement (mm) Displacement (mm) ‘31 third
0 traje
domly Figure
chosen6. from
Typical
the hand
set (O”,trajectories at theand
45”,. . ., 315”) rightatworkspace
a distance inofa 10 null Figure 7. Performance during initial exposure to a force field. Shown _I 5cm feed
cm was force field duringAfter
presented. no-visual feedback
the subject hadconditions.
moved to Dotsthe are 10 msec
target, apart.
the next are hand trajectories to targets at the right workspace while moving in
center-out movements
target, again chosen at a random direction and at 10 cm, was presented.
A target set consisted of 250 such sequential reaching movements. All
the force field shown in Figure 3. Movements originate at the center.
velocity dependent
All trajectories shown are under no-visual feedback condition. Dots are a reaching movement, with practice the subjects tended to con- practiced in the force field. This
targetscentripetal
were keptforceswith that
before force-field
make
in the up the of
confines G matrix
the 15 canx 15be cm
derived from the
workspace. 10 msec apart.
center-out movements computation of a correlation c
verge upon this straight-line trajectory. This recovery of the

6. Asforce-field = zero at rest

inertia represented
The targets tensor; cf. Slotine and Li, 1991,
a pseudorandom p 400). For example, the dif-
walk. jectories (see Appendix). We fou
original unperturbed response constitutes a clear example of an
ferential
In some equation
cases, describing the dynamics
the manipulandum of the armto and
was programmed the controller
produce forces 50
on theforhand
movements
adaptation
These forces, indicated
in theasforce
of the subject
by the
field of performed
the subject
vectorf,
Equation 1reaching
were computed
were movements.
as a function of
observed in previous reports (Morasso, 1981; Flash and
E- 1985), the hand path was essentially along a straight
Hogan,
adaptive behavior.
at four stages duringbyincreased
between a trajectory in the null
with the amount of
Further evidence of motor adaptation is offered by the sig-
kd 4 + G(a 4) + JW B J(q) 4 = C(a 4. t). (11) line .c.to the target. The velocity profile (see Fig. 1OA) had one the subject in the force field.
nificant change that occurred in the hand velocity profile at the
force-field adaptationefficients
the velocity of the hand:
where Cis defined in Equation 9. Values for joint stiffness and viscosity peak,E with approximately equal times spent to accelerate and onset of exposure to the force field, and after completion of the
for trajectories perform
(K and IJ’)were chosen based f= onB%,measurements of Mussa-Ivaldi et(1)al. E 0. hand.
decelerate Figure 11. Remarkably, all th
practice trials. Figure 10A shows the hand tangential velocity
where(1985)
X wasand
the Tsuji
hand and Goto (1994).
velocity The desired
vector, and B was atrajectory
constant q*(t)
matrixwas 8 ourthesubjects
m
Once were familiar with the task of reaching traces obtained when the hand was moving in a null field (cor- monotonic evolution of the cor
assumedviscosity
representing to be minimum jerk in hand-based
of the imposed environment coordinates lasting
in end-point 0.65
coor-
sec. Values used for these variables are summarized in Table 1. .-:: the null field, we began to introduce a force field in
within responding to the hand position traces of Fig. 6). Consistent Our subjects did not seem to
 Shifting the equilibrium point is
necessary, but is it also sufficient?
=> r-command must still shift from initial to final
posture, but must also generate the forces to
compensate for the force field during the
movement
that probably takes the form of non-monotonic,
“complex” time courses…
are such temporally complex (e.g,. non-monotonic)
r-commands necessary during unperturbed
movement
Estimating the descending signal
(~virtual trajectory)
descending
activation oint angle

sensor
activation

sensor time
activation muscle
muscle
activation activation
descending
activation
length o agonist

oint angle time

(1) Estimate the descending activation by
inverting a neuro-muscular model

simplified version Hill type mode:[Gribble, Ostry et al.,

98] .. 4 muscles

[Hummert, Zhang, Schöner]

 articular elbow joint muscles (red), two monoarticular shou
Kinematics
nd two biarticular muscles that span elbow and shoulder joi

two joint limb with 4 muscles

ng of Muscle Lengths
= 2 pairs of mono-articulatory m.

span neglect: bi-articulatory

along joints musclesor contract to change the ang
and stretch
hus induce a movement. Thus, the muscles resting length an
ion determine the actual length of a muscle. We modeled t
first-order
musclepolynomial
length link todepending
joint angles on the elbow and the should

li = ci + c0i,s ✓s + c0i,e ✓e i 2 [1, 6]. (3.

nd c0i indicate the resting length and the moment arms of t
es and can be found in table 3.1. The resting lengths are tak
Woo] and the moment arms from [Kistemaker et al., 2006]. Equ
 to the rate of⇢2length change 4.52 l˙i of the muscle.
Equation
⇢3 ⌧ 2 M̈ +4.52
of motions:2⌧ Ṁ + M = [Gribble
[2] M̃ et al 98]+
shoulder momentEach armof[m] the⇢4 six muscles 3.87is elbow Ai =moment
modeled [lseparately
i
arm ˙ [m]
and
i + µ · li ]
in t
GribbleHere, constant
M̃i is thetion level
actual arm
of amuscle
muscle
⇢5 force Gribble
is proportional
obtained
1.29 to the length
in equation constant
3.4 and Mi le
di↵eren is
mef muscle force. -0.03
with length
The value
·0.2 li and
⇢6 ⌧the
for -0.014length
is taken3.87-from
0.0079202
of the descending
[Gribble (✓et signal -0.0
2 al., 1998] and i. F

mee that the response to ·0.2

thethe
of
Ungraded
0.03 rate offorce:
muscle length change
corresponds
0.025 l˙i +of
to
-0.0043202 thethe muscle.
x,
empirical
✓ if x findings.
> 0 0.0T
Table 3.3: PCSA ⌧ canofbe
thefound
six modeled
in table 3.4. muscle groups. The [x]ability = of force generat
2 .
msf -0.03 -0.03 0, if x -0.023 0 ˙ +l
can be calculated Therebyismultiplying
strong evidence the PCSA for a sigmoidal
with N cm 2
relation
. The Aibetween
=values
[li muscle
are µ · lforce
i + taken i ] fr
mse et0.03
[Winters al.,lengthening
2012] and0.03shortening [Joyce et al., 1969]. Furthermore we0.023 assum
bef -0.03 force linearlyThe parameter
with
Muscle
depending µ relates
activation:
on the -0.016the
di↵erence musclesthe
- between
0.01146 threshold
✓2( actual muscle lengthlent
muscle activation
Figure from ingcifactor the representing the strength of proprioceptive
x, if Eachx>0 f
bee 2.2:0.03 The mechanism
resting length ofof force generation
muscle. The0.03 as- used
0.00636
resulting by
force +Gribble.
✓F yields:
[x] =2 force developm.
i
The descending
Calcium
2.2 represents
box inverseone command
kinetics acts
dynamics like a low pass filter
equation of the msucle model. Picture taken on the muscle l0,
3 if x from
0
[Gribble et
[Gribble et al.,
al., 1998]
1998] and is thus modeled by a critically damped second order
forcespass
from
Equation Table
muscle
of 2.4:
motion Muscle
activation The
moment
Calcium
actual
F = M
arm
dynamics:
force
[(f + of
Ha· for
f muscle
the M̃
arctan(f
six
+ is
f
lumped˙
approximated
· l )] +
muscles.
k(l ci ).by an
filter with a single parameter i ⌧:
The parameter
i 1 2 3 i 4 i
µ relates the muscles threshold length
i

Mi is the graded ing force from

factor
¨ equation
representing
˙ ˙ 3.5.
the The
strengthfactorsoff1proprioceptive
, fsA2 , if3 , f4 and
T =⌧ IM̈2 · ✓++ 2⌧ C(✓,
Ṁ ✓)+ ✓M = M̃ M̃i = ⇢i · (e (3)l4 1). cs,i (3
In the equation of motion
from [Gribble et al., the 1998]torque
and thenis used
adjusted to calculate
such that they the accelera
reproduc
¨ from the external
✓Here, of [Gribble Force-velocity
torque,
et al., which
1998]. is
The
10 here relationship
set
adjusted to zero,
parameters and
the passive
coriolis
can be found in sti
forcetabl
C
M̃ is the actual The
muscle factor
force s is constant
obtained in for all
equation muscles
3.4 andand
M is
is estimate
the grad
the torques
i
inertia from
matrix forces
I [GomiTheT actual
i =
and H i⇤force
Kawato,F 1996].
i (4)
i of a muscle M̃i is approximated by a
muscle force.
✓ The ◆ value cat’s
0
for ⌧ gastrocnemius
is taken from muscle [Feldman
[Gribble
Parameter et Value andand
al., 1998] Orlovsky,
chosen19 s
thatH~the c
responseiE the muscles
of the muscle ability of
corresponds to force-generation
thef1empirical and
findings. issAassumed
The value
ij = ¨ , -˙ M̃
0.82 = ⇢ · (e i
1).
⌧ canmotion cfrom torques (2
0 1 i i
be found iE in table 3.4. ✓ = I (T
physiological Text fC
cross-sectional ✓) (PCSA). Thus, every
2 , -area 0.57
since ciF =◆ ciE
0 0
There is ✓strong evidence value a [Winters
forThe factor setisrelation
sigmoidal al., 2012].
constant , -The
f3for
between all exponential
muscles
muscle force
0.43 andrelation
isand b
estimat
mus
TheF~i Coriolis
=
lengthening
F iF force is a force that acts in a direction perpendicular to
and shortening cle cat’s
[Joyce gastrocnemius
activation with
et al., 1969]. fmuscle
constant s and
4 ,Furthermore
s/m [Feldman
⇢ 58 weand
values Orlovsky,
is shown
assume a in 1
[F
pass
rotation FiE which [Hummert,
axis, is on
the axis Zhang,
aligned Schöner]
with
force linearly depending the
Thethe muscles
used values
di↵erence ability
of ofthe
s and
between ⌧⇢,arm
the scan inbethis
0.015
force-generation
actual case.
found
muscle and in The
istable
length Cor
assumed
3.3
and
k, - 17.35
 Comparing data to movements predicted data: red/magenta
from estimated descending activation (for the two directions)
perimental paradigm model: green/cyan
experimental data for the estimation of descending commands, we slow fast
0.55 & Flash, 1982].
d a standard movement experiment in the manner of [Hollerbach
d at naturalistic movements, restraining movement to planar, two-
on by instruction rather than constraining the arm mechanically by
landum. Twelve participants performed the eight movements illus-
Figure 1. (A thirteenth participant was excluded due to failure to
enough valid trials, see below.) All participants (8 female, 4 male,
0.5 and
= 25.67 years, SD= 3.80, age range: 22-35) were right handed
informed and written consent. The participants were compensated
Euro for the one hour experimental session.

4
Y-coordinate
1
2
0.45
6 8

0.4

5 3 7
0.4

0.2

0.35

0 S 0.2

0.3
Top view of the experimental set-up.
-0.2
The participant is seated in
0 0.2 -0.2 0 0.2
horizontal monitor with the center of his shoulder at a distance X-coordinate [m]
o the monitor. Participants moved their right arm between targets
on the monitor. Infrared markers (red asterisks) were attached to
 green: Results
muscle 0.02
MSF MSE
0.01
MEF MEE

activation 0 0

vation -0.02 -0.01

o blue:
0.02

zero 0
0.02
0
descending
vement
-0.02
-0.02
activation 0.02
0.02
central 0 0
-0.02

shifted
-0.02
0.02 0.02

Both 0 0
activation [rad]

-0.02 -0.02

more
-3 -3
#10 #10
0.02
5

fast
0 0
-5
-0.02

id line)
#10 -3 #10 -3
5 0.02

0 0
slow -5 -0.02

(dotted 0.02
0
0.02

0
-0.02 -0.02

solid: fast 0.02 0.02

dashed: 0
-0.02
0

-0.02
slow 0 50 100 0 50 100 0
time [sec]
50 100 0 50 100
resembling Latash's[4] N-shape in the fast.
time course of descending activation
The virtual trajectory changes gradually with an increasing
cocontraction (C= {0.2, 0.4, 0.6, 1.0}) and the N-shape is
… as a virtual trajectory
most pronounced for smaller C-commands

NE = 0.0953 rad NE = 0.165 rad

NS = 0.0565 rad NS = 0.095 rad

 how detailed the description
combinations of the movement
of the individual muscles. should be. In the scope of thi
e structure
nce ~ (t). The of
These descending
boundary conditions guarantee ~ that the
~ desired
(2) Estimate
state vector. Tominimal
boundary
X~f the
0 to for
conditions
is executed
were
descending
divided
in aat biologically
state vector
into
the beginning
produce realistic
three
plausible
classes.
second class of boundary conditions are the✓(t
range
and at the
movements,
0 )
activation
The
✓ first
endofof X(t)
~ ) ✓~ angle
✓(t
joint
0,class
start =and
upper arebou
lower
~the movement.
and
and joint ang
= 0.
co
T
f final
n of appropriate
condition is, boundary
that to
are restricted conditions
thea joints for
starta movement
shouldplausible
biologically the initialdepends
atrange: onend at
state and
escription Furthermore,
state: of the movement the velocity
should be. In at the
the
“minimal” change of descending~ activation scope
~
of The
beginning
this Virtu
ofproject
the movement
movement should be zero: ✓(t) < ✓max ,
ns were divided into three classes. The first class ~˙areAt
imally conditions
~˙ ~˙ ✓(t0 ) = 0,any mome
or at the beginning and at the end ~ of) the✓(t)
~ movement.
<=✓max . The first
of the ✓(t 0 ✓ start 0,
~˙ fatdetermined fro
the joints should start at the initial state and end ✓(t ) =the
0. final
nd u This not only prevents unrealistic ~ f ) angles
✓(t ✓~final =and
0. anglejointvelocities but also
torques
The third condition of this classforces.
is, that the joint angle accelera
om an to
he graded
muscle
bring
force
Furthermore, of the . velocity
lengths,
about
each
and
length
the
muscle
at the
changes
movement
but
at
end
and
oflimits
the the
hence
the space
beginning
movements of the and
ofmovement
shouldpossible .bounds
zero:and
be lower at thefor
en
The third class of boundary conditions are upper and
individual
tion movement muscles.
at signals should be zero:
~ . Within the given model,˙ control signals are defined as equilibr
~ ~ ~¨
f boundary length ✓(t
conditions
and 0 )should
✓start
are the =upper
0, always
therefore ~ lower
and✓(t ) = 0,
0have bounds ✓(ton
positive ) =the
0values. In in moveq
0, (3.13)
addition
eds.
oduce realistic
shows movements,
that
~✓(tf )if the joint angle
activation
~✓final = 0. A and
is ~˙ joint
zero, a angle velocity
growing ¨
~ does not have an e↵
0.2(3.14)
biologicallyactivation.
plausible Thus, ✓(t ) =
range:upper and lower bounds for ~ were
f 0. ~
✓(t ) = 0.
f introduced:
velocity
The at thecondition
third beginning
~✓(t) The ~
of of acceleration
this
angle the
classmovement
is, that and
theno
does joint at theappear
angle
explicitly end of the state
acceleration
-0.2 in the at theve
b
< the
✓maxmovements
,  ~ (t) be t 2 [tthe
0 , tf ].
and at the end of
be zero: extensor
depended should
on it.minEquation zero:
max
3.10 describes -0.2
(3.19) angle acceleratio
˙ ~˙ ~˙
~✓(t) ~˙ ~ and
✓(t
< ✓
angle) =
0 max
The listed boundary . 0,
velocity ✓, whichconstrain
conditions ~¨ is directlythe
contained in (3.15)
state-vector X ~ on
X(t) the tj
such
0 torque arises as a ✓(t0 ) = 0,
consequence
end-e↵ector~˙ moves from an start state X ~of
0
interacting
to a muscle
0.2
predefined forcesstate
target whi
ents unrealistic angles
thefand
✓(t ) = angle
0.
graded velocities
force (see sec. but
3.1).also
It isunrealistic (3.16)
important to note that con
biomechanical limits. ¨
[Ramadan,
gth changes Hummert,
and hence ✓(t
forces. Jokeit, Schöner ) = 0. revision)] 0.6 ~
~ f(under
However, there is an infinite number of control sequences (t) and resu
n of this class is, that the joint angle acceleration at the beginning
 Why “lambda” rather than “r”?

Figure 3: For a symmetric pair of muscles (thin black lines), their threshold lengths,
1 , and 2 uniquely determine the equilibrium posture of joint, Rsym = ✓,2 ✓,1
irrespective of co-contraction. (Note that threshold lengths are transformed from
 articular elbow joint muscles (red), two monoarticular shou
Kinematics
nd two biarticular muscles that span elbow and shoulder joi

two joint limb with 6 muscles

ng of Muscle Lengths
= 2 pairs of mono-articulatory m.

span +along
1 pair joints
of bi-articulatory
and stretchm. or contract to change the ang
hus induce a movement. Thus, the muscles resting length an
ion determine the actual length of a muscle. We modeled t
first-order
musclepolynomial
length link todepending
joint angles on the elbow and the should

li = ci + c0i,s ✓s + c0i,e ✓e i 2 [1, 6]. (3.

nd c0i indicate the resting length and the moment arms of t
es and can be found in table 3.1. The resting lengths are tak
Woo] and the
[Ramadan, moment
Hummert, arms
Jokeit, fromunder
Schöner, [Kistemaker
revision] et al., 2006]. Equ
T [sec] 0.455 0.448 0.460 0.443 0.400 0.495 0.482 0.479
SD [sec] 0.025 0.018
Paths data vs. model
0.030 0.023 0.018 0.019 0.019 0.019

le 1: Movement time (T ) and its standard deviation (TSD ) across participants

each of the eight movements in the two movement conditions, slow vs. fast.

2
0.5 8
1
Y in [m]

0.4

7
3
5 4
6
0.3

-0.2 -0.1 0 0.1 0.2

X in [m]

ure 4: EEF path for the simulations (black) and the experiment (blue) for the
blue: experiment
black: model
w condition (800ms). The paths are labeled according to the movement numbers
me conventions as in figure 1).
 , initial to a target position at

Minimal descending activation

di erent movement speeds.

exor extensor
re ex
Shoulder 0
mmand
o the
h,
0

0
u-u0 [-]

u0
slow: 0
Elbow
dotted
0 0
medium:
dashed
time
fast:
Biarticular
solid
0 20 40 60 80 100 0 20 40 60 80 100
Movement time [%] Movement time [%]
 Minimal lambda trajectories
Shoulder lambda Elbow lambda Biarticular lambda
flexors 0.18 0.26 0.26

0.14 0.22 0.25

0.18 0.26 0.26

0.24
0.14 0.22
0.18 0.26 0.28

0.14 0.22 0.24

0.18 0.24 0.26

0.16 0.22 0.22

[m]

0.18 0.26 0.26

slow: 0.14 0.22

0.24

dotted 0.18
0.24
0.26

0.14 0.22 0.22

medium: 0.18 0.26
0.26
dashed 0.22
0.14 0.22
0.2 0.25 0.28
fast:
0.24
solid 0.15
0 20 40 60 80 100
0.2
0 20 40 60 80 100 0 20 40 60 80 100
time [%]
 Minimal lambda trajectories
Shoulder lambda Elbow lambda Biarticular lambda
extensors 0.2 0.26 0.38

0.18 0.24 0.34

0.22 0.26 0.38

0.18 0.24 0.34

0.22 0.28 0.38

0.18 0.26 0.34

0.2 0.28 0.38
0.18 0.26 0.36
[m]

0.22 0.28 0.38

slow: 0.18 0.24 0.36

dotted 0.22 0.28 0.4

0.18 0.26 0.36

medium: 0.2 0.38
0.28
dashed 0.18 0.26 0.34
0.2 0.3 0.38
fast:
0.25 0.34
solid 0.15
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
time [%]
 shaped” for some movements, which may reflect compensation for increasing inter-
Hodgson-Hogan attractor trajectories
action torques.

in movement direction orthogonal to movement direction

0.2
in end- 0.6
0.4
effector -0.2
-0.2
space! 0.6
0.2 0.4

0.6 0.2

0.2 -0.2
0.6 0.2

0.2 -0.2
[m]

0.4 0.4
0.2 0.2
slow:
0.2 0.4
dotted 0.4 0.2

medium: 0.4 -0.2

dashed 0.2 -0.4

0.2 -0.2
fast: 0.4 -0.4
solid 0 20 40 60 80 100 0 20 40 60 80 100
[%]
 in Movement Direction
0.2

attractor trajectory in hand-

0

-0.2
0 20 40 60 80 100

space
in Movement Direction
-0.2

0.2
0 20 40 60 80 100
in Movement Direction
0.6

0.4

0.2
0 20 40 60 80 100
in Movement Direction
0.2

0.4

0.6
0 20 40 60 80 100

at higher speeds (solid line),

in Movement Direction

0.4

m
attractor trajectories are 0.2

0 20 40 60 80 100

temporally structured “just 0.2

in Movement Direction

right” for the hand to reach the 0.4

target
0 20 40 60 80 100
in Movement Direction

0.4

0.2

0 20 40 60 80 100
in Movement Direction

0.2

0.4

0 20 40 60 80 100

Movement
Time in [%]
 a linear time compression, and slow movements can be obtained from fast referen
commands by linear time dilation. Figures 14 and 16 illustrate that this scaling l
Do the time courses of lambda matter?
fails for ramps and for the minimal reference commands.
First, we constructed a ramp with constant rate of change in end-e↵ector spa
which we simply gave a very short duration of 0.1 s, shorter than the fast ram
we found for the minimal reference commands. Does such a short ramp lead to
faster making
movement? a slow lambda
Figure (ramp
14 shows in hand
that this is notspace)
the case: the movement th
fast =>
is produced is, doesn’t makethan
in fact, slower movement fast
the fastest movements we modeled with t
minimal reference command. In the simulation shown, co-contraction was set to
N. Increasing co-contraction further did not make the movement faster.

EEF-path Angles Lambda

2.2 0.4

0.55 2
0.35
0.5 1.8

1.6
angle [rad]

0.45 0.3
y [m]

[m]
1.4
0.4
1.2 0.25

0.35 1
0.2
0.3 0.8

0.6 0.15
-0.04 0.04 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
x[m] t[s] t[s]
Docommand
the
Second, time
we probedcourses of
the reverse scaling inlambda
which we tookmatter?
the minimal referenc
obtained for fast movement of 0.4 s movement time and rescaled it lin
early by a factor of two, nominally for a movement time of 0.8 s. Figure 16 show
the resulting path, joint trajectories, and reference commands. The hand’s pat
ismaking a fast lambda
clearly unrealistic in shapeslow: doesn’t
for slow make
movements witha an
good slow hook at th
extraneous
end of the movement. This is reflected by the joint angles’ overshooting their ta
movement
gets. Essentially, at this slow rate, the joint angles track the N-shaped referenc
commands! So, clearly, slow movements are not scaled down fast movements.

EEF-path Angles Lambda

2.2 0.4
0.6

2
0.55 0.35
1.8

0.5 1.6 0.3

angle [rad]

0.45 1.4
y [m]

[m]
0.25
1.2
0.4
1 0.2

0.35
0.8
0.15
0.3 0.6

0.4 0.1
-0.05 0 0.05 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
x [m] t [s] t [s]
ands and
Normalization of Model Parameters T
(3) Estimate
transfer
When
le lengths
0 descending
and muscle is active: 𝐴 activation
𝑘 from EMG
Assuming EMG is linearly related to MN activation, the normalization w
Then we parameters can be determined with data sets of 𝐸𝑀𝐺, .
ch stretch 𝐸𝑀𝐺 1 2 1 /𝑘 2
ommands Ex
Experiment: Sudden unloading of
preloaded wrist muscles elicits fle
motion to a new wrist position.
unloading experiment
We assume to
subjects maintain
determine linearbefore and after
thus also
relationship
unloading:between EMG
and descending
A series of
activation
, , were
obtained with four torque levels
(by estimating threshold
(0.4, 0.6, 0.8, 1Nm), each with 5
lengthtrials;
in unloading)
unloading was done for
both flexor and extensor,
separately.

[Zhang, Feldman, Schöner]

ed to
 Descending motor command thus also before and after

(3) Estimate descending activation from EMG

unloading:
Afferent inputs
(depending on , A series of , , were
α MN obtained with four torque levels
(0.4, 0.6, 0.8, 1Nm), each with 5
Activation
trials; unloading was done for
(𝐴
both flexor and extensor,
Muscle separately.
The threshold property of MN:
𝐴

of descending
unloading experiment to motor command
: intrinsic electrical MN threshold

We assume that afferent input is linearly related to

determine linear
muscle length and velocity: , 𝑘
is the activationInstitut für Neuroinformat
relationship between EMG
𝐴 ,where
threshold of stretch reflex, which is associated with a

and descending activation

muscle length change of /𝑘

Rewrite the equation:

𝐴 𝑘
Normalization of Model Parameters
(by estimating𝑘threshold
lengthProcedure
in unloading)
Unloading experiments0
1: When and muscle
for normalizing is active: 𝐴 𝑘
model parameters
2: Voluntary movements for estimating time structure of
Assuming
Recording: surface EMGEMG is (FCR)
of wrist flexor linearly
and extensorrelated For
toflexor:
MN activation,
0.0011 ; the normaliza
0.0015
(ECR) muscles; wrist position. For extensor: 0.0014 ; 0.0028
Dataparameters
analysis: for simplicity, canand be determined
are defined in units of with data sets
Discussions
of 𝐸𝑀𝐺,
1. It is possible to have a reasonable estimation o
joint angle (positive for extension). EMG was band-pass
𝐸𝑀𝐺
filtered (10 – 500 Hz) and1rectified. 2 1 /𝑘 2
2. The time profile of descending motor comman

[Zhang,
Experiment: Sudden Feldman,
unloading of Schöner]
 bot
Muscle sep
The threshold property of MN:
𝐴 ∗
time profile during fast movem
ast movementslowhad multiple peaks, but
: intrinsic electrical MN threshold
ectory during slow movement. fast
similar shape to joint trajectory dur

We assume that afferent input is linearly related to

muscle length and velocity: , 𝑘 flexor

𝐴 ,where is the activation

extensor
threshold of stretch reflex, which is associated with a
muscle length change of /𝑘

not affect Rewrite

Slow the equation:
movement: u profile Fast movement: does not affect
0.0011
peak ;
value, 0.0015 profile
of0.0014
profile
;
was
𝐴 not𝑘 sensitive to
0.0028 𝑘 before the first peak value
but modulates later parts of profile
ds from EMG signals.
It is possible to have a reasonable estimation of descending motor commands from EM
uring fast movements. Procedure
The time profile of descending motor commands may be non-monotonic during fast m
1: Unloading experiments for normalizing model parameters
2: Voluntary movements for estimating time structure of
 Why is this important ?

quasi-postural picture
target is an attractor….

optimal control picture

a precise time course of a motor command must be computed
and generated to move to the target and reach zero velocity
there

=> demands on the neural computations

=> demands on learning
Conclusion: Human motor control

Human movement uses “soft” muscles that have

nonlinear muscle dynamics
Postures are stabilized by reflexes, whose
thresholds must be shifted during movement
Those shifts by descending commands so solve
the “optimal control” problem = the right time
course so that the effector arrives at the target
in the desired time with small velocity and a
smooth temporal shape