J Appl Physiol 96: 1978–1987, 2004.
First published December 19, 2003; 10.1152/japplphysiol.00872.2003.
HIGHLIGHTED TOPIC Neural Control of Movement
Simultaneous control of hand displacements and rotations
in orientation-matching experiments
Elizabeth B. Torres1 and David Zipser2
1
Division of Biology, California Institute of Technology, Pasadena 91125; and 2Department of
Cognitive Science, University of California, San Diego, La Jolla, California 92093-0515
Submitted 18 August 2003; accepted in final form 9 December 2003
Torres, Elizabeth B., and David Zipser. Simultaneous control of
hand displacements and rotations in orientation-matching experi-
ments. J Appl Physiol 96: 1978–1987, 2004. First published Decem-
ber 19, 2003; 10.1152/japplphysiol.00872.2003.—In reach-to-grasp
movements, the interaction between the hand changes in position and
those in orientation is poorly understood. A theoretical approach
previously proposed (Torres EB and Zipser D. J Neurophysiol 88:
1–13, 2002) assumes that motion strategies are resolved in space
independently from the temporal dynamics of the motion and predicts
the coarticulation of the hand transport and rotation along the path.
The model implies that this simultaneous control is independent of
variations in speed and initial posture and required matching orienta-
tion. This paper presents experimental data from human subjects that
confirm the model’s predictions in the context of realistic, uncon-
strained, orientation-matching motions. Speed independence is quan-
tified in the similarity of the postural and endpoint position-orientation
paths obtained under three different speeds. Significant differences in
hand and joint kinematics are shown in response to changes in initial
posture and target orientation. The robustness of coarticulation under
all three experimental conditions supports the idea of an intermediate
stage that resolves the geometry of the motion independent of its
temporal dynamics.
temporal dynamics; geometric stage; redundant degrees of freedom;
sensorimotor transformation
COGNITIVELY GUIDED ACTIONS such as reach-to-grasp are compu-
tationally difficult to solve. One of the main problems is to
translate sensory input into motor execution since they “speak”
different languages. Sensory input is expressed in terms of an
object’s location in space, orientation, shape, etc., rather than
the set of postures (or muscle configurations) required to grasp
it successfully. A translation is needed between these disparate
representations. This translation must resolve such problems as
the difference in the number of degrees of freedom (DOF) in
the space where goals are specified and the space where they
are resolved, and the choice of one path among the infinity of
possible paths.
To address this problem, we have previously proposed (25)
a geometric stage between the representation of perceived
goals and the execution of movements. This geometric stage
computes paths without regard to the temporal dynamics of
motion. It outputs a signal containing postural information of
the same kind used by the motor execution.
Address for reprint requests and other correspondence: E. B. Torres, Cali-
fornia Institute of Technology, Division of Biology, 333 Beckman Behavioral
Biology Labs, M/C 216-76, Pasadena, CA 91125 (E-mail: etorres@vis.
caltech.edu).
1978 8750-7587/04 $5.00 Copyright © 2004 the American Physiological Society
Decoupling the geometry from the temporal dynamics of the
motion enables us to simulate realistic behavior and generate
testable predictions. In the context of reach-to-grasp move-
ments, the model yields an important prediction: the coarticu-
lation between hand transport and orientation along the path.
This prediction challenges the intuitive belief that the system
may control the excess DOF by a “division of labor,” whereby
proximal and distal joints have independent control of the
transport and orientation phases, respectively. This scenario is
quite possible because, due to redundancy, changes in any of
the many joint angles of the arm may result in such an
independent control.
To test coarticulation, we need an experiment different from
pointing and not as complex as grasping that requires a change
in the orientation of the hand while transporting it to the target
location. Pointing alone is insufficient to specify orientation.
Grasping adds the problem of obstacle avoidance that arises
toward the end of the motion when the fingers are about to
come in contact with the object, which unnecessarily con-
founds the issues.
The need for a cleaner assessment of coarticulation is best
evidenced in the seemingly contradictory results from the grasp-
ing-related experimental literature. On the one hand, previous
work involving prehension motions suggests that, in reach-to-
grasp movements, distal DOF, such as those spanned by the wrist,
are primarily used to shape and orient the hand, whereas proximal
DOF, such as those related to the shoulder and elbow, are used
independently to transport the hand to the target location (2, 15,
17). On the other hand, data from humans and nonhuman primates
in response to orientation matching suggest that both the orienta-
tion and the location of targets in space affect the posture of the
arm as a whole (14, 22).
Present computational models of motor control cannot simulate
unconstrained motions in three dimensions with redundant DOF
(11, 13, 23, 26), so there is no ground for comparison in this arena,
and experimental data asking the specific coarticulation question
that our model raises do not exist. This motivated the design of an
experimental task that requires the kind of orientation-matching
movement that we are able to simulate with geometry. In addition,
we systematically manipulate important movement parameters
such as speed, initial arm posture, and target orientation. Under
these conditions, we question the robustness of coarticulation and
compare human behavior to the model predictions concerning
The costs of publication of this article were defrayed in part by the payment
of page charges. The article must therefore be hereby marked “advertisement”
in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
http://www.jap.org
 COARTICULATION OF HAND TRANSPORT AND ORIENTATION
postural and endpoint paths. Of special interest is the behavior of
the postural paths under these conditions, because they have not
been systematically studied. The data we present in this paper
provide another layer of information concerning parameters that
might be useful to other models of motor control, as well as to the
field of robotics devoted to the modeling of arm manipulators.
Theory and background. The theoretical aspects of this
particular task have been presented elsewhere (25). The cost
optimized is the general notion of distance (the norm) in a
nonflat (Riemannian) space X defined by the goals of the task
冑兺 兺
m m
(r ⴰ f) task ⫽ g␣ij 关xitarget ⫺ fi共q兲兴关xjtarget ⫺ fj共q兲兴
i⫽1 j⫽1
The target for action determines the dimensions of the points
in this space. They are defined as functions of postural-related
parameters such as joint angles in Q space. One of the forms
that this geometric cost may adopt in reach-to-grasp with
orientation matching has been described (25), but other new
goal-directed behaviors have been simulated and compared
with data with new related geometric costs (24). This general
notion of distance defined by the task in X is preserved in
joint-angle space by a geometric pullback action on the metric
tensor G␣x of X into Q defined as G␮ T ␣
q ⫽ J Gx J where J is the
Jacobian matrix of the f map (the metric tensor is the first
fundamental form, a symmetric-positive-definite matrix whose
coefficients, the gij’s, are used in the general definition of the
line element for curved spaces). Preserving the notion of the X
distance in Q makes the differential map of the reach-to-grasp
example df: TqQ 傺 R7 3 Tf(q)X 傺 R4 an isometry and
guarantees that the vector flows generated in both tangent
spaces TqQ and Tf(q) X are locally the shortest-distance seg-
ments between the two current points. The proof of this
statement is beyond the scope of this report, but its significance
is that we can generate the straight line paths (the geodesics) of
any task space X defined by a set of goals and transfer this
notion to the higher dimensional space of joint angles Q.
The fact that these paths are geodesics in both spaces makes
them reparameterizable in time (9) and thus speed independent.
This is how we propose that it is possible to learn new motions
in the cognitive-strategy space before mastering the optimal
temporal profiles that ballistic/automatic motions show later.
One distinction between our approach and the rest of the
field is that this geometric construction enables us to identify
the subspace of joint-angle space Q where the motion dictated
by X is resolved. Thus we can isolate the DOF that are relevant
to the task in question from those that are redundant. The
significance of this will be highlighted later in the data from the
orientation-matching motions.
Other models of path generation (11, 26) assume the system
already “knows” the best path (or trajectory) before movement
initiates. Movement duration is also assumed to be known (for
a review, see Ref. 28). The only information needed in our
approach is the target of the task and the starting configuration
of the system (initial posture in the example studied here). This
information autonomously generates the solution strategy. Op-
timal control models of the kind above may capture the nature
of automatic movement, i.e., movements that have already
been learned, as in ballistic reaching. In contrast, our method
aims at characterizing the ability of a system to develop the
spatial solution to a given task online during learning.
J Appl Physiol • VOL
1979
Unlike other models (20, 27), this one does not need any a
priori information or built-in heuristics to generate a solution
path. There is no need to know the final posture as in Ref. 20,
nor is there a need to remember or store any information about
posture space.
Another key distinction is that the proposed framework does
not enforce a strict separation between kinematics and dynam-
ics in a serial way, whereby dynamics accommodates a previ-
ously computed kinematics trajectory. Kinematics models have
a time profile to resemble the bell-shaped profiles of ballistic
motions. Our approach does not produce a temporal profile.
Actual data show that, as the system tries to uncover a strategy
in response to a new task, the speed profiles change signifi-
cantly, whereas the hand paths remain similar (24). Unfortu-
nately, most present data come from overlearned behaviors, so
this fundamental feature of motor learning, which our theory
predicts, goes unnoticed.
Both the kinematics and the dynamics of motion have
parameter spaces with certain geometry. Some tasks emphasize
kinematics goals, whereas others emphasize dynamics-related
goals. What we provide is a mechanism to build geometric
strategies, which translate the cognitive conceptualization of
such goals into motor-like signals. We propose that this form
of geometric simulation stands between perception and action.
METHODS
Apparatus and experimental design. We needed to record the
position and orientation of the hand to validate the model’s predic-
tions. This was achieved by using a Polhemus Fastrak Motion Track-
ing System. This system uses electromagnetic fields to determine
three-dimensional position and orientation of up to four markers
relative to the stationary system at an update rate of 30 Hz each. Other
than the hands, we also recorded the forearm and upper arm motions.
This enabled us to recover the seven joint angles used to represent
postures and extend the analysis to paths in this space.
All receivers except the shoulder’s were mounted on a piece of
Plexiglas and affixed to the arm with tape. The shoulder receiver
located on the acromion was mounted on plastic and attached to a
harness that kept it from sliding during the motions. The locations of
the markers were ⬃5–7 cm from the corresponding joint, depending
on the anatomic measurements of each subject (Fig. 1A).
The apparatus we built for these experimental sessions is shown in
Fig. 1B. Each target was labeled with a number next to it and
consisted of a wooden cylinder 2 cm in diameter and 18 cm in length.
The orientations used were those which were comfortable to match,
ranging from 0 to 90°, where 0° corresponds to the principal axis of
the cylinder being horizontally oriented and 90° to the principal axis
of the cylinder being vertically oriented. The principal axis and the
center of mass of the targets were marked with a red stripe and a red
dot, respectively. The position and orientation distances of the targets
relative to the hand-starting configuration are shown in Table 1. The
basic task was to match the given position and orientation of a target
in front of the body. To this end, subjects held a cylinder in their right
hand and displaced it to match the given target. This required the
displacement and rotation of the hand toward the target describing
both a position and an orientation path.
Our model predicts that the direction of movement should be
decoupled from its speed. To question the consistency of behavioral
data with this idea, we investigated coarticulation under different
speeds. The model also predicts an effect on the resulting motion due
to different initial postures or to the required orientation. We varied
these parameters as well and questioned the robustness of coarticula-
tion under these experimental conditions.
96 • MAY 2004 • www.jap.org
1980 COARTICULATION OF HAND TRANSPORT AND ORIENTATION

Fig. 1. A: positions of the markers. B: exper-

imental apparatus used to hold the target
cylinders at different locations and orienta-
tions in space. C: target-holding device used
to rotate the cylinders at desired orientations
and to move them away from the frontal
plane.

Speed condition. The experimental session began with a recorded
instruction indicating the target number. This command was imme-
diately followed by a beep, which served both as a cue to remind the
subjects of the desired movement duration and to indicate that at the
end of the beep the motion was to begin. The length of the beep was
proportional to the desired movement time: fast (100 ms), normal (400
ms), and slow (700 ms). However, subjects adjusted the duration of
the movement to their own comfortable pace for each speed kind. A
summary of mean movement duration is shown in Table 1. There was
enough time between the recorded instruction and the beep so that
subjects could visually locate the target. The target number-speed
combinations were called at random by a computer. At the end of the
movement, subjects waited at the target for a short beep, which
indicated the recording was over and they could move their arm back
to the starting position. All movements were performed within 110 cm
of the stationary system’s origin. There was only one practice trial so
that subjects could become familiar with the instructions. We used six
targets positioned and oriented differently, six subjects, three different
speeds, and six repetitions per speed-target combination. For statisti-
cal analysis, we divided the data set for each target (108 paths per
target) into three sample groups, corresponding to the three speed
groups. Each sample group had 36 trajectories. Each target was
analyzed independently.
Initial posture. The question of how the initial posture affects
coarticulation was assessed with the same paradigm, but in this case
subjects started the movement from one of two postural configurations
corresponding to the same initial hand position. In one posture
(normal), the subject’s upper arm was at a comfortable configuration
and the palm of the hand was facing toward the left with the thumb
pointing upward (Fig. 2A). In the other posture (abducted), the upper
arm was abducted and the palm of the hand was facing toward the
right with the thumb pointing downward (Fig. 2B). These movements

Table 1. Summary of distance and movement time for the speed experiment
Target 1 Target 2 Target 3 Target 4 Target 5 Target 6

Pos dist, cm 101.99 99.30 96.29 83.02 39.43 60.03

Or dist, deg 105.83 107.79 126.12 101.76 65.49 93.60
Slow, ms 1,471.3⫾74.4 1,427.8⫾73.1 1,503.7⫾91.8 1,376.9⫾82.9 1,292.6⫾94.1 1,219.4⫾75.8
Normal, ms 1,037.0⫾22.6 991.7⫾66.1 1,021.3⫾27.8 925.0⫾28.8 952.8⫾81.0 854.2⫾22.2
Fast, ms 683.9⫾18.36 680.2⫾64.7 665.6⫾25.8 638.3⫾24.3 628.0⫾31.7 614.8⫾11.7
Values are means ⫾ SE. Straight-line distance from the hand to the targets (Pos dist) was obtained using the Euclidean norm in 3-dimensional space.
Orientation distance (or dist) was obtained from the discrepancy between the hand and the object’s principal orientation axis as described in Ref. 25. Target
locations and orientations were similar across experiments.

J Appl Physiol • VOL 96 • MAY 2004 • www.jap.org

 COARTICULATION OF HAND TRANSPORT AND ORIENTATION 1981
Table 2. Effect of speed, initial posture, and target c ⫽ Vref/Vmax and a ⫽ dref/d to compute the normalized velocity
orientation on the hand position and orientation paths vNorm(t) ⫽ c 䡠 vData[(c/a)t], from the hand sensor data velocity vData where
Vmax is the maximum speed, d is the experimental movement distance, t
Experiment Target 1 Target 2 Target 3 Target 4 Target 5 Target 6 is time, and Vref and dref are arbitrary reference velocity and distance
values, respectively, used to scale the data set uniformly. Figure 3A shows
Speed
Pos ⌳mean 0.8928 0.8773 0.9303 0.9700 0.9282 0.9132 the raw speed profiles for three sample motions to one target and their
Or ⌳mean 0.9674 0.9114 0.9383 0.9440 0.9672 0.9345 corresponding profiles normalized for time and distance.
Init Post For statistical comparison, the sensor position and orientation
Pos ⌳mean 0.2388 0.4406 0.5291 0.2477 0.4118 0.2058 outputs were resampled at equal time intervals with a fixed number of
Or ⌳mean 0.4061 0.6150 0.3944 0.3704 0.4372 0.3413 points per path (50 points). Figure 3B shows this for the hand sensor
Targ Or output of three representative paths, one per speed.
Pos ⌳mean 0.4440 0.4606 0.4985 0.3852 0.3485 0.3770 For each experimental condition, we performed standard multivar-
Or ⌳mean 0.2795 0.3325 0.2663 0.3248 0.3243 0.2906
iate analysis of variance on the position, orientation, and postural
Values are from the Wilk’s lambda test, critical values ⌳␣,p,vH,vE Experiment
: paths to test for significant differences as a function of speed, initial
⌳0.05,3,2,105
Speed
⫽ 0.885, ⌳0.05,3,1,70
InitPosture
⫽ 0.889, ⌳0.05,3,1,70
Target Or
⫽ 0.889, vH ⫽ k ⫺ 1, and posture, or required target orientation.
vE ⫽ k(n ⫺1) are the degrees of freedom (DOF) for hypothesis and error, We used the Wilks’ test statistic (19) for the individual path points.
respectively (k ⫽ 3, n ⫽ 36). This statistic is similar to the univariate F test, which compares the
between sum of squares (expressed in matrix H) to the within sum of
were performed at a normal, comfortable pace. As in the speed squares (expressed in matrix E). However, E and H convey multivar-
condition, we used 6 targets, 6 (different) subjects, and 6 repetitions iate information. The ratio ⌳ ⫽ determinant of E/determinant of (E ⫹
for a total of 72 paths/target. For consistency, the initial posture was H) used by this test is written in terms of the within sum of squares
enforced at the beginning of each trial in two ways: 1) arm posture and products matrix E and the total sum of squares and products
was reconstructed from the sensor data and represented with MAT- matrix (E ⫹ H). The ⌳ ratio uses the determinants of E and H to
LAB three-dimensional graphics we developed and 2) physical mark- examine the separation of mean vectors. This reduces their multivar-
ers in the setup were used to assist the subjects each trial. This was iate information to a single scalar.
particularly useful in the abducted posture. We have evidence to reject the null hypothesis when ⌳ ⫽
Target orientation. In a separate experimental section, we in- ⌳␣,p,vH,vE, where ␣ ⫽ 0.05 is the level of confidence, p is the number
structed the subjects to match the orientation of a target cylinder in of variables or dimensions (e.g., p ⫽ 3 for paths in 3 space or p ⫽ 7
two different ways coming from the same position. These motions for paths in 7 space), and vH ⫽ k ⫺ 1 and vE ⫽ k(n ⫺ 1) are the DOF
were carried out at a comfortable pace. For those targets whose for hypothesis and error, respectively. The corresponding critical
principal orientation axis was closer to the horizontal, we instructed values for⌳␣,p,vH,vE can be found in Table A. 9 in Ref. 19.
the subjects to match the principal orientation axis either as if they
were approaching it from above with the palm facing away from the RESULTS
subject (Fig. 2C) or from below with the palm facing toward the
subject (Fig. 2D). For those targets whose principal orientation axis was Changes in speed. In this section, we quantify the speed
closer to vertical, we instructed the subjects to match it either with the independence of orientation-matching paths for human arm
palm facing left and the thumb pointing up (Fig. 2E) or with the palm motion. Several aspects of the kinematics of movement are
facing right and the thumb pointing down (Fig. 2F). We obtained the known to be invariant to changes in speed for pointing move-
same number of paths per target as in the previous condition.
Proper written consent from each subject was obtained for these
ments to targets constrained to the saggital plane (3) and to
experiments, which was approved by the University Human Subjects targets distributed across the three-dimensional space (18)
Committee. Twelve right-handed subjects (5 men and 7 women) where the wrist was constrained. It is unknown how these
participated in the recording sessions (6 in the speed experiment and results extend to more complex orientation-matching motions
6 in the other 2 experiments). They were not aware of the purposes of toward targets positioned and oriented differently across the
the experiments. None of them had any motor or visual impairment. workspace, where the arm is unconstrained. It is also unknown
Data processing. To extract the motion from the sensors output, we how the postural paths behave under these conditions. Other
defined the beginning and end of the movement as 5% maximum than the end-point position and orientation paths, we analyzed
velocity along a speed profile. For each trajectory, we determined the
points where the velocity drops to 5% of the maximum and eliminated
the motion of seven joint angles of the arm. Two questions of
the data beyond those points. interest are whether, in orientation matching, the kinematic
Overall, movement speed ranged from 30 (minimum) to 240 cm/s aspects of movement still remain independent from changes in
(maximum) across subjects. Movement velocity was normalized for time speed and whether there is coarticulation between changes
and distance as in Ref. 3. We use their time and distance scaling factors: in position and orientation of the hand. In addition, we

Table 3. Effect of initial posture on the arm configuration

2 Init Postures Target 1 Target 2 Target 3 Target 4 Target 5 Target 6
⫺4 ⫺4 ⫺4 ⫺4
Arm plane (␪) 16.35 (5.7⫻10 ) 10.22 (0.05) 12.38 (6.4⫻10 ) 16.97 (8.3⫻10 ) 16.82 (1.5⫻10 ) 19.40 (2.8⫻10⫺4)
Final arm plane 2.32 (0.01) 4.87 (0.003) 1.85 (0.05) 4.50 (0.008) 2.24 (0.02) 3.31 (0.04)
Posture (7 DOF) 0.005⫾0.012 0.023⫾0.08 0.0137⫾0.003 0.0049⫾0.014 0.0093⫾0.005 0.0496⫾0.01
Final posture (7 DOF) 0.034 0.097 0.0064 0.047 0.0307 0.025
For the plane of the arm, values are from one-tail critical t-test for 2 groups. The t and corresponding P values shown in the first row were averaged across
the path. In the second row, the corresponding values for the plane of the arm in the final arm configuration are shown. For the whole posture, the third and fourth
rows show the results from the Wilk’s lambda test computed for 7 DOF averaged across the 50 points in the path and for the final arm posture, respectively
(⌳*0.05,7,1,70 ⫽ 0.75).

J Appl Physiol • VOL 96 • MAY 2004 • www.jap.org

1982 COARTICULATION OF HAND TRANSPORT AND ORIENTATION

Table 4. Effect of instructed target orientation on the arm configuration

2 Target Orient Target 1 Target 2 Target 3 Target 4 Target 5 Target 6
⫺4 ⫺4 ⫺4
Arm plane (␪) 12.07 (1.2⫻10 ) 7.04 (0.0015) 14.09 (1.3⫻10 ) 7.09 (0.03) 14.61 (2.9⫻10 ) 9.17 (0.001)
Final arm plane 10.38 (2.8⫻10⫺4) 10.5 (5.3⫻10⫺4) 8.11 (0.005) 11.1 (2.3⫻10⫺4) 5.47 (0.04) 8.24 (2.11⫻10⫺4)
Posture (7 DOF) 0.098⫾0.0011 0.0306⫾0.0016 0.016⫾0.02 0.049⫾0.015 0.0139⫾0.012 0.0049⫾0.004
Final posture (7 DOF) 0.027 0.097 0.0516 0.019 0.031 0.096
Data are as in Table 3.

investigated whether coarticulation is robust to the speed ma-
nipulations.
Under three different speed conditions, we found similar-
ities in the position-orientation hand paths and on the
postural excursions of the arm (Fig. 4). Of particular interest
in this study is the postural behavior, because it has not been
previously analyzed in the context of a model that generates
paths in posture space. If, as we conjecture, the system
computes these paths geometrically, they should be inde-
pendent of the speed of the motion, so our model can be

Fig. 2. Experimental conditions. A: normal (comfort-

able) posture corresponding to the initial hand position.
B: abducted posture corresponding to the same initial
hand position. C–F: instructed hand orientations used to
match the given targets for both near-to-horizontal and
near-to-vertical tilts. C: match the given target orienta-
tion as if coming to it from above. D: match the given
target orientation as if coming to it from below. A
near-to-vertical target can be matched with the palm of
the right hand facing left (E) or with the palm of the
hand facing right (F).

J Appl Physiol • VOL 96 • MAY 2004 • www.jap.org

 COARTICULATION OF HAND TRANSPORT AND ORIENTATION
used to generate such paths, as shown in the examples used
by Torres and Zipser (25).
To address the question of homogeneity of the mean paths
for each one of the three speed conditions, we used a ratio from
standard multivariate analysis of variance, the Wilk’s ␭ test.
The ␭ values obtained were well above the critical value:
⌳*0.05,3,2,105 ⫽ 0.885 for the three x, y, z parameters in the
positional paths; ⌳*0.05,4,2,105 for the four quaternion-compo-
nent parameters in the orientation paths. In addition, the anal-
ysis corresponding to the seven joint angles yielded similarity
of the three group means, ⌳*0.05,7,2,105 ⫽ 0.79, where ␯H ⫽ k ⫺
1 ⫽ 2 and ␯E ⫽ k(n ⫺ 1) ⫽ 105 are the DOF for hypothesis
and error, respectively, for k ⫽ 3, the number of speed
conditions, and n ⫽ 36 for six subjects and six repetitions.
Table 2 summarizes these statistics. The speed profiles of these
motions were unimodal across speeds and similar according to
the criterion used by Atkeson and Hollerbach (3). In all the
analyses we performed, we failed to reject the null hypothesis
concerning similarity of the means.
We also found that coarticulation of the hand changes in
position and orientation across the path for all speed conditions
(Fig. 5A). Together with the path similarities, this result ac-
counts for a single geometric strategy robust to changes in the
temporal dynamics of the motion.
Manipulation of the initial posture. The arm paths generated
by the modified gradient method resemble behavioral paths in
response to orientation-matching motions, and, as such, several
J Appl Physiol • VOL
1983
Fig. 3. A: representative raw speed profiles
from the hand sensor output of 1 subject’s
motion to 1 target. Their corresponding pro-
files normalized for time and distance are
shown at right. B: raw hand sensor positional
data for 3 representative trajectories of the
slow, normal, and fast movements. Right: the
same spatial paths resampled at equal time
intervals with 50 points each.
of its features compare to those known from actual arm
movements (25). For example, simulated motions of this kind
reflect the dependence of final posture from the starting posture
observed in pointing (21) and reach-to-grasp movements (8).
Here, in an orientation-matching task, we used two different
initial postures (normal and abducted) to address two ques-
tions: 1) whether, besides the known effect on the final posture,
the initial posture has a significant effect on the whole postural
and hand paths; and 2) whether coarticulation is robust to this
manipulation.
The use of an abducted initial posture requires more effort to
solve the movement. Because of redundant DOF, there are
several possible strategies for coping with this difficulty. For
instance, subjects could first reset the posture to a normal
starting configuration and then move to the target. This first
part of the movement would require traversing a path in
posture space that does not result in motion of the hand. The
experimental data, however, show a strategy that is consistent
with the model’s output: subjects’ movement from the given
starting configuration in posture space results in hand motion
that continuously decreases the remaining distance to the target
(Fig. 5B).
The statistical analysis of the kinematics data due to the two
different starting postures shows a significant effect on the
resulting hand positional and orientation paths (Table 2), as
well as on the postural paths and on the plane of the arm (as
defined in Ref. 18) (Table 3). In contrast, consistently for each
96 • MAY 2004 • www.jap.org
1984 COARTICULATION OF HAND TRANSPORT AND ORIENTATION

Fig. 4. Graphical representation of the averaged data for se-

lected targets. Cylinders at the beginning and end of the paths
indicate the starting and target position orientation, respec-
tively. A: 95% confidence region around the mean hand position
path for each speed group. The averaged paths for the other 2
speed groups are plotted inside the confidence region corre-
sponding to a given speed group. The idea is to identify the
areas along the path where the groups differ significantly. The
region around each mean path is built pointwise in the follow-
ing way. The inequality for the Hotelling’s distribution {[(v ⫺
p ⫹ 1)/vp]a⬘Sa ⬍ Fp,v ⫺ p ⫹ 1,␣}, where v ⫽ n ⫺ 1, p ⫽ 3, n ⫽
36 at the ␣ level used to estimate the region where S ⫽ E/(nk ⫺
k) is an estimate for the sample variance-covariance matrix, E is
the within sum of squares and products matrix used in the
computation of the ⌳ ratio (19), K is the number of experimen-
tal conditions, and F refers to the F-distribution. The term a⬘Sa
is the Mahalanobis distance from a sample point to the mean.
We generate N random vectors a around each point in the mean
path of a group to simulate errors in a normal trivariate distri-
bution. The radius around the mean is chosen as the maximum
distance from all points in the sample to the sample mean ⫾⑀
close to (0, 0, 0). The inequality above is evaluated, and values
above the F3,34,0.05 critical value are rejected. The maximum of
the values we accept is set as the at least 95% confidence
circular region around the mean point in the speed-group path.
Each one of the circles is connected to form the region, and the
mean paths from the other 2 groups is plotted. As the figure
shows, for each one of the groups, the paths from the other 2
speed groups fall inside the confidence region. B: the 7-joint-
angle mean paths per speed condition, which reconstruct the
sensor data. They correspond to movements of 1 target for 1
subject averaged across 6 repetitions. On the horizontal axis,
there are 50 resampled points per path (see METHODS). On the
vertical axis, joint-angle measurements are in degrees.

subject, the coarticulation of hand changes in position and
orientation holds across the movement paths to all targets
independent of changes in the initial posture. Figure 5B shows
this result for one subject’s motions to one target. A similar
result was obtained across all subjects. This suggests a unique
cognitive strategy based on extrinsic goals, such as those
related to distance.
Change in target orientation. Previous data from humans (22)
and nonhuman primates (14) have shown that both the orientation
and the location of targets in space affect the final posture of the
J Appl Physiol • VOL
arm. This raises the question of whether these results extend to the
entire postural path. According to our model, motions that begin
at the same hand configuration but require matching different
orientations at the same space location should change the postural
and hand paths but not the coarticulation of the changes in hand
position and orientation along the path.
Statistical analysis on the subject’s endpoint paths (all ␭ well
below critical value at the 0.05 ␣ level; see Table 2) and
postural paths, as well as the plane of the arm (see Table 4),
reveal differences as a function of required target orientation.
96 • MAY 2004 • www.jap.org
 COARTICULATION OF HAND TRANSPORT AND ORIENTATION 1985

Fig. 5. The ␣ ratio or coarticulation parameter plotted across the path for resampled motion to a target under each experimental
condition (see METHODS for explanation of the resampling procedure): 3 speeds (A), 2 initial postures (B), and 2 required target
orientations (C). Each point represents the remaining orientation squared distance (x-axis) vs. the remaining squared positional
⫺1
distance (y-axis). The orientation distance on the x-axis is computed at each time step using the matrix A ⫽ Otarget 䡠 H, where O
and H are the orientation matrices for the target object and the present hand location, respectively. They come from the sensor data.
The matrix A thus computed reflects the discrepancy between these 2 orientations at each time step. The parameter ⌽ comes from
cos⌽ ⫽ 1⁄2Trace (A) ⫺ 1 in the angle-axis parameterization of A. The remaining positional distance on the y-axis is obtained using
the usual Euclidean norm between the hand xyz position and the target location at each time step. We use the square of the distances
in each axis. At 0, the hand is on the target at the correct orientation. The slope of 1 in each plot is the coarticulation ␣ parameter.
D: parsing the degrees of freedom relevant to this task: the joint angle increment vector at each step in time is obtained from the
postural data paths. This is the data gradient vector, an element in the tangent space to Q, TqQ. The matrix Gx␣ is estimated from
the data and the model using a least-squares method similar to that explained by Torres and Zipser (25) to generate the
posture-dependent symmetric-positive-definite metric tensor G␮ T ␣
q ⫽ J Gx J using also the Jacobian coordinate transformation matrix
from the map f in the model. The orthonormal basis from the singular value decomposition of G␮ ␮
q , SVD(Gq ) ⫽ V DV is then used
T

so that the dot product between the data gradient vector and each column vector Vj, j ⫽ 1 . . . 7 from the orthonormal basis V is
computed at each time step. The 4 smooth lines capture the path in the 4-dimensional subspace of Q (which is a copy up to distance
of X) were the degrees of freedom relevant to the task change. This is the solution subspace for this particular task. The space X
has been (isometrically) embedded in Q so that the originally noninvertible many-to-one map becomes a 1-to-1 linear map operating
on the tangent spaces to both manifolds, thus solving the redundancy problem. The 3 nonsmooth remaining paths reflect the
subspace of redundant parameters.

In addition, coarticulation generalized across target locations
and orientations (Fig. 5C), even for those motions requiring
target orientations that resulted in an uncomfortable final
posture.
DISCUSSION
We have previously proposed a novel approach to move-
ment control whereby a geometric stage mediates the transfor-
J Appl Physiol • VOL
mation from sensory input to motor execution. The motivation
stems from two empirical sources: 1) neurophysiological data
from posterior parietal cortex (PPC), which suggests a dynam-
ics-free representation of movement (16) along with evidence
for multisensory integration and correlates of coordinate trans-
formations (1); and 2) experiments involving humans as evi-
dence for speed-independent paths during pointing in two (3)
and three (18) dimensions.
96 • MAY 2004 • www.jap.org
1986 COARTICULATION OF HAND TRANSPORT AND ORIENTATION
Our solution is based on a differential approach, so at each
time step, it requires input regarding present target1 informa-
tion combined with present posture to compute the hand-target
discrepancy. This input can but does not need to be continu-
ously monitored. Offline integration of the gradient vector for
n steps is also possible to compute the full or partial path to the
target for action. In addition, estimation of target information is
feasible without actual access to it. There is evidence that in
PPC neurons can be activated by the plan to use a specific
effector without spatial information (6).
The hand-target discrepancy vector, for which the motor
system needs to code, must arise from the mixture of very
different representations, both extrinsic and intrinsic in nature.
The consideration of arbitrary coordinate systems to represent
movement goals and parameters, along with the invariance of
the brain solution to arbitrary coordinate transformations, inti-
mately links this problem to the concepts and methods of
Riemannian geometry. It is in this sense that the approach we
propose addresses all of these issues and gives a general
equation
dq task ⫽ (G qtask)䡠ƒ (rtask ⴰ f )⌬t
which provides a principled way to compute the autonomous
flow of geometric motion as a function of the task in both the
task and the posture space. This principle is the same regard-
less of the coordinate representations of choice.
Initially, the job of the geometric stage would be to find
optimal strategies to solve the task geometrically (as when an
athlete imagines a motion without performing it), but later, as
the movement becomes more automatic through practice and
repetition, the role of the geometric stage would switch to that
of a “supervisor,” who is able to correct for errors or pertur-
bations online. In this way, the execution system would not be
tied up with the problems of learning a new task, adapting to
sudden changes in the goals of a task, or responding to changes
in the task’s context.
This framework not only endows the motor system with the
ability to simulate motion but also makes possible the ex-
change of information between the subsystem linked to the
geometric stage and that concerned with the execution of these
geometric strategies. This is because the output signal of the
geometric stage is expressed in parameters of the same kind
as those needed by the execution system to compute the
dynamics of motion, and yet it carries task-dependent cognitive
information.
The separation that we propose between learning (geometric
strategy formation) and automaticity (temporal-dynamics op-
timization) cannot be captured in the present neurophysiolog-
ical data. These data come from overlearned and too simplistic
reaching movements that have already become automatic (bal-
listic), as shown in their bell-shaped speed profiles. A fair test
to this theory will come from new experiments simultaneously
recording activity from PPC and M1 either as the animal learns
a new task or in the presence of a sudden perturbation that calls
for new strategies. In these cases, the prediction of this theory
is that the PPC signal will temporally advance that of M1,
1
Bear in mind that the target in this approach is used in a more general sense
than the location of a point in space. The target has as many dimensions as the
task requires, and these may be linked to different sensory modalities, not just
to vision.
J Appl Physiol • VOL
canceling or reversing this effect as time progresses and the
motion becomes automatic. We know from recent transcranial
magnetic stimulation experiments that PPC but not M1 is
critical to adjust to sudden target displacements during reach-
ing movements (7).
In the context of arm movement, which is modeled by using
joint angles to represent postures, the output signal of the
geometric stage is a directional unitary vector that indicates
how much each joint angle in the arm should change to
incrementally bring the hand closer to the target. This joint-
angle velocity vector is required in the computation of kinetic
and potential energy (5) necessary to take into consideration
gravity, inertia, changes of mass, etc. In particular, the kinetic
energy, T ⫽ 1⁄2mv2 ⫽ 1⁄2m(ds ៮/dt)2 involves explicitly the line
element ds៮, hence it depends on the geometry of space, which
in the general case that interests us may be non-Euclidean:
៮2 ⫽ ¥ni, j gijdqidqj. The quantity ds
ds ៮2 is important in our
distance-based formulation because it does not change regard-
less of the coordinates employed (12). It is invariant to any
particular reference system. The quantities of gij on the other
hand are covariant; they change as a function of the coordinate
representation chosen to better capture the task at hand, which
in turn changes as a function of the posture q. The special
metric tensor G plays a fundamental role in our formulation. It
allows the development of a complete geometric analysis, not
only of the three-dimensional space, but also of the space of n
dimensions. This is significant because, as we have empha-
sized, a given task may span many dimensions, which in turn
have to be translated into those of the higher-dimensional
posture space.
Parameters of both disparate worlds are intimately related
and seem to be mutually optimized (10). In addition, a recent
prismatic-adaptation study suggests that “what is learned is not
simply a new endpoint trajectory but pertains to the movement
of the whole arm” (4). Such relationships are captured in our
formulation in three ways: 1) in the form of a pull-back action,
a geometric operation (r ⴰ f ) that expresses parameters of one
space as functions of parameters in another space; 2) in the
optimization of this pullback, which produces a solution in
both spaces; and 3) in the action to preserve in Q the notion of
distance in X given by G␮ T ␣
q ⫽ J Gx J to build a local isometric
embedding of X into Q and give a one-to-one correspondence
between the tangent space to X and that tangent to the subspace
of Q, which the embedding defines.
This general definition enables in-depth study of the geom-
etry of various realistic goal-oriented behaviors in systems with
redundant DOF. The specific example presented in this paper
tests predictions concerning reach-to-grasp with orientation-
matching motions. One of the predictions concerns the parallel
control of the hand transport and orientation phases along the
movement. Such coarticulation naturally emerges from the
form of the proposed objective function, where two terms
account for the position and orientation differences from the
hand to the target. This scalar function amounts to the measure
of distance in a four-dimensional space spanned by the three
components of the positional vector and the orientation-dis-
crepancy parameter. Because the geometry of task space is
preserved, these four dimensions are transferred to the space of
the arm, where seven DOF have to be controlled. Out of these
seven DOF, only those corresponding to a four-dimensional
96 • MAY 2004 • www.jap.org
 COARTICULATION OF HAND TRANSPORT AND ORIENTATION
subspace are recruited, leaving the other dimensions redundant
for this task. Figure 5D illustrates this for one target.
The position and orientation differences depend on all of the
arm joints in a redundant way. Thus changes in any of the
individual joint angles could potentially decouple the control of
the hand changes in position from its changes in orientation
along the path. The data, however, agree with the model’s
predictions of coarticulation and show that this general out-
come is robust to dynamics-related manipulations. This is
consistent with the proposed idea that the transformation from
extrinsic goals to motor commands may be governed by
geometric principles.
Across all three experimental conditions we tried, the posi-
tion and orientation discrepancies decreased at the same rate.
This contradicts the intuitive belief that, in reach-to-grasp
movements, distal DOF spanned by the wrist and proximal
DOF related to the shoulder act in a decoupled way. A
simplifying strategy of this kind, whereby DOF at the wrist
would be primarily concerned with the rotation of the hand,
whereas those at the shoulder would mainly control the
transporting of the hand to the target location (2, 15), is not con-
sistent with the subjects’ performance in response to this task.
The “division-of-labor” strategy can be achieved under cog-
nitive control when the task is different. In fact, we have
observed two instances in which subject’s motions show an
interesting outcome. Under specific instructions, they are able
to break down the movement toward a target into discrete
segments where the rate of change of transport and orientation
differs. Subjects can also alter coarticulation when obstacles
are positioned on the way to other targets (Torres EB, unpub-
lished observations). In both instances, the speed profiles are
not unimodal and the coarticulation parameter ␣ is no longer
constant along the entire movement path. In the model, it is
also possible to manipulate the ␣ ratio by changing the weight
placed on either phase of the motion. This feature permits the
simulation and empirical assessment of more interesting tasks.
However, human subjects show simultaneous rather than serial
coordinated control of hand position and orientation across all
three experimental conditions.
Present models cannot simulate realistic paths with redundant
DOF in the arm. Up to this point, in complex motions of this kind,
it has been very difficult (if not impossible) to engineer the
coarticulation result we report here. Likewise, it has been chal-
lenging to computationally reproduce experimental pointing mo-
tions where the initial posture of the arm affects the final posture
(21) or where changing the required target position and orientation
affects the arm posture as a whole (22).
Our treatment of geometry independent of temporal dynam-
ics makes it possible to account for all of these behaviors with
a computational model. The observed similarity of postural and
endpoint paths under three different speeds is consistent with
the idea of our proposed intermediate geometric stage. It extends
former results from constrained pointing motions (3, 18) to more
complex orientation-matching movements. Furthermore, we pro-
vided postural data in agreement with the model simulations to
account for the effect of changes in initial posture and required
target orientation. Any model of motor control that resolves the
detailed temporal profiles of movement must account for the
experimental observations presented here.
J Appl Physiol • VOL
1987
GRANTS
This study was supported by 3 F31NS10109-0551 (to E. Torres) and
Systems Developments Grant ICS9820 (to D. Zipser).
REFERENCES
1. Andersen RA and Buneo CA. Intentional maps in posterior parietal
cortex. Annu Rev Neurosci 25: 189–220, 2002.
2. Arbib MA. Programs, Schemas, and Neural Networks for Control of
Hand Movements: Beyond the RS Framework. Hillsdale, NJ: Erlbaum,
1990.
3. Atkeson CG and Hollerbach JM. Kinematics features of unrestrained
vertical arm movements. J Neurosci 5: 2318–2330, 1985.
4. Baraduc P and Wolpert DM. Adaptation to a visuomotor shift depends
on the starting posture. J Neurophysiol 88: 973–981, 2002.
5. Benati M, Gaglio S, Morasso P, Tagliasco V, and Zaccaria R. Anthro-
pomorphic robotics. I. Representing mechanical complexity. Biol Cybern
38: 125–140, 1980.
6. Calton JL, Dickinson AR, and Snyder LH. Non-spatial, motor-specific
activation in posterior parietal cortex. Nat Neurosci 5: 580–588, 2002.
7. Desmurget M, Epstein CM, Turner RS, Prablanc GE, Alexander GE,
and Grafton ST. Role of the posterior parietal cortex in updating reaching
movements to a visual target. Nat Neurosci 2: 563–567, 1999.
8. Desmurget M, Grea H, and Prablanc C. Final posture of the upper limb
depends on the initial position of the hand during prehension movements.
Exp Brain Res 119: 511–516, 1998.
9. Do Carmo M. Riemannian Geometry. Boston, MA: Birkhäuser, 1992.
10. Flanders M, Hondzinski JM, Soechting JF, and Jackson JC. Using arm
configurations to learn the effects of gyroscopes and other devices.
J Neurophysiol 89: 450–459, 2003.
11. Flash T and Hogan N. The coordination of arm movements: an experi-
mentally confirmed mathematical model. J Neurosci 5: 1688–1703, 1985.
12. Gray A. Modern Differential Geometry of Curves and Surfaces with
Mathematica (2nd ed.). London: CRC Press, 1998.
13. Harris CM and Wolpert Daniel M. Signal-dependent noise determines
motor planning. Nature 394: 780–784, 1998.
14. Helms Tillery SI, Ebner TJ, and Soechting JF. Task dependence of
primate arm postures. Exp Brain Res 104: 1–11, 1995.
15. Jeannerod M and Decety J. The Accuracy of Visuomotor Transforma-
tion: An Investigation into the Mechanisms of Visual Recognition of
Objects. Norwood, NJ: Ablex, 1990.
16. Kalaska JF, Cohen DAD, Hyde ML, and Prud’homme M. Parietal area
5 neuronal activity encodes movement kinematics, not movement dynam-
ics. Exp Brain Res 80: 351–364, 1990.
17. Lacquaniti F and Soechting JF. Coordination of arm and wrist motion
during a reaching task. J Neurosci 2: 399–408, 1982.
18. Nishikawa KC, Murray. ST, and Flanders M. Do arm postures vary
with the speed of reaching? J Neurophysiol 81: 2582–2586, 1999.
19. Rencher AC. Methods of Multivariate Analysis. New York: Wiley, 1995.
20. Rosenbaum DA, Loukopoulos LD, Meulenbroek RGJ, Vaughan J,
and Engelbrecht SE. Planning reaches by evaluating stored postures.
Psychol Rev 102: 28–67, 1995.
21. Soechting J, Buneo C, and Flanders M. Moving effortless in three
dimensions: Does Donders’ Law apply to arm movement? J Neurosci 15:
6271–6280, 1995.
22. Soechting J and Flanders M. Parallel, interdependent channels for
location and orientation in sensorimotor transformations for reaching and
grasping. J Neurophysiol 70: 1137–1150, 1993.
23. Todorov E and Jordan M. Optimal feedback control as a theory of motor
coordination. Nat Neurosci 5: 1226–1235, 2002.
24. Torres EB and Andersen RA. How Do Monkeys Learn to Avoid
Obstacles During Reaching? New Orleans, LA: Society of Neuroscience,
2003.
25. Torres EB and Zipser D. Reaching to grasp with a multi-jointed arm. I.
A computational model. J Neurophysiol 88: 1–13, 2002.
26. Uno Y, Kawato M, and Suzuki R. Formation and control of optimal
trajectory in human multijoint arm movement. Minimum torque-change
model. Biol Cybern 61: 89–101, 1989.
27. Wang X and Verriest JP. A geometric algorithm to predict the arm reach
posture for computer-aided ergonomic evaluation. J Visual Comp Anim 9:
33–47, 1998.
28. Wolpert D. Computational approaches to motor control. Trends Cogn Sci
1: 209–216, 1997.
96 • MAY 2004 • www.jap.org