Professional Documents
Culture Documents
Modelo Hoff
Modelo Hoff
Modelo Hoff
71,481-488 (1994)
Biological
Cybe
9 Springer-Verlag 1994
Abstract. In modeling human and primate motor mized during the movement, he derived a position func-
behavior, optimization has been used to mathematically tion of time given by a fifth-order polynomial, uniquely
describe hand trajectories during reaching movement, specified by the initial and final values of position, velo-
with the duration of the movement given as part of the city, and acceleration. If the target has zero velocity and
boundary conditions. As duration is an input to the acceleration at the start and end of the movement, the
model rather than an output, a description is lacking of velocity profile is symmetric and bell-shaped, much like
how the duration depends upon the circumstances of the the low-accuracy pointing movements performed by the
movement. In the present work, we extend a minimum subjects studied.
jerk model of reach trajectory planning to include a pen- Flash and Hogan (1985) examined subjects perform-
alty for duration and show that it can be used to quantify ing unconstrained arm movements in the horizontal
the trade-off between quickness and effort in reaching plane, holding a light-weight manipulandum. The room
movements. We then show that a given trade-off for was darkened, removing visual feedback of arm location.
a nominal reaching movement can predict the duration Targets were indicated by light-emitting diodes at distan-
of related movements, specifically reaching movements ces of 20-40 cm. Among other experiments, they had
subject to perturbation of the target during their course. subjects move between points in the plane without ob-
The mathematical model is tested against several inde- stacles. It was found that the hand's path was approxi-
pendent bodies of experimental data. mately a straight line (as predicted by the minimum jerk
criterion), regardless of the start and end points of the
movement. Also, the trajectory of the hand was predicted
well by the minimum jerk hypothesis, yielding character-
1 Introduction istic, symmetric, bell-shaped speed profiles. Thus, the
principle that explains elbow rotation also explains
It is generally accepted that voluntary, goal-directed, whole arm movements.
reaching movements involve central programming of Hoff and Arbib (1992, 1993) used the minimum jerk
motor patterns with, in many cases, some form of feed- model to reproduce reaching trajectories from three
back modulation. For example, where the equilibrium bodies of experimental data. In each of the experiments,
point hypothesis (Feldman 1986) predicts that at the reaching movements were not only made in a point-to-
beginning of a voluntary movement a joint's equilibrium point fashion in response to a single target step, but also
position shifts at once to the desired final position for in response to a perturbed, or double-step, target. The
that joint with the limb following according to the mech- behavior under the perturbation condition was modeled
anical properties of the limb and its musculature, Bizzi et in the following way: It was assumed that when the target
al. (1984) showed that instead of a simple step to a new of reach is perturbed, there is some delay while the visual
value the virtual equilibrium position follows a time target location signal reaches the neural motor circuitry
varying trajectory [this idea being dubbed the equilib- which generates the movement toward the target. When
rium trajectory hypothesis by Flash (1987)]. In exploring the new target location reaches this circuitry, a new
the nature of such motor programs, Hogan (1984) optimal movement is initiated from the current hand
modeled elbow rotations in pointing movements of mon- position toward the secondary target. This view of the
keys toward a visually located target. The movements perturbation response is discussed in detail in Hoff and
were in the horizontal plane, about 60 ~in magnitude, and Arbib (1992). In the experimental data, the duration of
of intermediate speed (about 700 ms in duration). He the two-step movement varies greatly, seemingly depen-
proposed the minimum jerk hypothesis to describe the dent on factors such as the distance and direction of
kinematics of such movements. By applying the calculus perturbation. P61isson et al. (1986) used a perturbation
of variations, using the optimization criterion that the paradigm in which at movement onset the target was
mean squared jerk (third derivative of position) be mini- occasionally and unexpectedly perturbed further away
482
from the subject. A small change was seen in movement the arm are activated to their maximum tension in order
time. In contrast, in the target reversal experiment car- to produce the movement. This is clearly not the case,
ried out with monkeys by Georgopoulos et al. (1981), the since we experience much more smooth, relaxed move-
movement time was more than doubled after target per- ments in our day-to-day reaching. Second, we might
turbation. What is needed is a single model of movement suggest that smoothness, effort, or energy is minimized,
time determination which predicts a variety of such data. without regard to duration. For the system described
In the two-step model, as in the original minimum jerk above, this would lead to the degenerate case of a move-
model, the duration of the movement is an input, set to ment that takes infinitely long to occur. Thirdly, we
that which is observed experimentally. The question re- might suggest that duration is purely a result of mus-
mains as to what determines the movement duration in culoskeletal mechanical properties. If this were the case,
unperturbed and perturbed reaching. In the following, we one would have no control over the duration of one's
take a step toward answering that question by extending reaching movements, which clearly not the case. Instead,
the minimum-jerk cost function to include a penalty for the minimum jerk/time model implies that both efficiency
duration, such that duration may emerge from the opti- and expediency are important in movement, and that the
mization process. We base the cost function on the dura- emergent duration is based on the trade-off between the
tion of a nominal movement and show how the duration two. This is palatable to our intuition, and the task of this
of perturbed movements emerges from subsequent op- paper is to test the model quantitatively, by comparing it
timizations. to several bodies of experimental data.
The trajectory is to go from the given initial state to
the given final state, while minimizing (2). The solution to
2 Modeling duration: the minimum jerk/minimum time this optimization problem is given in Appendix A and
model yields the complex relationship between t:, the boundary
o o f f
conditions Xl, x2, xl, x2, and R given in (A14). In Appen-
We begin by defining a dynamic system for the move- dix B, the solution is given for the particular case where
ment of the hand in space, along with a cost measure to there are static boundary conditions, i.e.
be optimized. We then perform the optimization and
compare the simulation results to empirical data. We first t: = (60D) 1/3 R 1/6 (3)
define a state vector for each dimension of a two dimen-
This says that the duration of a movement (t:) is propor-
sional movement: xi(t), i = 1, 2, are 3 x 1 vectors of posi-
tional to the cubed root of the distance moved (D), and
tion, velocity, and acceleration. Taking the driving input,
the constant of proportionality is the 'arbitrary' weight-
ui(t) to be the jerk, each dimension of the movement is
ing of smoothness versus time, R. This dependence on
described by
R is consistent with our intuition: If smoothness is penal-
X1 = Axl + BUl (la) ized more than duration, a long, slow movement will
result, with longer duration, i.e. increasing R increases t:.
X2 = Ax2 + Buz (lb)
This presents both a problem and a solution for
where modeling duration as an emergent property of an optim-
ization model. The problem is that duration is still
essentially a chosen parameter, since the modeler chooses
A= 0 , B= R. However, when movement distance varies in an ex-
periment, we can extrapolate from one movement time to
0 the movement times of other distances. We do so by
xi(O) = x ,.~
, x,(t:) = x{, i = 1, 2 solving (3) for R
The cost function consists of terms penalizing duration R = t~/(60O) z
and jerk:
t=tf
Then for a chosen movement distance and time, we find
I = tf + R S (u~ W u 2)dt (2) R, which is used along with other movement distances in
t=0 (3) to find their associated durations. Further, the dura-
tion of an unperturbed movement can be used to predict
or, equivalently,
the duration of a perturbed movement. As in Hoff and
t=ty
Arbib (1992, 1993), perturbed movement is modeled as an
I= ~ (Ru~ + Ru~ + 1)dt optimal trajectory interrupted and replaced by a new
t=O
trajectory connecting the current hand state to the new
where R is the (positive and constant) relative weighting target. There is some delay while the visual target loca-
between the two factors in the cost and is to be deter- tion signal reaches the neural motor circuitry which
mined later. Equation (2) gives a formal description of the implements the movement, after which a new optimal
determination of movement duration which is intuitively movement is initiated toward the secondary target. The
acceptable. Consider the following three alternative duration of this movement is predictable by solving the
models for duration determination. First, we might sug- above optimal control problem, but with novel boundary
gest that time alone is minimized. This leads to saturated conditions. The new terminal condition is given by the
control values, or 'bang-bang' control, i.e. the muscles in perturbed target location. The new initial condition is the
483
:::I /\
state of the system, xi(t), i = 1,2, at the time the
new trajectory begins. To find the duration of the pertur-
bed movement, we use t I and D from the non-perturbed
movement to find R. Then we apply R to (A14) and 4 41
solve the polynomial equation to find the duration of I ; I I 0
-2118 -4 260 524 7118 -21111 2 270 5311 E06
the second section of the perturbed movement. Note
that xi(0) in (A14) is the current, nonstatic state at TIME (MSEC) TIME (MSEC)
3 Modeling perturbation data -2EE -4 260 524 788 -2ee 2 1TO 538 806
a 9 . . . . . . ,.= . . . . . .
9 .?
~,o A :b -b,
. \ -
Y
\ ,, ,,,
FIRST L l Gt4T
STRYS ON r ~10 I00 1~0 200 2~0 300 400 ~OHTROt. AXIS
T
W
a
b X 50 m m
800 "-
g 800
g 600 -
-- Measured duration
g'|
"0 400 -
" . . . . . . . . . . . . . . . . . . . . . .
200 -
oJ Modeled duration (ms) "~ s0o
0 ----" Measuredduration [ma)
0 9 ) ! I
0 100 2OO 300 40o
ISI (ms)
PL C20 PR
Fig. 2. a Macaque hand trajectories in target reversal task of Georg-
opoulos et al. (1981). As duration of presentation of primary target
b Target
increases, so does total movement time. (Reprinted from Georgopoulos
et al. 1981, with permission from The American Physiological Society, Fig. 3. a Target location perturbation experiment of Paulignan et al.
Bethesda.) b Movement time data and model prediction for the target (1991). Object to be grasped is unexpectedly moved left to right. Paths
reversal experiment of Georgopoulos et al. (1981). Dotted line indicates of thumb (T), index finger (I), and wrist (W) are shown for each of the
the duration of non-perturbed reaching movement. Measured duration three conditions, (Reprinted from Paulignan et al. 1991, with per-
is available for one case; the model predicts the durations for the other mission from Springer-Verlag Publishers, New York.) b Movement
perturbation cases in the experiment. ISI, interstimulus interval, the time data and model prediction for the target perturbation experiment
time before the target is switched. Note that the model predicts shorter of Paulignan et al. (1991). C20, Unperturbed movement. Modeled
overall movement time when the target is switched sooner duration is identical to measured duration by definition. PL/PR, pertur-
bed left and perturbed right
Thus, one would expect that the optimum would be the xl and x2. Similarly, the costate is the concatenation of
same as that for the classic minimum-jerk. We formally Pl and Pa. Thus, the Hamiltonian is
show this equivalence in Appendix C. The added feature
H=L + pr21 + pT22
of the optimization of this model is that it gives not only
the optimal trajectory, but also the movement duration. or
Lastly, note that even though the dimensions of hand
movement are dynamically decoupled, as defined by (1), H = (Ru 2 + Ru~ + 1) + pW(Axl + Bul)
the minimization criterion yields a movement time t I
which depends on the boundary conditions for each + p r ( A x 2 + Bu2) (A4)
dimension, and which in turn affects the trajectory of Since the formulation is a linear system with a quadratic
each dimension. For example, if the target location com- cost functional, we can minimize the Hamiltonian by
ponent is changed in dimension 1 only, then t I will be differentiating with respect to the control, and setting the
affected, as will be the trajectory in dimension 2. This has derivative equal to zero.
implications for other multidimensional movements,
such as oculomotor saccades, in which one dimension of OH
eye rotation slows for the other when an oblique move- - - = 0 = 2Ru~ + prB, u* = - prB/(2R) (a5a)
C3Ul
ment is being made.
~H
= 0 = 2Ru* + pTB, u~ = pTB/(ZR) (A5b)
Acknowledgements. This work was done in the Universityof Southern
California's Center for Neural Engineering under a grant from the t?u2
Human Frontier ScienceProgram, and under the advice and guidance Now (A1), (A3), and (A5) define a system of differential
of Professor Michael Arbib, to whom I owe thanks.
equations which may be solved with suitable boundary
conditions to yield the optimal trajectory. While the
boundary conditions x ~ x{, i = 1, 2, are given, since tz is
Appendix A being optimized it is not given a priori. Instead, we have
one additional constraint, that the Hamiltonian at the
Given the dynamic system defined by final time is zero (Bryson and Ho 1975)
22 = Ax2 + Bu2 (alb) Since the system dynamics and cost function integrand
are autonomous (i.e. not explicit functions of time), we
where have that the time derivative of the Hamiltonian is zero
A=
[ilil 0
0
,
and simplifying,
0 = - - p L23 / ( 4 R ) - - p 2 ,23 / ( 4 R ) +
Also, we may use the definition of A,
1 + p~Axl + p ~ A x 2 (A7) cb21(t, O) =
l~[: 0
0
0
A= 0 , Axi = 3 , pT Axi = Pi, lXi,2 + Pi,2xi,3 (Allb)
~22(t, 0 ) = --89 2 --tl
0
and enter it into (A7) to obtain Note that this result gives the solution trajectory for
0 = - pZ,3/(4R ) - P 2 ,2
3/(4R) + 1
(A10) if the initial conditions, x~(0) and pi(0), are known.
Two problems remain: We do not know pi(O), and we
+ Pl,lX1,2 Jr- P1,2X1,3 + P2,1X2,2 + P2,2X2,3 (A8) have still to find the duration tl, using (A8). Given the
initial and final values for the state vector, x~(0) and
which, we remind ourselves, applies at every time t.
x~(tl), we can find pi(0) as a function of ts: We assign
We return now to the problem of finding the solu-
t = t I in (A10),
tions for x~(t), p~(t), i = 1, 2. Entering u*(t) into (A1) we
have ?,(04
2, = Axi + B( - Brp,/(2R)) pi(tf)J Lq~zl(tl, 0) cl)22(tI, 0)J Lpi(0)J
Xi = Axi -- BBTpi/(2R) the first row of which is
for i = 1, 2. Combining (A9) with (A3) yields xi(tl) = @ l l ( t f , O)xi(O) + ~12(tf, O)pi(O)
Solving for p~(0),
[ ; : ] - - [ A - BBT J lPi]Xi
pi(O) = ~[21(tf, O)xi(tj-) - cb[1 (re, 0)q~ll(tl, O)xi(O) (h12)
Entering A and B, we obtain the following differential where
equation: -240 - 120t 20t 2]
-0 1 0 0 0 0 ~biz~(t, 0 ) - - - t56R 120t - 56t 2 8t3 / (A13)
0 0 1 0 0 0 Lzot 2 _ st 3 t4 J
This formula for pi(0) provides the information needed
for finding the solution trajectory. It also allows us to
find the duration. Given boundary conditions xi(0) and
000 -1 0 0 xi(tfl, we find p~(0) using (A12) and then enter it, along
000 0 -1 0 with x~(0), into (A8), the equation for the free end time
condition. (Note we are arbitrarily and conveniently
The next step is to find the state transition matrix, choosing to solve this equation at time t = 0.) This yields
defined by the following sixth-order polynomial in t I
xi(t) cb(t,~,[xi(O)~ 0=
p,(t)J = - 3600R(Ax~,l + Ax 2,1
For convenience, we define four 'sub-matrices' of the + tr 2800R(Axl. 1AXl, 2 + Ax2, lAx2.2 + 2Axl, lx], 2
state transition matrix,
+ 21x2, ~x~ + t}R( - 5761x 2, 2 - 360Axl, ~Axl, 3
[ xi(t)l = [~ll(t'O)
pi(t)J ~2,(t,O)
~12(t'O'][xi(O)l
~'~2(t,O) Lp,(0)J
(A10)
- 576Ax22,2 - 360Ax2.1Ax2, 3 - 2160 Axl. 2 x~ 2
- 2160x]21 - 2160Ax2,zx~ 2 - 2160x~22)
The solution consists of matrices of monomials of order
up to 5, + t}R(144Ax1,2Axl,3 + 144Ax2,2Ax2,3
+ 240Axx,3x],2 + 48AXl,2X~ + 240Ax2,3X~
~b11(t, O) = 1 + 48Ax2,zX~ + t } R ( - 9 1 x 2 3 - 91x~,3
0 - 12AxL3x~ -- 12x]23 -- 12Ax2,3x~
_ 5t 4 20t 2] - 12x~23) + t} (114)
-- t St 3 -- 20t 2 60t / (Alla) where Ax~ = x { - x ~ This polynomial may then be
9 ~z(t, 0) = 240---R L20t2 solved numerically, given values for the boundary
- 60t 120 J conditions x], x2, x{, x{, and a value for R.
487
X~ = , X =
Entering these into (A19)
for i = 1, 2. Then taking (A8) at t = 0, the last four terms
vanish, leaving
0 = -- p2.3/(4R) -- p2, 3/(4R) + 1
l-o
4R : P l , 32 + P 2 ,23 (A15) Now entering the boundary conditions and this expres-
sion for p(0) into (A20)
Entering xi(0) = 0 in (A12),
pi(O) = crp;21(tf, O)xi(tf)
Introducing the static x~(ty), using (A13), and calculating
only the third component of p~(0),
Entering in the matrix definitions from (A11) and (A13)
P~,3 (0) = - 6 R t f 5 2 0 t } x f
= -- 1 2 0 R t f 3 x f
--t
5t 3 -- 2 0 t 2 60t l
Combining this with (A15) yields
x(t) = ~ [_2~ _ 60t 120_]/
4 R = 1 2 0 2 R e t f 6xf~ + 1202 R 2 tf 6 X2f2
t~ = 60ZR(XYl~ + x{ ~)
Appendix C •
I L20t} -St}
56t}
2olli]
t) J
We wish to show that the cost function defined in
this work yields the same trajectory as that given by
the classic minimum-jerk formulation. Following Hogan I240 1
(1984), we take the one-dimensional (I-D) case and --t
- 240R[t
4~ -- 5t 3 20el - 6__
R |12%| xf
assume static boundary conditions. For the 1-D system,
we can drop the 'i' subscripts, rewriting (A10) and t~ [2ot~ J
(A12) as,
= x f t / t } ( 6 t 4 -- 15t3tf + 10tzt})
[x(t)~-~-~ ~ l l ( t ' O )
p(t)J [_~z,(t, 0)
~ 1 2 ( t ' 0)~ ~ X(0) ]
q~z2(t,O)J [_p(0)J
(A18) = a{6(t/d) 5 - 1 5 ( t / d ) 4 + lO(t/d) 3}