Smoothness Maximization Along a Predefined Path Accurately

Predicts the Speed Profiles of Complex Arm Movements

EMANUEL TODOROV AND MICHAEL I. JORDAN

Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139

Todorov, Emanuel and Michael I. Jordan. Smoothness maximi- etc. This impressive ability to take into account a diverse set
zation along a predefined path accurately predicts the speed profiles of task constraints and rapidly ‘‘fill in’’ the missing details of
of complex arm movements. J. Neurophysiol. 80: 696–714, 1998. the movement suggests that the motor system has very gen-
The speed profiles of arm movements display a number of regulari- eral and efficient mechanisms for solving ill-posed problems.
ties, including bell-shaped speed profiles in straight reaching move-
ments and an inverse relationship between speed and curvature in
Consequently, a fruitful line of research in the field of motor
extemporaneous drawing movements (described as a 2/3 power control has been to identify, mostly through observational
law). Here we propose a new model that simultaneously accounts studies, regularities in biological movement that are not in
for both regularities by replacing the 2/3 power law with a smooth- any way implied by the task. Once identified, such regulari-
ness constraint. For a given path of the hand in space, our model ties can be used to infer the structure of the underlying
assumes that the speed profile will be the one that minimizes the motor system that produces them. Examples of this approach
third derivative of position (or ‘‘jerk’’). Analysis of the mathemati- include: Listings law, Fitts law, the isochrony principle (Viv-
cal relationship between this smoothness constraint and the 2/3 iani and Schneider 1991), linearly related joint velocities
power law revealed that in both two and three dimensions, the (Soechting and Terzuolo, 1986), piece-wise planar three-
power law is equivalent to setting the jerk along the normal to the dimensional (3D) hand paths (Soechting and Terzuolo
path to zero; it generates speed predictions that are similar, but
clearly distinguishable from the predictions of our model. We have
1987), straight paths and bell-shaped speed profiles of reach-
assessed the accuracy of the model on a number of motor tasks in ing movements (Morasso 1981), and the inverse relation-
two and three dimensions, involving discrete movements along ship between instantaneous tangential speed and curvature
arbitrary paths, traced with different limb segments. The new model of the hand path (Jack 1895; Lacquaniti et al. 1983).
provides a very close fit to the observed speed profiles in all cases. In this work, we focus on the relationships observed be-
Its performance is uniformly better compared with all existing tween the hand paths and speed profiles of arm movements
versions of the 2/3 power law, suggesting that the correlation that are spatially constrained. We begin with a brief summary
between speed and curvature may be a consequence of an underly- of existing versions of the minimum-jerk model (Flash and
ing motor strategy to produce smooth movements. Our results Hogan 1985) and the 2/3 power law (Lacquaniti et al. 1983)
indicate that the relationship between the path and the speed profile and compare their implications for the biological mecha-
of a complex arm movement is stronger than previously thought,
especially within a single trial. The accuracy of the model was
nisms underlying trajectory formation. We then develop a
quite uniform over movements of different shape, size, and average new model that essentially combines the appealing features
speed. We did not find evidence for segmentation, yet prediction of these prior models; it simultaneously accounts for both
error increased with movement duration, suggesting a continuous the bell-shaped speed profiles of straight reaching move-
fluctuation of the ‘‘tempo’’ of discrete movements. The implica- ments and the inverse relationship between curvature and
tions of these findings for motor planning and on-line control are speed of curved hand movements. In a series of experiments,
discussed. we demonstrate that the new model provides an accurate
description of a number of discrete arm movements with
arbitrary hand paths in both two and three dimensions. Fi-
INTRODUCTION nally, we discuss the implications of our observations for
the generation of complex hand trajectories.
The majority of human arm movements produced in ev-
eryday life are underconstrained: the desired effect on the
MODELS OF TRAJECTORY FORMATION
external environment can be achieved in a large number of
ways, thus the exact movement executed at a particular time In this section, we distinguish three groups of models
is not completely determined by the motor task. We effort- (Fig. 1) based on the levels of biological processing they
lessly plan and execute underconstrained arm movements parallel. Our classification emphasizes the implications of
that achieve the desired effect and are at the same time different models for motor control rather than mathematical
smooth and graceful. Furthermore, we can produce appro- similarities among them.
priate movements in tasks that involve almost arbitrary com-
binations of constraints: spatial locations that the hand Complete models of trajectory formation
should pass through, objects that have to be avoided, forces
that have to be exerted with very accurate timing, require- The most ambitious group of models predicts the outcome
ments on the speed at certain points along the path, desired of the processing in the entire CNS in a restricted set of tasks.
durations of the entire movement or smaller segments of it, Such models provide a complete ‘‘recipe’’ for trajectory

696 0022-3077/98 $5.00 Copyright q 1998 The American Physiological Society

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 697

FIG . 1. Schematic representation of our

model classification scheme. Grey boxes
correspond to groups of models, – – – ,
predictions these models make. , the
actual flow of information in the CNS (the
internals of the ‘‘CNS’’ box are of course
hypothetical). Observed trajectory can be
further separated into a priori independent
features, in this case speed and path.

generation, starting with the externally specified task de- cause this model is purely kinematic, it predicts invariance
scription and leading to an observable trajectory. It would with respect to translation, rotation, and spatial or temporal
be ideal if one could construct a model of this kind that scaling. It has been shown that planar rotation of the desired
agrees closely with a wide range of experimental data; unfor- movement can significantly affect the hand path in via-point
tunately, this has proven to be difficult. (Uno et al. 1989), as well as obstacle avoidance (Sabes
A general method for constructing such models is pro- and Jordan 1997) tasks. We show more examples of this
vided by optimization theory: one defines a ‘‘cost’’ func- phenomenon later.
tional and attempts to show that the movements typically
produced in a given task are the ones that minimize cost, Models of intrinsic regularities
subject to the constraints imposed by the task. A number of
models based on optimization theory have been considered On the other extreme, we have models that do not explic-
(Flash and Hogan 1985; Nelson 1983; Uno et al. 1989; Wann itly assume anything about the motor task or the flow of
et al. 1988), the cost being energy, force, impulse, peak information processing in the CNS. They simply quantify
speed, peak acceleration, duration, or torque change. The intrinsic relationships between features of the observed tra-
model that seems to account for the largest body of data, jectories that are not related a priori (in this case, between
while at the same time being attractively simple, is the ‘‘min- the path and speed profile). Note that a causal relationship
imum-jerk’’ model (Flash and Hogan 1985). The cost is between the two features is not necessarily implied: such
defined as the squared jerk (3rd derivative of hand position models are consistent with any proposed mechanism of tra-
over time), integrated over the entire movement; it is mini- jectory formation that produces the observed relationship
mized over all trajectories that have the same initial point, either through an explicit computation or as an emergent
final point, and possibly pass through a small set of interme- property of some other, possibly unrelated process.
diate (via) points specified in the task itself. In reaching It has been proposed (Lacquaniti et al. 1983) that for a
tasks, the minimum-jerk model predicts straight paths and large class of movements with curved hand paths, the speed
bell-shaped speed profiles that match experimental observa- n(t) is related to the curvature k(t) through a power law:
tions rather well (Flash and Hogan 1985). There are cases, n(t) Å gk(t) 01 / 3 , where g is a constant gain factor. The
however, where subjects systematically produce nonstraight validity of this law is typically demonstrated through a corre-
reaching movements (Atkeson and Holerbach 1985). Be- lation between n and k 01 / 3 or between log( n ) and log( k ).

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
698 E. TODOROV AND M. I. JORDAN

Although the resulting correlation coefficients are rather high profiles) that can be scaled and positioned as to form a
(Viviani and Cenzato 1985), it is not immediately obvious desired trajectory. The parameters of the model (timing,
from such analyses how much of the details of the speed scale, and position of each segment) initially were extracted
profiles are captured, apart from the general trend that speed from the experimental data and then adjusted iteratively until
decreases with increasing curvature. It has been argued, for a good fit to the observed trajectories was obtained.
example, that the actual speed profiles can be skewed relative Viviani and Flash (1995) proposed a modification of the
to the power law prediction, and the correlation between minimum-jerk model, which assumes that the movement is
log( n ) and log( k ) still remains high (Wann et al. 1988). represented internally as a set of salient points and the veloc-
The analysis that could resolve this issue, which thus far ity (speed direction and magnitude) at these points. The
has been absent from the literature, is a direct comparison authors extracted the velocities from the observed hand tra-
superimposing the actual and predicted speed profiles. In its jectories and used standard optimization methods to solve
original form the power law has two limitations: first, it does for the resulting minimum-jerk path and speed profile (the
not apply to paths that have inflection points or straight solution is a concatenation of 5th order polynomial splines).
segments (i.e. for k Å 0 the predicted speed is infinite), and This model also has been studied by Yalov (1991), who
it is inaccurate at low curvature; and second, the fact that found that overall it is more accurate than versions of the
n(t) decreases toward 0 at the end points of discrete move- minimum-jerk model using only the position (and tangent
ments cannot be captured. To fix the first problem, Viviani to the path) of a via-point.
and Schneider (1991) suggested a modified power law: Another possible model in this category is ‘‘post hoc seg-
n(t) Å g[ k(t) / e] b where e depends on the average speed mentation’’: start with any model that predicts a local speed
and b is a free parameter estimated from the actual data. profile given a portion of the path, define a segmentation
Despite these limitations, the power law seems to provide rule that somehow selects segment boundaries given the
a surprisingly good description of the speed profiles of com- complete path, and apply the model to each segment. Be-
plex movements, which has motivated recent attempts to cause the local speed profiles generated by this procedure
find an explanation of this regularity. Within the framework will have to be scaled appropriately, one has to measure the
of equilibrium-point control (Bizzi et al. 1992), Gribble and actual time it takes the subject to traverse each segment of
Ostry (1996) have shown that if the equilibrium point is the path and use those times in fitting the model.
moved along an ellipse with a constant speed, the hand traces Because the hypothetical internal representation is not di-
a similar ellipse obeying the power law. A potential problem rectly observable and cannot be derived from the task de-
with this explanation is that the validity of the equilibrium- scription (otherwise the model would belong to the 1st cate-
point hypothesis itself recently has been questioned in stud- gory), it has to be inferred from the observed trajectory
ies arguing that the required stiffness is much larger than itself. Thus the above models are validated by extracting
experimental measurements under dynamic conditions (Ben- their parameters from the experimental data and demonstra-
nett et al. 1992; Gomi and Kawato 1996). Another proposed ting that these parameters are sufficient to reconstruct the
explanation (Pollick and Sapiro 1996) is the mathematically same data. Strictly speaking, this procedure shows that the
interesting fact that a point moving according to the power family of trajectories produced by the model contains the
law maintains constant ‘‘affine velocity’’; it remains to be family of human arm trajectories; in itself it is not a proof
clarified how affine velocity is related to arm movements that the proposed internal representation actually is being
and why it might be advantageous to keep it constant. An used by the motor system. To support the latter claim and
alternative suggestion, which we will consider in the present make detailed comparisons to models in the other two cate-
paper, is that maximally smooth (i.e. minimum jerk) move- gories, we need a way of quantifying how much of the
ments have speed profiles described by the power law (Wann information present in the data is being extracted in fitting
et al. 1988). Viviani and Flash (1995) tested this empirically the parameters of the model. This may appear to be a prob-
by applying both the 2/3 power law and their modified lem with the power law as well because it predicts a speed
minimum-jerk model (see further) to the same experimental profile given a path extracted from the observed trajectory.
data. These authors concluded that the speed-curvature rela- The crucial difference is that the path of a trajectory is not
tionship described by the power law is not implicit in the a priori related to the speed profile, whereas in the preceding
minimum-jerk hypothesis. examples, the parameters extracted from the observed trajec-
tory contain information about all aspects of the trajectory
Partial models of trajectory formation (i.e., both spatial and temporal information).

Finally, a number of intermediate models have been pro- CONSTRAINED MINIMUM-JERK MODEL
posed, attempting to bridge the gap between providing a
complete recipe for trajectory formation and simply quanti- The main advantage of the power law is that in principle
fying intrinsic regularities of observed trajectories. Such it can be applied to almost any movement: given an arbitrary
models assume that the motor task is transformed (via some path it always predicts a speed profile. The problems with
unspecified planning process) into a compact spatio-tempo- it are related to the particular formula used to predict speed
ral representation, involving a small number of parameters from path. In contrast, all known problems with the original
that are sufficient to generate the observed trajectory. minimum-jerk model are in the path prediction: it can fail
Morasso and Mussa-Ivaldi (1982) suggested that complex to predict the hand path given the configuration of via-points,
hand trajectories are composed of partially overlapping lin- but there are no known cases where the path was predicted
ear ‘‘strokes’’ (modeled as B splines with bell-shaped speed accurately while the speed profile was not.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 699

Our model combines the appealing features of the two r * Å t *, r 9 Å kn, r - Å k*n / k( tb 0 kt)
models: it preserves and extends the generality of the power We can now substitute in Eq. 2
law while replacing the particular relationship between cur-
vature and speed with a smoothness constraint. We propose L Å \n( k*sh 3 / 3ksh sF ) / t(sf 0 k 2sh 3 ) / b(sh 3kt ) \ 2 (3)
that for an observed path of the hand in space, the speed and because the tangent, normal, and binormal are orthogo-
profile along that path will be the one that minimizes jerk nal
(in general ‘‘smoothness’’ is synonymous to ‘‘having small
high-order derivatives’’; the measure of smoothness we use L Å ( k*sh 3 / 3ksh sF ) 2 / (sf 0 k 2sh 3 ) 2 / (sh 3kt ) 2 (4)
is one of many possibilities). This constrained principle, like
the power law, quantifies an intrinsic relationship between Relation to the power law
the path and speed of the observed trajectory rather than
suggesting a concrete mechanism for trajectory formation Now that we have an explicit relation between curvature
(i.e., we do not suggest that speed is being computed explic- and jerk, we analyze the constrained minimum-jerk principle
itly as a function of hand path). Note that we use the path to find how it relates to the power law. We define a function
observed on a particular trial to predict the speed profile on [dependent on s(t)] that is equal to the term multiplying
the torsion t in Eq. 4
that same trial, which is a natural way to model the signifi-
cant variability in speed profiles in a block of identical trials As Å sh 3k(s)
(see further). In contrast, the original minimum-jerk model
Taking derivatives with respect to time on both sides, we
predicts a single optimal movement for a given configuration
obtain
of via-points and thus can only be applied to average data.
k* (s)sh 4 / 3sh 2sF k(s) Å A s* sh
Formal definition k* (s)sh 3 / 3sh sF k(s) Å A s*

Let r(s) Å [x(s), y(s), z(s)] be a 3D curve describing Note that the term on the left is exactly the term multiplying
the path of the hand during a particular trial, where s is the the normal vector in Eq. 3. We now substitute in Eq. 3
curvilinear coordinate (distance along the path), and the
L Å \n(A s* ) / t(sf 0 Ask ) / b(Ast ) \ 2
tangential speed is sg (t) (sg is a time derivative, r * is the
derivative with respect to s, and boldface signifies vector L Å A s* 2 / (sf 0 Ask ) 2 / A 2s t 2 (5)
quantities). Our model assumes that the temporal profile of
the movement is the scalar function s(t) that minimizes From the definition of As , we have

* ZZdtd ZZ
T 3 2 sh (t) Å A s1 / 3 k 01 / 3
JÅ 3
r[s(t)] dt (1)
0
The power law states that A 1s / 3 Å const or A *s Å 0. Thus the
This differs from the original formulation of the minimum- power law is equivalent to setting the normal component of
jerk hypothesis in that here the path r is given, and we are the instantaneous jerk to 0, and the binormal component
only minimizing over the space of speed profiles. Flash and proportional to t, everywhere along the movement.
Hogan (1985) computed both the path and the speed given For two-dimensional (2D) curves r(t) Å [x(t), y(t)],
a set of via points. Let us denote the term inside the integral this can be shown more directly. The power law n(t) Å
by L const.k(t) 01 / 3 in 2D can be written as
ZZ ZZ ZZ ZZ ZZ ZZ S D
q
d3 2
d2 2
d 2
(x *y 9 0 y *x 9 ) 2 01 / 3
LÅ r(s) Å r * (s)sh Å (r 9 (s)sh 2 / r * (s)sF (x * 2 / y * 2 ) 1 / 2 Å const
dt 3 dt 2 dt (x * 2 / y * 2 ) 3 / 2

L Å \r - (s)sh 3 / 3r 9 (s)sh sF / r * (s)sf \ 2 (2) x *y 9 0 y *x 9 Å const

This minimization problem is invariant to rotations and Taking derivatives, and canceling terms, we get
translations of the path r. Our goal is to transform the above
x * /y * Å x - /y - , r *\r -
expression into a form that makes that invariance explicit.
We know from differential geometry (e.g., Prakash 1981) The jerk vector points along the tangent, which is orthogonal
that a 3D curve can be defined uniquely, up to translations to the normal; therefore the jerk along the normal is zero.
and rotations, by its curvature k(s) and its torsion t(s). It Before analyzing experimental data, we can ask whether
is not possible to simply express r in terms of k and t and the power law and the constrained minimum-jerk model pre-
substitute in Eq. 2. However, the derivatives of r can be dict similar speed profiles for discrete movements along arbi-
expressed as functions of the curvature and torsion. The path trary synthetic paths. In Fig. 2, we have plotted three syn-
r satisfies Frenet’s formulas thetic paths, for which we have computed the speed profiles
t * Å kn, n * Å tb 0 kt, b * Å 0 tn predicted by our model ( ) and the power law (– – –).
Also shown is the magnitude and direction of the instanta-
where t(s) is the unit tangent, n(s) is the unit normal, and neous jerk vector at several positions along the path. Note
b(s) is the unit binormal vector to the curve r(s) at point that in most cases the normal jerk is close to zero, and the
s. Using the fact that the derivative of the path with respect two speed profiles are rather similar (the deviation near the
to the curvilinear coordinate s is the unit tangent vector, we end points is due to the fact that the power law cannot predict
obtain movements that start and stop). There are differences, how-

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
700 E. TODOROV AND M. I. JORDAN

FIG . 2. A–C: examples of the con-

strained minimum-jerk ( ) and power
law ( – – – ) speed profile predictions for the
synthetic paths on the left. Jerk vector is plot-
ted at several points along the path (for the
minimum-jerk speed profile). D: tangential
and normal jerk components for the middle
portion of example C. Two small perturba-
tions are applied to the minimum-jerk speed
profile ( ) and the corresponding effects
on the jerk components are shown.

ever: in example B where the curvature is almost constant, optimal speed profile increase both terms, implying that in
the power law predicts a speed profile that is flatter than the this case the optimal speed profile simultaneously minimizes
minimum-jerk prediction. For most of the paths we have both terms and not just their sum (this happens around points
studied, the two predictions are similar in the sense that they of high curvature). In summary, the constrained minimum-
go up and down together, but clearly distinguishable (i.e., jerk model and the power law predict similar speed profiles
example A is not typical). for a given path. There seem to exist a family of paths for
Why is it that the power law, which predicts instantaneous which the two are exactly equivalent, and identifying that
speed only using the local curvature, comes close to solving family may provide further insights, but this is outside the
the global minimization problem (Eq. 1)? One intuitive an- scope of the present paper.
swer is the following: because the total jerk is a sum of two
(in 3D of 3) nonnegative terms and the power law minimizes EXPERIMENT 1
one of them (normal jerk is set to 0), it is an efficient
approximation to the solution of Eq. 1 (suggesting that in Methods
3D the difference between the 2 speed predictions will be To test our model and compare it with the power law, we
larger). It seems, however, that this is not always the com- studied four tasks in which a path was specified and subjects
plete answer. In Fig. 2D we have plotted the tangential and were asked to execute an arm movement along that path.
normal components of the jerk over the middle portion of We used discrete movements of short duration and paths of
example C. We can see that small perturbations around the varying degree of complexity. To avoid drawing conclusions

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 701

from an isolated movement, we used several paths in each viewed in a semitransparent mirror. The system was cali-
task. The same eight right-handed subjects participated in brated so that the projected images appeared to be on the
all four tasks. Short breaks were given between tasks. The table (see Wolpert et al. 1995 for details). Subjects held a
entire experiment lasted Ç45 min. End-point position was small cursor (tracked by the infrared cameras), which had
recorded at 100 Hz with an Optotrak 3020 infrared system. a light-emitting diode attached to it that provided end point
TASK 1 — HAND / FINGER MOVEMENTS. Subjects stood in front
visual feedback during the movement. The room was dark
of a horizontal table to which we had taped a white sheet so subjects could not see their arm. The same eight templates
of paper (letter format) with the templates shown in Fig. 3 from task 1, scaled up by a factor of 8, were projected one
printed on it. Subjects held a pointer in their right hand; the at a time. The task was again to position the cursor at the
task was to position the pointer near the circle on the current starting point of the template, wait for a beep confirming a
template, wait to hear a beep signaling that the tip of the correct initial position, and then trace the path. Each path
pen was in the starting area, and then trace the specified was traced 10 times. After each trial, we displayed a posi-
path. After 10 trials, a beep of different pitch signaled the tional error score, which was the maximum deviation (in
subject to move to the next template. No instructions were millimeters) between the desired and actual paths. Again,
given regarding the duration or speed of the movement. We subjects were allowed to make movements of any speed.
observed that subjects used predominantly hand and finger This task involved predominantly shoulder and elbow move-
movements in this task. ments.
TASK 2 — 2D ARM MOVEMENTS. Subjects sat on an adjustable TASK 3 —VIA POINTS. We used the same setup as in task 2.
chair and made arm movements on a table positioned slightly The display contained a start point, an end point, and four
below shoulder height so that the movements (made by slid- intermediate points, numbered 1–4. The subject was asked,
ing the arm on the table) occurred in a horizontal plane at after moving to the start point and hearing a confirmation
shoulder height. We used a VGA projector to display com- beep, to make a movement that passed through the interme-
puter generated images on a tilted screen, which the subject diate points in the specified order (without stopping), and
finished at the end point. The configurations we used are
shown in Fig. 4. There were a total of 6 configurations, every
other configuration being a 907 rotation of the preceding
one. Subjects executed 10 trials in each block. As before,
movement speed was not specified. Note that the configura-
tions used do not correspond to the continuous paths from
the previous two tasks.
TASK 4 — 3D ARM MOVEMENTS. Subjects stood in front of the
table, holding the cursor in their right hand and a 3D wire-
frame model in their left hand. The models were Ç10 1 10 1
10 cm and represented smooth 3D curves. The task was to
first study the model and then try to make a movement with
the right hand that had the same shape (not necessarily the
same size). The starting point on the model was marked.
There was no specified starting position for the arm because
the movement could not be superimposed on the wireframe
model. At the beginning of each trial, subjects had to stop
moving the cursor and wait for a beep. The experimenter
waited for the subject to reposition and then manually acti-
vated a procedure that checked for zero speed before move-
ment onset. Each wireframe model was used in 10 trials,
there were a total of three models. Movement speed and
duration were not specified.

Data analysis
The data from each trial for each subject and task were
analyzed separately, i.e., no averaging over trials was per-
formed. The first step was to smooth the raw data. We used
cubic spline smoothing (De Boor 1978) applied to x(t),
y(t), and z(t). The parameter l (which determines the
amount of smoothing) was set adaptively for each trial, so
that the maximum deviation between any smoothed sample
point and the corresponding raw data point was within a pre
specified radius. That radius was set to 0.002 of the extent
FIG . 3. Templates used in tasks 1 and 2, drawn to scale. s, starting of the path—the high accuracy of the Optotrak system obvi-
point on each template. Order of presentation is row-1st. ated the need for substantial smoothing. Once we obtained

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
702 E. TODOROV AND M. I. JORDAN
Maximum deviation MAX( n1, n2) Å max [Én1(s) 0 n2(s)É]
Mean absolute deviation MAD( n1, n2) Å »Én1(s) 0 n2(s)É…
q
Standard deviation STD( n1, n2) Å » [ n1(s) 0 n2(s)] 2 …
Unexplained variance VAR( n1, n2) Å 1 0 r[ n1(s), n2(s)] 2

The first three quantities indicate deviations between in-

stantaneous speed predictions in units relative to the average
speed. Thus they are sensitive to the amount of fluctuations
in the observed speed profiles (i.e., relatively flat speed pro-
files will be easier to fit according to these measures). If
the speed profiles are not scaled to have unit average speed,
then these measures will be sensitive to the average speed.
The last measure corresponds to the percent variance (fluc-
tuations around the mean) in the observed speed profile not
explained by the predicted speed profile and is affected by
difference in shapes rather than scaling factors.
For each model, we computed the predictions over the
entire movement and over the middle 60% of the move-
ment ( i.e., starting at 20% and ending at 80% along the
path ) . The latter was done to assess the accuracy of the
power law away from the end points, where it is likely to
be inaccurate because it doesn’t have a mechanism for
enforcing 0 speed at the initial and final points of discrete
movements. We examined a number of speed profiles to
ensure that the 20% of the movement eliminated at each
end was enough to cover the parts of the speed profile
that seemed affected by initial acceleration and final decel-
eration.
The prediction for the original power law was obtained
directly from the equation n( s ) Å gk( s ) 01 / 3 . For paths
with inflection points, we added a small positive constant
to the curvature, which was held constant for all trials.
Thus the original power law has no free parameters: it
takes the path, and the duration of the movement, and
predicts a speed profile. The modified power law n( s ) Å
FIG . 4. Via point configurations used in task 3, drawn to scale. s, g[ k( s ) / e] b has two free parameters ( e, b ) . We opti-
starting point; *, end point. Every other configuration (right) is a 907 mized them ( with a nonlinear simplex method ) separately
rotation of the preceding one. for each trial and each measure of deviation defined above.
Although in the original formulation e depended on aver-
the cubic spline representation of the data for each trial, all age speed, we estimated it from the data to obtain the
subsequent steps (i.e., extracting speed and curvature) were most accurate fit the modified power law could provide.
applied directly, without any further filtering. The possibility of movement segmentation ( Viviani and
We then computed a speed profile prediction from the origi- Cenzato 1985 ) is addressed later.
nal power law, the modified power law, and the constrained The prediction of the constrained minimum-jerk model
minimum-jerk method (see further). To compute the differ- was computed using nonlinear optimization methods for
ence between each prediction and the actual speed profile, we variational problems ( see APPENDI X ). In its present form,
expressed all speed profiles as a function of are length s and the minimization requires the speed and acceleration at
scaled them so that the average speed was one and the path the end points ( in principle this could be avoided, by
lengths were equal. Scaling of path lengths was necessary minimizing over those parameters as well ) . For a discrete
because subjects did not follow the specified templates ex- movement, both the speed and acceleration at the end
actly, thus the paths on different trials had slightly different points are zero; because this is known a priori, and not
length. Alignment along s does not affect the power law pre- extracted from the experimental data, the model has no
dictions, which are local; it is necessary for our model because free parameters when applied to the entire movement. To
it is a global method and prediction errors early in the move- apply the model to the middle 60% of the movement, we
ment accumulate and cause misalignment towards the end if used the actual end-point speed and acceleration measured
speed is expressed as a function of time. experimentally. Note, however, that we are predicting
We defined four measures of deviation between two speed rather complex speed profiles, thus the two end points of
profiles n1(s) and n2(s) that are already scaled to have unit a submovement contain relatively very little information.
average speed; õrú signifies averaging over s, r(. , .) is In contrast, the modified power law uses the entire speed
a correlation coefficient. profile to extract the values of its free parameters.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 703

Results and discussion and the rotated configuration of via-points. The minimum-
jerk path ( ) remained invariant to rotation. Note that
In the first two tasks, for which the complete path was the deviations are not just isolated trials but stable strategies
specified and the movement was superimposed on it, subjects used by the subjects; also, the actual paths depended on the
could follow the specified paths accurately. In the 3D task, orientation of the template, contrary to the model prediction.
the majority of subjects could not replicate the specified Although some of these deviations may be the result of arm
shapes well and did not converge to a very stable movement dynamics affecting the path, the large differences (e.g., Fig.
strategy in 10 trials. After the experiment, subjects reported 5A) between the paths adopted by different subjects for the
that they were aware of their poor performance but did not same via-point configuration suggest that multiple via-point
know how to perform the task better. Although this is an tasks may have a significant cognitive component, i.e., sub-
important issue to consider, both the power law and the jects have to study the configuration and decide what general
constrained minimum-jerk model predict a speed profile path they want to take. In fact, we observed that on the initial
given the actual, not the desired path of the hand, so we still trials in a new configuration, subjects sometimes stopped and
can apply them and assess the accuracy of their predictions. corrected their movement, which is unlikely to happen if
It also was observed that the torsion t varied rather smoothly they already had planned a smooth path that passes through
instead of being nonzero only at a discrete set of points as the via-points in the specified order. Whatever the reason
suggested by Morasso (1983) and Soechting and Terzuolo for these discrepancies, we conclude that the original mini-
(1987). Deviations from piecewise-planar drawing in 3D mum-jerk model cannot account for the significant between
also have been observed by Gusis (1995); we did not ana- subject variability on multiple via-point tasks.
lyze this discrepancy further because it is not the focus of We now turn to a quantitative comparison of our model
the present study. and the power law. Figure 6 shows summary error statistics,
In the via-point task, we specified four intermediate points, for all tasks and all error measures, over the complete ( left)
and subjects were free to choose any path that passed through and middle 60% of the movement (right). For each subject,
them in the correct order. Our constrained minimum-jerk we computed the median prediction error in each block of
method cannot make a prediction about the actual path sub- 10 trials and then averaged over all templates in a given task.
jects choose in this task, but the original model (Flash and The constrained minimum-jerk model performed uniformly
Hogan 1985) predicts that the path will be the one that better than the modified power law, which in turn was uni-
minimizes jerk. While this is often the case, we found numer- formly better than the original power law (the latter obvi-
ous systematic deviations. Figure 5 shows data from six ously has to be the case). Analysis of variance showed that
different subjects; each pair of plots represents the last 5 of in the majority of the comparisons (the 3D task/complete
the 10 trials (rrr) that the subject made, on the original movements on the MAD and STD measures being the excep-

FIG . 5. Examples of paths traced in task

3. Each pair of plots is data from 1 subject
(6 different subjects shown). , mini-
mum-jerk path (which was not shown to the
subjects); rrr, actual paths on the last 5
trials of the 10 trial block. ●, end point.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
704 E. TODOROV AND M. I. JORDAN

FIG . 6. Summary error statistics with

standard error bars for the 3 models, com-
puted over the entire movement (left) and
the middle 60% (right) for all tasks and error
measures. For each subject, the median error
score in each block of 10 trials was com-
puted, and the results were averaged over all
templates in the task.

tion), the differences are significant at the P õ 0.05 level. difficult tasks the performance of the model changes with
Furthermore, this ordering holds for every subject we tested. experience, but there was simply very little learning in our
The difference is equally large over the complete movement experiments. Indeed the spatial error did not change signifi-
and the middle 60%, therefore it is not explained by the cantly over trials.
discrepant predictions of the power law at the end points of There are some differences between the four error measures
discrete movements. we defined. The maximum deviation penalizes the original
To analyze the dependence of the model performance on power law on templates with inflection points because the
the template being traced, we averaged the data over subjects predicted speed at 0 curvature is very large (it would be infinite
and compared the three prediction methods for each tem- if we didn’t add a small constant to the curvature). The mean
plate. Figure 7 (arm task—complete movements) shows absolute deviation and the standard deviation are very similar
that the ordering holds for all templates we tested and not up to a scaling factor. The unexplained variance seems to be
just the ones that include inflection points. This was true for the most sensitive measure—it is the only one that detects
all tasks. It is interesting that the constrained minimum-jerk significant differences in the behavior of our model on different
model shows a rather uniform level of performance, both tasks and templates. Thus it will be the measure used in the
over different templates and over different tasks. We ex- remaining part of the paper, where we study how the model
pected that there might be an effect of trial number, resulting depends on the details of the task. All subsequent analyses
from learning the pattern in the course of the 10 trials, but also have been performed with the mean absolute deviation
there was no evidence of that. It is still possible that in more measure, yielding very similar results.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 705

FIG . 7. Prediction errors for all error

measures and all templates in task 2—com-
plete movements, averaged over subjects.

Analysis of speed profiles tion for the average trial (which would be less meaningful
because both prediction methods are nonlinear).
Thus far we have considered statistics based on the aver- It is evident that the modified power law prediction has
age prediction errors. In this section, we perform a more significant systematic error, which is not restricted to the end
detailed comparison of the actual speed profiles. Figure 8 points of the movements. The flat speed profiles it predicts in
shows speed profiles of all eight subjects tracing templates template 2 correspond to a region of constant curvature (as
1, 2, 3, and 5 from Fig. 3. These data are from the next was noted in Fig. 2). We conclude that the power law cap-
experiment, where a short movement time (1.5 s) was en- tures very well the maxima and minima of the actual speed
forced—these are the conditions under which both our profiles but otherwise it is not a satisfactory model of the
model and the modified power law are most accurate (see details present in the experimental data. Although some of
further). We have scaled the last five speed profiles (com- the errors are due to skewing (as Wann et al. 1988 argued),
plete movements) in each block for each subject and aver- the predominant error is in the predicted values at the speed
aged them; the solid line shows the actual speed profiles; the extrema.
dashed line shows the constrained minimum-jerk predictions; The prediction of the constrained minimum-jerk method
and the dotted line shows the modified power law predictions. is much more accurate, and most of the remaining error
Because scaling does not align the speed profiles perfectly, seems to be due to between-subject variability. Still, there
averaging is not necessarily a meaningful operation, but it is some systematic error left. The first speed maximum on
helps identify systematic prediction errors. Note that we show template 1, and the second maximum on template 3 were
averages over the predictions for individual trials not a predic- both slightly underestimated for all subjects. Overall, the

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
706 E. TODOROV AND M. I. JORDAN

FIG . 8. Average speed profiles for all sub-

jects, small 1.5-s movement in experiment 2
(4 of the 5 templates shown, numbers corre-
spond to template positions in Fig. 3). Solid
line, actual speed profile; dashed line, con-
strained minimum-jerk prediction; dotted line,
power law. All speed profiles were scaled be-
fore averaging.

predictions of the model are rather close to the experimen- in Fig. 8) —although the modified power law adds a positive
tally observed speed profiles, and in some cases (last 2 sub- constant to the curvature (optimized for the particular trial),
jects in template 2, for example) it is hard to imagine that this mechanism is apparently not sufficient to deal with in-
any model would be more accurate. flection points. We might expect that if the paths included
Next we examine data from individual trials because the straight segments, the performance of the modified power
averaging procedure is likely to obscure some details. For law would be even worse. The inflection points are not the
templates 1, 2, 3, 5, and 8, we pooled all trials together and only problem, however, note the deviations in the Fig. 9,
found the one for which both error scores were as close as top right, which correspond to a path with strictly positive
possible to the corresponding medians. The actual ( ), curvature.
modified power law (rrr), and constrained minimum-jerk
( – – – ) speed profiles for the selected trials (middle 60%) Tempo fluctuations and segmentation
are shown in Fig. 9. We can see that there are large errors
in the power law prediction that are not related to end point Both the constrained minimum-jerk model and the modi-
acceleration (deceleration). Note the unnatural sharp peaks, fied power law assume that the tempo, or gain, or ‘‘psycho-
corresponding to inflection points (they were smoothed out logical speed’’ of the movement is held constant, and the

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 707

FIG . 9. Typical examples of single trial speed

profiles ( ), constrained minimum-jerk pre-
diction ( – – – ), and power law (rrr) on single
trials— experiment 2, middle 60% of the move-
ment.

variation in speed is only related to the path. Clearly that ject tracing template 1. We then applied the segmented version
doesn’t have to be the case; e.g., we can ask the subjects to of the modified power law to the data from all subjects on
speed up toward the end of the movement. It is conceivable the arm task—subset of templates used in experiment 2 (see
that even in the absence of explicit instructions to control further). Each movement was segmented at the speed maxima
speed, the tempo of the movement is variable. For each (corresponding to curvature minima). The portions between
model, we define the gain (or tempo) as the ratio between the end points and the closest segment boundary were elimi-
the actual and predicted speed profiles at each point in time nated. We applied the modified power law with segmentation
(the definition is model-dependent by necessity). Viviani with either fixed a Å 0.25 (which was found to be the optimal
and Cenzato (1985) suggested that the gain is a piecewise- fixed value) or a being optimized for each individual trial of
constant function of time: in each segment it is proportional each subject (making the comparison to our model somewhat
to a power function of the corresponding path length: g Å unfair). Figure 10 (bottom) shows the speed prediction error,
KP a ( a here corresponds to the parameter g defined by relative to the prediction error of the nonsegmented modified
Viviani and Cenzato). power law, for the two different segmentation methods. The
We computed the gain factor for a number of discrete improvement in all cases is much smaller compared with the
movements for both models and did not find any case in improvement of the (nonsegmented) constrained minimum-
which it was convincingly piece-wise constant. Figure 10 jerk method over the modified power law. This segmentation
(top) shows a typical plot five consecutive trials for one sub- method also was applied at speed minima (corresponding to

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
708 E. TODOROV AND M. I. JORDAN

FIG . 10. Top: gain fluctuations for the

modified power law and our model (consec-
utive movements of 1 subject tracing tem-
plate 1). Vertical lines correspond to points
of maximum and minimum speed. Bottom:
for each segmentation method (fixed a Å
0.25 or a optimized separately for each trial)
and our model, we show the prediction error
relative to the prediction error of the unseg-
mented modified power law applied over the
same portion of the path as the 1 selected
by the segmentation scheme.

curvature maxima—as in the model of Morasso and Mussa- speed profiles ( a weaker correlation was observed for the
Ivaldi 1983) yielding very similar results. Note that the opti- modified power law ) . Figure 11 shows a scatter plot of
mal fixed value of a Å 0.25 is significantly smaller than the mean movement duration ( over all templates ) and mean
value of a Å 0.6 observed by Viviani and Cenzato (1985). prediction error for our model in all tasks. There are several
Because a Å 0 corresponds to a lack of segmentation, this is possible interpretations of this result. It may be that the
another indication that segmentation is simply not present in tempo fluctuations accumulate with movement time, thus
our data, and the tempo fluctuations are continuous and at movements that take longer have less predictable speed
present unpredictable. The fluctuations in Fig. 10 are system- profiles. It is also possible that at high average speed, move-
atic because the trials are from the same subject, but they ments are for some reason more regular. Because we did
become unpredictable across subjects (as can be inferred from not vary the size of the templates systematically, duration
Fig. 8). was correlated with speed [ Wann et al. ( 1988 ) also ob-
Although the details of the gain fluctuations cannot be served that the power law holds better for faster movements
explained, their magnitude ( directly related to prediction and attributed the effect to speed instead of duration ] . An-
error ) could vary systematically with some other variable. other possibility is that the effect is a spurious between-
We already saw that it is not very sensitive to the task or subject correlation, i.e., some subjects naturally produce
the template being traced. A comparison of prediction error more predictable speed profiles, and the same subjects tend
over subjects, however, revealed a strong correlation be- to make faster movements when duration is not specified.
tween how rapidly subjects moved on the average and how The next experiment was designed to distinguish between
well the constrained minimum-jerk model predicted their these possibilities.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 709

FIG . 11. Scatterplot of movement dura-

tion vs. prediction error for the constrained
minimum-jerk model, all tasks. Each data
point represents 1 subject, the results were
averaged over all trials the subject made in
the corresponding task.

EXPERIMENT 2 (1, 2, 3, 5, and 8 in Fig. 3), each presented in three blocks

of 10 identical trials. In block 1, the template was large (same
To eliminate the possibility that between-subject variabil-
size as in experiment 1), and desired movement duration was
ity can account for this correlation, the experiment had a
3 s. In block 2, the template was also large, and duration
within-subject design. We now specified the duration of the
was 1.5 s. In block 3, the template was scaled down by a
movement and also varied the spatial scale, which allowed
us to distinguish between the effects of movement duration factor of 2, and desired duration was 1.5 s. Thus blocks 1
and average speed (and size). A new group of eight subjects and 2 were matched for size, blocks 2 and 3 for movement
was recruited for this experiment. duration, and blocks 1 and 3 for average speed (assuming
of course that subjects produced movements of the specified
durations, which they did) —see Fig. 12, top. At the begin-
Methods
ning of each trial, the subject positioned the end-point in the
The setup was the same as in task 2—subjects were asked starting area and waited for the start circle to disappear,
to make arm movements on a horizontal table by tracing a confirming a correct initial position. The computer sounded
continuous curve projected on the table. To reduce experi- a beep when the subject started moving and a second beep
mental time, we used an arbitrary subset of five templates when the subject stopped moving. After each trial, the actual

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
710 E. TODOROV AND M. I. JORDAN

FIG . 12. Top: schematic representation

of the 3 conditions in the experiment. Bot-
tom: summary error statistics for experi-
ment 2, complete movements (left) and
middle 60% (right). Results are computed
in the same way as in Fig. 6. Last prediction
method is the average speed profile over a
block of 10 consecutive trials used to pre-
dict each individual trial.

duration was shown (in units of 0.01 s). The desired duration Analysis of variance revealed significant differences for the
was printed on the screen throughout the experiment. Sub- constrained minimum-jerk predictions, and the modified
jects were told what the numbers meant and were instructed power law applied to the middle 60% of the movements;
to trace the paths for the specified durations. the prediction error for the large 3-s movement was larger
than the error for both the large 1.5-s and the small 1.5-s
Results and discussion movements (P õ 0.01). This result indicates that the accu-
racy of the prediction depends on the duration not the aver-
The data were analyzed in exactly the same way as in age speed or spatial scale of the movement. We also can
experiment 1. All subjects traced the paths accurately and conclude that the correlation in experiment 1 was not entirely
could adjust the duration of their movements in the first due to systematic between-subject variability because the
two to three trials in each block. Prediction error statistics, second experiment yielded the result within subjects. The
averaged over templates, are shown in Fig. 12. Again, we higher accuracy of our model when applied to the middle
found that the constrained minimum-jerk method predicts 60% instead of the complete movement may be due to the
speed profiles closer to the ones observed experimentally, fact that the model is being applied to a movement of shorter
compared to both versions of the power law. In this case, duration. However, this also may be caused by using the
the constrained minimum-jerk prediction was about three experimentally observed values of speed and acceleration at
times better than the modified power law and four times the end points of the middle segment.
better than the original power law. The difference is still It is possible that the effect of movement duration on
present when speed profiles are computed for the middle model accuracy is due to the fact that observed speed profiles
60% of the movements. in short-duration movements are somehow different/sim-
It can be seen in Fig. 12 that the prediction errors for the pler, and any model would yield better results. To assess this
large fast movements are always similar to the small ones alternative, we have included in Fig. 12 another prediction
and different from the errors for the large slow movements. method related to the amount of trial-to-trial variability pres-

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 711

ent in a block of identical trials. We computed the mean the speed profiles of reaching movements of different length
speed profile over each block of 10 trials (separately for and duration can be aligned almost perfectly after straightfor-
each subject), and used it to predict the speed profile for ward rescaling, and complete models of trajectory formation
each of the individual trials. The pattern of prediction errors such as the original minimum-jerk model or the minimum
we see in Fig. 12 is very different: it appears that trial-to- torque-change model (Uno et al. 1988) have been applied
trial variability decreases with speed rather than duration to the average trajectory.
and is essentially the same over the entire movement and Under these assumptions, we can attempt to identify the
the middle 60%. Thus the higher model accuracy for shorter level of processing where the relationship between path and
durations results from a stronger coupling between path and speed emerges. In particular, the smoothness observed in
speed rather than a difference in the speed profiles them- experimental data may be present already in the plan, in
selves. which case, our model should apply better to the average
trajectory than to single-trial trajectories (note that averaging
Trial-to-trial variability by itself does not necessarily produce smoother trajectories,
i.e., if we average several optimal trajectories, we will obtain
It often is assumed that the motor system constructs a a suboptimal one—because the model is nonlinear). Fur-
plan of the desired trajectory for the current task, and in- thermore, the speed profile predicted from the average path
stantiates that plan on each particular trial. Although there should be more accurate in explaining the single trial speed
is no widespread agreement on what exactly such a plan profiles than the predictions obtained from the single trial
might contain, it seems reasonable to assume that the actual paths. In the notation of Fig. 13 [where MJ (r) is the speed
trajectories produced on consecutive trials on the same task predicted by the model for a given path, õrú signifies
are variations around the same plan, where the variability is averaging over a block of trials, and : is the error between
due to unspecified sources of noise. Thus the average trajec- actual and predicted speed profiles). MJ(path) : speed ú
tory over a block of identical trials corresponds to the move- MJ( » path … ) : speed ú MJ( » path … ) : õspeedú. Alterna-
ment that would result from the plan in the absence of vari- tively, if smoothness is inherent in the (unknown) sources
ability. In support of this notion, it has been observed that of variability, we should expect the single trial predictions

FIG . 13. Comparison of prediction errors of

our model (on all 1.5-s movements in experiment
2) and possible reconstructions of a ‘‘central
plan.’’ MJ(path) : speed—our model, applied on
a trial-by-trial basis ( j ). õspeedú : speed—the
average speed profile over a block of 10 trials,
used to predict each individual speed profile.
MJ( » path … ) : speed—the speed profile generated
by our model for the average path and used to
predict each individual speed profile. MJ( » path … ) :
õspeedú —the speed profile generated by our
model for the average path and used to predict the
average speed profile. MJ(template) :
õspeedú —the speed profile generated by our
model for the specified template and used to pre-
dict the average speed profile. MJ(template) :
speed—the speed profile generated by our model
for the specified template, and used to predict each
individual speed profile.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
712 E. TODOROV AND M. I. JORDAN

to be more accurate than both average methods: MJ(path) : minimum-jerk model yields predictions that are generally
speed õ MJ( » path … ) : speed õ MJ(; » path … ) : õspeedú. similar to the power law, its performance is significantly
Because this analysis is most likely to yield clear results better than both the original and modified versions of the
when the model works best, we focused on the data from power law for all tasks and movement templates we studied.
all 1.5-s movements in experiment 2. As Fig. 13 shows, none Furthermore the model naturally extends the path-speed rela-
of the above orderings hold. It is true that the constrained tionship previously observed in extemporaneous movements
minimum-jerk model describes the average trajectory to include the acceleration and deceleration phases of dis-
better than individual trajectories: MJ(path) : speed ú crete movements, which are much more common. Thus it
MJ( » path … ) : õspeedú, therefore smoothness is present in can be argued that the local speed-curvature relationship
the plan and adding variability decreases the planned expressed by the power law, to the extent that it fits experi-
smoothness. However, the speed profile predicted from the mental data, can be derived from a more global smoothness
average path explains the individual speed profiles a lot constraint that relates the path to the speed profile.
worse than the single trial predictions: MJ(path) : speed õ In comparing any two models, we have to look not only
MJ( » path … ) : speed. In fact, even the average speed profile at their competence in accounting for experimental data, but
(which is the best possible constant predictor) performs also at the model complexity (Occam’s razor). We empha-
worse than the single trial prediction method: MJ(path) : size that model complexity has nothing to do with the com-
speed õ õspeedú : speed. Therefore, the trial-to-trial vari- plexity of the mathematics involved: instead it is related to
ability observed in speed profiles is partially predictable how many and what free parameters are used in fitting the
given the trial-to-trial variability in the path. Note that speed data. In that sense, the constrained minimum-jerk model and
predictions based on the template rather than the average the original power law are equivalent—they both postulate
path are a lot less accurate, suggesting that the spatial devia- that the speed profile of a movement is a well defined func-
tions from the specified template may be present in the plan. tion of hand path, and do not require any extra parameters.
In summary, the picture that emerges from this analysis The modified power law has higher model complexity be-
is the following: the relationship between speed and path is cause it involves a free parameter ( b ) extracted from the
present in the plan and is stronger than that observed on observed trajectories. Furthermore, all versions of the power
individual trials. However, the trial-to-trial variability is not law require extra mechanisms for dealing with regions of
at all ‘‘random’’; the path and the speed profile represented zero curvature and with zero speed at the end points of a
in the central plan are modified together in a way that pre- discrete movement. Thus we argue that the statistical com-
serves the planned smoothness. It is of course possible that plexity of the model presented here, when applied to discrete
a strict separation between planning and execution does not movements with inflection points or straight segments, is no
exist, and thus the above analysis yields mixed results. greater than any version of the 2/3 power law. We also have
to compare the difficulties involved in the possible biological
GENERAL DISCUSSION implementation of each model. From that point of view,
the power law may appear more appealing because it is
This paper has presented a constrained minimum-jerk
model of the relationship between hand path and speed pro- analytically tractable (i.e., speed can be obtained directly
file of complex arm movements. We applied the model to from curvature, without any numerical approximation). If
a wide range of spatially constrained motor tasks involving we believe that the biological system actually uses the con-
discrete movements of short durations and arbitrary paths. crete mathematical formula, this is certainly a big advantage.
The new model proposed here unifies two sets of experimen- However, we are dealing with models that only quantify
tal observations: the characteristic bell-shaped speed profile intrinsic relationships between path and speed instead of
of straight reaching movements, and the path-speed relation- specifying a recipe for trajectory formation. For example, it
ship in extemporaneous movements with complex paths. The has been suggested that a 2/3 power law may emerge out
speed predictions of our model are equivalent to those of of the complex nonlinear interactions of a large number of
the original minimum-jerk model when the path is straight direction-tuned neurons in primary motor cortex (Lukashin
and more accurate when the Flash and Hogan (1985) model and Georgopoulos 1993) —such an implementation does not
fails in the path prediction. Although our model uses the use the analytical tractability of the power law, eliminating
entire path (i.e., a continuous curve) to predict speed, it is any advantage it may have over the model presented here.
possible instead to represent the path as a large number of Contrary to previous investigations (Viviani and Cenzato
intermediate points and possibly tangents at those points— 1985) we did not find evidence for segmentation; instead
if such a representation contains sufficient information to the ‘‘tempo’’ seemed to fluctuate rather smoothly and unpre-
reconstruct the continuous path through some interpolation dictably. We think this is due to the different nature of the
method, this becomes equivalent to our model. In practice, movements studied here. Previous work on segmentation has
for the family of templates studied here, we have found that focused on repetitive movements of long duration, which
the entire path can be reconstructed from Ç10 points sam- implied (at least visually) a rather obvious segmentation
pled at equal distances. This does not imply any internal pattern. Instead we used movement durations of 1–2 s with
representation consisting of 10 discrete points—it is simply templates that are not obviously composed of a small number
a consequence of the mathematical fact that curves as smooth of simple curves concatenated together. It is possible that
as the ones used here can be reconstructed easily through discrete movements of such short duration are not internally
interpolation. segmented even when they have complex paths. Alterna-
We showed that although mathematically the constrained tively, the decomposition into segments may be variable

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
 SPEED PROFILE PREDICTIONS OF COMPLEX ARM MOVEMENTS 713

from trial to trial, preventing us from finding any evidence problem, which can be approached using a number of techniques.
for it. It turns out that the standard techniques for solving variational
The high accuracy of the constrained minimum-jerk problems do not work very well in this case.
model allowed us to make two important observations that From the Euler-Lagrange principle, the solution s(t) satisfies
the 6th-order nonlinear ordinary differential equation ÌL/ Ìs 0
have not received previous attention: the relationship be- d/dt ÌL/ Ìs * / d 2 /dt 2 ÌL/ Ìs 9 0 d 3 /dt 3 ÌL/ Ìs *9 Å 0, subject to
tween path and speed is stronger for movements of shorter the constraints on s(t) and its derivatives at the two end points of
duration and is not affected by spatial scale or average speed the movement. This problem is very difficult numerically because
(it is also rather uniform across tasks and movement paths) the solution depends on high order derivatives of the path, which
and the path of the hand on a particular trial contains signifi- is recorded experimentally and thus has some noise in it. We found
cant information about the speed profile on that same trial; that both shooting and relaxation methods for solving boundary
this information is lost if we only study the average path value problems (Press et al. 1992) fail.
over a block of identical trials. Another possibility is to minimize Eq. 1 directly by representing
Presently the model is applied to the complete path, yield- s(t) as a linear combination of N fixed basis functions fi (t): s(t)
ing a complete speed profile. It is unlikely that the motor Å ( iÅ1
N
ci fi (t) and using gradient descent with respect to the coef-
system maintains the path-speed relationship globally, espe- ficients ci . We used sixth-order B splines as basis functions and a
preconditioned conjugate gradient method to find the optimal coef-
cially over movements of very long duration. Our finding ficients (Gill et al. 1981). This minimization always converged
that prediction error increases with movement duration (and but frequently found local minima, and it was necessary to restart
not average speed or spatial scale) suggests that the ‘‘sliding it a number of times before a good solution could be found. An
window’’ over which the model applies best may be rather additional problem is that it is not clear what a good solution is.
small, i.e., Ç1 s. A more local relationship between path The method we used here is less direct but has the advantage
and speed is also consistent with observations that the 2/3 that it is faster and seems to converge to the same minimum regard-
power law applies in tracking tasks where little advance less of the initial conditions. Rather than using the complete path
planning of maximally smooth movements is possible. Thus r(s), we chose a set of 10 intermediate points, equally spaced along
it is desirable to modify the procedure for fitting the model the path (i.e., their positions contain no temporal information). We
so that it can be applied to segments of the path, without then found the minimum-jerk movement (in the sense of Flash
requiring end-point speed and acceleration (only duration and Hogan 1985) through those points, given the velocity and
acceleration at the two end points. If the resulting minimum-jerk
for that segment). This can be done by minimizing jerk over movement has a path very close to r(s), we are guaranteed that
these parameters as well. its speed profile is the solution to our original problem (i.e., we
Our examination of trial-to-trial variability suggests that are computing a minimum over a given set by minimizing over a
smoothness is centrally planned, and the variability of the superset and ensuring that the solution is in the original set). We
hand path and the speed profile are coupled in a way that describe the minimization procedure in the following text.
maintains the planned path-speed relationship. Because the It has been shown (Flash and Hogan 1985) that for given passage
movements we studied are not executed at maximum speed, times T, positions x, velocities n, and accelerations a at the end
one likely source of variability is the effect of on-line correc- points of one segment, the minimum-jerk trajectory is a 5th-order
tions. Such on-line corrections, however, have to be incorpo- polynomial in t, the coefficients of which can be determined easily
rated smoothly in the movement in a manner that is consis- using the end-point constraints. It is then possible to integrate the
tent with the overall path-speed relationship described by the squared jerk analytically, and sum it over all segments. The jerk
of the optimal trajectory is
constrained minimum-jerk model. Evidence for such error N01
correction mechanisms is provided by the double-step para- JT,x,n,a (T 1. . .N , x1. . .N , n1. . .N , a1. . .N ) Å ∑
digm, where a reaching movement is smoothly corrected to iÅ1

end at a perturbed target location if subjects are unaware of 3a 2i 0 2ai ai/1 / 3a 2i 3ai/1ni 0 2ai/1ni / 2ai ni/1 0 3ai/1ni/1
3 / 24
the perturbation (Pellison et al. 1986). It is also possible (Ti/1 0 Ti ) 4 (Ti/1 0 Ti ) 3
that low-level correction/stability mechanisms (e.g., some
4n 2i / 7ni ni/1 / 4n 2i/1 (ai 0 ai/1 )(xi 0 xi/1 )
version of equilibrium point control) combined with a source / 48 / 120
(Ti/1 0 Ti ) 2 (Ti/1 0 Ti ) 2
of noise may give rise to variations around the desired trajec-
tory that preserve its intrinsic properties. The desired trajec- ( ni / ni/1 )(xi 0 xi/1 )
/ 720 / 720 (xi 0 xi/1 ) 2
tory itself does not have to be generated by a servo control Ti/1 0 Ti
mechanism, but only the low-level on-line corrections to it. Note that the above expression is quadratic in all n and a and thus
Finally, we cannot rule out the possibility that the trial- can be minimized with respect to those parameters by setting the
to-trial variability we studied actually was planned. It is gradient to 0 and solving (numerically) the resulting system of linear
quite difficult to identify the levels of processing on which equations. Using this minimization, for given passage times T (and
different regularities emerge based on observational studies intermediate point positions x), we can compute the jerk for the
alone. The problem is that once a movement is executed, it optimal trajectory: JT1. . .N,x1. . .N Å (xÅX,Y,Z minn1. . .N,a1. . .N JT,x,n,a (T 1. . .N ,
is impossible to determine whether the subject had planned x1. . .N , n1. . .N , a1. . .N ). Finally, the function JT1. . .N,x1. . .N is minimized
to produce exactly that movement or the outcome was sig- over all possible T 2. . .N01 using a nonlinear simplex method. We
nificantly affected by on-line processes. Perturbative experi- found that the resulting path is equivalent to r(s), and the speed
mental manipulations will be required to address these issues profiles obtained in this way were indistinguishable from the speed
directly. profiles obtained with the direct method above, when it was run
enough times to find a global minimum.
APPENDIX. CONSTRAINED MINIMIZATION
Minimizing the integral of L Å ( k*sg 3 / 3ksg s̈) 2 / (s̈ 0 k 2sg 3 ) 2 We thank T. Flash, P. Viviani, E. Bizzi, and Y. Matsuoka for helpful
/ (sg 3kt ) 2 with respect to the temporal profile s(t) is a variational suggestions on the manuscript.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.
714 E. TODOROV AND M. I. JORDAN

This work was supported by Grant N00014-94-1-0777 from the Office MORASSO, P. Three dimensional arm trajectories. Biol. Cybern. 48: 187–
of Naval Research. E. Todorov is a Howard Hughes Predoctoral Fellow. 194, 1983
M. I. Jordan is a National Science Foundation Presidential Young Investiga- MORASSO, P. AND MUSSA IVALDI, F. A. Trajectory formation and handwrit-
tor. ing: a computational model. Biol. Cybern. 45: 131–142, 1982.
Address for reprint requests: E. Todorov, MIT, E25-526, 45 Carleton NELSON, W. L. Physical principles for economies of skilled movements.
St., Cambridge, MA 02139. Biol. Cybern. 46: 135–147, 1983.
PELLISON, D., PRABLANC, C., GOODALE, M. A., AND JEANNEROD, M. Visual
Received 29 September 1997; accepted in final form 20 April 1998. control of reaching movements without vision of the limb. II. Evidence
of fast unconscious processes correcting the trajectory of the hand to the
final position of a double-step stimulus. Exp. Brain Res. 62: 303–311,
REFERENCES 1986.
POLLICK, F. E. AND SAPIRO, G. Constant affine velocity predicts the 1/3
ATKESON, C. G. AND HOLLERBACH, J. M. Kinematic features of unrestrained power law of planar motion perception and generation. Vision Res. 37:
vertical arm movements. J. Neurosci. 5: 2318–2330, 1985. 347–353, 1997.
BENNETT, D. J., HOLLERBACH, J. M., XU, Y., AND HUNTER, I. W. Time- PRAK ASH, N. Differential Geometry. An Integrated Approach. New York:
varying stiffness of human elbow joint during cyclic voluntary movement. McGraw-Hill, 1981.
Exp. Brain Res. 88: 433–442, 1992. PRESS, W., TEUKOLSKY, S., VETTERLINC, W., AND FLANNERY, B. Numerical
BIZZI, E., HOGAN, N., MUSSA-IVALDI, F. A., AND GISZTER, S. Does the Recipes in C. New York: Cambridge Univ. Press, 1992.
nervous system use equilibrium-point control to guide single and SABES, P. AND JORDAN, M. Obstacle avoidance and a perturbation sensitivity
multijoint movements. Behav. Brain Sci. 15: 603–613, 1992. model for motor planning. J. Neurosci. 17: 7119–7128, 1997.
DE BOOR, C. A Practical Guide to Splines. New York: Springer-Verlag, SOECHTING, J. F. AND TERZUOLO, C. A. An algorithm for the generation of
1978. curvilinear wrist motion in an arbitrary plane in three dimensional space.
FLASH, T. AND HOGAN, N. The coordination of arm movements: an experi- Neuroscience 19: 1393–1405, 1986.
mentally confirmed mathematical model. J. Neurosci. 5: 1688–1703, SOECHTING, J. F. AND TERZUOLO, C. A. Organization of arm movements in
1985. three-dimensional space. Wrist motion is piecewise planar. Neuroscience
GILL, D., MURRAY, W., AND WRIGHT, M. Practical Optimization. New 23: 53–61, 1987.
York: Academic, 1981. UNO, Y., KAWATO, M., AND SUZUKI, R. Formation and control of optimal
GOMI, H. AND KAWATO, M. Equilibrium-point control hypothesis examined trajectory in human multijoint arm movement. Biol. Cybern. 61: 89–101,
by measured arm stiffness during multijoint movement. Science 272: 1989.
117–120, 1996. VIVIANI, P. AND CENZATO, M. Segmentation and coupling in complex move-
GRIBBLE, P. AND OSTRY, D. Origins of the power law relationship between ments. J. Exp. Psychol. Hum. Percept. Perform. 11: 828–845, 1985.
movement velocity and curvature: modeling the effects of muscle me- VIVIANI, P. AND FLASH, T. Minimum-jerk, two-thirds power law, and iso-
chanics and limb dynamics. J. Neurophysiol. 76: 2853–2860, 1996. chrony: converging approaches to movement planning. J. Exp. Psychol.
GUSIS, A. Arm Trajectory Planning in 3D Space (Masters thesis). Rehovot, Hum. Percept. Perform. 21: 2–53, 1995.
Israel: Weizmann Institute of Science, 1995. VIVIANI, P. AND SCHNEIDER, R. A developmental study of the relationship
HOGAN, N. AND FLASH, T. Moving gracefully: quantitative theories of motor between geometry and kinematics of drawing movements. J. Exp. Psy-
coordination. Trends Neurosci. 10: 170–174, 1987. chol. Hum. Percept. Perform. 17: 198–218, 1991.
JACK, W. R. On the analysis of voluntary muscular movements by certain WANN, J., NIMMO-SMITH, I., AND WING, A. M. Relationship between veloc-
new instruments. J. Anat. Physiol. 29: 473–478, 1895. ity and curvature in movement: equivalence and divergence between a
LACQUANITI, F., TERZUOLO, C., AND VIVIANI, P. The law relating the kine- power law and a minimum-jerk model. J. Exp. Psychol. Hum. Percept.
matic and figural aspects of drawing movements. Acta Psychol. 54: 115– Perform. 14: 622–637, 1988.
130, 1983. WOLPERT, D., GHAHRAMANI, Z., AND JORDAN, M. I. Are arm trajectories
LUK ASHIN, A. V. AND GEORGOPOULOS, A. P. A dynamical neural network planned in kinematic or dynamic coordinates? An adaptation study. Exp.
model for motor cortical activity during movement: population coding Brain Res. 103: 460–470, 1995.
of movement trajectories. Biol. Cybern. 69: 517–524, 1993. YALOV, S. Computational Models of Hand Trajectory Planning in Drawing
MORASSO, P. Spatial control of arm movements. Exp Brain Res. 42: 223– Movements (Masters thesis). Rehovot, Israel: Weizmann Institute of Sci-
227, 1981. ence, 1991.

/ 9k2b$$au17 J796-7
Downloaded 07-14-98 08:06:29
from [Link]/journal/jn neupa on February
([Link]) LP-Neurophys
23, 2025.