Proceedings of the 26th Annual International Conference of the IEEE EMBS
San Francisco, CA, USA • September 1-5, 2004
A Nonlinear Tracking Method of Computing Net Joint Torques for
Human Movement
Ajay Seth1and Marcus G. Pandy1
1
Department of Biomedical Engineering, University of Texas, Austin, Texas
Abstract – Determining individual muscles forces from
human performance has greatly depended on the quality of
inverse dynamics solutions, as muscle force decomposition
remains the only feasible approach for determining muscle
forces non-invasively in human movement. However,
legitimate questions about the accuracy of inverse dynamics
arise, with resultant torques/forces failing to drive a forward
model through the observations from which they were derived.
While optimization of forward dynamics to match
experimental data is considered more accurate, the simplicity
and low computational costs of inverse methods are favored
over the large computing requirements of optimization. In this
paper, an evolution in the inverse methods for computing
accurate and reliable torques is presented, whereby the
relative speed of inverse dynamics is combined with the desired
accuracy of forward dynamics. This method is based on
developing a nonlinear tracker that determines the net muscle
torques which accurately follow clinically observed kinematics
and ground reaction forces. The results show that the method
is robust and can produce accurate estimates of the joint
torques during movement. The method outlined here is a
necessary first step to solving the muscle force indeterminancy
problem more efficiently.
Keywords – inverse dynamics, nonlinear control, computed
torques, feedback linearization, jumping
I. INTRODUCTION
Muscle forces are critical for quantifying muscle action
in human performance and also for determining realistic
loading conditions for prosthetic devices such as hip or knee
replacements. Although direct measurements of muscle
force from human performance would be ideal, this
approach is obviously not feasible from an ethical
standpoint. Therefore, external observations such as marker
positions, ground reaction forces (GRFs) and muscle EMG
are often used with computer models of the body to estimate
muscle forces non-invasively. The most common approach
is to use inverse dynamics to compute net joint torques and
to then use decomposition methods to estimate individual
muscle forces. Inverse dynamics requires the kinematics
(including accelerations) and a link-segment model, from
which the net joint torques can be calculated. Specifically,
force/torque balances are performed on the segments of a
model from the ground up, where the GRFs and center-of-
pressure are known from force-plate data.
In recent years, studies such as those by Neptune et al.
[1] and Zajac et al. [2] have employed a forward dynamics
0-7803-8439-3/04/$20.00©2004 IEEE 4633
model actuated by multiple muscles within an optimal
tracking problem. In this approach, muscle excitation inputs
to a musculoskeletal model are adjusted via an optimization
algorithm in order to minimize performance differences
between the model and the human. This approach, however,
requires thousands of forward simulations to be executed in
order to minimize the objective function, which is a total
measure of tracking error. A comparison of a static
decomposition approach based on inverse dynamics and a
large-scale forward optimization for determining muscle
forces in gait [3] concluded that the two methods produced
similar results. Thus, based on inverse dynamics torques,
researchers have begun to include muscle dynamics into the
muscle force decomposition process [4], while others [5]
employ EMG-driven muscle models to estimate muscle
forces directly from EMG. In either case, torques computed
from inverse dynamics are used at some point to affirm the
muscle forces obtained from the decomposition approach.
In studies reported by Anderson and Pandy [3], Thelen
et al. [4], and Lloyd and Besier [5], the quality of the
inverse dynamics torques is not addressed. For example,
Anderson and Pandy [3] used the torques from their forward
optimization results to circumvent inverse dynamics, and
thus their similarity conclusions are, in effect, based on
perfect inverse dynamics data. In reality, this is far from the
case, when one considers skin marker error, noise enhancing
effects of double differentiation, and the mechanical
differences between the human and the model. Inverse
dynamics is therefore inherently imprecise. Although low-
pass filtering and least squares estimations of segment
rotations can mitigate some of the errors, torques computed
from inverse dynamics still may not reproduce measured
joint movements and GRFs when applied in a forward
simulation [6]. A simple reliability check would be to drive
the model forward using inverse dynamics torques and to
compare the resultant kinematics and GRFs against
measured values. Unfortunately, such checks are rare even
when the torques computed from inverse dynamics are the
basis for estimating muscle forces.
The purpose of this study is to develop a more robust
tracking method for computing net joint torques based on
inverse dynamics. A tracking method is required to ensure
that the torques computed from inverse dynamics will drive
a model in a stable forward simulation and reproduce the
observed kinematics and ground reaction forces. The
method is demonstrated by computing the torques needed to
produce a maximum height jump. Even though the model
used to simulate jumping includes nonlinear ground contact,
the tracking method is able to reproduce the kinematics
(segment rotations) and GRFs observed during jumping.
II. METHODOLOGY
A six degree-of-freedom planar model of vertical
jumping (Fig. 1) was developed using the same parameters
and ground contact equations described by Anderson and
Pandy [7]. However, the model excludes the toes and fuses
the pelvis and trunk segments together. The peak isometric
torques for the ankle, knee and hip reported by Anderson
and Pandy [7] were used as the maximum available torques.
The optimal control problem was solved to determine the
joint torques needed to produce a maximum height jump.
These torques and simulated kinematics and GRFs are the
reference values against which the results from inverse
dynamics and the tracking method are compared. The
simulated kinematics and GRFs were sampled at a
frequency of 120Hz and 1080Hz, respectively. Gaussian
noise was added to both angular positions and GRFs. A
standard deviation of 5% of peak displacement was used for
the angular positions and 1% of peak force for the GRFs.
These data represented experimental observations and were
used as the inputs to both the inverse dynamics and tracking
methods. Data were filtered with a second order lowpass
filter with a cutoff frequency of 8Hz for the kinematics and
64Hz for the GRFs. The filtered segment positions were
used to estimate joint velocities and accelerations via
(central) finite-difference numerical differentiation.
The tracking approach can be viewed as stabilizing
inverse dynamics via feedback, so that errors due to noise
and sampling effects can be sufficiently rejected. Runge et
al. [8] described a method, which linearized the dynamics
for posture control and enabled the determination of the
torques necessary for rejecting external perturbations in
maintaining an upright posture. Linearizing the dynamics,
however, is only valid for motion bounded about a
particular state, which is not the case in jumping, where the
joints move through large ranges of motion. Therefore,
feedback linearization (FBL), which is the basis of the
computed torques method (see, for example, [9]), is also
used here. In addition, the present approach builds upon the
computed torques method by enabling GRFs to be tracked
simultaneously with joint kinematics.
The system equations of motion (EOMs) can be written
as:
q&& = M −1 [ f (q, q& ) + S (q, q& ) + Tτ ] (1)
where q is a vector containing the angles of the foot, shank,
thigh, and HAT (head, arms, and trunk) as well as the
horizontal and vertical translations of the foot; M is the mass
matrix; f is the vector of generalized forces resulting from
gravity, centripetal and Coriolis effects; S is the vector of
ground contact force contributions; and T is the coefficient
matrix of the control torques τ. In first order form:
4634
Q& = F (Q ) + G (Q )τ (2a)
y = [q out , s ]
T
(2b)
where Q = [ q, q& ]T, F = [Q1..6, M-1{f+S}]T, and
G = [06x3, M-1T]T. The system output vector, y, is made up of
two components: 1) qout, the generalized coordinates minus
the horizontal and vertical positions of the foot; and 2) s, the
net horizontal and vertical ground contact forces and the
resultant contact moment about the foot. Experimentally, s
corresponds to force plate data acting in the sagittal plane.
The translations of the foot are excluded from the output
because the contact dynamics are extremely sensitive to
small displacements. GRFs are included instead because
these data can be measured accurately in an experiment.
The basis of the FBL method is to appropriately feed
back the outputs (or their derivatives) such that the
nonlinearities of the system can be eliminated. This is
accomplished by differentiating the outputs until the
controls (torques) appear explicitly. Twice differentiating
the outputs corresponding to qout yields the rotational EOMs
in (1):
&y&q = ∇z ⋅ F + ∇z ⋅ Gτ , (3)
where ∇z = ∇(∇q out ⋅ F ) = [0 4×6 , I 4×4 ,0 4×2 ] .
The ground contact forces, s, are functions of the foot
coordinates and their velocities, due to ground springs and
dampers in the contact model. Thus, differentiating s once
leads to the acceleration of the foot and its respective
EOMs. Substituting the foot EOMs into s& provides the
explicit output as a function of τ,
y& s = ∇s ⋅ F + ∇s ⋅ Gτ (4)
where ∇s is the gradient of the ground contact forces with
respect to the states, Q.
We now introduce a new control vector,
v = [ &y&q , y&s ]T
such that,
∇z ∇z
v =   ⋅ F +   ⋅ Gτ
∇s ∇s
and solving for the controls,
−1
∇z   ∇z 
τ =   ⋅ G v −   ⋅ F (5)
∇s   ∇s 
yields the feedback linearizing control law. Note that the
matrix being inverted is not square and a pseudo-inverse can
be used. Applying this control law results in the output
dynamics (3) and (4) being transformed to the linear system,
 those obtained from inverse dynamics. The computed
 &y&q  torques were virtually identical to the reference torques (Fig.
 =v (6)
4) up until the time when the heel left the ground (0.21s);
 y& s  after heel lift-off, the hip and knee torques became erratic.
which is easily stabilized with linear feedback and the
appropriate gains on the tracking errors.

v q = q&ˆ& − 2λq (q& out − q&ˆ ) − λ2q ( qout − qˆ ) (7a)

v s = sˆ& − λ s ( s − sˆ) (7b)

Thus, at any instant in time, the control vector, v, is

only a function of the observed kinematics, q̂ , the ground
reaction forces, ŝ , (along with their estimated derivatives),
and the model outputs. The λ’s are the magnitudes of the
real left-hand-plane (i.e. stable) poles that form the gains for
the kinematics (q) and contact forces (s) errors (7). The
torque is then computed from the FBL control law (5) and
applied to the model at the initial observed state. From then
on the torques are determined for the entire performance Fig. 2: Tracking (solid), inverse dynamics (dashed), observed (shaded), and
time via integration of the system EOMs (1). reference (light dotted, superimposed on solid) segment angular trajectories
The contact force equations provided by Anderson and
Pandy, [7], are applied to two points on the foot: the heel
and toe ends of the single foot segment. The gradient of s
was computed analytically using the MATLAB™ Symbolic
Toolbox. Transitions from the horizontal spring force to an
upper-bound frictional force caused a “jump” in the gradient
due to the instantaneous change in the governing equations.
Even though this caused a transient in the control, it did not
significantly affect the tracking performance (Figs. 2 and 3).
The torques computed from inverse dynamics were
applied to the model to perform a forward simulation. The
resulting kinematics and GRFs were compared to the same
quantities obtained from the tracking method.

III. RESULTS
Fig. 3: Tracking (solid), inverse dynamics (dashed) and observed (shaded)
GRFs

Fig. 1: Model performing a maximal height jump

Consistent with model and experimental results

reported in the literature ([7], [10], [11]), the model (Fig. 1)
jumped to a height of 37.24 cm. The tracking results
reproduced the desired kinematics (Fig. 2) and GRFs (Fig. Fig. 4: Tracking (thin solid), inverse dynamics (dashed) and reference
3) with good accuracy and were significantly better than (shaded) joint torques

4635
 IV. DISCUSSION
The nonlinear tracker performed well in reproducing
the performance of the model that matched the input
observations as illustrated in Figs. 2 and 3. An unexpected
outcome was that the kinematics obtained from tracking
more closely matched the reference kinematics than did the
observed kinematics. This occurred because both the joint
kinematics and the GRFs were tracked simultaneously. The
feedback gains, therefore, ultimately dictate the trade-off in
tracking performance between the kinematics and the GRFs.
The fact that measurements of GRFs are much more
accurate than measurements of joint kinematics is fortuitous
because the performance of the tracking method can be
enhanced by having relatively higher gains on GRFs errors.
In this way, the kinematics act to guide the model, whereas
the accuracy can be derived from the GRFs. Thus, the
kinematics determined from the tracking method are more
reliable than the observed kinematics because the
dependence of the kinematics on GRFs is enforced by virtue
of the model.
Even though forward simulations using the torques
from inverse dynamics follow the general trends of the
kinematics (Fig. 2), the resulting GRFs are erroneous (Fig.
3). The torques from inverse dynamics over estimate the
peak output of all the joints, especially the ankle before 0.1s.
In contrast the tracking torques reproduce the input torques
well for most of the jump. When the heel leaves the ground
(0.21s), the hip and knee torques computed by the tracking
method become erratic (Fig. 4). The reason is that the heel
contact point at that time is no longer governed by spring-
damper equations, but rather switches to a frictional limit
dependent on the vertical GRF. The hip and knee torques
are unable to fully recover from this transition for a few
reasons. First, the ankle torque dominates the interaction
with ground just prior to takeoff [10], and so the input GRFs
predominantly influences the ankle control, which also
explains the accurate ankle torque throughout the tracking.
The tracker, therefore, must depend more on the joint
kinematics for determining the knee and hip torques.
Second, the kinematics, which are typically less accurate,
are even more error prone closer to takeoff due to position
sampling when the jumper is moving at its peak velocity.
Thirdly, the heel leaving the ground essentially changes the
controllability of the system. Specifically, as the foot
translates upward the contact damping diminishes (at 5mm
above the ground) thereby rendering the controller less
effective, because it is through the velocity term (4) that the
feedback linearization was obtained.
V. CONCLUSION
Overall, simultaneous tracking of GRFs and joint
kinematics achieved accurate results despite the addition of
4636
noise and sampling error for a highly dynamic task such as
jumping. Consequently, the tracking method determined
the required input torques more accurately than
conventional inverse dynamics. Controller design, however,
can be a challenging task, especially with ground contact
models that include hard nonlinearities (i.e., discontinuous
dynamics or gradients), which can inadvertently change the
controllability of the system and thus the requirements on
feedback linearization. In such cases, it may be necessary to
design multiple controllers for different phases of a single
task.
The results of this study should prove useful for those
who wish to acquire quantitatively accurate net joint torques
from musculoskeletal models, especially for the purpose of
decomposing joint torques into individual muscle forces.
ACKNOWLEDGMENT
Supported in part by the National Science Foundation,
Engineering Research Centers Grant EEC-9876363. A. Seth
thanks the International Society of Biomechanics for the
support of a 2004-2005 ISB Dissertation Grant.
REFERENCES
[1] Neptune, R. R., Kautz, S. A., and Zajac, F. E. (2001). Contributions
of the individual ankle plantar flexors to support, forward progression
and swing initiation during walking. Journal of Biomechanics,
34:1387-98.
[2] Zajac, F. E., Neptune, R. R., and Kautz, S. A. (2002). Biomechanics
and muscle coordination of human walking. part i: introduction to
concepts, power transfer, dynamics and simulations. Gait and Posture,
16:215-32
[3] Anderson, F. C. and Pandy, M. G. (2001). Static and dynamic
optimization solutions are practically equivalent. Journal of
Biomechanics, 34:153-61.
[4] Thelen, D. G., Anderson, F. C., and Delp, S. L. (2003). Generating
dynamic simulations of movement using computed muscle control.
Journal of Biomechanics, 36:321-328.
[5] Lloyd, D. G. and Besier, T. F. (2003). An emg-driven
musculoskeletal model to estimate muscle forces and knee joint
moments in vivo. Journal of Biomechanics, 36:765-776.
[6] Risher, D. W., Shutte, L. M., and Runge, C. F. (1997). The use of
inverse dynamics solutions in direct dynamics simulations. Journal of
Biomechanical Engineering, 119:417-422.
[7] Anderson, F. C. and Pandy, M. G. (1999). A dynamic optimization
solution for vertical jumping in three dimensions. Computer Methods
in Biomechanics and Biomedical Engineering, 2:201-231.
[8] Runge, C. F., Zajac, F. E., Allum, J. H. J., Risher, D. W., Bryson, A.
E., and Honegger, F. (1995). Estimating net joint torques from
kinesiological data using optimal linear system theory. IEEE
Transactions on Biomedical Engineering, 42(12):1158-1164.
[9] Seth, A., McPhee, J. J., and Pandy, M. G. (2003). Multi-joint
coordination of vertical arm movement. Journal of Applied Bionics
and Biomechanics, 1(1):45-56.
[10] Bobbert, M. F. and van Ingen-Schenau, G. J. (1988). Coordination in
vertical jumping. Journal of Biomechanics, 21(3):249-262.
[11] Pandy, M. G., Zajac, F. E., Sim, E., and Levine, W. S. (1990). An
optimal control model for maximum-height human jumping. Journal
of Biomechanics, 23(12):1185-1198.