Vous êtes sur la page 1sur 14

1 Predi

tions of models for arm movements


In an experiment parti ipants were asked to either make pointing movements
with a stret hed arm to targets or to tou h targets with a bent arm. In this
se tion predi tions by some models for the path of the hand and the posture
of the arm will be onsidered.

1.1 Minimum-Jerk

The minimum-jerk prin iple was onsidered be ause hand paths were often
found to be straight in extrinsi oordinates, and the velo ity pro le of the
hand was found to be bell-shaped.
The minimum-jerk prin iple states that the following fun tion of extrinsi
oordinates is minimized:

8 !2 ! 9
Z tf < d3 x d3 y 2 =
CJ = + dt (1)
0 : dt3 dt3 ;
By minimizing the fun tion in equation (1) a hand path in 2D is obtained,
for example, a movement of the hand a ross a table surfa e as in a writing
movement. The parameter tf spe i es the duration of the movement.
It an be shown that the following fun tions satisfy the riterium in equa-
tion (1), when the movement is assumed to start at position xs ; ys and to
end at posision xe ; ye, subje t to the onstraint that the velo ity of the hand
is zero in the begin and the end position.

0 !3 !4 !5 1
t t t
xJ = xs + (xs xe )  10 + 15 6 A (2)
tf tf tf
0 !3 !4 !5 1
t t t
y J = y s + (y s xe )  10 + 15 6 A (3)
tf tf tf
As is shown in Figure 1 the minimum-jerk prin iple results in straight
hand paths and bell-shaped velo ity pro les.
In the 2D ase the arm posture an be omputed from the hand position,
if two joints are onsidered, as is shown in Figure 2.
The joint angles relate to the lo ation of the nger tip by the following
equations:

1
10 0.5

9
0.4

Horizontal speed
Vertical position
8

7 0.3

6 0.2
5
0.1
4

3 0
4 6 8 10 0 10 20 30 40
Horizontal position Time

Minimum jerk
10 Minimum angular jerk

0
0 5 10 15
Arm posture min−angular jerk

Figuur 1: Hand path and velo ity pro le predi tions by minimum-jerk and
minimum-angular jerk. The posture of the arm at the starting and ending
position for the minimum-angular jerk were obtained by means of a grid
sear h. The lower left panel shows the arm postures for the minimum-angular
jerk prin iple.
finger tip

L2
x−axis beta

y−axis

alfa
L1

Figuur 2: De nition of angles in 2D arm movements.

x = L1 os + L2 os( + ) (4)
y = L1 sin + L2 sin( + ) (5)
The hand position equation an easily be extended to the 3D ase. The
fun tion to be minimized hanges into:

2
8 !2 !2 ! 9
Z tf < d3 z 2 =
d3 x d3 y
CJ = + + dt (6)
0 : dt3 dt3 dt3 ;
whi h results in an additional equation for the z-dimension:

0 !3 !4 !5 1
t t t
z J = z s + (z s z e )  10 + 15 6 A (7)
tf tf tf
The path along whi h the hand moves remains straight and all the velo ity
pro les for all three dimensions are bell-shaped. However, omputing the arm
posture from the hand path be omes a problem. If the arm is extended only
targets at the end of the arm an be rea hed, whi h are all lo ated on the
surfa e of a sphere with a radius equal to the length of the arm. Now onsider
the task in whi h the parti ipant is asked to point with his arm extended in
the dire tion of a target. The arm angles involved in this task are illustrated
in Figure 3, in whi h the torsion of the forearm is assumed to be equal to
the torsion of the upper arm.

Top View
Side View
torsion
(gamma)
z−axis
alpha
x−axis
y−axis
x−axis

beta

Figuur 3: De nition of angles in 3D pointing movements.

By asking to point in a ertain dire tion, instead of tou hing a target, the
target lo ations an be represented by two oordinates. The arm illustrated
in Figure 3 has three possible joint rotations. The pointing task asks to sele t
the appropriate and angle, where the torsion an, in theory, take any
value. The question is whether the minimum-jerk model onstrains the angle
in any way. This is not the ase. The tip of the nger will travel along
the shortest path on the surfa e with a radius equal to the arm length. The
amount of torsion of the arm is not of importan e to the lo ation of the nger

3
tip. Spe ifying a start or an end posture won't onstrain the posture of the
arm during the movement a ording to the minimum-jerk model. The model
allows for torsion hanges during the movement, sin e these torsion hanges
don't a e t the jerk as de ned in equation (6). Only by spe ifying a fun tion
for the torsion will allow for a unique predi tion of the arm movement by the
minimum-jerk prin iple.
Instead of pointing movements one might want to onsider tou hing move-
ments in 3D instead. To be able to rea h ea h point in spa e the arm should
be allowed to bend at the elbow joint. The situation with a bent elbow is
illustrated in Figure 4.

Top View
Side View

delta
z−axis

x−axis

alpha
x−axis
y−axis
torsion delta
(gamma)
beta

Figuur 4: De nition of angles in 3D tou hing movements.

In the tou hing task situation the target lo ation an be spe i ed by


three oordinates. There are four joint rotations whi h an be used to bring
the hand to the target position. This means that there is one 'redundant'
degree of freedom. The minimum-jerk prin iple does not spe ify how to
onstrain this degree of freedom. To see this, onsider the hand being at
some intermediate position on it path to the target position. If the lo ation
nger tip an be xed at this intermediate position the arm an still move
around by hanging the upper and forearm orientations simultaneously. All
the postures obtained with the same nger tip lo ation are allowed by the
minimum-jerk prin iple, sin e it only onstrains the position of nger tip.

1.2 Minimum-Angular-Jerk

In the arti le by Wada et al. (2001, Neural Networks) a prin iple simi-
lar to the minimum-jerk is des ribed. The important di eren e between
minimum-jerk and minimum-angular-jerk, is the oordinate spa e in whi h

4
arm movements are assumed to be planned. The minimum-jerk prin iple as-
sumed that arm movements are planned in extrinsi oordinates, whereas the
minimum-angular-jerk prin iple assumed arm movement planning in intrinsi
joint oordinates.
For the situation illustrated in Figure 2, the minimum-angular-jerk prin-
iple states that the following fun tion is minimized:
8 !2 !2 9
Z tf < =
d 3
d
3
CAJ = + dt (8)
0 : dt3 dt3 ;
The solution of this minimization problem is obtained in a similar to the
solution for the minimum-jerk riterium:
0 !3 !4 !5 1
t t t
AJ = s + ( s e)  10 + 15 6 A (9)
tf tf tf
0 !3 !4 !5 1
t t t
AJ = s + ( s e)  10 + 15 6 A (10)
tf tf tf
As shown in Figure 1 the path of the hand predi ted by the minimum-
angular-jerk riterium is slightly urved.
For 3D pointing movements with a extended arm the minimum-angular-
jerk prin iple predi ts that the torsion won't hange during the movement,
be ause any hange in torsion will in rease the angular-jerk. A sequen e of
movements whi h demonstrates that people are not likely to x the amount
of torsion during movements, is illustrated in Figure 5. In the drawings the
square at the end of the subje t shows the orientation of the hand. First
the subje t rotates his arm in the beta-angle dire tion shown in Figure 3.
Then a rotation around the alpha-angle orientation dire tion is made. Finally
an addition beta-angle dire tion orientation is made, resulting in a di erent
orientation of the hand. If the sequen e in Figure 5 is repeated several times,
torsion rotations are very likely due to restri tions on the angles a ross whi h
joints an be rotated.
By onstraining the arm posture at the beginning and the end of the
movement, tou hing movements an be predi ted by the minimum-angular-
jerk model. For ea h joint angle a separate equation is obtained:
0 !3 !4 !5 1
t t t
iAJ = is + (is ie )  10 + 15 6 A (11)
tf tf tf
(12)

5
2 3 4
1

Figuur 5: Illustration of a sequen e of arm movements.

where i is equal to either ; ; or Æ . If no starting and ending pos-


ture is provided, the starting and ending posture an be sear hed for whi h
minimized the angular-jerk of the movement.
The hand elbow position based on the joint angles in the shoulder is given
by:

xe = L1 sin sin (13)


ye = L1 os sin (14)
ze = L1 os (15)
where L1 is the length of the upper arm, and:

xw = xe L2 fsin Æ [ os sin os + sin os ℄ + os Æ sin sin g (16)


yw = ye + L2 fsin Æ [ os os os sin sin ℄ + os Æ os sin g (17)
zw = ze + L2 fsin Æ os sin os Æ os g (18)
for the position for the wrist, where L2 is the length of the forearm.
In Figure 6 arm postures are illustrated. In ea h subplot one of the joint
angles was varied while the values of the other angles were xed. When xed,
; and were xed at 20 degrees, and Æ was xed at 80 degrees.
To plot Figure 7 the starting and ending posture were sele ted at random
and the minimum-angular-jerk path was omputed. The rst subplot shows
the posture of the arm during the movement, and the other three subplots
show the velo ity pro les in ea h of the extrinsi spa e oordinates.
If only the starting posture is prede ned, and the ending posture has to
be sear hed for, the sear h spa e should be de ned. As a qui k solution to
this problem a riterium is de ned whi h is based both on the distan e of the
hand to the target lo ation and the total angular jerk whi h takes the arm

6
Alpha Beta

0 0

−5 −5

−10
−10
−15

10 0 0
10 −2
5 −5 −4
−10 −6
0 0 −8

Gamma Delta

0 0

−5
−10

−10
−20

10 10
0 0
−2 −2
5 −4 5
−6 −4
0 −8 0

Figuur 6: Joint rotations in 3D.


0.25

0 0.2
Speed x−direction

−5
0.15
z

−10
0.1
−15
20 0.05
0
10
−5
0
y 0 −10 x 0 10 20 30 40
Time

0.25 0.07

0.06
0.2
Speed y−direction

Speed z−direction

0.05
0.15 0.04

0.1 0.03

0.02
0.05
0.01

0 0
0 10 20 30 40 0 10 20 30 40
Time Time

Figuur 7: Movement in 3D predi ted by minimum-angular-jerk.

to the nal posture. By means of a simplex sear h the minimum is sear hed
for of the fun tion:
q
C AJep = a (dx2 + dy 2 + dz 2 ) +
Tf n
X o
d3 i2 + d3 i2 + d3 i2 + d3 Æi2 (19)
i=0
7
in whi h a an be sele ted to vary the importan e of the distan e to the
target in the sear h for the end posture. The notation d3 indi ates that three
times the di eren e with the previous values has to be taken (dis rete version
of the ontinuous ase shown in equation (8)). A path sele ted on the basis
of the simplex sear h of the minimum of the fun tion in equation (19) with
a set to 10, is shown in Figure 8.

−2

−4

−6

−8

−10
0
−2
8
−4
6
−6
4
−8
2 −10
0

Figuur 8: Path sele ted by the minimum-angular-jerk riterium.

A problem with the algorithm is that not always the same end posture is
obtained: The sear h spa e has several lo al minima. This is illustrated in
Figure 9. This gure was generated by running the optimalization pro edure
100 times, and by plotting the nal posture obtained in ea h run.
Be ause several optimal nal postures are obtained by minimizing the
riterium in equation (19) no lear predi tions an be made about whi h
nal postures would be obtained by minimizing the angular jerk for di erent
starting postures. A better way for sele ting the nal posture is needed. A
di erent pro edure was proposed by Rosenbaum et al. in 1995 on the basis
of stored nal postures. This pro edure, known as the knowledge model, will
be dis ussed in the following se tion.

1.3 Knowlegde model

The model des ribed by Rosenbaum et al. (1995) assumes that postures
are stored in memory. Before making a movement all stored postures are
evaluated and a weighted posture is sele ted. Be ause this weighted posture

8
0
−1
−2
−3
−4
−5
−6
−7
−8
−9

8 0
6
−2
4
2 −4
0 −6
−2

Figuur 9: Optimal nal postures obtained on the basis of the riterium shown
in equation (19).

won't bring the hand exa tly to the target position an iterative sear h is
performed, alled feed-forward orre tion.
Ea h posture is assigned a value whi h indi ates how inattra tive this
posture is as a nal arm posture. This value is de ned by:

! !
Sp Vp
Cp = ws + wv (20)
maxp Sp maxp Vp
in whi h Cp is the ost of a posture, Sp is the spatial ost, Vp is the travel
ost, and ws = 1 wv . The spatial ost is de ned to be:

rn o
Sp = (xt xw )2 + (yt yw )2 + (zt zw )2 (21)

that is, the Eu ledian distan e between the nger tip and the target position.
The travel osts are de ned by:

n
X  
Vp = (kj j ) 1 + (Tj kj log( j + 1))2 (22)
j =1
in whi h aj is the angular displa ement, and kj is the joint's expense. The
nal posture is sele ted by Gaussian weighting of postures with their osts:

9
Xm G(C )
P~ = Pm p P~p (23)
p=1 q=1 G(Cq )
in whi h:

!
1 Cp2
G(Cp) = p exp
2(min(Cp ))2
(24)
min(Cp) 2

20

15

10

−5
−20 −15 −10 −5 0 5 10

Figuur 10: Illustration of the feedforward orre tion me hanism. The ir le


between the two arrows shows the target position, the asterisk shows the new
target position.
In Figure 10 the feedforward me hanism is illustrated. The ir les in
the plot denote the stored postures, whi h were sele ted at random at the
start of the simulation. The dotted onne ted lines show the initial posture.
The solid onne ted lines represent the posture sele ted on the basis of the
stored postures. The ir le to whi h the two arrows point is the target hand
position. The sele ted posture does not bring the hand to this target position.
To orre t for this a new target position is sele ted, whi h is represented by
asterisk. This new target is at half the distan e of the hand to the target,
but opposite to the hand position. A new planning pro ess is performed with
this new target position, with the previous sele ted posture as the starting
posture.

10
In Figure 11 a sequen e of 10 feedforward orre tions is shown for a 2D
arm movement. For this simulation all joint stifness parameters (kj ) were
set to 100, and the weight ws was set to 0.7. For the sele ted target (at
position (5,5)), and the sele ted number of stored postures (20) the orre tion
me hanism doesn't seem to onverge to the target position. This might be
be ause large angular depla ements are needed to rea h the target, whi h
outweight the spatial osts.
25

20

15

10

−5

−10
−20 −15 −10 −5 0 5 10 15

Figuur 11: Development of the sele ted arm posture for 10 feedforward or-
re tions.
How mu h the weight assigned to the angular displa ement is a e ting
the nal posture of the hand, is investigated by running the simulation with
either a large weight assigned to nal posture or a large number of stored
postures. The out ome of a simulation run is shown in Figure 12. The plots
show that in reasing the number of postures de reases the spatial distan e
between the hand and the target position. By in reasing the number of stored
postures, the sele ted end posture depends less on the random sele tion of
postures stored in memory. In the arti le by Rosenbaum et al. (1995) a
solution is proposed for the dependen e of the sele ted end posture on the
stored postures is memory. A learning algorithm is introdu ed. Ea h stored
posture is assigned a value based on its usefulness in previous rea hes. If the
value of a stored posture drops below a ertain, prede ned value the posture
is repla e by a random posture. The amount of learning depends both on
this prede ned value, and the learning rate, whi h determines how mu h a
value is updated after ea h movement.

11
20 stored postures, w=1 100 stored postures, w=1

20
15
15
10
10

5 5

0 0

−5
−5

−20 −10 0 10 −20 −10 0 10 20

100 stored postures, w=0.7


20

15

10

−5

−10
−20 −10 0 10

Figuur 12: Results of a simulation whi h investigated the e e t of the weight


assigned to the nal hand position, and the e e t of the number of stored
postures. The diamond represents the target position, the dotted lines show
the starting posture, and the solid lines show su essive planned end postures.

The e e t of learning on the lexi on of postures was studied by means of


a simulation. Figure 13 shows the distribution of stored postures over time
when repeated movements are made to a restri ted part of the workspa e. For
this simulation a high learning rate was sele ted: The value of stored postures
depended mu h on the urrent movement and less on previous movements.
The initial value of the postures was set to 1, and postures of whi h the
value dropped below 0.001 were repla ed. With these parameters about 50%
of the postures were repla ed after ea h movement. Figure 13 shows that the
distribution of stored postures ( ir les) shifts towards the part in workspa e
in whi h target (diamonds) were presented. However, this shift doesn't result
in mu h more a urate movements, as shown in lower right panel of the gure.
The posture sele tion pro ess of the knowlegde model an easily be ex-
tended to three dimensions. In the simulations of whi h the results will be
shown below, the wrist was assumed to be xed, to make the number of
redundant degrees of freedom in the 3D situation equal to the number in the
2D ase. Figure ?? shows the sele tion pro ess of a posture in 3D.
The 3D version of the posture sele tion model su ers from the same
problems as the 2D version. The target hand position is not always rea hed
espe ially when the target is in the part of workspa e opposite to the initial
posture. The sele ted posture depends on the stored postures, espe ially

12
First movement After 20 movements

20 20

15 15

10 10

5 5

0 0

−5 −5

−20 −10 0 10 −20 −10 0 10

Distance towards target


After 50 movements Proportion of replaced postures
5
20
4
15

10 3

5
2
0
1
−5
0
−20 −10 0 10 0 10 20 30 40 50
Number of movements

Figuur 13: Results of a simulation whi h investigated the e e t of learning


on the distribution of stored postures, and on spatial a ura y. In the rst
three subplots ir les show the stored postures, and diamonds the target
lo ations. In the lower right panel the spatial error and the proportion of
repla ed postures is plotted over time.

when the number of stored postures is low (around 30).


To over ome the problems of the 1995-version of the model, a new version
of the model was proposed in 2001. The 2001-version of the model repla es
the weighted osts by a onstraint hierar hy.

13
20

15

10

−5

−10

−15

−20
20

10

0
10
5
−10 0
−5
−10
−20 −15
−20

Figuur 14: Posture sele tion in 3D. The dashed line represents the initial pos-
ture, the solid lines the sele ted postures at several steps during the feedfor-
ward orre tion pro ess. The ir les denote stored postures, and the asterisk
the the target position.

14

Vous aimerez peut-être aussi