See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269099032

Optimal Pattern Generator For Dynamic Walking in humanoid Robotics

Article · December 2014

CITATIONS READS

4 277

4 authors:

David Galdeano Ahmed Chemori

Altran Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier…
12 PUBLICATIONS   23 CITATIONS    161 PUBLICATIONS   576 CITATIONS   

SEE PROFILE SEE PROFILE

Sébastien Krut Philippe Fraisse

Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier… Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier…
104 PUBLICATIONS   1,369 CITATIONS    196 PUBLICATIONS   1,279 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Aerial robotics View project

ANR-ARROW: Accurate and Rapid Robots with large Operational Workspace View project

All content following this page was uploaded by Ahmed Chemori on 04 December 2014.

The user has requested enhancement of the downloaded file.

 c 2014 TSSD
1861-5252/ Transactions on
Systems, Signals & Devices
Vol. 9, No. 1, pp.1-20

F
Optimal Pattern Generator For Dynamic
Walking in humanoid Robotics

D. Galdeano, A. Chemori, S. Krut and P. Fraisse

OO
LIRMM, Université Montpellier 2
LIRMM - UMR 5506 - CC 477, 161 rue Ada, 34095 Montpellier Cedex 5, France.

Abstract This paper deals with an optimal Zero Moment Point (ZMP)
based pattern generator for stable dynamic walking in humanoid
robotics. The proposed method is based on a Three-Mass Linear
Inverted Pendulum Model (3MLIPM), used as a simplified model
of the biped robot. The 3MLIPM considers the biped robot as
a three point masses and two-link system. A ZMP based per-
formance index is then used in an optimization problem whose
solution gives the best values of the model’s parameters w.r.t.
dynamic walking stability. Numerical simulations are presented
to show the effectiveness of the proposed optimal pattern gener-
ator for the case of the biped walking robot SHERPA.

Keywords: Biped walking robot, Pattern generation, Three-mass linear in-

verted pendulum model, Optimization.
PR
1. Introduction
A humanoid robot is a robot with an appearance based on that of the
human body. Humanoids walking is a very challenging field of research
due to the complexity of synthesizing stable gaits for these systems.
Indeed, within this field, humanoid robot control needs sophisticated
control schemes to deal with their complexity including:

the high-order nonlinear dynamics,

the variable structure model according to the different phases of
the walking cycle,
the contact constraints with the ground that should be managed,
2 D. Galdeano et al.

the hybrid character of the dynamics due to rigid impacts between

the robot’s foot and the ground,
the stability during walking that should be ensured.

F
Most of the proposed control schemes in the literature are based on
the use of some reference trajectories that should be tracked in real-
time. This fact shows the importance of a pattern generator in humanoid
walking control. Several types of pattern generators have then been
proposed in the literature. However, those who guarantee an a-priori
walking stability are often based on one of the following stabilization
criteria:

OO
The Center Of Mass (COM)
The Zero Moment Point (ZMP)
The Foot Rotation Indicator (FRI)

The COM [1] is the mean location of all masses of the robot links. It
is usually used as a static stability criterion. The ZMP [2, 3] is the point
of junction between the center of the vertical reaction forces and the
ground. It is the most used dynamic stability criterion. The FRI [4] is a
point on the foot/ground-contact surface where the net ground-reaction
force would have to act to keep the foot stationary. It is an indication
of postural stability and, in case of instability, indicates how the robot
will fall. It is used as a dynamic stability criterion.

2. Related works
PR
In the literature several methods have been proposed for generating
walking trajectories in humanoid robotics.
A method based on a motion capture of human walking has been
proposed in [5]. This method can gives human-like motions, however its
main drawback lies in the captured data that can be hard to adapt to
the humanoid robot.
Another method of trajectory generation for stable dynamic walking
is proposed in [6]. It consists in using a 3rd order spline function to
generate feet and hip trajectories. The foot trajectories can be adapted
to the ground variations to generate a stable dynamic walking on a rough
terrain.
In [7] another pattern generation method has been proposed. It con-
sists in using Fourier series to generate stable walking, with an iterative
Optimal Pattern Generator For Dynamic Walking 3

procedure to guarantee the stability. The main drawback of such a

method lies in the computation time that does not enable a real-time
implementation.
Another interesting method is the so called Inverted Pendulum Model

F
(IPM) [8, 9] that considers the robot as a single point mass and massless
legs. This method simplifies the dynamics of the robot to an inverted
pendulum with a point mass linked by a telescopic leg to a spherical
ground/leg joint.
The Linear Inverted Pendulum Model (LIPM) [10, 11] is an extension
of the IPM where the height of the torso is considered to be constant
leading to a more natural movement.

OO
The IPM ignores the dynamics of the legs since it considers the stance
leg as an inverted pendulum with a point mass. This fact can involve
a loss of walking stability since the considered model is considerably
simplified and consequently not sufficiently accurate. For instance, this
is the case when the robot has heavy legs or has no torso (case of our
robot SHERPA). In order to approximate the robot with a more accurate
model, it would be necessary to consider the whole-body dynamics by
using a preview control to correct the error between the IPM and the
real robot as in [12] or to use a multiple point mass model instead of
considering a one point mass model.
For instance, a Two-Mass Inverted Pendulum Model (TMIPM) and
Multiple Mass Inverted Pendulum Model (MMIPM) extending the IPM
were proposed in [13]. In the MMIPM, the considered model is composed
by more than one mass. For instance these masses are located at the hip
and along the swinging leg in the case of a two-mass model.
PR
The Gravity-Compensated Inverted Pendulum Model (GCIPM) [14]
uses also one mass to represent the body of the robot and an additional
mass to represent the swing leg.
A three-mass model using the concept of ZMP has been studied in
[15]; however, it doesn’t use any LIPM to generate trajectories and the
torso of the biped robot moves up and down.
A generation of walking trajectories using a three point mass model
to calculates center of mass trajectories from footstep locations has been
proposed in [16]. This approach use offline optimization of some free
geometrical parameters like the trunk angular motion w.r.t. the speed
of the robot. These parameters are then used in a real-time fast planning
to compute the reference torque patterns to apply on the robot.
Another three-mass model using the inverted pendulum concept is
proposed in [17]. However, it has one important drawback related to
4 D. Galdeano et al.

the geometrical parameters, such as the position and the location of

masses, which have been chosen arbitrarily. One good idea would then
be to tune these parameters at their best values in order to enhance the
walking stability of the robot, and this is the scope of the present paper.

F
The proposed solution is an extension of the method proposed in [17]
to deal with dynamic stability of walking in the generated trajectories,
as well as changes in direction during walking. As in [17], the robot’s
dynamics will be approximated by a Three-Mass Linear Inverted Pen-
dulum. The hip and feet trajectories are generated by the movement of
masses. The walking movement of the robot is generated in the sagit-
tal and frontal planes separately. The joint trajectories are computed

OO
using inverse kinematics of the biped robot. In order to ensure a sta-
ble dynamic walking, the parameters of the model must be well tuned.
The best way to perform such a tuning is through an optimization of
the generated joints’ trajectories to minimize the ZMP excursion within
the footprint which will increase the stability margins. The second con-
tribution of this work is to propose online change of walking direction,
where new optimal trajectories are generated to ensure a-priori stability
during walking while turning.
This paper is organized as follows: in the next section, the proto-
type of our demonstrator SHERPA is introduced. Section 4 introduces
the simplified model that will be used in the generation of the reference
trajectories and how it was used in the pattern generator proposed in
[17]. In section 5, the proposed extension of this method is presented,
where our contributions are highlighted. Numerical simulations are pre-
sented and discussed in Section 6, where the effectiveness of the pro-
posed method is shown and compared with the original one for the case
PR
of SHERPA robot. The paper ends with some concluding remarks.

3. Description of SHERPA prototype

The SHERPA walking robot (cf. Fig. 3) is a French biped robot
developed at the LIRMM laboratory [18] within the framework of the
national project ANR-06-BLAN-0244 SHERPA. The name of this robot
comes from the so-called Sherpa, who are members of a people of Tibetan
stock living in the Nepalese Himalayas, and who often serve as porters
on mountain-climbing expeditions. Indeed, this robot is built to carry
loads while walking in a human environment [19].
SHERPA is composed of a hip linking two legs together. Each leg has
six degrees of freedom (dof), and the robot is equiped with 12 actuators,
which is enough to reproduce a human gait.
Optimal Pattern Generator For Dynamic Walking 5

F
OO
Fig. 1. SHERPA kine-
matics model
Fig. 2. CAD model of
SHERPA
Fig. 3.
robot
The SHERPA

These dofs are distributed on the different articulations of the robot

as follows: three dofs at the hip, one dof at the knee and two dofs at
the ankle as it is illustrated on Fig. 1. The geometric parameters of the
robot are summarized in Tab. 1.

Table 1. Geometric parameters of SHERPA biped robot.

Description Parameter Value

Length of thigh Lthigh 0.5033 m
Width of the hip Ls 0.31 m
Total weight of the robot mass 47.1 kg
PR
This robot differs basically from the other walking robots by its actu-
ation system, which is transparent (backdrivable, with low inertia), and
organized in modules (as illustrated in Fig. 4–5).
Each actuation module includes two actuators acting in parallel on
two dof simultaneously. The mechanical transmission of these modules is
such that when the two actuators work together on the joint, they cause
the movement of the first dof; and when they act in opposite directions
they cause the movement of the second dof. These modules are equipped
with custom-made electric motors.
The transmission of movement is based on the use of a ball screw
transforming the rotation of the hollow shaft electrical actuator in a
translational movement. This last one is then transmitted to pulleys
6 D. Galdeano et al.

using cables to produce the desired rotational movement. This basic

principle of motion transmission is illustrated in Fig. 6.

F
moduleOO
Fig. 4. CAD view of an actuator Fig. 5. View of a real manufactured
module
PR
Fig. 6. Cable/Pulley transmission system used in actuation module of
SHERPA.

This technology gives the robot’s actuators remarkable characteristics

such as no backlash, low friction, reversibility of the chain of transmission
and low inertia.
The kinematic model of SHERPA robot is created through the link
representation proposed in [20]. This model uses the 12 joints’ values
as well as the position and orientation of the hip to generate the Carte-
sian position of each articulation in the operational space. The forward
kinematics model writes:
X = f (q) (1)
Optimal Pattern Generator For Dynamic Walking 7

where X ∈ R39×1 includes Cartesian positions of articulations and the

hip’s center position, q ∈ R18×1 includes a 12 × 1 vector of joints posi-
tions, and a 6 × 1 vector of the position and orientation of the hip.
The inverse kinematic model is obtained by solving an analytical equa-

F
tion to find the articular positions expressed in terms of operational space
positions of the hip and feet. The inverse kinematic model can then be
written as:
q = g(Xr ) (2)
where Xr ∈ R9×1 represents the operational space position of the hip
(3 × 1) and the feet (6 × 1). The orientation of the hip and the feet are

OO
kept constant and equal to zero.
The computation of the ZMP and CoM of the robot is based on
the formalism presented in [20], in which the ZMP evaluation uses the
angular momentum with the overall dynamic model of the robot.

4. Three-mass linear inverted pendulum model

The three-mass Linear Inverted Pendulum Model (3MLIPM) as intro-
duced in [17] simplifies the biped robot to a three-link system (as shown
in Fig. 7) with a point mass on each link. The three masses represent
the torso and the two legs, unlike the single mass model (as in IPM or
LIPM) where a unique mass is located at the hip of the robot. The three
links are connected together at the hip.
This model is more accurate than the single mass model, especially
for biped robots without torso, where the position of the CoM can be
very different from the hip position.
PR
Z

m2

m1
m3

Y
X

Fig. 7. Graphical representation of the Three-Mass Linear Inverted Pendulum

Model.
8 D. Galdeano et al.

From Fig. 7, the equations of moment applied on the supporting

ankle can be formulated as follows:
3
X 3
X
τx = mi (g + z̈i )yi − mi ÿi zi (3)

F
i=1 i=1
3
X 3
X
τy = mi ẍi zi − mi (g + z̈i )xi (4)
i=1 i=1

where τx and τy are the torques applied on the ankle, mi represents the
value of the ith point mass, (xi , yi , zi ) are the Cartesian positions of the

OO
ith point mass (i = 1, 2 or 3 represent respectively the stance leg, the
torso and the swing leg), (ẍi , ÿi , z̈i ) are their corresponding accelerations.
For simplification, the height of the masses is assumed to be constant
and the torque applied on the supporting ankle is assumed to be zero
[17]. With these assumptions, (3) and (4) can be simplified into:
3
X

i=1
mi ÿi zi =
3
X

i=1
mi gyi and
3
X

i=1
mi ẍi zi =
3
X

i=1
mi gxi (5)

These equations are decoupled, so the movement can be generated in

sagittal and lateral planes separately.

4.1 Assumptions and notations

The following assumptions are considered to simplify the calculations
[17]:
PR
The ground is flat and horizontal.
The height variation of each mass can be neglected.
The double support phase is considered instantaneous.
There is no energy loss during impact.
The standing foot has no overturn when touching and leaving the
ground.
The swing foot is parallel to the ground.
The torso is upright.

Besides, the following notations are used:

Optimal Pattern Generator For Dynamic Walking 9

2T : The whole stepping cycle.

2Ds : The steeping length.

Ls : The width of a step.

F
4.2 Motion study in the sagittal plane
The movement in the sagittal plane and the associated notations are
illustrated in Fig. 8.

OO
Fig. 8. Illustration of movement in the sagittal plane.
PR
With the above assumptions, heights and masses can be normalized
as follows:

z2 = γz1 , z3 = z1 , x2 = λx1 , m2 = km1 , m3 = m1 (6)

where γ, λ and k are the parameters extracted from the robot’s model.
Considering (6), equation (5) can be simplified into:

bẍ1 + dẍ3 = ax1 + x3 (7)

z1 z1
a = kλ + 1 , b = (1 + kγλ) , d= (8)
g g

In order to generate walking gaits, a trajectory for the swing foot

(mass m3 ) is necessary. If the trajectory of this last one is known, then
those of the two other masses can be computed. According to [17], a
10 D. Galdeano et al.

sinusoidal function is chosen to represent the trajectory of the mass m3 ,

as follows:

x3 = A cos(ωt + φ) (9)

F
where parameters φ and A are given in the sequel by equation (13).
Here, ω is not the step frequency, it is a parameter that must be chosen
to keep the foot position and velocity positive in order to ensure a forward
movement of the foot during the step.
Using (7) and (9), x1 can be expressed by:
√a √a
x1 = C1 cos(ωt + φ) + C2 e( b )t + C3 e−( b )t

OO
With the above assumptions given in section 4.1, initial conditions
can be fixed as:

x1 (0) = −

x1
T
2
Ds
 2λ

= 0 , x3
Ds
, x3 (0) = − (2λ − 1)
 2λ
T
2
=0
(10)

(11)

(12)

The equations’ coefficients can then be identified using initial condi-

tions:
π T Ds (2λ − 1)
φ=− − ω , A=− (13)
2 2 2λ cos φ
2
A(1 + dω ) √a
C1 = − 2
C3 = −e b t C2 (14)
a + bω
 
PR
1 Ds
C2 = √a − − C1 cos φ (15)
1 − e bt 2λ

With all trajectories of point masses computed, the trajectories of

the hip and ankles can then be determined using geometrical constraints
from Fig. 8.
xst (t) = 0 for 0 ≤ t ≤ T , xh (t) = λx1 (t) for 0 ≤ t ≤ T
(16)
zst (t) = 0 for 0 ≤ t ≤ T , zh (t) = λz1 for 0 ≤ t ≤ T

where xst , zst represent Cartesian positions of the stance foot along x
and z axis respectively. xh , zh are the Cartesian positions of the hip
along x and z axis respectively.
During the step, the vertical position of the swing foot needs to be
higher than the floor. A sinusoidal shape can then be an appropriate
Optimal Pattern Generator For Dynamic Walking 11

function for the swing foot trajectory along the z axis. However, due
to the previous assumptions, the influence of this change of height is
neglected in computations.
λx3 (t) − λxh (t)
xsw (t) = for 0 ≤ t ≤ T

F
λ−1
√ s   (17)
2Hs 2π π
zsw (t) = 1 + sin t− for 0 ≤ t ≤ T
2 T 2

where xsw , zsw are the positions of the swing foot along the x and z
axis respectively, and Hs is the maximal height of the swing foot during

OO
walking.

4.3 Motion study in frontal plane

The movement in the frontal plane and the associated notations are
illustrated in Fig. 9.

m2 Z

Hip

Knee
PR
m1 m1 m3

Y
0 Ls

Fig. 9. Illustration of movement in the frontal plane.

The 3MLIPM equations are now derived in the frontal plane.

uÿ1 − vy1 = w (18)

Where the coefficients u, v and w are given by:

 
z1 Ls
u = (2 + kλγ) , v = (2 + kλ) , w = Ls + k (19)
g 2
12 D. Galdeano et al.

These equations are solved considering the following initial conditions:

 
T
y1 (0) = 0 , ẏ1 =0 , y1 (T ) = 0 (20)
2

F
The trajectory of the mass m1 along the y axis is then computed and
expressed by:
√v √v w
y1 (t) = C1 e u t + C2 e− u t − (21)
v

Where:
√v

OO C1 =
w 1 − e uT
√v
v 1 − e2 u T
√v
, C2 = e u t C1

Once the trajectory of the mass m1 is computed, those of m2 and

m3 can be easily determined using geometrical constraints illustrated on
Fig. 9. From these trajectories and the inverse kinematics of the robot,
the trajectories of all the joints of the robot can then be computed.

4.4 Adaptation of the general model for SHERPA

(22)

robot
The SHERPA robot is a biped robot without a torso (cf. Fig. 3). In
order to simplify the model, the torso point mass is set on the hip. This
modification results in γ = λ.
It is worth to note that the basic principle of the proposed method
in [17] and summarized above uses some geometrical parameters of the
PR
robot (such as m1 , m2 , m3 and z1 , z2 , z3 ) but doesn’t give a method to
calculate them. In order to take into account the dynamic stability of
the resulting walking trajectories, we propose to tune these parameters
using optimization to enhance the walking performance. The idea of
such contribution is detailed in the following section.

5. Reference trajectories optimization

The proposed pattern generator in [17] computes joint reference tra-
jectories using arbitrary defined parameters. This can be improved by
choosing the best values of some of them (namely m1 and z1 ) to ensure
walking stability. Since we are interested in dynamic walking, the posi-
tion of the Zero Moment Point (ZMP) [2] is then used as an indicator of
stability.
Optimal Pattern Generator For Dynamic Walking 13

5.1 Dynamic stability margins

The stability margins are defined by the distance to the limit of sta-
bility (i.e. the boundary of the footprint), they are illustrated in Fig.
10.

F
OO
Fig. 10. ZMP displacement inside the footprint and stability margins.

The dashed interior rectangle explicits the boundaries of the ZMP

displacements. Based on Fig. 10, the stability margins can be mathe-
matically expressed as follows:
α α
M x1 = +min(dZMPx (t)), M x2 = −max(dZMPx (t))
2 2
β β
M y1 = +min(dZMPy (t)), M y2 = −max(dZMPy (t)) (23)
2 2
∀ t ∈ [tsi , tsf ]
M x = min(M x1 , M x2 ), M y = min(M y1 , M y2 )
PR
were dZM Px (t) and dZM Py (t) are the deviations of the ZMP trajectory
with respect to the center of the stance foot along x axis and y axis
respectively, tsi and tsf are the time instants of landing and lift-off of
the stance foot respectively. The duration of the step is T = tsf − tsi .

5.2 Optimization w.r.t. dynamic walking stability

The first main contribution of this paper is to improve the pattern
generator proposed in [17] by considering an optimization criterion in
order to find the best values of z1 and m1 to enhance the dynamic
walking stability.
In order to do that, the ZMP corresponding to the generated joint
trajectories should be computed with the overall dynamics of the robot
and compared with the desired ZMP, set to the center of the stance foot.
14 D. Galdeano et al.

This chosen ZMP desired position corresponds to a maximum stability

margins.
The following objective function is then proposed to be optimized
w.r.t. the parameters z1 and m1 :

F
  r 
ẑ1 1 1
= arg  min max (xzmp − xdzmp )2 + (yzmp − ydzmp )2
m̂1 z α β
 1 
m1
(24)

OO
where xzmp and yzmp are the positions of the computed ZMP, xdzmp
and ydzmp are those of the desired one. z1 and m1 are the optimization
parameters, α and β are the length and the width of the foot (i.e. along
x axis and y axis respectively, cf. Fig. 10).
By minimizing the maximum normalized deviation of the ZMP tra-
jectory as cost function, the objective function computes the mass dis-
tribution to ensure the largest stability margins.

5.3 Change of direction and stability optimization

One key feature of a pattern generator is to allow the biped robot to
change the walking direction using the kinematic model like in [21], or
the dynamic model as in [22] with a two point mass system model.
As proposed in [17], the original 3MLIPM pattern generator is de-
signed to make the robot walking only in a straight line, our second
PR
contribution is then to modify it to allow a change of direction while
walking. The change of direction is obtained through the application of
a rotation at the hip of the stance leg, as follows:

R πt
Ω=− cos (25)
2 T

with Ω is the angle of rotation and R is its amplitude.

The change of direction alters the stability of the biped robot. There-
fore, the proposed optimization criterion (24) is used again to improve
dynamic walking stability. The solution of this optimization problem
gives the best values for z1 and m1 which allow a better dynamic stabil-
ity of the robot walking and turning.
Optimal Pattern Generator For Dynamic Walking 15

6. Simulation results
A simulator for SHERPA biped robot was developed using the Graph-
ical User Interface of MatlabTM software. Its graphical interface is shown
in Fig. 11.

F
OO
Fig. 11. View of the Graphical User Interface of SHERPA biped robot simu-
lator.

This interface enables the tuning of some parameters of the robot as

well as those of the optimization criterion. A graphical animation of the
robot or the three mass model can also be displayed in the simulator to
PR
show the obtained movements.
The parameters z1 and m1 are computed by optimization, the other
parameters are constant and summarized in Table 2.
Using the developed simulator, four simulation scenarios are proposed
to validate the proposed optimal pattern generator, namely:

Simulation 1 : optimal trajectories generation for straight walking,

Simulation 2 : trajectories generation for walking with change of
direction without optimisation,
Simulation 3 : trajectories generation for walking with a change of
direction and optimization.
Simulation 4 : Effects of walking parameters
16 D. Galdeano et al.

Table 2. Parameters of the 3MLIPM pattern generator.

Description Parameter Value

Weight of one leg m1 Optimized

F
Weight of the torso m2 mass − 2m1
Height of the mass m1 z1 Optimized
Height of the hip z2 0.95(Lshin + Lthigh )
Step time T 0.7 s
Step length Ds 0.4 m
Calculus variable ω π/(5T )

OO Calculus variable λ=γ z2 /z1

These simulations will be detailed and commented in the following.

6.1 Simulation 1: straight walking

The optimization criterion given in (24) is used to find the best values
for z1 and m1 using the fminsearch algorithm proposed within Matlab
software. The optimization algorithm uses a simplex search method de-
scribed in [23]. The obtained optimization results are coherent with
their physical meanings: the masses are found to be positive and the po-
sitions of the three masses are inside the convex envelope of the robot.
The obtained optimal solution is summarized in Table 3, and the sim-
ulation results are illustrated, for six walking steps, through curves of
Figs. 12-18.
PR
Table 3. Resulting optimized parameters.

Parameter Without optimization With optimization

z1 0.6 m 0.2598 m
m1 6 kg 0.4442 kg

Figures 12 and 13 represent respectively the evolution of the joints’

positions and velocities, where it can clearly be seen that the obtained
trajectories are periodic. Furthermore the trajectories of one leg are
symmetrical w.r.t. those of the other one.
Optimal Pattern Generator For Dynamic Walking 17

F
OO
Fig. 12. Joints’ positions generated by the proposed optimal pattern generator.
PR

Fig. 13. Joints’ velocities generated by the proposed optimal pattern generator.

Figures 14, 15 and 16 represent the evolution of the ZMP and the
CoM positions with respect to the footprints of the biped robot on the
ground generated respectively by the linear inverted pendulum model
18 D. Galdeano et al.

(LIPM), the original (3MLIPM) pattern generator proposed in [17] and

the optimal one proposed in this paper.

F
Fig. 14. Evolution of ZMP and CoM trajectories with the LIPM.

OO
Fig. 15. Evolution of ZMP and CoM trajectories with the original 3MLIPM.

Fig. 16. Evolution of ZMP and CoM trajectories with the proposed optimal
3MLIPM.

The ZMP position calculation is issued from an angular momentum

PR
evaluation based on Kajita’s formulation [20]. This computation is not
base on the simplified three-mass model; indeed, it uses the dynamic
model of the biped robot to produce a realistic ZMP evaluation.
For the LIPM as well as the original 3MLIPM, the ZMP moves inside
the footprint with a big variation, however when the optimal parameters
are used, the ZMP is more concentrated in the center of the footprint
and the stability margins are better. Therefore, the walking dynamic
stability is clearly improved with the proposed optimal 3MLIPM pattern
generator due to the increase of the stability margins.
Fig. 17 displays the evolution of the position of the CoM along the
x and y axes during walking. The obtained trajectories are cyclic along
the y axis with a reduced amplitude thanks to the optimal parameters.
Fig. 18 shows the trajectories of movement of the swing foot for one
step.
Optimal Pattern Generator For Dynamic Walking 19

F
OO Fig. 17. CoM evolution versus time with optimization.
PR
Fig. 18. Evolution of the swing foot trajectories.

6.2 Simulation 2: walking with a change of

direction
The objective of this simulation is to evaluate the stability margins
of the pattern generator proposed in [17] in case of a change of direction
during walking. The obtained result is depicted in Fig. 19 where it is
worth to note that the change of direction during walking has not been
yet considered in the original 3MLIPM, therefore a simple simulation
shows a loss of stability during walking while turning.
20 D. Galdeano et al.

F
OO
Fig. 19. Evolution of ZMP and CoM trajectories with change of direction.

Indeed, when the robot changes the walking direction, the position
of the ZMP is moving within footprints, being sometimes outside of the
stance footprint, therefore the robot becomes unstable. Consequently, a
computation of optimal parameters in this scenario will be necessary to
improve the dynamic walking stability. This is the objective of the next
simulation scenario.

6.3 Simulation 3: change of direction with

optimization
PR
In the proposed optimal pattern generator, the optimization process
computes the best values of the parameters z1 and m1 and uses them to
generate the joints’ trajectories.
Figures 20 and 21 represent respectively the evolution of the new
joints’ positions and velocities for this scenario. Theses trajectories allow
a rotation of the biped robot during walking.
Figure 22 represents the evolution of the ZMP and the CoM positions
as well as the footprints of the biped robot on the ground level when
turning with an angle of 20 degrees at each walking step.
With the proposed optimization, as illustrated in Fig. 22, the robot
remains stable during walking while turning since it keeps the ZMP
always inside the footprint of the supporting leg.
According to the obtained results, from simulation 1 and 3, it is clearly
shown that the proposed optimal pattern generator has significantly
Optimal Pattern Generator For Dynamic Walking 21

improved the dynamic stability of the biped robot for both scenarios:
straight walking and change of direction.

F
OO
Fig. 20. Joints’ positions with change of direction generated by the proposed
optimal pattern generator.
PR

Fig. 21. Joints’ velocities with change of direction generated by the proposed
optimal pattern generator.
22 D. Galdeano et al.

F
OO
Fig. 22. Evolution of ZMP and CoM trajectories in case of change of direction
generated by the proposed optimal pattern generator.

6.4 Simulation 4: effects of walking parameters

The optimal values summarized in Table 3 are used to compute stabil-
ity margins M x and M y for straight walking with different step lengths
and the objective is to analyze the sensitivity of the model w.r.t. step
length variation.
PR
Fig. 23. Evolution of ZMP and CoM trajectories with Ds = 0.4

Fig. 24. Evolution of ZMP and CoM trajectories with Ds = 0.2

The optimal values of mass distribution have been calculated with

a step length of 0.4 m. These values are then used with different step
Optimal Pattern Generator For Dynamic Walking 23

lengths Ds. The corresponding stability margins are summarized in Ta-

ble 4. It is worth to note that the modification of the step length don’t
induce a loose of stability during walking. The optimal mass distribu-
tion make the proposed pattern generator less sensitive to step length
variations.

F
Table 4. Step length influence on stability margins

Ds 0.1 0.2 0.3 0.4 0.5

OO Stability
Mx
My
stable
0.105
0.028

7. Conclusions and future work

stable
0.108
0.029
stable
0.110
0.030
stable
0.112
0.030
stable
0.115
0.029

The objective of this work was to generate stable dynamic walking

for SHERPA biped robot. To deal with this problem, the original pat-
tern generator proposed in [17] based on a Three-Mass Linear Inverted
Pendulum Model has been extended with a ZMP-based optimization to
improve dynamic walking stability. The optimization use a ZMP based
performance index in an optimization problem whose solution gives the
best values of the model’s parameters w.r.t. dynamic walking stability
computed using a whole body model. The case of change of direction
PR
during walking has also been studied and stability in this case has been
enhanced thanks to the optimization.
Several numerical simulations have been presented to show the effec-
tiveness of the proposed optimal pattern generator for the case of the
biped walking robot SHERPA.
Future works can include real-time implementation of the proposed
method on the prototype of SHERPA biped robot and the creation of a
motion database to be able to change the walking speed of the robot in
real-time.

Acknowledgments
This research was supported by the French National Research Agency,
within the project R2A2 (ANR-09-SEGI-011).
24 D. Galdeano et al.

References
[1] S. Ma, T. Tomiyama and H. Wada. Omnidirectional static walking of a
quadruped robot. IEEE Trans. on Robotics, 21(2):152–161, 2005.
[2] M. Vukobratovic and B. Borovac. Zero-moment point-thirty five years of

F
its life. Int. J. of Humanoid Robotics, 1(1):157–173, 2004.
[3] P. Sardain and G. Bessonnet. Forces acting on a biped robot. Center of
pressure-zero moment point. IEEE Trans. on Systems, Man and Cyber-
netics, Part A: Systems and Humans, 34(5):630–637, 2004.
[4] A. Goswami. Postural stability of biped robots and the foot-rotation in-
dicator (FRI) point. Int. J. of Robotics Research, 18(6):523–533, 1999.
[5] K. Harada, K. Miura, M. Morisawa, K. Kaneko, S. Nakaoka, F. Kane-

OO
hiro, T. Tsuji and S. Kajita. Toward human-like walking pattern gener-
ator. IEEE/RSJ Int. Conf. on Intelligent robots and systems (IROS’09),
(Albuquerque, NM, USA), :1071–1077, 2009.
[6] Q. Huang, S. Kajita, N. Koyachi, K. Kaneko, K. Yokoi, H. Arai, K. Ko-
moriya and K. Tanie. A high stability, smooth walking pattern for a biped
robot. IEEE Int. Conf. on Robotics and Automation (ICRA’99), :65–71,
Detroit, Michigan, USA, 1999.
[7] J. Yamaguchi, E. Soga, S. Inoue, and A. Takanishi. Development of
a bipedal humanoid robot-control method of whole body cooperative
dynamic biped walking. IEEE Int. Conf. on Robotics and Automation
(ICRA’99), 1:368–374, Detroit, Michigan, USA, 1999.
[8] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Yokoi and
H. Hirukawa. Biped walking pattern generation by a simple three-
dimensional inverted pendulum model. Advanced Robotics, 17(2):131–
147, 2003.
[9] Z. Tang and M. Er. Humanoid 3D Gait Generation Based on In-
verted Pendulum Model. IEEE 22nd Int. Symp. on Intelligent Control
(ISIC’07), :339–344, Singapore, 2007.
PR
[10] S. Kajita, F. Kanehiro, K. Kaneko, K. Yokoi and H. Hirukawa. The 3D
Linear Inverted Pendulum Mode: A simple modeling for a biped walk-
ing pattern generation. IEEE/RSJ Int. Conf. on Intelligent Robots and
Systems (IROS’01), :239–246, Maui, Hawaii, USA, 2001.
[11] S. Feng and Z. Sun. A simple trajectory generation method for biped
walking. 10th Int. Conf. on Control, Automation, Robotics and Vision
(ICARCV’08), :2078–2082, 2 Hanoi, Vietnam, 008.
[12] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi
and H. Hirukawa. Biped walking pattern generation by using preview
control of zero-moment point. IEEE Int. Conf. on Robotics and Automa-
tion (ICRA’03), 2:1620–1626, 2003.
[13] A. Albert and W. Gerth. Analytic Path Planning Algorithms for Bipedal
Robots without a Trunk. J. of Intelligent and Robotic Systems, 36(2):109–
127, 2003.
[14] J. Park and K. Kim. Biped robot walking using gravity-compensated
inverted pendulum mode and computed torque control. IEEE Int. Conf.
Optimal Pattern Generator For Dynamic Walking 25

on Robotics and Automation (ICRA’98), :3528–3533, Leuven, Belgium,

1998.
[15] J. Kanniah, Z. Lwin, D. Kumar and N. Fatt. A ZMP management
scheme for trajectory control of biped robots using a three mass model.
Proceedings of the 2nd Int. Conf. on Autonomous Robots and Agents

F
(ICARA’04), :458–463, Palmerston North, New Zealand, 2004.
[16] T. Buschmann, S. Lohmeier, M. Bachmayer, H. Ulbrich and F. Pfeiffer. A
collocation method for real-time walking pattern generation. IEEE/RAS
Int. Conf. on Humanoid Robots, :1–6, 2007.
[17] S. Feng and Z. Sun. Biped robot walking using three-mass linear inverted
pendulum model. Proceedings of the Int. Conf. on Intelligent Robotics
and Applications (ICIRA’08), :371–380, Springer-Verlag, Wuhan, China,

OO 2008.
[18] http://www.lirmm.fr [Online].
[19] I. M. C. Olaru, S. Krut and F. Pierrot. Novel mechanical design of biped
robot sherpa using 2 dof cable differential modular joints. Proceedings
of the 2009 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems
(IROS’09), :4463–4468, St. Louis, MO, USA, 2009.
[20] S. Kajita, H. Hirukawa, K. Harada and K. Yokoi. Introduction à la
commande des robots humanoides. Translated in French by Sakka, S.
Springer, 2009.
[21] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Yokoi and
H. Hirukawa. A realtime pattern generator for biped walking. IEEE Int.
Conf. on Robotics and Automation, ICRA’02, 1:31–37, IEEE, 2002.
[22] J. T. Kim and J. H. Park. Quick change of walking direction of biped
robot with foot slip in single-support phase. 11th IEEE-RAS Int. Conf.
on Humanoid Robots (Humanoids’11), :339–344, 2011.
[23] J. C. Lagarias, J. A. Reeds, M. H. Wright and P. E. Wright. Convergence
properties of the nelder–mead simplex method in low dimensions. SIAM
PR
J. on Optimization, 9(1):112–147, 1998.

Biographies

David Galdeano received the M.S. degree in Robotics

and Automation from the University Montpellier 2,
France, in 2010. Currently he is a Ph.D. student in the
Laboratory of Informatics, Robotics, and Microelectron-
ics (LIRMM). His current research interests include hu-
manoid robotics, whole-body posture control.
 26 D. Galdeano et al.

Ahmed Chemori received his MSc and PhD degrees

respectively in 2001 and 2005, both in automatic control
from the Grenoble Institute of Technology. He has been
a Post-doctoral fellow with the Automatic control labo-
ratory of Grenoble in 2006. He is currently a tenured re-

F
search scientist in Automation and Robotics at the Mont-
pellier Laboratory of Informatics, Robotics, and Micro-
electronics. His research interests include nonlinear, adap-
tive and predictive control and their applications in hu-
manoid robotics, underactuated systems, parallel robots, and underwater ve-
hicles.

OO Sébastien Krut received the M.S. degree in mechani-

cal engineering from the Pierre and Marie Curie Univer-
sity, Paris, France, in 2000 and the Ph.D. degree in auto-
matic control from the Montpellier University of Sciences,
Montpellier, France, in 2003. He has been a Post-doctoral
fellow with the Joint Japanese-French Robotics Labora-
tory (JRL) in Tsukuba, Japan in 2004. He is currently
a tenured research scientist in Robotics for the French
National Centre for Scientific Research (CNRS), at the
Montpellier Laboratory of Computer Science, Microelectronics and Robotics
LIRMM), Montpellier, France. His research interests include design and con-
trol of robotic systems.

Philippe Fraisse received M.Sc degree in Electrical En-

gineering from Ecole Normale Superieure de Cachan in
PR
1988. He received Ph.D. degree in Automatic Control
in 1994. He is currently Professor at the University of
Montpellier, France. He is the head of robotics depart-
ment (LIRMM) and co-chair of French National Work-
group (GDR Robotique) working on Humanoid Robotics
(GT7). He is also member of JRL-France scientific board
(Japanese- French joint Laboratory for Robotics, AIST-JRL) and member of
IEEE. His research interests include modeling and control applied to robotic
and rehabilitation fields, including humanoid robotics, robotics for rehabilita-
tion.

View publication stats