You are on page 1of 14

Cogn Comput (2012) 4:549–562

DOI 10.1007/s12559-012-9160-5

Physiological LQR Design for Postural Control Coordination


of Sit-to-Stand Movement
Asif Mahmood Mughal • Kamran Iqbal

Received: 21 May 2011 / Accepted: 26 June 2012 / Published online: 10 July 2012
Ó Springer Science+Business Media, LLC 2012

Abstract The neurophysiological mechanisms involved awareness of the existence of optimizing controllers in the
in postural stabilization are not well understood. Active and central nervous system.
passive mechanisms at muscle and spinal levels as well as
visual and vestibular processes are known to contribute Keywords Biomechanical modeling  Optimal control 
toward postural stabilization and coordination of voluntary Physiological optimization  Postural stabilization 
movement. The motivation for this research is to use a Movement coordination
modeling–simulation framework to achieve two aims:
(a) to ascertain viability of a physiologically motivated
optimal controller design in the maintenance of posture and Introduction
coordination of voluntary movement and (b) to study the
relative contribution from active (feedforward) and passive The role and function of the central nervous system (CNS),
(feedback) mechanisms in the execution of said movement. and the interaction of different orientation senses in pos-
We employ a multi-segment sagittal model built on ana- tural stabilization are not well understood [1, 2]. In the
tomical proportions with three degrees of freedom, context of this study, postural stabilization refers to our
including rotation at the ankle, knee, and hip joints. The ability to maintain standing posture amid internal and
behavior of the biomechanical model is controlled by an external perturbations. Human sagittal biomechanics for
optimal linear quadratic regulator whose state and control posture control have often been compared to an inverted
weights are derived from physiological considerations. pendulum [3, 4]. This unstable structure is supported over a
Representative postural and voluntary movements are small base of support (BOS), that is, the area under two
simulated to illustrate the analysis–synthesis framework of feet. Further, postural stabilization is aided by the presence
biomechanical movement. Our analytical and simulation of somatosensory and proprioceptive processes in the body.
results support an active–passive model of postural stabil- Investigators have reported that standing humans, when
ization and movement coordination. Besides expanding our perturbed by forward–backward translation of a moving
understanding of the physiological stabilization processes support surface and instructed not to move their feet, typ-
in the body, the insight gained from this study promotes ically respond by using a combination of ‘ankle’ and/or
‘hip’ strategies in the sagittal plane [5, 6]. Further, the
ankle strategy is adopted for small perturbations and may
be combined with hip strategy for larger perturbations; still,
A. M. Mughal (&)
larger perturbations elicit stepping and grasping responses
Department of Applied Science, University of Arkansas at Little
Rock, Little Rock, AR, USA [7].
e-mail: asifmahmoodmughal@gmail.com Winter et al. [8] have studied the relationship of the
center of mass (COM, controlled variable) and the center
K. Iqbal
of pressure (COP, controlling variable) during body sway
Department of Systems Engineering, University of Arkansas
at Little Rock, Little Rock, AR, USA through the use of a lightly damped inverted pendulum
e-mail: kxiqbal@ualr.edu model. The authors hypothesized that the close phase

123
550 Cogn Comput (2012) 4:549–562

relationship between COM and COP negates the presence Sit-to-stand (STS) task offers an example of an ordinary
of reactive processes in balance control. This prompted the but skilled movement that has been extensively studied by
authors to consider a stiffness-only model, whose stiffness researchers [19–21]. Analysis of laboratory data shows that
constant is set by CNS and is sufficient to control the large STS movement comprises two main phases: the forward
inertial load against the gravitational forces. Their view of thrust phase and the extension phase. Kerr et al. [19] have
postural stabilization as a purely passive mechanism is further divided the kinematics of movement into regions of
strengthened by the consideration that the latencies in the flexion momentum, momentum transfer, lift off, and
motor servo loop with the low-pass characteristics of the extension phases. They collected data from seat off to
biological muscle tend to limit the effectiveness of active termination from 50 healthy males and females (age, 2–78)
and reactive control mechanisms. Existence and involve- and provided mean time of 1.907 ± 0.057 s. STS move-
ment of optimal control processes in postural stabilization ment terminates when hip velocity reaches zero and set-
and movement coordination has been widely acknowl- tling of movement in this time requires higher torque.
edged by researchers [9–12]. Todorov and Jordan [9] have Researchers have also discussed strategy-related issues of
stated that emergent properties of optimal feedback control STS transfer for movement coordination of joints and the
produced a certain desirable behavioral richness in the case constraints involved in movement coordination [20, 21].
of a multi-joint arm model. The authors advanced a mini- We have previously studied the regulation of STS move-
mal intervention principle, under which deviations from the ment using a fuzzy biomechanical model controlled by a
average trajectory were corrected only when they inter- fuzzy optimal linear quadratic regulator (LQR)-based
fered with task performance. Sachaal and Schweighofer controller using Gaussian and triangular membership
[10] have similarly proposed a stochastic optimal control functions [22]. Further, we have compared and contrasted
model of human brain. According to the authors, noise is the performance of H2 and H? robust and optimal control
predominant in the nervous system, and to generate smooth techniques in the coordination of STS movement [23].
movements, CNS movement strategy minimizes the effect However, this modeling scheme takes more time for set-
of noise on the motor output. Kuo [12] has explicitly tlement ranging from 4 to 10 s for movement termination.
considered the control of balance based on optimal control A similar sagittal plane modeling scheme was studied by
principles. The author postulates that the optimal control Jo and Massaquoi [24] and bipedal walking model in [25].
model for the control of balance appears to serve as an In this study with recurrent PID controller for terminal
ideal framework to study CNS responses with ambiguous movement, researchers achieved energy efficiency with
sensory input. Further, parameterization of state and con- muscular co-activation and proprioceptive force feedback.
trol weightings may be used to implement physically The controllers used in those studies were, however, not
meaningful objectives subject to constraints. derived from physiological considerations. It can be
Biomechanical models have been frequently employed hypothesized that optimal controllers based on physiolog-
by researchers in support of laboratory studies to test the- ical considerations would be better representative of the
ories of motor control and learning. Models with various underlying CNS decision-making processes. The aim of
complexities and detail have been used in the literature to this paper is to employ a modeling–simulation paradigm to
simulate postural control and coordination of voluntary study the performance of an optimal controller derived
movement [8, 12–18]. According to Kuo [12], control of from physiological cost consideration for regulation of
balance in human upright standing is particularly well postural and voluntary STS movement. We based upon two
suited for modeling and is also a popular experimental physiological functions, kinematic and kinetic variables for
paradigm. Furthermore, modeling in this paradigm (dis- derivations of control gains. In addition, we aim to study
turbances to posture in the sagittal plane) is particularly the involvement of active (feedforward) and passive
convenient, as the musculoskeletal dynamics can be line- (feedback) torques in the performance of said movements.
arized without significant impact on the simulation accu- We presented initial results in [26] with active and passive
racy. Barin [16] has developed a multi-segment model of feedback for postural balance and STS stabilization. We
human postural dynamics. Pai and Iqbal [17] have used an note that these can be individually varied to study their
inverted pendulum over moving foot model to investigate relative contribution toward postural stabilization and
slipping, sliding, and falling behavior. The authors have movement coordination. Further, we now introduce phys-
demonstrated how overlapping feasible stability region iological cost optimization for controller synthesis in pas-
(FSR) in the COM position-velocity phase plane could be sive torques, whereby we use sensitivity functions of
exploited for postural control in unfamiliar environment. physiological variables as weights in the optimization
Iqbal and Pai [18] have employed a four-segment biome- algorithm.
chanical model to study optimization of movement termi- Some researchers [5, 6, 27, 28] have preferred to use
nation and the role of knee joint in postural stabilization. simpler one and/or two segment models for investigation of

123
Cogn Comput (2012) 4:549–562 551

postural control in humans. On the basis of these studies, In the next section, we provide mathematical formula-
researchers have proposed the existence of discrete ‘ankle’ tion of the sagittal plane model with dynamic equations.
and/or ‘hip’ strategies to maintain postural balance. This Then, we provide the mathematical equations for control-
view of postural stabilization appears to discount the role ling the system with kinematic variables, that is, CoM, and
of knee joint in postural stabilization. Iqbal and Pai [18] kinetic variables, that is, GRF. Later, we showed the sim-
have earlier studied the role of knee joint in postural ulation results of the scheme followed by discussion and
reactions and determined that motion at the knee joint conclusion.
delivered a small but recognizable benefit to the termina-
tion of total body forward movement. Keshner and Allum
[29] have similarly determined that in many postural per- Formulation of the Multi-segment Biomechanical
turbations, the body moves as a multi-link structure with Model
extensive knee movements.
Optimal control of human postural and voluntary move- The mechanics of the human body in the performance of
ment has been considered by other researchers. For example, postural and voluntary movement is modeled as a multi-
Kuo [12] has proposed optimal control of postural balance segment structure with joint torque actuation at the ankle,
following BOS perturbations. According to the author, linear knee, and hip joints. More specifically, a four-segment
quadratic control provided a means to use relatively few rigid-body model connected by single-degree-of-freedom
parameters to describe a large number of gains, and it joints is used to represent sagittal plane mechanics (Fig. 1).
guaranteed a large degree of robustness to time delays. Kuo
[13] further addressed combining optimal state estimation
with feedback control in a model of human postural balance.
According to authors, inertial coupling between body seg-
ments discounted the possibility of using purely local feed-
back for postural balance. Furthermore, sensing of kinematic
positions and velocities of all body segments appeared to l3
constitute the minimal informational set necessary for pos-
tural stabilization. Thus, model-based studies have tended to I3
favor full state feedback for computer simulation of postural m3g
control and voluntary movement.
θ3
In this context, we also note that since both the joint
angular positions and velocities are being used as state
variables, the full state feedback control successfully rep- τ3
resents viscoelasticity at individual joints. In the control l2
theoretic literature, this is termed as proportional-deriva- I2
tive (PD) control. It is well known that such control imparts
m2g
a certain phase-lead to the closed-loop system, thus making θ2
its behavior anticipatory. We may, therefore, conclude that τ2
the optimal PD type control used in our study implicitly
includes functional elements of the motor servo loop that
promote postural stability, though these were not explicitly
modeled in the mathematical formulation. Jo and Massa- I1
l1
quoi [24, 25] provided results of upright balance and m1g
walking with cerebellum stabilization and cerebrocer-
ebello-spinomuscular interaction for sagittal plane models. θ1
In these studies muscular interactions and proprioceptive
feedbacks motivate toward an energy efficient control τ1
x
scheme. In this paper, we are studying an optimal control
scheme with optimality condition reaches through qua- x
dratic indices, that is, appropriate selections of Q and R Fy Fx
matrices. The selection of Q and R matrices based upon
lf
kinematic variables validates the finding in [20, 21] and
also confirms the controller scheme shows better results Fig. 1 Multi-segment biomechanical model comprising foot, leg,
with kinematic variables than with kinetic variables. thigh, and head, arm, and trunk (HAT) segments

123
552 Cogn Comput (2012) 4:549–562

 !
The segments represent a bilateral symmetrical arrange- 
€ 1 oG 0
ment of the feet, legs, thighs, and head–arm–trunk (HAT). Dh ¼ ½Dðhe Þ Dh þ s ð3Þ
ox x¼xe
The CNS commands and active force generation in the
muscle are represented by torques applied at the ankle, In this study, we are using standing position as a stale
knee, and hip joints. The body segment parameters given in equilibrium point at which Eq. (3) is linearized. Equation 4
Table 1, that is, the segment length, mass, center of gravity, shows the state vector for linearization; it has all three
and moment of inertia, are based on average anatomical angles at p=2 with zero velocities and zero joint torques.
proportions [27]. The length of the stationary foot segment
defines the BOS in the anterior–posterior direction, where xe ¼ ½ p=2 p=2 p=2 0 0 0  ð4Þ
it may be noted that for static postural stability, COM of The linearized system of dynamic equations is
the body must be restricted to the BOS [18, 19]. represented in the usual state space form given as:
_ ¼ AxðtÞ þ BuðtÞ
xðtÞ
   
Dynamic Equations of the Model 0 I 0 ð5Þ
A¼ B¼
A1 0 B1
The dynamical equations of the multi-segment biome-  T
where x ¼ Dh1 Dh2 Dh3 Dh_ 1 Dh_ 2 Dh_ 3 and A
chanical model are nonlinear and may be symbolically
and B matrices are obtained by differentiating (2) with
represented as:
  respect to the state and input variables, respectively, and
DðhÞ€h þ H h; h_ h_ þ GðhÞ ¼ s
*
ð1Þ computed at the equilibrium point. The linearized model,
(3), represents the behavior of the system in the small
where h is a vector representing the DOF of the model; neighborhood of the defined equilibrium posture. For the
DðhÞ€ _ h,
h, Hðh; hÞ _ and GðhÞ are, respectively, the inertial, 3-DOF presented here, A is a 6 9 6 matrix, B is a 6 9 3
0
Coriolis, and gravitational components of the generalized matrix, and uðtÞ ¼ s represents the input torque variations
torques acting on the model (see ‘Appendix’). Let si , with respect to the equilibrium torques. Feedback control
i = 1–3, define the joint torques acting at the ankle, knee, of the posture is required for stable stance and coordination
and hip joints; then, the input torque vector in (1) is defined of body movement. The pair (A,B) have full-rank con-
as s ¼ ½ s1  s2 s2  s3 s3 T where ‘T’ denotes the
*
trollability matrix, which means that all three joint torques
matrix transpose operation. The joint torques each are controllable through feedback of state vector.
comprise a feedforward and a feedback component, that
is, s ¼ sfb þ sff , where the feedforward component, sff ;
represents the descending commands that are task specific, Physiologically Motivated Optimal Controller Design
and the feedback component, sfb ; represents the passive
*
viscoelasticity in the musculo-tendon structures surround- The controlled input, s, in (1) above represents the torques
ing the joints. This feedback torque is generated by CNS as generated at the ankle, knee, and hip joints in response to
a result of muscular stretches required for movement. The the muscle commands issued by the CNS that are regulated
feedforward component, sff , also ensures that static equi- through spinal reflexes in the physiological system [28].
librium at the standing posture and at discrete points along The corresponding control input uðtÞ in (4) excludes the
a nominal reference trajectory, href , is maintained. Using descending commands responsible for trajectory generation
joint angular positions, h; and angular velocities, h; _ as the due to linearization and represents purely the reactive
state variables, the dynamic equation (1) can be recast into processes and passive stiffness. However, for the purposes
the state variable form as: of this study, we assume the spinal reflexes to have a
h   i minimal contribution in the control effort uðtÞ; which is

h ¼ D1 ðhÞ H h; h_ h_  GðhÞ þ s
*
ð2Þ assumed to result purely due to passive stiffness of the
muscle-joint structures. It is noted that the physiological
Eqs (1) and (2) represent a nonlinear system of equations in neural transmission delays are irrelevant in this paradigm.
angular and velocity variables. This system can be linearized In the modeling environment, this passive torque is gen-
by Taylor series approximation around an equilibrium point as erated via feedback of state variables through suitable
follows: Let he represent the equilibrium position, Dh ¼ feedback gains. We select LQR-based controller design for
h  he represent the perturbation from the equilibrium, and the control of biomechanical model due to its optimizing
* *
s ¼ s  se represent the torque perturbation from the static energy for the system through quadratic indices. The state
*
equilibrium torques, where se ¼ Gðhe Þ; then, the linearized weighting Q and input weighting R matrices for the LQR
system of dynamic equations is given as: design can be chosen to get required performance on

123
Cogn Comput (2012) 4:549–562 553

control effort, initial and final positions, and the desired Scholz, and Schöner [20, 21] who have studied the role of
dynamic and steady-state responses. We note that the COM-based constraints in the coordination of STS move-
optimal steady-state LQR design naturally develops into a ment, we will explore the use of the COM equations to
proportional-derivative (PD) controller that has been sug- generate state weighting matrix Q. In the first attempt, we
gested by some researchers for control of postural move- suggest using the sensitivity functions of the COM position
ments [18]. In the general model of a PD controller, the with respect to the degrees of freedom (joint angles) as the
feedback torque sfb is represented as sum of two compo- weights. The horizontal (anterior–posterior) COM position
nents: a position-dependent component and a velocity- and velocity in the case of multi-segment model are given
dependent component, that is, as:

sfb ¼ Kp h  href  Kv h_ ð6Þ P


3
fi cos xi  cmf
i¼1
where we note that the above input torques represent the xCOM ¼ lf  a þ
P
3
passive (feedback) component of the overall joint torques m þ mf
applied to the biomechanical model. Two feedback gain i¼1
ð11Þ
matrices Kp and Kv represent the respective gains for posi- P
3
fi ðsin xi Þ  xiþ3
tion and velocity terms. Jo and Massaquoi [24, 25] repre- i
vxCOM ¼
sented the same equation for passive muscular force, but in P
3
this case, we are representing it for joint torques applied by m þ mf
i¼1
CNS. Since the reference velocities at equilibrium point are
zero, then there is no separate need for representation of h_ e . where lf and mf represent the length and mass of the foot,
The active (feedforward) component is computed such as to and a, b, c represent the foot parameters related to the
balance the gravitational torques at the ankle, knee, and hip position of the ankle [18], and fi are corresponding grav-
joints, along the desired reference trajectory. itational components in each segments(see ‘Appendix’).
The optimal control effort for infinite horizon LQR The sensitivity derivates of the COM position and
design is computed from the unique symmetric positive velocity corresponding to joint angles Dh; and joint
definite solution matrix M to the following algebraic velocities Dh, _ computed at the standing posture,
xe ¼ ½ =2 0 p=2 0 0 0  , produce the state
p T
Riccati equation (ARE):
weights in Q. Let qii represent the diagonal terms in the
MA þ AT M  MBR1 BT M þ Q ¼ 0 ð7Þ Q matrix; then, the weights for position and velocity states
The corresponding LQR controller gain matrix is given are assigned as follows:
 
as: oxCOM 2

qii ¼   ði ¼ 1; . . .; 3Þ
KLQR ¼ R1 BT M ð8Þ oxi x¼xe ;u¼ue
  ð12Þ
The optimal joint torques for the biomechanical model ovxCOM 2

qii ¼   ði ¼ 4; . . .; 6Þ
in (4) are computed as: ox  i x¼xe ;u¼ue

uðtÞ ¼ KLQR  xðtÞ ð9Þ The cross-diagonal terms in the Q matrix are selected as
Finally, the closed-loop system dynamics are given as: zero for independent minimization of position and velocity
errors. We note that the COM sensitivity functions provide
_ ¼ ðA  BR1 BT MÞxðtÞ
xðtÞ ð10Þ weights for the state variables, but not for the input
Physiological LQR Controller Design variables. In this case, the input weighting R can be
arbitrarily selected to minimize maximum joint torque.
The state weighting matrix Q and the input weighting Another variable that has the potential to be useful for
matrix R are the key elements of the optimal controller physiological cost optimization is the GRF. The expression
design. These are normally selected as diagonal matrices, for the horizontal GRF is given as:
where the diagonal terms can be arbitrarily selected for the X
3

best response based on trial and error. In the case of the Fx ¼ fi x€i sin xi þ x_2i cos xi ð13Þ
i¼1
biomechanical model, however, we desire that the selection
of the optimal controller weights should relate to the Differentiating and linearizing the horizontal GRF at the
physiological phenomena. Two physiological criteria, that standing posture generates the desired sensitivity functions.
is, the COM and the ground reaction force (GRF), are The diagonal terms of Q and R matrices are accordingly
considered for this purpose. Taking a cue from Resiman, assigned as:

123
554 Cogn Comput (2012) 4:549–562

 2 ct
 x as: xR ðtÞ ¼ 1e
1þect , or alternatively xR ðtÞ can be obtained
qii ¼ oF
oxi  ði ¼ 1; . . .; 3Þ;
 x¼x e ;u¼ue from the output of a stable open-loop linear system. In ref
 x 2
qii ¼ oF ði ¼ 4; . . .; 6Þ ð14Þ [30], we have proposed a third-order linear model, whose
oxi 
 2x¼xe ;u¼ue output resembles the experimental data for the STS
 x
rii ¼ oF
osi  ði ¼ 1; . . .; 3Þ maneuver. This reference trajectory model is given as
x¼xe ;u¼ue
follows: x_ R ¼ AR xR , where AR is selected as open-loop
where we note that the sensitivity functions in the case of stable matrix. The following AR matrix was proposed in
GRF produce the diagonal elements in Q as well as R [30] to generate the reference trajectories for postural
matrices for optimization. Finally, a hybrid LQR design stabilization.
scheme may be considered that includes a combination of 2 3
0 1 0
COM- and GRF-based optimization for state and input
AR ¼ 4 0 0 1 5 ð17Þ
weighting matrices. The hybrid state weighting matrix, Qh ; 0 2 1
and input weighting matrix, Rh ; will be selected as:
In this study, we will use their model for reference
Qh ¼ QCoM þ QGRF ; Rh ¼ RGRF ð15Þ
trajectory generation to construct segment trajectories for
both the postural and voluntary STS movements. As t ! 1
Simulation of Postural and Voluntary Movement then xR ðtÞ ! 1 by employing initial conditions, trajectories
will be generated accordingly as given in [30].
We conducted computer simulations of representative
postural and voluntary movements to ascertain the viability
of our physiological LQR controller for the multi-segment Simulation of Postural Movement
biomechanical model. The simulation results are based on
the nonlinear dynamic model in (1–2) with linear optimal Typically, in a postural adjustment initiated in response to
controller in (9). The overall block diagram for simulation BOS perturbation, the COM of the body is located behind
of postural and voluntary movement is shown in Fig. 2. the heel at movement initiation. The COM would normally
have a nonzero anterior velocity at movement initiation
Reference Trajectories for Movement Simulation that could be translated into nonzero individual joint
angular velocities. However, in order to include the static
In order to simulate a point-to-point movement, it is positions in the evaluation of controller performance, we
desired that a tracking model with reference trajectory choose to have zero initial COM anterior velocity at
input be employed. The feedback in a tracking model movement initiation. This posture represents a person
ensures that errors with respect to an assumed reference leaning with the wall in a static position, that is, no
trajectory go to zero with time. We propose the following velocities at the joint. The upright standing posture is a
general form of the individual joint reference trajectories stable equilibrium position, at which model is linearized,
for postural and voluntary movement: and we follow with the assumption that initial and final
velocities are zero in our simulation. Introducing initial
hRef ðtÞ ¼ x0 þ ðxf  x0 Þ  xR ðtÞ ð16Þ velocities means to accommodate the reaction forces into
where x0 and xf define the initial and final posture in terms the model (such as seat reaction force). The intended
of joint angles, and xR ðtÞ is either a sigmoid function given postural movement terminates with the COM positioned
over the BOS, and with zero COM anterior velocity.
Without loss of generality, we assume the terminal position
θ(t) to be the stance position with the COM positioned over the
Nonlinear ankle joint. This position is also picked to minimize the
Model Reference
+ control effort at movement termination. The selected initial
Trajectories
and final positions for postural movement are sketched in
u(t) xr(t) Fig. 3. For illustration, the starting position for postural
Passive
movement is selected as [1.37, 1.47, 1.42] rad. The COM at
Controller x(t)
movement initiation is located at about 35 % (of foot
+
length, see Fig. 7) behind the heel. The movement termi-
Active nates with final joint angular positions at [1.57, 1.57, 1.57]
Controller rad.
Simulation of the nonlinear model (2), with active and
Fig. 2 Simulation block diagram for the biomechanical model passive torques generated using (8), shows a smooth stable

123
Cogn Comput (2012) 4:549–562 555

case of the hybrid controller. The knee and hip torques start
as extensor torques that later switch to flexor torques to aid
in movement termination. Finally, we note that the observed
maximum torque values are all within their physiological
limits.
On the basis of the simulation results for the postural
movement, COM-based optimal controller generated an
overall smoother (i.e., shape of profile angle) dynamic
response in the angular profiles for ankle, knee, and hip
joints. The GRF-based optimal controller generated rela-
tively faster response, which was less smooth at least in
Fig. 3 a Initial position, x0, and b final position, xf, for the postural
movement case of knee joint. The hybrid design scheme demonstrated
the similar results like COM-based design in the angular
dynamic response from initial position x0 to final position xf. profiles, while also minimizing the torques at the ankle and
Figure 4 shows the joint angle profiles for 4-s simulation of knee joints. Overall, the optimal controller with the Q
the postural movement using optimal controllers derived matrix derived from COM sensitivity functions and the R
from COM, GRF, and hybrid design schemes. The joint matrix derived from GRF sensitivity functions appears to
angular profiles are plotted in radians and are measured with give the best results in the case of postural movement.
respect to the horizontal. There is observed initial opposite Figure 6 shows the relative contribution from the active
direction in the hip and ankle movement; this is due to the and passive torques toward the execution of the postural
simplification of modeling and not introducing initial reac- movement. We note that both active feedforward and
tion forces into the modeling. This phenomenon was studied passive feedback torques settle to zero at movement ter-
in ref [22] with initial velocities, which produces results that mination. This is because no support is required in the
follow reference trajectories very well but controller output assumed stance position when the weight of the body is
beyond the physiological limits. Figure 5 shows the net joint supported over the ankle joint. While there is little varia-
torques generated during simulation of the assumed postural tion in the ankle torque profiles, the passive knee torque is
movement. The maximum joint torques generated during about 20 % lower in the case of hybrid controller. The hip
movement simulation were found to be [106, 46, 28] N-m in torque exhibits maximum variation among the COM- and
the case of COM-based controller, [110, 47, 25] N-m in the GRF-based controllers. The passive torque in the case of
case of GRF-based controller, and [104, 43, 28] N-m in the COM-based controller is about 80 % higher than the GRF

Ankle Knee Hip


1.65 1.64 1.58

1.62 1.56
1.6
1.54
1.6
1.55 1.52
1.58
Angular Profiles

1.5
1.5 1.56 CoM
1.48 GRF
1.54 Hybrid
1.45
1.46
1.52
1.4 1.44

1.5
1.42
1.35
1.48 1.4

1.3 1.46 1.38


0 2 4 0 2 4 0 2 4
time time time

Fig. 4 Joint angles in radians for the postural movement with COM-based design (solid), GRF-based design (dash), and hybrid design scheme
(dash dot). The joint angles are in radians measured from horizontal

123
556 Cogn Comput (2012) 4:549–562

Ankle Knee Hip


120 50 30

25
100 40

CoM
20
80 30
GRF
Hybrid
Net Torques

15

60 20

10

40 10
5

20 0
0

0 -10 -5
0 2 4 0 2 4 0 2 4
time time time

Fig. 5 Net joint torques (N-m) for the postural movement with COM-based design (solid), GRF-based design (dash), and hybrid design scheme
(dash dot)

controller and also reaches its peak later in time, which Figure 7 shows the horizontal COM/COP position
shows that the relatively large deviation in the hip angular (normalized by foot length) and the horizontal GRF
profile (Fig. 4) takes longer to settle. The hybrid design accompanying the postural movement. The normalized
entails relatively lower passive torque contribution in the COM and COP anterior positions show a smooth variation
case of knee joint and closely follows the COM design for from initial 35 % behind heel to the final ankle position
both ankle and hip joints. Finally, as observed from the (20 % ahead of the heel). It is noted that the initial COP is
figure, significant passive torques were generated under the placed outside of the BOS due to zero initial conditions
various controller schemes used for postural movement. assumed on joint angular velocities at movement initiation.

Ankle Knee Hip


100 50 25

CoM
90 45
GRF
Active (feed-forward) torque component

80 40 20 Hybrid

70 35

60 30 15

50 25

40 20 10

30 15

20 10 5

10 5

0 0 0
0 50 -10 0 10 20 30 -10 0 10 20
Passive (feedback) torque component

Fig. 6 Active versus passive torques for postural stabilization

123
Cogn Comput (2012) 4:549–562 557

Fig. 7 Normalized COM, Fx


x-CoM x-CoP
normalized COP, and horizontal 1.5 1.5 20
GRF accompanying postural CoM
movement for the three GRF
controller design schemes 1.4 1.4
Hybrid 15
CoM
GRF

Ground Reaction Forces


1.3 1.3
Hybrid
10

CoM Profiles
1.2 1.2
5
1.1 1.1

0
1 1

-5
0.9 0.9

0.8 0.8 -10


0 2 4 0 2 4 0 1 2 3 4
time time time

We further note that following 5 % increase at the begin- lift off had taken place earlier and no chair and/or hand
ning, COM and COP trajectories stay in phase over the reaction forces were taken into consideration.
course of postural movement, which is in conformity with As discussed above, coordination of voluntary movement
the physiological observation [8]. The GRF excursions required specification of reference trajectories for the joint
were slightly higher in the case of GRF-based design as angles. Figure 9 shows the resulting angular profiles from
compared to the COM controller, while the hybrid scheme simulation of STS movement for the three controller designs,
produces lower peaks in the GRF variable. In general, the where the adopted joint reference trajectories for STS task
COM minimization scheme appears to be more effective in are also shown in the figure. As seen from the figure, COM-
controlling both the COM and the GRF behavior. based controller design produces joint trajectories that are in-
phase with the reference trajectories in the case of ankle and
Simulation of Voluntary Movement knee joints, but not so in the case of hip joint. The joint
trajectories for the GRF-based design are in-phase with the
We used sit-to-stand (STS) movement as an example of reference trajectories for both ankle and hip joints. We note
point-to-point voluntary movement to assess the perfor- that the peak ankle joint variation is higher (2.5 times the
mance of the various controller design schemes. The assumed reference trajectory value in the case of COM-
desired initial [(1.57, 0, 1.57) rad] and final [(1.57, 1.57, based design and about 4 times in the case of GRF design),
1.57) rad] positions for the assumed STS movement are resulting in higher ankle settling times in the case of GRF
sketched in Fig. 8. Similar to the postural movement, we design. The same, however, is 50 % lower than the assumed
assumed zero initial joint angular velocities at the initiation reference trajectory peak in the case of hip joint. Also, the
of STS movement. Further, it was assumed that the chair COM-based controller generated a hip trajectory that
appears to be out of phase with respect to the assumed ref-
erence trajectory for that joint. The hybrid design scheme
appears to provide the better responses between the COM-
and GEF-based schemes.
Figure 10 shows the net joint torques for the STS move-
ment. These results are recast into the active–passive torque
comparison in Fig. 11, where we observe that the active hip
torque for STS movement stays negative, that is, an extensor
torque; however, net hip torque shows initial flexion torque
followed by extension torque, which is required to facilitate
the movement in upward extension phase. We note that
though passive feedback torques are high for the STS
movement, their relative contribution toward overall effort is
Fig. 8 a Initial and b final position for the STS movement comparable to the postural movement studied above. Also,

123
558 Cogn Comput (2012) 4:549–562

Fig. 9 Angular profiles (rads) Knee Hip


Ankle
for STS movement with COM- 1.6 1.6 2.1
based design (solid), GRF-based CoM
design (dash), hybrid scheme
1.5 1.4 2 GRF
(dash dot), and the reference
trajectories (light solid) Hybrid
1.4 1.2 Ref
1.9

1.3 1
1.8

Angular Profiles
1.2 0.8
1.7
1.1 0.6

1.6
1 0.4

1.5
0.9 0.2

0.8 0 1.4

0.7 -0.2 1.3


0 2 4 0 2 4 0 2 4
time time time

Fig. 10 Net joint torques (N-m) Ankle Knee Hip


for the STS movement with 450 350 60
COM-based design (solid), CoM
GRF-based design (dash), 400
300 GRF
hybrid scheme (dash dot) 40
Hybrid
350
250
300 20

200
250
Net Torques

200 150

-20
150
100

100 -40
50
50
-60
0
0

-50 -50 -80


0 2 4 0 2 4 0 2 4
time time time

from Fig. 11, the hybrid controller seems to generate smooth Discussion and Conclusion
transitions between the active and passive joint torques.
Figure 12 shows the COM/COP variation and the horizontal This study investigated the physiological cost optimization
GRF during STS movement. Finally, we observe that while in the synthesis of LQR controller for simulation of pos-
the COM-based controller does a good job of minimizing the tural and voluntary movement in the case of a four-segment
COM variation during STS movement, the GRF-based sagittal computational dynamic model. Further, we ana-
controller appears to do a poor job of minimizing the hori- lyzed the relative contribution from active control (feed-
zontal GRF. Once again we observe that following a slight forward processes) and passive stiffness (feedback
initial dip in COP, the COM and COP trajectories stay in processes) in the performance of the said movements. The
phase over the course of STS movement. sagittal model used in this study comprised four rigid-body

123
Cogn Comput (2012) 4:549–562 559

Ankle Knee Hip


250 250 10 CoM
GRF
Hybrid
0
Active (feed-forward) torque component

200 200
-10

-20
150 150

-30

100 100
-40

-50
50 50

-60

0 0 -70
0 200 0 100 0 50 100
Passive (feedback) torque component

Fig. 11 Active versus passive torques for the STS movement

Fig. 12 Normalized COM Fx


x-CoM x-CoP
position and the horizontal GRF 2.6 4.5 400
for STS movement CoM
2.4 4 GRF 300
Hybrid
2.2
3.5 200 CoM
Ground Reaction Forces

GRF
2
3 100 Hybrid
CoM Profiles

1.8
2.5 0
1.6
2 -100
1.4

1.5 -200
1.2

1 1 -300

0.8 0.5 -400


0 2 4 0 2 4 0 1 2 3 4
time time time

links representing stationary foot, leg, thigh, and HAT. The segments was represented via torque actuators located at
applied torques at the ankle, knee, and hip joints included a the joints. The limited bandwidth effect of the muscle
feedforward component representing muscle activation via contractile dynamics was largely ignored in the simplified
descending commands from higher centers of the CNS, and mathematical model used in this study. Linearization of
a feedback component representing passive viscoelasticity model dynamics around standing posture was undertaken
at the joints. We did not explicitly model afferent feedback purely for the purpose of linear optimal LQR controller
from spindle, Golgi receptors, spinal reflexes, and/or long- synthesis to settle at steady-state end point of the move-
loop responses, eliminating the need to consider the neural ment. Standing posture is an equilibrium point in this
transmission delays in the motor servo loop. The function model with negligible feedback torques to keep the upright
of biological muscle actuators in positioning the body standing posture. In our previous studies [22], we presented

123
560 Cogn Comput (2012) 4:549–562

TSK fuzzy model with linearization at sitting and standing movement termination. In the case of STS movement, the
posture, and this scheme can also be extended further with passive torques peaked twice for the knee joint, showing
more linearizing points in the trajectory. However, there somewhat jerky movement at that joint. We note that
are also some controller design problems for multiple lin- though the relative contribution of passive torques toward
earizing points and combing the controller for full trajec- postural stabilization and movement coordination was
tory. Full nonlinear model was employed for the simulation found to be significant, overall torques in all cases were
of postural and voluntary movement in order to evaluate well within physiological limits. In this regard, it is also
the effectiveness of the controller. Sensitivity derivates of worth noting that the passive joint torques developed dur-
the COM anterior position and the horizontal GRF were ing simulation of postural and voluntary STS movement
employed as weights in the physiological cost optimization were physiologically feasible, and the corresponding joint
leading to LQR design. Specifically, we separately con- stiffness values, as represented in the physiologically
sidered LQR designs arising from minimization of sensi- optimized controller gain matrix, were not excessive. In
tivity functions representing the COM anterior position and other words, movement simulation by the simulation model
the horizontal GRF variables with respect to joint positions and the controller used in this study did not generate
and velocities. A hybrid scheme combining the two excessive passive torques as was the case in modeling
approaches was also considered. study in [8]. In addition, we similarly observed a close
Simulation results were presented for two kinds of phase relationship between COM and COP with physio-
movements: (a) a representative postural adjustment of the logically feasible joint torques as opposed to [8], which
type elicited following a perturbation to the BOS and suggested excessive ankle torque for their stiffness-only
(b) voluntary STS movement as a representative of skilled model. This may be due to the additional DOF considered
voluntary movements. We used the same physiological in our study and further reinforces the role of knee joint in
LQR controller gains for both movements. The reference postural stabilization. Finally, we note that in modeling
trajectories for the STS movement were generated from a studies, the overall joint torques underlying a particular
third-order linear model [30]. No reference trajectories in movement can be abridged through a movement optimi-
the case of postural movement were considered. Our sim- zation process [18]. In that study, the authors have pro-
ulation results for both movements indicate that, in general, posed a multi-criteria optimization where the cost function
the COM-based controller produced smoother movement included feedback torques generated during movement
profiles with comparable settling times (about 50 % longer simulation in addition to residual dynamics and/or position
settling time in the case of hip trajectory for postural errors at movement termination. Other physiologically
movement). Since the COM represents a kinematic vari- relevant cost functions can be used for this purpose. For
able, as opposed to GRF, which is a kinetic variable, this example, energy optimal control of disturbance rejection
study shows that the sensitivity functions of kinematic has been proposed by [31].
variables are more relevant toward optimization of postural The effectiveness of active and/or passive mechanisms
and voluntary movement. These results thus corroborate of postural control has been a topic of intense debate
the experimental results of Resiman, Scholz, and Schöner among researchers [8–14]. Motivated by the experimental
[20, 21] regarding applicability of kinematic variables. In observation that the oscillation on the support surface of
that study, the authors also found that subjects preferred a the COM appeared to be in phase with the COP, and the
range of state–space combinations that did not appreciably theoretical consideration that such phase lock was incom-
disturb the COM linear and angular momentums during the patible with active control of balance, Winter et al. [8]
performance of STS movement. This may indicate that advanced a simpler stiffness-based model to solve the
movement learning resulted in the selection of optimal complex postural stabilization problem. According to these
COM momentum for the STS task. However, since our researchers, stabilization of quiet standing might be
biomechanical model did not possess the many redundant achieved by the stiffness of ankle muscles alone without
DOF considered in [20, 21], the results in the two studies any significant active or reactive component except for
may not be directly comparable. background setting of the stiffness parameters. Their view
The relative contributions of the active feedforward and of postural stabilization drew further support from the well-
passive feedback joint moments in the case of postural and known equilibrium point hypothesis [32], which states that
voluntary STS movements were studied in Figs. 6 and 11. well-coordinated multi-joint movements made extensive
In all cases, passive torques were assumed to be zero at use of spring-like properties of the musculo-tendon appa-
moment initiation. As the intended movement progressed, ratus and peripheral neural feedback loops, thus alleviating
passive torques build up as a function of deviation from the computational burden on the CNS. The role of CNS in
assumed reference trajectory reaching a peak value fol- these movements was relegated to background setting of
lowed by a recession and finally going back to zero at the threshold lengths of the muscles involved. In addition,

123
Cogn Comput (2012) 4:549–562 561

movement trajectories could be simply constructed through Table 1 Definition of biomechanical four link model terms
a gradual shift in the equilibrium point without explicit Inertial components [d11 d12 d13 [11.2087 9.6933 5.9652
compensation for dynamics [33]. Equilibrium point (N s2) d22 d23 d33] 9.3593 5.9652 11.8305]
hypothesis, though well established in the motor control
Gravitational [f1 f2 f3] [26.45 22.54 13.87]
community for over two decades and in particular, advance-
components (Nm)
ment in stiffness measurement has enabled researchers
Foot mass (kg) mf 1.91
[34, 35] to conclude that computer simulations presented in
Foot length lf 0.27
support of the hypothesis may have been based on faulty
Ankle Heel (m) a 0.05
assumptions about stiffness in the joints. The research on
Ankle Height (m) b 0.07
this topic thus remains inconclusive. Our simulation results
Ankle–foot COM (m) c 0.08
support an active–passive model of postural and voluntary
Gravity g (m/s2) g 9.8
movement stabilization.
Mass of model (kg) m 64.09
Researchers in [36, 37] presented the movement coor-
dination of biologically inspired robots derived from
human cognition. These movements should be optimally s1 þ bFx  cmf g
controlled for energy minimization as well as maintaining xCOP ¼ lf  a þ ðA4Þ
Fy
the coordination between the limbs (or extremities).
Movement coordination in STS task is dependent upon References
physiological variables as reported by experiments in [20,
21]. Our model emphasizes the movement coordination 1. Peterka RJ. Postural control model interpretation of stabilogram
based upon these experimental results for computation of diffusion process. Biol Cybern. 2000;82:335–43.
control mechanism, which is both cognitive in nature and 2. Peterka RJ. Sensorimotor integration in Human Postural Control.
J Neurophysiol. 2006;88:1097–118.
optimal for amount of effort to reach its task objective. In 3. Van Soest AJ, Haenen WP, Rozendaal LA. Stability of bipedal
future, applications such as rehabilitation or robotic engi- stance: the contribution of cocontraction and spindle feedback.
neering for prostheses will use such frameworks that are Biol Cybern. 2003;88:93–301.
cognitive controlled and optimal in nature for more natural 4. Maurer C, Mergner T, Peterka RJ. Multisensory control of human
upright stance. Exp Brain Res. 2006;171:231–50.
movement. This framework will be extended with fuzzy 5. Nashner LM, McCollum G. The organization of human postural
and neural modeling of movement coordination with movements: a formal basis and experimental synthesis. Behav
physiological optimization in order to compute better Brain Sci. 1985;8:135–72.
cognitive control for entire voluntary motion for human or 6. Horak FB, Nashner LM. Central programming of postural
movements: adaptation to altered support-surface configurations.
biologically inspired robotic applications. J. Neurophysiology. 1986;55:1369–81.
7. Maki BE, McIlroy WE. The role of limb movements in main-
Acknowledgments This work was supported in part by a grant from taining upright stance: the change-in-support strategy. Phys Ther.
the Arkansas Science and Technology Authority. 1997;77:488–507.
8. Winter DA, Patla AE, Prince F, Ishac M, Gielo-Perczak K.
Stiffness control in quiet standing. J Neurophysiol. 1998;80:
Appendix 1211–21.
9. Todorov E, Jordan MI. Optimal feedback control as a theory of
motor coordination. Nat Neurosci. 2002;5:1226–35.
The inertial, Coriolis, and gravitational matrices appearing 10. Schaal S, Schweighofer N. Computational motor control in
in the nonlinear dynamic model (1) are given as [18]: humans and robots. Curr Opin Neurobiol. 2005;15:675–82.
2 3 11. Harris CM, Wolpert DM. Signal-dependent noise determines
d11 d12 cosðh1  h2 Þ d13 cosðh1  h3 Þ motor planning. Nature. 1998;394:780–4.
6 7
D ¼ 4 d12 cosðh1  h2 Þ d22 d23 cosðh2  h3 Þ 5 12. Kuo AD. An optimal control model for analyzing human postural
balance. IEEE Trans Biomed Eng. 1995;42:87–101.
d13 cosðh1  h3 Þ d23 cosðh2  h3 Þ d33
13. Kuo AD. An optimal state estimation model of sensory integra-
ðA1Þ tion in human postural balance. J Neural Eng. 2005;2:235–49.
14. Pandy M. Computer modeling and simulation of human move-
H¼ ment. Annu Rev Biomed Eng. 2001;3:245–73.
2 3
0 d12 h5 sinðh1  h2 Þ d13 h6 sinðh1  h3 Þ 15. Van der Helm FCT, Rozendaal L. Musculoskeletal Systems with
6 7 Intrinsic and Proprioceptive Feedback. In: Winters JM, Crago PE,
4 d12 h4 sinðh1  h2 Þ 0 d23 h6 sinðh2  h3 Þ 5
editors. Biomechanics and neural control of movement and pos-
d13 h4 sinðh1  h3 Þ d23 h6 sinðh2  h3 Þ 0 ture. New York: Springer-Verlag; 2000. p. 164–74.
ðA2Þ 16. Barin K. Evaluation of a generalized model of human postural
dynamics and control in the sagittal plane. Biol Cybern. 1989;
G ¼ g½ f1 cos h1 f2 cos h2 f3 cos h3  ðA3Þ 61:37–50.

123
562 Cogn Comput (2012) 4:549–562

17. Pai YC, Iqbal K. Simulated movement termination for balance 27. Winter DA. Biomechanics and motor control of human move-
recovery: can movement strategies be sought to maintain stability ment. New York: Wiley; 1990.
even in the presence of slipping or forced sliding? J. Biome- 28. Brooks V. The neural basis of motor control, Chapter 4. Oxford:
chanics. 1999;32:779–86. Oxford University Press; 1986.
18. Iqbal K, Pai YC. Predicted region of stability for balance 29. Keshner E, Allum J. ‘‘Muscle activation patterns coordinating
recovery: motion at knee joint can improve termination of for- postural stability from head to foot’’ In Multiple muscle systems:
ward movement. J. Biomech. 2000;33:1619–27. biomechanics and movement organization, New York: Springer-
19. Kerr KM, White JA, Barr DA, Mollan RAB. Analysis of sit-to- Verlag; 1990. pp. 481–497.
stand movement cycle in normal subjects. Clin Biomech. 30. Mughal AM, Iqbal K. ‘‘Synthesis of angular profiles for bipedal
1997;12:236–45. sit-to-stand movement’’, 40th southeastern symposium on system
20. Resiman DS, Scholz JP, Schöner G. Coordination underlying the theory, New Orleans, LA, 2008.
control of whole body momentum during sit-to-stand. Gait Pos- 31. Mihelj M, Munih M, Ponikvar M. Human energy: optimal control
ture. 2002;15:45–55. of disturbance rejection during constrained standing. J Med Eng
21. Scholz JP, Resiman DS, Schöner G. Effects of varying task Technol. 2003;27:223–32.
constraints on solutions to joint coordination in a sit-to-stand task. 32. Feldman AG. Functional tuning of the nervous system with
Exp Brain Res. 2001;141:485–500. control of movement or maintenance of a steady posture. ii.
22. Mughal AM, Iqbal K. ‘‘A fuzzy biomechanical model with H2 controllable parameters of the muscle. Biophysics. 1996;11:
optimal control of sit-to-stand movement’’, Proceedings of the 565–78.
2006 American control conference, Minneapolis, MN; 2006. 33. Shadmehr R. The equilibrium point hypothesis for control of
pp. 3427–3432. movement. Baltimore, MD: Department of Biomedical Engi-
23. Mughal AM, Iqbal K. ‘‘A comparison of fuzzy model based neering, Johns Hopkins University; 1998.
optimal control systems for biomechanical sit-to-stand move- 34. Gottlieb GL. Rejecting the equilibrium-point hypothesis. Mot
ment’’, Proceedings of the IEEE international conference on Control. 1998;2:10–2.
engineering of intelligent systems, Islamabad, Pakistan; 2006. 35. Gomi H, Kawato M. Equilibrium-point control hypothesis
24. Jo S, Massaquoi SG. A model of cerebellum stabilized and examined by measured arm stiffness during multijoint movement.
scheduled hybrid long-loop control of upright balance. Biol Cy- Science. 1996;272:117–20.
bern. 2004;91:188–202. 36. Cecilia L, Roland J. Bio-inspired sensory-motor coordination.
25. Jo S, Massaquoi SG. A model of cerebrocerebello-spinomuscular Auton Robots. 2008;25:1–2.
interaction in the sagittal control of human walking. Biol Cybern. 37. Cecilia L, Gioel A, Eugenio G, Giancarlo T, Roland J, Hitoshi K,
2007;95(3):279–307. Zbigniew W, Maria C, Paolo D. A bio-inspired predictive sen-
26. Mughal AM, Iqbal K. ‘‘Active control vs. passive stiffness in sory-motor coordination scheme for robot reaching and pre-
posture and movement coordination’’, IEEE International con- shaping. Auton Robots. 2008;25:85–101.
ference on systems, man and cybernetics, Montreal, 2007.
pp. 3372–3376.

123

You might also like