International Journal of Advanced Robotic Systems

ARTICLE

Identifying Ground-Robot Impedance to

Improve Terrain Adaptability in Running
Robots
Invited Feature Article

Juan C. Arevalo1*, Daniel Sanz-Merodio1, Manuel Cestari1 and Elena Garcia1

1 Centre for Automation and Robotics, CSIC, La Poveda, Madrid, Spain
* Corresponding author(s) E-mail: juan.arevalo@csic.es
Received 21 February 2014; Accepted 14 October 2014
DOI: 10.5772/59888
2015 The Author(s). Licensee InTech. This is an open access article distributed under the terms of the Creative Commons Attribution License
(http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.

Abstract
To date, running robots are still outperformed by animals,
but their dynamic behaviour can be described by the same
model. This coincidence means that biomechanical studies
can reveal much about the adaptability and energy effi
ciency of walking mechanisms. In particular, animals
adjust their leg stiffness to negotiate terrains with different
stiffnesses to keep the total leg-ground stiffness constant.
In this work, we aim to provide one method to identify
ground-robot impedance so that control can be applied to
emulate the aforementioned animal behaviour. Experi
mental results of the method are presented, showing welldifferentiated estimations on four different types of terrain.
Additionally, an analysis of the convergence time is
presented and compared with the contact time of humans
while running, indicating that the method is suitable for
use at high speeds.

42]. Nevertheless, these robots are still outperformed by

animals. Thus, there remains a lot to learn from their
biological counterparts, particularly in the case of adapta
bility and endurance.
It is generally accepted that a spring-loaded inverted
pendulum (SLIP), i.e., a point mass on a simple Hookean
spring, models the basic dynamics of running remarkably
well. This model explains the mechanics of running of
several animals, including humans [1, 8, 9, 11, 14, 17, 35].
Moreover, it has been applied to describe the dynamic
behaviour of running robots [2, 19, 41]. This coincidence
implies that the conclusions obtained from studies of
animal-running dynamics should also be applicable to
robots.

1. Introduction

The stiffness of the leg spring (the spring used to model the
energetics of running) is an important parameter in the
dynamics of running because it influences contact time and
stride frequency [15, 36]. Biomechanical studies of running
reveal that animals, including humans, can adjust the
stiffness of their leg spring. Experimental data show that
this adjustment depends on velocity [3], stride frequency
[15] and surface stiffness [16, 28, 37].

Although there are only a few agile legged machines,

research into faster robots is increasing [2, 20, 22, 26, 38, 39,

Changes of stiffness to increase velocity are small, and the

same effect can be achieved by increasing the leg stride

Keywords running robot, adaptability, ground-robot

impedance

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

(angle swept by the leg) while maintaining the same stride

frequency [14]. Moreover, stiffness changes depending on
stride frequency are made to preserve a certain velocity
[15]. In contrast, stiffness changes made to accommodate
different surface stiffnesses have more important conse
quences. First, if the same leg stiffness is used on every
surface, the body can be put out of balance whenever an
abrupt change in the surface stiffness occurs; however,
adjusting the stiffness offers a smooth transition between
surfaces [17]. Second, adjusting the leg stiffness to accom
modate different terrains can lead to more efficient running
because increasing the leg stiffness as the terrain stiffness
decreases leads to less mechanical work being done by the
leg and more mechanical work being done by the terrain
[18]. In other words, by adjusting the leg stiffness to
accommodate terrain adaptability, energy efficiency can be
enhanced.
To extrapolate this conclusion to legged robots, one
important consideration is how to know when the stiffness
adjustment must be performed. Animals, or at least birds
[12] and humans [17], change their leg stiffness before
stepping onto a new surface when they are knowingly
changing surfaces. Nonetheless, the question remains as to
how they adjust leg stiffness one the first occasion, when
the surface is unknown. This question leads to the idea that
some sort of information about ground stiffness must be
acquired directly from the surface (on at least) the first time
they step on it while running.
The aim of this study is to provide a method to obtain
ground stiffness information with a legged robot rather
than to propose methods to adjust leg stiffness, although
much research effort is being put into solving the latter
problem [5, 7, 24, 25, 27]. Nevertheless, this application is
kept in mind throughout the paper. The reason for this
approach is that identifying the leg-ground stiffness while
running must be treated as an independent problem due to
the various application restrictions (these restrictions will
be clarified in the next section). Additionally, the ultimate
goal of our identification method is not to estimate the
ground stiffness as accurately as possible but rather to
estimate the robot-ground interaction properties, which are
the control target.
Another important reason for this study is that, in realworld conditions, it is not possible to predict every contact
surface that will be encountered. Nevertheless, it might be
possible to store the information of ground stiffness
beforehand in order to adjust the robots leg stiffness using
some type of perception, although perception (e.g., vision)
has its limits in detecting different surfaces; for example,
two different terrains might be hidden under a given cover.
For this reason, in order to obtain information about the
terrain, we propose to use an event-driven system identi
fication method. This method is based on the well-known
recursive least squares (RLS) method [23, 33] and aims to
estimate the parameters of the ground as modelled as a
spring dashpot. The RLS method was chosen because it is
2

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

computationally efficient, allowing online computation,

and because it has shown good convergence and accurate
results in similar applications.
This paper is organized as follows. Section 2 presents the
contact model we used as well as a brief summary of the
detail, and the necessary considerations for its application
are established. Section 4 describes the experimental setup
and the implementation of the algorithm on a leg proto
type. In Section 5, the experimental results are presented,
and finally, the conclusions are provided in Section 6.
2. Background
Most of the biomechanics research on the adaptability of
humans and other animals to different terrains has focused
on leg stiffness [3, 12, 1417, 28, 3537]. However, there are
studies that show that the centre of mass (COM) of humans
running on dissipative materials (such as sand) follows a
bouncing motion with the same stride frequency as on hard
surfaces [31, 43]. For this reason, to incorporate an estimate
of the energy dissipation, the contact model chosen in this
work was the spring dashpot model [7, 21]:
Fc = kd + bd&

(1)

where Fc is the contact force, is the displacement of the

ground, and k and b are the stiffness and damping con
stants of the surface, respectively. This particular model
was also chosen because it is linear and thus relatively easy
and quick to identify. Other more complex models that
describe the contact dynamics more accurately are nonlin
ear and require special techniques of system identification,
more execution time, and frequency-richer data than the
data that can be obtained using a legged robot (while
walking or running). The relatively poor frequency
information of the data obtained using a legged robot is
because the types of signals that can be given to excite the
ground mechanical system using a legged robot are
limited; other specialized signals that allow for the extrac
tion of frequency-rich information can cause the robot to
lose balance, which is undesirable.
As stated above, the identification procedure was devel
oped based on the well-known RLS method for system
identification. This choice is justified by the fact that it offers
a low modelling error (defined in the next section) when
used in similar applications, e.g., identifying the contact
properties of a determined surface [4]. Here, a brief
summary of the RLS method is provided for the sake of
completeness. However, the interested reader is referred to
the specialized literature [10, 23, 33].
2.1 RLS method
The RLS method consists of applying the least squares
criteria to minimize the modelling error of a process

expressed as a linear regression, and it can be summarized

as follows:

qk =qk -1 + Lk ( yk - j KT qk -1 )

(2)

d=

1
1
1
1
Fc - d& - verr d = j2T q 2 - verr
k
b
k
k

1 b
1
1
d& = - d - verr d& = j3T q3 - verr
b k
b
b

(10)

(11)

Pk -1j k
l + j KT Pk -1j k

(3)

P jkjT P
1
Pk -1 - k -1 T k k -1
l
l + j K Pk -1j k

(4)

q1 - q1 = M (j1 ) w

(12)

^ is the vector of the parameters describing the

were
k

w
q2 - q 2 = M (j2 )
k

(13)

w
q3 - q3 = M (j3 )
b

(14)

Lk =

Pk =

system, Lk is a matrix related to the inverse of the measured

covariance matrix of the samples, Pk is the inverse of the
measured covariance matrix of the samples, is a vector
grouping the inputs and outputs of the system, yk is the
present output of the system, is a design parameter
termed the forgetting factor , and the sub-index k indicates
the present iteration of the algorithm.

It is important to note that this algorithm can cause matrix

P to become negative because of round-off errors; to avoid
this situation, we have used Biermans U-D factorization
algorithm [32] to compute matrix P and update matrix L.
Note that equation (1) is a linear differential equation, not
an autoregressive process. Nonetheless, it can be trans
formed into a linear regression by simply rewriting the
equation as follows:
T

d = j q

(5)

where:

j T = -d&, F
q T = [b / k , 1 / k ]

(7)

(8)

It is important to note that equation (1) can be written in

different forms. To clarify the choice of equation (7), let
verr denote the unknown modelling errors and the noise in
the measurements of the regression (i.e., force, velocity and
position). As such, we can write the estimations as:
Fc = kd + bd& + verr Fc = j1T q1 + verr

where M(i) is a matrix that depends on the regressor and

w is a vector containing the modelling errors and the noise
in the measurements. These last equations can be used to
find and upper bound in terms of the largest singular value
of the matrix M(i). Let us denote this singular value as max
(M(i)). Thus, we can write the upper bound as:

q1 - q1 = s max ( M (j1) ) w

(15)

w
q2 - q 2 = s max ( M (j 2))
k

(16)

w
q3 - q3 = s max ( M (j 3))
b

(17)

(6)

Thus, the modelling error is defined as:

E = d - j Tq

Now, if we apply a least squares estimation to the above

equations, this yields [33]:

(9)

The previous analysis suggests that if the maximum

singular values are equal, then equation (??) provides the
best estimate. Moreover, the results of a comparison
between the results applied to the data will be provided
further below.
Now, using the linear regression form of equation (1), it is
possible to apply the RLS method directly to the identifi
cation problem. However, because our intention is to use
the algorithm in a legged robot (namely, to identify the
contact properties of the ground using legs), a few concerns
arise: When should the identification start? How are the
samples taken? Is it necessary or even practical to use the
system identification during the whole gait cycle? The next
section is devoted to answering these questions and it
describes in detail the system identification procedure
proposed in this work.

Juan C. Arevalo, Daniel Sanz-Merodio, Manuel Cestari and Elena Garcia:

Identifying Ground-Robot Impedance to Improve Terrain Adaptability in Running Robots

3. Identification procedure
Before describing the identification procedure, it is useful
to review the literature concerning haptic interfaces. In that
field, several studies have focused on recognizing the
dynamic properties of manipulated objects [13, 29, 30, 34].
These studies have mainly used two approaches: active
sensing methods [34], which offer precise and accurate
estimations of the properties in the presence of noise, but
which require specialized control techniques; and passive
sensing methods [13, 29, 30], in which the identification
procedure is independent of control but which are less
efficient [29].
3.1 Basic considerations

Leg in
swing phase

Identify
Ground
Impedance

Idle

Leg in
stance phase

1. State diagram
diagram of the
identification
method
FigureFigure
1. State
of system
the system
identification
method.

aproduce
given asurface
might
enhance
the accuracy
quantization
error
in the identification
algoof the
Although the application addressed in this paper with
is
estimation.
Nevertheless,
we
chose
to
reset
the
algorithm
rithm.
These
errors
might
lead
the
system
identification
different and some necessary considerations may differ,
algorithm
to
fail
because
it
might
consider
the
system
to be
here in order to prepare the robot for a new surface,
as
some of the conclusions of the previous work on haptic
in
equilibrium.
Thus,
the
sampling
frequency
should
be point
interfaces still apply. Active sensing, although more
using the estimates of another surface as the starting
in order to reflect the changes in the
precise, is not suitable for use in legged robots because
mightchosen
affectaccordingly
convergence.
system
response.
A
more detailed discussion of the
using a specialized control method for the identification
sampling
frequency
will
be addressed
in Sectionthe
5. sampling
At this point, it is important
to discuss
would disrupt the robot running behaviour, which might
cause destabilization. Moreover, active sensing methods
frequency. When controlling a system, the sampling
3.2 Robot
structure be
considerations
continuously preform the system identification. On the
frequency
should
as high as possible so as to
other hand, the system identification literature establishes
maintain
good
tracking
error
the
reference
trajectories,
Another
important
point
is thefor
effect
of the
robots structure
that identifying a system in equilibrium is useless because
on the of
identification.
Let method
us considerused.
a robot Nevertheless,
during a
regardless
the
control
information about its dynamic behavior cannot be ob
bouncing gait in contact with the ground. The behaviour of
tained. This implies that the identification should start in
as this case the control loop frequency (2,000 kHz) may
the
robot and
shouldproduce
resemble a
a SLIP.
Nevertheless,
manyin the
be too high
quantization
error
soon as the foot touches the ground and cease when the foot
robots have rigid links, and the spring compression at the
leaves the ground or the contact has been stabilized
identification algorithm. These errors might lead the
lowest part of the stance phase could be achieved by flexing
Therefore, passive sensing is the logical choice.
system
identification
algorithm
fail
because
it might
a rigid
leg. In this case,
there is no to
need
to consider
defor
the
system
beandintheequilibrium.
Thus, the
Considering this last remark, one possible state diagram consider
of
mation
on the
robots to
links,
state diagram described
sampling
should
besufficient.
chosen accordingly in order
the identification procedure can be envisaged (see Figure 1).
in thefrequency
above section
might be
to reflect the changes in the system response. A more

This state machine consists of two basic states: (a) the "idle"
In contrast, the robot might not be constructed entirely of
detailed
will be
state, in which all of the matrices and vectors of the
rigiddiscussion
links, as is theof
casethe
withsampling
many robots frequency
designed to run
addressed
in
Section
5.
identification method are reset to their original values and
[20, 22, 26, 3840, 42]. In this case, when computing the
wait until the leg contacts the ground (in this case, it goes
ground penetration (), the compression of the robots links
to the next state); and (b) the "identify ground impedance"
must be considered to obtain an accurate measure.
3.2. Robot
structure considerations
state, which performs the system identification according
With this last consideration, a modified version of the
to the RLS method described in the previous section. Another important point is the effect of the robots
block diagram in Figure 1 with a wider range of applica
structure
identification.
LetIn us
a robot
Note that the "idle" state is said to reset the system identi
bilityon
can the
be designed
(see Figure 2).
this consider
diagram, two
fication algorithm. This feature is not strictly necessaryduring
if
a
bouncing
gait
in
contact
with
the
ground.
new states have been included, an "initialization" state
the robot steps on the same surface more than once. In fact,
and a "compute
position"should
state. The
initialization
The behaviour
of COM
the robot
resemble
a SLIP.
using the estimates of a prior identification with a given
state computes
initial
leg rigid
position,
its and
maximum
Nevertheless,
manythe
robots
have
links,
the spring
surface might enhance the accuracy of the estimation.
length and
position
when
algorithm
starts.
compression
at the
theCOM
lowest
part of
thethe
stance
phase
could be
Nevertheless, we chose to reset the algorithm here in order
Next,
the
"compute
COM
position"
state
is
included
to need
achieved
by
flexing
a
rigid
leg.
In
this
case,
there
is
no
to prepare the robot for a new surface, as using the esti
compute the COMs vertical displacement to pass the
to consider deformation on the robots links, and the state
mates of another surface as the starting point might affect
information to the next identification state at the begin
diagram
described in the above section might be sufficient.
convergence.
ning of the stance phase. Following this, the "identify
ground the
impedance"
state cannot
compute
the COMs vertical
At this point, it is important to discuss the sampling
In contrast,
robot might
be constructed
entirely of
displacement from the moment at which the foot makes
frequency. When controlling a system, the sampling
rigid links, as is the case with many robots designed to
contact with the ground. Thus, the ground penetration
frequency should be as high as possible so as to maintain
run [20, 22, 26, 3840, 42]. In this case, when computing
will be equal to the difference between the COMs
good tracking error for the reference trajectories, regardless
the ground
penetration (), the compression of the robots
of the control method used. Nevertheless, in this case the
displacement and the legs compression.
links
must
be
considered to obtain an accurate measure.
control loop frequency (2,000 kHz) may be too high and
4

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

With this last consideration, a modified version of

the block diagram in Figure 1 with a wider range
of applicability can be designed (see Figure 2).
In
this diagram, two new states have been included, an

F
c

t
p
C

N
t
t
r
T
r
b
t
i
r

I
i
i
t

T
t
m
l
(
j
s
a
a
a
5
a
0

T
m
o
i
t

sampling
sampling
so as to
ajectories,
vertheless,
kHz) may
or in the
lead the
e it might
Thus, the
y in order
A more
y will be

he robots
er a robot
e ground.
e a SLIP.
the spring
e could be
is no need
d the state
sufficient.

entirely of
esigned to
omputing
he robots
measure.

ersion of
der range
e 2).
In
uded, an
ion" state.
position,
when the
tion" state
placement
tion state
wing this,
mpute the
at which

Leg in
swing phase

Identify
Ground
Impedance

Compute
COM
Position

Leg in
stance phase

Figure
2. diagram
State diagram
of the
system identification
method
Figure
2. State
of the system
identification
method considering
the
considering the robot structure.
robot structure

the foot makes contact with the ground. Thus, the ground
Note
that to will
increase
the adaptability
to thebetween
terrain, the
the
penetration
be equal
to the difference
ultimate
goal of the identification
is not to estimate the
COMs displacement
and the legs compression.
ground properties as accurately as possible but rather to
Note that to increase the adaptability to the terrain,
estimate the robot-ground interaction properties. Thus, the
the ultimate goal of the identification is not to estimate
leg stiffness can be adjusted to hold the total robot-ground
the ground properties as accurately as possible but
impedance
constant,
which is desirable
because
biome
rather to estimate
the robot-ground
interaction
properties.
chanical
in humans
18] show
that the
humans
Thus, thestudies
leg stiffness
can be[17,
adjusted
to hold
total
adjust
their leg impedance
stiffness to maintain
total impedance
as
robot-ground
constant,thewhich
is desirable
constant,
leading to more
energy-efficient
because biomechanical
studies
in humansrunning.
[17, 18] show
that humans adjust their leg stiffness to maintain the total
In
the next section,
the system
procedure is
impedance
as constant,
leadingidentification
to more energy-efficient
implemented
on
a
robotic
leg
prototype
[5,
20]
and used to
running.
identify the impedance of the leg in contact with different
In the next section, the system identification procedure is
types of ground.
implemented on a robotic leg prototype [5, 20] and used to
identify the impedance of the leg in contact with different
4.
Experimental
types
of ground.setup

The HADE leg prototype (see Figure 3) is a planar, three4. Experimental setup
degree-of-freedom
(DOF), under-actuated mechanism
featuring
the characteristics
a horse
leg [20].
It isa actuated
The HADE
leg prototypeof(see
Figure
3) is
planar,
three-degree-of-freedom
by
two series elastic actuators(DOF),
(SEAs), oneunder-actuated
at the hip and
mechanism
featuring
characteristics
of driven
a horse
one
at the knee,
while thethe
ankle
joint is passively
by
legextension
[20]. It spring
is actuated
two series
elasticKactuators
=
3,
113
an
with by
a stiffness
constant
s
(SEAs), one at the hip and one at the knee, while the ankle
N/m. The hip and knee actuators have a spring with an
joint is passively driven by an extension spring with a
encoder attached that acts as a force sensor with a resolu
stiffness constant Ks = 3, 113 N/m. The hip and knee
tion
of 1.2have
N per
count and
measuring
the
actuators
a spring
withan
anencoder
encoder for
attached
that acts
position
with
a
resolution
of
50.8e-6
m
per
count.
The
ankle
as a force sensor with a resolution of 1.2 N per count and
has
a linear encoder
with a resolution
ofwith
224e-6
m per count,
an encoder
for measuring
the position
a resolution
of
which
0.69ankle
N perhas
count.
50.8e-6ismequivalent
per count.asThe
a linear encoder with
a resolution of 224e-6 m per count, which is equivalent as
The
is fixed
0.69 leg
N per
count.to the wall by a sliding rail, so horizontal
movements of its COM are restricted but it is capable of
The leg is
fixed to the
wall by athe
sliding
rail, movement
so horizontalis
moving
vertically.
Moreover,
vertical
movements of its COM are restricted but it is capable
limited by a mechanical top at the bottom in order to
of moving vertically. Moreover, the vertical movement
prevent falling when the leg is not in contact with the
is limited by a mechanical top at the bottom in order
ground.
Incremental
encoders
areis used
to contact
measurewith
the
to prevent
falling when
the leg
not in
position
of each
joint and encoders
to indirectly
the joint
the ground.
Incremental
are measure
used to measure
torque
by measuring
spring
deflection
at each
actuator
the position
of each the
joint
and to
indirectly
measure
the
joint the
torque
by measuring
spring
deflection
at each
and
fetlock
spring. Thetheleg
implements
a position
actuatorscheme
and the
fetlock
spring. kinematics
The leg and
implements
control
based
on inverse
propor
a position control scheme(PID)
based
on inverse
kinematics
tional-integral-derivative
controllers
at each
active
and (hip
proportional-integral-derivative
(PID) is
controllers
at
joint
and knee), while ground contact
detected by
each active joint (hip and knee), while ground contact
monitoring the ankle spring torque. The control was
is detected by monitoring the ankle spring torque. The
implemented using a National Instruments PXI-1024Q
control was implemented using a National Instruments

Figure 3. The HADE

Leg Prototype.
Conceptual
design (left).
Real prototype
Figure
Conceptual
design
(left).
Real
Figure3.3.The
TheHADE
HADE Leg
Leg Prototype.
Prototype. Conceptual
design
(left).
Real
(right).
prototype
(right).
prototype
(right).

1.2

The complete experimental setup is shown in Figure 4. The

Thecomplete
complete experimental
experimental setup
Figure
4.
The
setupisisshown
showninin
Figure
foot
waswas
placed
over
two
different
surfaces
and
then 4.
Thefoot
foot
placed
over
two
different
surfaces
and
then
The
was
placed
over
two
different
surfaces
and
then
commanded to
to move
move upwards
upwards until
until it
it was not
not in
in contact
contact
commanded
commanded
to move
upwards
until itwas
wastonot
inincontact
with
the
surface.
Next,
it
was
commanded
stay
that
with the surface. Next, it was commanded to stay in that
with
the surface.
Next,
it wasitcommanded
to stay
in that
position
for
0.5
s.
Afterwards,
was
commanded
to
return
position for 0.5 s. Afterwards, it was commanded to return
position
for 0.5
s. Afterwards,
itwith
was the
commanded
to stay
return
to
to its
its initial
initial position
position (in
(in contact
contact with
the ground)
ground) and
and stay
tothere
its
initial
position
(in
contact
with
the
ground)
and
stay
there for
for another
another 0.5
0.5 s.
s. This
This procedure
procedure was
was repeated
repeated twice
twice
there
for
another
0.5
s.
This
procedure
was
repeated
twice
for
for each
each experiment.
experiment.

x1

for each experiment.

0.8

0.6

0.4

0.2

Mean squared modelling error

acy of the
algorithm
urface, as
ting point

Initialisation

Mean squared modelling error

method.

Leg in
swing phase

PXI-1024Q (Austin, TX, USA) and had an execution time

4.1. Im
TX, USA)
hadthe
an identification
execution time algorithm
of 0.5 ms per
of(Austin,
0.5 ms per
cycle,and
while
had
HADE
and had an
execution
time
4.1. Implem
cycle,
while(Austin,
the identification
had
an execution
anPXI-1024Q
execution
time
of TX,
5 msUSA)
per algorithm
cycle.
Prior
of
0.5ofms
perper
cycle,
while the identification algorithm had
HADE
leg
time
5 ms
cycle.
an execution time of 5 ms per cycle.
were
Prior to th
Hip
setup
were
perf
Hip
value
setup
pub
Hip
frequ
values
fo
SEA
error
Hip
frequency
Knee
SEA
SEA
river
error
aga
Knee
SEA
river
sand
were
were
coll
metal
metal
boxN
Knee
a 60
Knee
a 60NNaro
fo
N around
squar
SDF
square w
SDF
this t
this time,
more
more; and
Fetlock
Fetlock
lastlast
10 s 10
o
resam
resamplin
25 and
Hz
25 Hz
ex
the the
exper

Figure 5.

Figur

In Fig. 5
frequency
Fig
the In
mode
frequ
frequenci
thedecm
error
frequ
error
intr
error
position.
to the
riv
error
the positi
samp
because
th
to the
variations
the sa
the becau
quan
Nonethel
variat
of magni
the q
and 7. T
for None
why t
of ma
normalize
andbec7
larger
for w
in positio
in the
cas
norm
thislarger
type
identified
in po

Figure 4.
Experimental setup.
Leg stepping over the
variable-impedance box (right) and sand (left).

Figure 4. Experimental setup. Leg stepping over the variable-impedance box

Figure
4. sand (left).
Experimental setup.
Leg stepping over the
(right) and
Moreover,
the experiments were performed on four
variable-impedance box (right) and sand (left).

different types of terrain: (A) rigid ground; (B) an

Moreover, the experiments
were
performed
on four
adjustable-impedance
metal box
[4], designed
to resemble
Moreover,
the
experiments
were
performed
on(see
four
different
types
of
terrain:
(A)
rigid
ground;
(B)
an
adjusta
the behaviour of the spring dashpot contact model
different
types
of
terrain:
(A)
rigid
ground;
(B)
ble-impedance
metal
box
[4],
designed
to
resemble
the
equation (1)); (C) river sand; and (D) rocks. The last two an
adjustable-impedance
metal
box [4],
designed
towith
resemble
types
of terrain
werespring
used
bydashpot
filling
a wooden
box
the
behaviour
of the
contact
model
(see
the
behaviour
of
the
dashpot
contact
(see
4). rocks.
corresponding
material
(see
equation
(1)); (C)
riverspring
sand;Figure
and
(D)
Themodel
last two
equation
(C)
river
sand;
and (D)
rocks.box
The
lastthe
two
types of (1));
terrain
were
used
by filling
a wooden
with
types
of terrain were
used
byFigure
filling4).
a wooden box with the
corresponding
material
(see

corresponding material (see Figure 4).

Juan C. Arevalo, Daniel Sanz-Merodio, Manuel Cestari and Elena Garcia:

Identifying Ground-Robot Impedance to Improve Terrain Adaptability in Running Robots
www.intechopen.com
www.intechopen.com

in the
this ty
identi

Identifying ground-robot impe

4.1 Implementing the system identification algorithm on the

HADE leg

Knee

etlock

x 10

Mean squared modelling error

0.6

0.4

0.2

0.005

0.01

0.015
0.02
0.025
Sampling Period (s)

0.03

0.035

0.005

0.01

0.015
0.02
0.025
Sampling Period (s)

0.03

0.035

0.04

Asthe
thealgorithm
algorithm
is intended
toused
be used
with different
As
is intended
to be
with different
types
Figure 6. Mean square error vs sampling time for river sand.
types
of terrain,
havea chosen
a frequency
sampling frequency
of
terrain,
we havewe
chosen
sampling
of 200 Hz,
ofAs200
Hz,
partly
because
it apresented
a near-minimum
partly
because
it presented
near-minimum
modelling
the
algorithm
is intended
to be used
with
different
modelling
error
usedchosen
inthat
a terrain
thatour
resembles
typeswhen
of terrain,
a sampling
frequency
error
used when
inwea have
terrain
resembles
model.
our
model.
Moreover, itwhen
the aalgorithm
was
of 200
Hz,
partly
near-minimum
Moreover,
when
thebecause
algorithm presented
was implemented
with the
implemented
with
the
PXI-1024Q,
higher
sampling
modelling higher
error when
usedfrequencies
in a terraininterfered
that resembles
PXI-1024Q,
sampling
with
frequencies
interfered
with the
execution
of the control
our model.
Moreover,
when
the algorithm
was
the
execution
of
the
control
loop.
The
problems
explained
loop.
The problems
here could
be solved
by
implemented
with explained
the PXI-1024Q,
higher
sampling
here
could
be solved
bymore
using
a computer power
with more
using
a
computer
with
computational
and
frequencies interfered with the execution of the control
computational
power
with the
use of
higher
with
theThe
use of
higherand
resolution
encoders.
loop.
problems
explained
here
could
beresolution
solved by
encoders.
using
a
computer
with
more
computational
power
and
To choose the forgetting factor, the algorithm was applied
with
the
use
of
higher
resolution
encoders.
to choose
the same
as the factor,
sampling
experiment,
To
thedata
forgetting
the frequency
algorithm was
applied
using
values
of forgetting
as
between
0.89 and
1.
Figs. 7 experiment,
and
show
To
choose
factor,
thefrequency
algorithm
was8applied
to
the
samethe
data
the sampling
the
mean
square
modelling
error
vs.

for
the
metal
box
to the
sameofdata
as the sampling
experiment,
using
values
between
0.89 and frequency
1. Figs. 7 and
8 show
and
thevalues
river sand,
respectively.
using
of

between
0.89
and
1.
Figs.
7
and
8
show
the mean square modelling error vs. for the metal box and
theriver
mean
square
modelling error vs. for the metal box
the
sand,
respectively.
and thex 10river sand, respectively.

0.8

0.04

2.4

Figure
5. Mean
squaresquare
error vs
sampling
time fortime
the metal
boxmetal box.
Figure
5. Mean
error
vs sampling
for the

2.2

ed on four
nd; (B) an
to resemble
model (see
The last two
box with the

x 10

2.2
1.8

1.6

1.8
1.4
1.6
1.2
1.4
1

1.2
0.8

0.6

0.8
0.4
0.88
0.6
0.4
0.88

0.9

0.9

0.92

0.92

0.94

0.94

0.96

0.96

0.98

1
0.88
1.5

0.88
Figure 8.

Consideri
Figure 8.
have cho
error
valu
Consider
minimum
have cho

error the
val
Once
minimum
were
chos
for
differ
Once
the
implemen
were cho
have
been
for diffe
output
for
impleme
have bee
output fo

where e i
Yi is the m
from
thee t
where
better
tha
Yi is the
through
from the(
RLS
meth
better
tha
[4],
where
through
ofRLS
unmod
meth
terrains
d
[4], wher
the
of robot.
unmo
terrains d
the robot
Table 1. M

0.98

Figure 7. Mean square error vs for the metal box

errorthat
vs for
the metal
InFigure
Fig. 7.
7, Mean
it cansquare
be seen
values
of box.
between

Table 1. M

Figure 7. Mean square error vs for

the metal box.

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

:
Identifying ground-robot impedance to improve terrain adaptability in running robots

2.4
2

Mean Square Modeling Error

Mean Square Modeling Error

ing over the

In In
Fig.
5, it
shown
Fig.
5, is
it is
shownthat
that- -generally
generally- -as
as the
the sampling
sampling
frequency
decreases
(i.e.,
thethe
sampling
timetime
increases),
the
frequency
decreases
(i.e.,
sampling
increases),
the modelling
error increase
(except
for the sampling
modelling
error increase
(except for
the sampling
frequen
frequencies
kHzHz,
and
333.3theHz,
where
the
cies
between 2 between
kHz and 2333.3
where
error
decreas
error
decreases).
This
is
explained
by
the
quantization
es). This is explained by the quantization error introduced
error introduced by the position encoders measuring
by the position encoders measuring position. Neverthe
position. Nevertheless, when the algorithm was applied
less,
thesand,
algorithm
was applied
the river
sand, the
to when
the river
the modelling
errortoalways
increased
as
modelling
error frequency
always increased
as the sampling
frequen
the sampling
was reduced.
The difference
is
cy because
was reduced.
The
difference
is because
river
is
the river
sand
is less stiff,
and thusthe
there
aresand
larger
variations
thethere
position
for the
same force;
less
stiff, and in
thus
are larger
variations
in thetherefore,
position
the
quantization
error
has
less
influence
thehas
result.
for the same force; therefore, the quantizationon
error
less
Nonetheless,
in
Fig.
6
and
8,
the
errors
are
one
order
influence on the result. Nonetheless, in Fig. 6 and 8,
the
of magnitude larger than the errors shown in Figs. 5
errors are one order of magnitude larger than the errors
and 7. This has a twofold explanation: the first reason
shown
in Figs. 5 and 7. This has a twofold explanation: the
for why this occurs is that the errors shown here are not
first
reason for why
this occurs
is that
the the
errors
shown
normalized,
and thus
the errors
with
river
sandhere
are
arelarger
not normalized,
the errors
with
the river
sand
because it isand
lessthus
stiff and
there are
larger
variations
position;
the second
reason
that,
arein larger
because
it isand
lessmost
stiffimportant
and there
are islarger
in the case
the river the
sand,
the model
less accurate
as
variations
in ofposition;
second
and ismost
important
this
type
of
terrain
presents
nonlinear
behaviour,
and
the
reason is that, in the case of the river sand, the model is less
identified model is a linear approximation of the system.

2
2.5

1.5

Sampling Period (s)

x 10

3
3.5

2.5

Figure
6. 6.
Mean
square
error
vs 0.015
sampling
for
river
sand
0 Mean
0.005
0.01 error
0.02time0.025
0.03
0.035sand.
0.04
Figure
square
vs sampling
time
for
river

in Figure 4.
ces and then
ot in contact
stay in that
ded to return
nd) and stay
peated twice

3.5

x1

4 x
4.5

x 10

ign (left). Real

1.2

4.5

Mean Square Modeling Error

Mean Square Modeling Error

Hip
SEA

Prior to the implementation on the HADE leg, experiments

4.1. Implementing the system identification algorithm on the
were performed on data obtained using the experimental
HADE leg
setup published in [4] with the RLS algorithm to choose
Priorfor
to the
leg, experiments
values
theimplementation
forgetting factoron
the
andHADE
the sampling
frequen
on data
using the
experimental
cy.were
Fig. performed
5 and 6 show
theobtained
mean square
modelling
error
setup published in [4] with the RLS algorithm to choose
against the sampling time for the metal box and the river
values for the forgetting factor and the sampling
sand,
respectively. The data used in these simulations were
frequency. Fig. 5 and 6 show the mean square modelling
collected
by applying
a forcetime
square
to the
metal
error against
the sampling
forwave
the metal
box
andbox
the
and
the
river
sand;
the
force
wave
consisted
on
a
60
N force
river sand, respectively. The data used in these simulations
forwere
15 seconds;
next,
square wave
of square
10 N around
thethe
60
collected
by aapplying
a force
wave to
box and
sand; the the
force
wave wave
consisted
on
N metal
force offset
forthe
10 river
s; afterwards,
square
ampli
a 60was
N force
for 15 to
seconds;
a square
wave
of the
10
tude
increased
40 N next,
for 10
s; at this
time,
N around the 60 N force offset for 10 s; afterwards, the
frequency
was changed to 1.5 Hz for 10 s more; and finally,
square wave amplitude was increased to 40 N for 10 s; at
the reference force was set to 0 for the last 10 s of the
this time, the frequency was changed to 1.5 Hz for 10 s
experiment
The the
results
in Fig.force
5 and
6 by
more; and[4].
finally,
reference
was
set resampling
to 0 for the
thelast
data
using
frequencies
between
2
kHz
and
Hz 6and
5 and
by
10 s of the experiment [4]. The results in Fig.25
applying
the RLS
algorithm
to 10 repetitions
of the
experi
resampling
the data
using frequencies
between
2 kHz
and
25 Hz
the RLS algorithm to 10 repetitions of
ment
forand
eachapplying
type of terrain.
the experiment for each type of terrain.

Mean squared modelling error

Mean squared modelling error

ecution time
gorithm had

accurate as this type of terrain presents nonlinear behav

iour, and the identified model is a linear approximation of
the system.

0.89
and 0.94 lead to a more or less constant modelling error;
In Fig. 7, it can be seen that values of between 0.89 and
nevertheless,
lambda
increases
fromof0.94
to 0.99, the
In Fig. 7, itascan
be seen
that values
between
0.89
0.94 lead to a more or less constant modelling error;
modelling
errortodecreases
it reachesmodelling
its minimum
and 0.94 lead
a more oruntil
less constant
error;
nevertheless,
asaslambda
increases
0.94
to
around
= 0.98
and
then
increasesfrom
a little
between
=
nevertheless,
lambda
increases
from
0.94
to 0.99,
0.99,the
the
modelling
until
reaches
8, a similar
behaviour
is minimum
observed,
0.99
to =error
1.
Indecreases
Fig.
modelling
error
decreases
untilitit
reaches its
its
minimum
although
of thea modelling
error is=
around the=constant
0.98 andbehaviour
then increases
little between
from
0.97; then
the error
8, =
a similar
behaviour
is increases
observed,
0.99 to= 0.89
= 1.through
In Fig.
although
constant
behaviour
the modelling
error is
sharply
andthe
starts
decreasing
untilof
it reaches
its minimum
= 0.89
=
0.97; then
atfrom
=
0.996
andthrough
increasing
towards
= the
1. error increases
sharply and starts decreasing until it reaches its minimum
at = 0.996 and increasing towards = 1.

Once the
chosen, t
HADE
leg
Once the
tochosen,
use the
the
"initia
HADE
le
the
ankle
to use
the
obtain
an
the "initi
the
theincrem
ankle
that
thean
s
obtain
weight.
the increT
[6].
that the s
weight.
[6].

around = 0.98 and then increases a little between =0.99

to = 1. In Fig. 8, a similar behaviour is observed, although
the constant behaviour of the modelling error is from =
0.89 through = 0.97; then the error increases sharply and
starts decreasing until it reaches its minimum at = 0.996
and increasing towards = 1.

4.5

x 10

Mean Square Modeling Error

3.5

2.5

1.5

0.04

er sand.

h different
frequency
-minimum
resembles
ithm was
sampling
he control
solved by
power and

as applied
xperiment,
nd 8 show
metal box

ween 0.89
ling error;
o 0.99, the
minimum
ween =
observed,
ng error is
r increases

1
0.88

0.9

0.92

0.94

0.96

0.98

Figure
8. Mean
square
error vs.
for
sand
Figure
8. Mean
square
error
vs.the
river
for the
river sand.

Considering
Consideringthe
theresults
resultsofofthe
theprevious
previousexperiments,
experiments, we
we
havechosen
chosena value
a value
of 0.996,
as it presented
an
have
of ofof0.996,
as it presented
an error
error value
nearminimum
the minimum
metal
boxand
andthe
the
value
near the
withwith
the the
metal
box
minimumerror
errorvalue
valuewith
withthe
theriver-sand.
river-sand.
minimum
Oncethe
thevalues
valuesfor
forlambda
lambdaand
andthe
thesampling
samplingfrequency
frequency
Once
werechosen,
chosen,we
wetested
testedthe
thealgorithm
algorithmwith
withthe
thesame
samedata
data
were
for different forms of equation (1). The results of this
for different forms of equation (1). The results of this
implementation are shown in table 1, in which the errors
implementation
are shown
table 1,sum
in which
errors
have been divided
by the in
squared
of thethe
measured
have
been
divided
by
the
squared
sum
of
the
measured
output for meaningful comparison. That is:
output for meaningful comparison. That is:
iN=1 Ei2
i =1 Ei2iN=1 Yi2
e=
N

e=

Once the parameters of the identification algorithm were

chosen, to implement the system identification on the
HADE leg, given the leg characteristics, it was necessary to
use the state machine depicted in Figure 2. Here, during the
"initialization" state, both the total extended length of the
ankle spring and the offset force had to be computed to
obtain an accurate measure. The latter is a consequence of
the incremental encoders mounted on the leg and the fact
that the spring starts pre-compressed because of the legs
weight. This implementation has also been presented in [6].
To compute the ground penetration, and because there was
no compute
accelerometer
or inertial
measurement
unit (IMU)
To
the ground
penetration,
and because
there
mounted
on
the
leg,
we
used
the
COM
as
the
origin
of the
was no accelerometer or inertial measurement unit (IMU)
coordinateonframe.
Thus,
making
ZCOMas=the
0 origin
made of
it
mounted
the leg,
we used
the COM
necessary
to
consider
the
ground
penetration
as
the
the coordinate frame. Thus, making ZCOM = 0 made
difference
between
the foots
Zfoot and
it
necessary
to consider
thevertical
groundposition
penetration
as the
the
difference
between
the foots
vertical
position
and
the
ankles vertical
position
Zankle.
Moreover,
it Z
isf important
oot
ankles
position
Zankle . Moreover,
is important
to note vertical
that in the
implementation,
we not itonly
measure
to
thatpenetration
in the implementation,
we not onlyinmeasure
thenote
ground
but also the compression
the leg
the
penetration
but alsowethe
compression
the
due ground
to its weight.
This is because
only
have access in
to the
leg
due
to
its
weight.
This
is
because
we
only
have
displacement of the contact interface between the robot and
access to the displacement of the contact interface between
the ground. Therefore, in the algorithm we are performing
the robot and the ground. Therefore, in the algorithm
the identification on the combined robot-ground system,
we are performing the identification on the combined
which
for consistency
with for
theconsistency
rest of thewith
paper
we
robot-ground
system, which
the rest
call
ground
penetration
.
Nevertheless
the
reader
should
of the paper we call ground penetration . Nevertheless
be aware
hereafter,
when that
we say
groundwhen
penetration,
the
readerthat
should
be aware
hereafter,
we say
we
will
be
referring
to
the
combined
displacement
of the
ground penetration, we will be referring to the combined
leg and the ground
= ground
), which(is =
equal
to the
displacement
of the(leg
and +
therobot
ground
ground
+
difference
),
which
is
equal
to
the
difference
between
Z
between
Z
Z
(see
Figure9).
robot
f oot
foot
Ankle
Z Ankle (see Figure9).

(18)

where e is the
Ei are the residuals and
Yi is the measured output of the system. As can be seen
from the table, the chosen form of the recursion performs
where e is the normalized error, Ei are the residuals and
better that its counterparts, as suggested by equations (15)
Yi is the measured output of the system. As can be seen from
through (17). A detailed analysis of the residuals of the
the
table,
the chosen
form oftypes
the recursion
better
RLS
method
with different
of terrainperforms
can be found
in
that
its
counterparts,
as
suggested
by
equations
(15)
[4], where it is shown that the residuals exhibit evidence
through
(17). A detailed
analysis
of thenonlinear
residuals of
thelinear
RLS
of unmodelled
behaviour
in both
and
method
types oferrors
terrainand
cannon-linearities
be found in [4],
terrainswith
due different
to quantization
in
the robot.
where
it is shown that the residuals exhibit evidence of
unmodelled behaviour in both nonlinear and linear
terrains
tosquare
quantization
errors and
non-linearities
Table 1.due
Mean
error for different
forms
of equation (1).in
the robot. Parameterization Normalized Error

= Fc /k b/k
0.1587
Fc = k + b
0.4503
Parameterization
Normalized Error
= Fc /b k/b
0.3562
N
Yi 2
i =1 error,
normalized

= Fc /k b/k

0.1587

Fc = k + b

0.4503

Zfoot

One last
identifica
of contac
at the sol
spring. I
not oscil
to monit
algorithm
spring ha
of the m
is the ca
wheneve

Neverthe
contact c
threshold
oscillatio
COM po

In our c
left the g
incremen
caused a
was lifte

a1

(18)

touched
point eve

Zankle

5. Result
a2
b

k
robot
ground

a3

Figure
Identification
parameters.
Figure 9.9.
Identification
parameters

As can
can be
As
be seen
seen in
in Fig.
Fig. 9,
9,rrobot
obot can
can vary
vary only
only due
due to
to flexion
flexion
of
the foot,
foot, even
evenwithout
withoutobservable
observableground
groundpenetration.
penetration.
of the
Nevertheless, as our primary objective is to identify the
Nevertheless,
as of
ourthe
primary
objectivesystem,
is to identify
total
impedance
ground-robot
this is the
not
total
impedance
of
the
ground-robot
system,
this
is not
an
an issue. In other words, if the ground is perfectly
rigid
issue.
Instiff
other
words, ifthe
thealgorithm
ground is perfectly
rigidthe
(orlegs
else
(or
else
enough),
will identify
stiff enough),
the algorithm
identify
thetotal
legsstiffness
impe
impedance,
which
is a logicalwill
result
since the
will
converge
legsresult
stiffness.
is because,
dance,
which istoathe
logical
sinceThis
the total
stiffnesswhen
will
computing
equivalent
stiffnessThis
of a series
of springs:
converge tothethe
legs stiffness.
is because,
when
computing the equivalent stiffness of a series of springs:

Once the parameters of the identification algorithm were

= Fc /b k/b
0.3562
chosen, to implement the system identification on the
Table
1. Mean
error
different
forms of equation
(1) necessary
HADE
leg,square
given
theforleg
characteristics,
it was
lim (1/Keq ) = lim (1/K1 + 1/K2 ) = 1/K1
(19)
to use the state machine depicted in Figure 2. Here, during
K2
K2
the "initialization" state, both the total extended length of
C. Arevalo, Daniel Sanz-Merodio, Manuel Cestari and Elena Garcia:
the ankle spring and the offset force had to be computed to JuanThe
force Terrain
was computed
theRobots
force
Identifying Ground-Robot exerted
Impedancefoot
to Improve
Adaptability using
in Running
obtain an accurate measure. The latter is a consequence of
measurements on the ankle spring only. This force was a
the incremental encoders mounted on the leg and the fact
good approximation because the torque transmitted by the
that the spring starts pre-compressed because of the legs
spring to the ankle was the same as the torque transmitted
weight. This implementation has also been presented in

The resu
ground
Figure 1
the estim
line) and
Gaussian
box (con
scaling w

where f
n is the d
data, and

Scaled G
represen
quantitat
the estim
3,500 N/

Moreove
7 for both

of the cu
addition
Table 2 s
the estim

lim (1 / K eq ) = lim (1 / K 1 + 1 / K 2) = 1 / K 1

k2

(19)

k2

The exerted foot force was computed using the force

measurements on the ankle spring only. This force was a
good approximation because the torque transmitted by the
spring to the ankle was the same as the torque transmitted
by the ankle to the ground, as they were connected to each
other through a rigid link. Nevertheless, the arm from the
spring to the ankle was variable; therefore, this arm had to
be computed for each force measurement. Technically, the
arm from the foot to the ankle was also variable, as it
depended upon the area of contact of the foot; however,
this variation was small because the foot touched the
ground at approximately the same contact point every
time.
One last thing to note about the implementation of the
identification algorithm on the HADE leg is the detection
of contact. As the HADE leg has no force or contact sensor
at the sole of the foot, it was necessary to monitor the ankle
spring. If the spring is sufficiently rigid such that it does not
oscillate when the leg is in the swing phase, it suffices to
monitor only the spring and launch the identification
algorithm when its value is below zero (the force of the
spring has the opposite sign from the contact force because
of the manner in which it is attached to the leg). This is the
case for the HADE leg, and so contact is detected whenever
the value of the force becomes negative.

threshold. The value of the threshold will depend upon the

oscillation, and it can also be computed in the "compute
COM position" state.
In our case, the threshold was computed when the leg left
the ground. This is because the encoders used were
incremental and the leg started resting on the ground. This
caused an offset in the force measurements when the leg
was lifted.
5. Results
The results of the identification performed on the rigid
ground and variable-impedance box are presented in
Figure 10, which consists of the scaled Gaussian bell of the
estimated total stiffness (left) with the box (continuous line)
and rigid ground (dashed line), as well as the Gaussian bell
of the damping coefficient (right) for the box (continuous
line) and rigid ground (dashed line). The scaling was
performed as follows:

G=

( f )
Sigma (n)

(20)

where f is the Gaussian function obtained from the data, n

is the distribution of the values of the histogram of the data,
and G is the Gaussian bell to be plotted.

Nevertheless, note that if the spring does oscillate, then the

contact can be detected when the force reaches a certain

0.1

0.1
Metal Box
Rigid Ground

0.09
0.08

0.08

0.07

0.07

0.06

0.06

0.05

0.05

0.04

0.04

0.03

0.03

0.02

0.02

0.01

0.01

0
2000

2500

3000

3500

4000

4500

Metal Box
Rigid Ground

0.09

1.4

K (N/m)

1.6

1.8
2
B (Ns/m)

2.2

2.4

2.6

Figure 10. Scaled Gaussian bell of the estimation results for the stiffness (left) and damping (right) coefficients for the metal box (continuous line) and the rigid

ground
line)Gaussian bell of the estimation results for the stiffness (left) and damping (right) coefficients for the metal box
Figure
10.(dashed
Scaled
(continuous line) and the rigid ground (dashed line).

Gaussian
bellsforwere
chosen of
to the
offer
intuitive
TableScaled
2. Estimation
results
the stiffness
boxan
and
rigid
representation of the results. Though they do not offer a
ground.

quantitative result, the figure

shows thatvalue
the majority
of the
Estimated
Error
estimated
values lay
at approximately
2,750
N/mMean
and 3,500
Ground
Theoretical
(N/m)
Mean (N/m)
STD
(%)
N/m for the box and rigid ground, respectively.

Box Moreover, from

2,639the curves it can
2,801.5
68.45
9.35
be seen that the results for
Rigid
3,113
3,392.6
161.37
6.16
both types of ground are differentiated, as the majority of

the curves areas do not overlap with each other. In

Table 3. Estimation results for the damping coefficient of the box

and rigid ground.

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

Estimated value
Ground Mean (Ns/m) STD
Box

2.0533

0.0453

terrain were
the case of the
rocks, most
addition,
and relatively
in order toclose.
offer In
a quantitative
evaluation,
of the2estimated
200 Ns/m
Table
shows thevalues
mean were
value,approximately
the standard deviation
of (see
Figure
11
right),
while
in
the
case
of
the
sand
the
values
the estimated values, the theoretically calculated value, and
were
concentrated
at approximately
600 Ns/m. To provide
the
mean
of the estimation
error.
quantitative results, the mean and standard deviation of

Note
that in thevalues
case of
ground,
the are
identified
the estimated
forthe
therigid
rocks
and sand
presented
stiffness
corresponds
to
the
legs
stiffness,
as
the
series
in Table 4.
combination of two series springs when K1 >> K2 tends to
K2. Moreover, in the experiment, the final position of the
Table 4. Estimation results for the sand and rocks.

Estimated stiffness Estimated damping

Ground Mean (N/m) Std Mean (Ns/m) Std
Rocks
Sand

2,436.9
1,181.1

761.81
276.60

281.01
528.22

84.13
124.85

leg on the ground was the extended position, which meant

that the legs stiffness was approximately equal to the ankle
spring because the series elastic actuator s elastic constant
was 15 times higher than that of the ankle spring.
Estimated value
Ground

Theoretical (N/m) Mean (N/m)

Error

STD

Mean (%)

Box

2,639

2,801.5

68.45

9.35

Rigid

3,113

3,392.6

161.37

6.16

Table 2. Estimation results for the stiffness of the box and rigid ground

The theoretical values given in Table 2 were computed with

the stiffness values supplied by the manufacturer, so the
tolerance error was also included in the identification. The
damping constant was not included in the table because
this value was unknown for both the leg and the box, and
no theoretical value could be used to evaluate the perform
ance of the identification method. However, the estimated
values of the damping coefficient and their standard
deviation are shown in Table 3.
Figure 11 shows the results of the identification of the sand
and rocks. The figure shows the scaled Gaussian bells of the

0.1

estimated values of the stiffness (left) and damping

coefficients (right) for the rocks (dashed line) and the sand
(continuous line). In the curves, it can be observed that the
two types of ground have well-differentiated stiffnesses. In
the case of the rocks, the majority of the points were
concentrated at approximately 2,500 N/m, while in the case
of the sand they were concentrated at approximately 1,100
N/m. Note that both of these terrains generate nonlinear
behaviour, such as elastic deformation (they can be
compressed due to the geometrical rearrangement of the
rocks or the grains of sand). Thus, the dispersion standard
deviation of the estimation should be larger than in the
previous case.
Estimated value
Mean (Ns/m)

STD

Box

2.0533

0.0453

Rigid

1.7471

0.0862

Table 3. Estimation results for the damping coefficient of the box and rigid
ground

0.1
Riversand
Rocks

0.09

0.08

0.07

0.07

0.06

0.06

0.05

0.05

0.04

0.04

0.03

0.03

0.02

0.02

0.01

0.01
2000

2000
K (N/m)

4000

6000

8000

Riversand
Rocks

0.09

0.08

0
4000

Ground

0
400

200

200

400
600
b (Ns/m)

800

1000

1200

1400

Figure 11. Scaled Gaussian bell of the estimation results for the stiffness (left) and damping (right) coefficients for the rocks (continuous line) and the river-

Figuresand
11.(dashed
Scaled
Gaussian bell of the estimation results for the stiffness (left) and damping (right) coefficients for the rocks (continuous
line)
line) and the river-sand (dashed line).

environments
well modelled
equation
Regardingare
the reasonably
damping coefficients,
it can beby
inferred
from
(1). the curves that the damping coefficients of both types of
terrain were relatively close. In the case of the rocks, most

of the estimated
5.1. Convergence
time values were approximately 200 Ns/m (see
Figure 11 right), while in the case of the sand the values

Another
important
aspect
of identifying600
theNs/m.
terrain
were
concentrated
at approximately
Tocontact
provide
properties
with
a
legged
robot
is
the
convergence
time.
quantitative results, the mean and standard deviation
of
This parameter
is
important
because
the
identification
the estimated values for the rocks and sand are presented
should
performed
completely within the stance phase
in be
Table
4.
of any particular leg.
average
values ofEstimated
the components
Figure 12 shows the
Estimated
stiffness
damping
of the vector
of
parameters
used
in
the
identification
of
Ground
Mean (N/m)
Std
Mean (Ns/m)
Std
sand versus
time.
Note
that
the
figure
shows
the
values
of
Rocks
2,436.9
761.81
281.01
84.13
= [1/k,
b/k], which
was intentional
because
the vector
Sand
1,181.1
276.60
528.22
124.85
of the parameters was initialized with 1 = 1/k = 0,
Table 4. Estimation results for the sand and rocks
implying that k inf (which might have been confusing
for representation purposes).
14
12

x 10

The convergence time was approximately 0.15 s, which

was lower than the value measured by Kerdok et al. [28].
K sand +varies
std ( K sand
) < speed,
K rocks - std
K rocksproposed
)
(21)
This value
with
so (the
algorithm
can be used at relatively high speeds as long as the contact
time isKhigher
than
0.15 s.value of the samples and std(K)
where
indicates
the mean

indicates the standard deviation. In contrast, the damping

3
coefficientsx 10did
not appear useful in discriminating
6
between the terrains. Nevertheless, these coefficients
b/k did
1/k
indicate the energy losses during locomotion, so they
can
5
be used to compensate for that energy.
4

Juan C. Arevalo, Daniel Sanz-Merodio, Manuel Cestari and Elena Garcia:

Identifying Ground-Robot Impedance to Improve Terrain Adaptability in Running Robots
b/k
1/k

Parameter value

important
to note that
the
average
velocity
of legged
Table
4. demonstrates
that
the
standard
deviation
of all therobots
is
below
2
m/s
(with
the
exception
of
Boston
quantities was high compared to their respective Dynamics
mean
Cheetah).
values.
This result indicates unmodelled behaviour, which
was to be expected because the model used in this work
Moreover, the convergence time differed as the impedance
was linear, whereas the terrain behaviour is not. Nonethe
varied (i.e., with different systems). Figure 13 shows the
less, they had well-differentiated values of stiffness.
values of mathematically:
the estimated parameters with the rigid ground.
Expressed

0.1

0.1

Figure 11. Scaled Gaussian bell of the estimation results for the
stiffness (left) and damping (right) coefficients forRiversand
the rocks (continuous
Riversand
Rocks
Rocks
0.09 line).
0.09
line) and the river-sand (dashed
0.08

0.08

environments are reasonably

well modelled by equation
0.07
(1).
0.06
0.05

important
to note that the average velocity of legged robots
0.07
is below
2
m/s (with the exception of Boston Dynamics
0.06
Cheetah).
0.05

5.1. Convergence
timeof the 0.04
The value
standard deviation for the rigid ground Moreover,
Moreover,
the
convergence time
time differed
impedance
0.04
the
convergence
differedasasthe
the
impedance
and adjustable-impedance
box was approximately 2% of varied
varied
(i.e.,
with
different
systems).
Figure
13
shows
thethe
0.03
0.03
shows
(i.e., with different systems). Figure 13
Another important aspect of identifying the terrain contact
the mean, while 0.02
the values for the rocks and sand were values
values
of
the
estimated
parameters
with
the
rigid
ground.
0.02
of the estimated parameters with the rigid ground.
properties with a legged robot is the convergence time.
between 23% and
convergence
time
approximately
0.15 s,0.15
which
0.01 30%. The standard deviation was
0.01
TheThe
convergence
timewas
was
approximately
s, was
which
This parameter is important
because the identification
relatively high in both
cases because there was unmodelled was
lower
than
the
value
measured
by
Kerdok
et
al.
[28].
This
0
0
lower
than
the
value
measured
by
Kerdok
et
al.
[28].
4000
2000
0
2000
4000
6000
8000
400
200
0
200
400
600
800
1000
1200
1400
should be performed completely within the stance phase
behaviour due to motor transmissions,
joint friction and This
value
varies
withwith
speed,speed,
the so
proposed
algorithmalgorithm
can be
K (N/m)
bso
(Ns/m)
value
varies
the
proposed
of any particular
leg.
other factors that are intrinsic to the leg system. Therefore, canused
at relatively
high speeds
as long as
contact
is
be used
at relatively
high speeds
asthe
long
as thetime
contact
the
implemented
parameter
method
provides
higher
than
0.15
s.
shows
average
values
components
Figure 12
Figure
11. the
Scaled
Gaussian
bellestimation
of of
the the
estimation
results
for thetime
stiffness
(left)
and
damping
(right)
coefficients
for
the
rocks
(continuous
is higher than 0.15 s.
line)
and
the
river-sand
line).
more
accurate
results(dashed
for thein
rigid
and the adjust
of the vector
of
parameters
used
theground
identification
of
abletime.
impedance
box, as
environments
reasonably
sand versus
Note that
thethese
figure
shows the are
values
of
environments are reasonably well modelled by equation
important
to note that the average velocity of legged robots
x 10
6
well
modelled
by
equation
(1). because the vector
= [1/k,
b/k
]
,
which
was
intentional
b/k
(1).
is below 2 m/s (with the exception of Boston
Dynamics
1/k
of the parameters was initialized with 1 = 1/k = 0,
Cheetah).
5
Convergence
time
implying 5.1
that
k inf (which
might have been confusing
5.1. Convergence time
Moreover, the convergence time differed as the impedance
for representation purposes).
Another important aspect of identifying the terrain contact
4
varied
(i.e., with different systems). Figure 13 shows the
Another important aspect of identifying the terrain contact
properties with a legged robot is the convergence time. This
values
of
the estimated parameters with the rigid ground.
properties
with a legged robot is the convergence time.
x 10
3
14parameter is important because the identification should
The
convergence
time was approximately 0.15 s, which
This parameter is important because b/k
the identification
be performed completely within the stance
phase of any
was
lower
than
the
value measured by Kerdok et al. [28].
1/k
12
should
be performed completely within the stance phase
2
particular leg.
This
value
varies
with
speed, so the proposed algorithm
of any particular leg.
10
can
be
used
at
relatively
high speeds as long as the contact
Figure 12 shows the average values of the components of
1
12 shows the average values of the components
Figure
time is higher than 0.15 s.
8the vector of parameters used in the identification of sand
of the vector of parameters used in the identification of
versus time. Note that the figure shows the values of =
0
6
sand
versus time. Note that the figure shows the values of
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
x 10
[1/k, b/k], which was intentional because the vector of the
6
4 = [1/k, b/k], which was intentional because the vector
b/k
parameters was initialized with 1 = 1/k = 0, implying that
of the parameters was initialized
with 1 = 1/k = 0, Figure 13. Estimated values of the parameters 1/k (dashed line) and1/kb/k
(solid
line)
obtained
by
applying
the
identification
algorithm
to
rigid
ground
values of the parameters 1/k (dashed
k inf (which might have been confusing for representa Figure 13. Estimated
5
2
implying
that k inf (which might have been confusingline)vs and
time b/k (solid line) obtained by applying the identification
tion purposes).
for
representation purposes).
algorithm to rigid ground vs time.
0
Parameter value

Parameter value

0.05

x 10 0.15
0.1
14

0.2

0.25
0.3
Time (s)

0.35

0.4

0.45

0.5

12

b/k
1/k

Parameter value

Figure 12. Estimated values of the parameters 1/k (dashed

10 line) obtained by applying the identification
line) and b/k (solid
algorithm to sand versus time.
8

The above discussion is valid with an ankle spring with

K
= 3,discussion
1133 N/m andiswith
thewith
execution
time of
5 ms. with
If
The spring
above
valid
an ankle
spring
the
sampling
frequency
is
increased,
then
the
convergence
Kspring = 3, 113 N/m and with the execution time of
time can
be reduced.
However,
as was
stated in previous
5 ms.
If the
sampling
frequency
is increased,
then the
2
sections,
this
would
require
encoders
withasawas
higher
convergence time can be reduced. However,
stated
1
in resolution.
previous sections,
this would require encoders with a
Parameter value

higher resolution.
0
6. Conclusion
The parameters6 1/k and b/k reach convergence at
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
approximately t4= 0.3s. In contrast, biomechanical studies
6. Conclusion
According to biomechanical studies, humans and animals
show that the contact time of humans when running on
Figure
13.
Estimated
values of the
the same
parameters
1/k (dashed
adjust
their
leg stiffness
to maintain
mechanics
2
According
to biomechanical
studies, humans
and animals
rigid and soft elastic surfaces at a speed of 3.7 m/s is
line) and b/k (solid line) obtained by applying the identification
when
running
on
surfaces
different
This
adjust
their
legtostiffness
to with
maintain
the stiffnesses.
same mechanics
algorithm
rigid ground
vs
time.
approximately 0.2
s [28]. McMahon, in [37], reported
0
adjustment
allows
them
to
achieve
higher
energy
efficiency
when running on surfaces with different stiffnesses.
that the contact time for a human running over pillows
2
on adjustment
more flexible allows
terrains them
while maintaining
This
This
to achieve
higher
energy
0.05 was
0.1 over
0.15
0.2
0.3
0.35
0.4it 0.45
0.5
at a speed of 4 m/s
0.3 0.25
s. Although
is true
The above discussion
is valid
with anbalance.
ankle spring
with
Time (s)
behaviour
can
also
be
beneficial
for
robots,
as
their
behav
efficiency
on =
more
flexible
maintaining
Kspring
3, 113
N/m terrains
and withwhile
the execution
time of
that contact time decreases with increasing speed, these
iour during
can can
be explained
by the samefor
mathe
ThisIfrunning
behaviour
also
be beneficial
robots,
Figure 12. Estimated
values of the
parameters
1/k (dashed
line) and b/k balance.
5 ms.
the sampling
frequency
is increased,
then the
results show
algorithm
can
used
for running
Figurethat
12. the
Estimated
values
ofbethe
parameters
1/k (dashed
model
as
animals.
Following
this
idea,
we
(solid line) obtained by applying the identification algorithm to sand versus as matical
their
behaviourtime
during
running
canHowever,
be explained
convergence
can be
reduced.
ashave
wasby
stated
at moderate
speeds
(on
viscous
terrains).
Moreover,
it
is
line) and
b/k (solid
line) obtained
by applying
the identification
time
described
in detail
a system-identification
procedure
based with a
algorithm to sand versus time.
in previous
sections,
this would require
encoders
on higher
the RLSresolution.
method to be used in legged robots. The
The parameters 1/k and b/k reach convergence at approxi
algorithm
identifies
the combined impedance of the leg and :
www.intechopen.com
9
The
parameters
1/k and
b/k reachstudies
convergence
mately
t = 0.3s. In contrast,
biomechanical
show that at the ground simultaneously. Moreover, it considers several
Identifying
ground-robot
impedance
to
improve
terrain adaptability in running robots
approximately
t
=
0.3s.
In
contrast,
biomechanical
studies
6. Conclusion
the contact time of humans when running on rigid and soft
show
that the contact time of humans when running on aspects, such as ground modelling, the structure of the
elastic surfaces at a speed of 3.7 m/s is approximately 0.2 s
and the sampling
frequency. studies, humans and animals
According
to biomechanical
rigid
and soft elastic surfaces at a speed of 3.7 m/s is robot
[28]. McMahon, in [37], reported that the contact time for a
adjust
their
leg
stiffness
to maintain the same mechanics
approximately
0.2 s [28]. McMahon, in [37], reported Additionally, we experimentally
evaluated the HADE leg
human running over pillows at a speed of 4 m/s was over
when
running
on
surfaces
with different stiffnesses.
that the contact time for a human running over pillows (an under-actuated 3-DOF planar mechanism).
The results
0.3 s. Although it is true that contact time decreases with
This
adjustment
allows
them
achieve higher energy
at a speed of 4 m/s was over 0.3 s. Although it is true of the evaluation are encouraging, as to
the algorithm offers
increasing speed, these results show that the algorithm can
efficiency on more flexible terrains while maintaining
that contact time decreases with increasing speed, these good
convergence and tracking when applied to elastic
be used for running at moderate speeds (on viscous
balance. This behaviour can also be beneficial for robots,
results show that the algorithm can be used for running terrains
(a rigid ground and adjustable impedance box) in
terrains). Moreover, it is important to note that the average
as their behaviour during running can be explained by
at moderate speeds (on viscous terrains). Moreover, it is
velocity of legged robots is below 2 m/s (with the exception
of Boston Dynamics Cheetah).

10

www.intechopen.com

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

addition to unknown dissipative terrains (sand and rocks).

Despite the nonlinearity of the behaviour of these last two

:
Identifying ground-robot impedance to improve terrain adaptability in running robots

terrains, the algorithm is capable of producing welldifferentiated estimations of their respective stiffnesses and
produces an estimation of their damping properties.
Moreover, the convergence time of the algorithm was
analysed, comparing the convergence time with the
published data on contact time in running humans. This
comparison shows that the algorithm is faster than the
contact time of human running (3.7 m/s). Thus, it can be
used at relatively high velocities if the contact time is lower
than 0.15 s.
One major disadvantage of our approach is that if the
walking speed is high, the information can only be used in
the next step. Thus, in the worst case scenario (where the
impedance of the ground varies substantially between one
step and another) this makes the model-based approach a
better option. Nevertheless, model-based approaches
requires a model of the robot, while our approach does not.
Moreover, this method can be used to construct a map of
such types of terrain to be used later for other robots to
adjust their impedance. Another possible solution is to use
a combined approach, using the proposed method to
compute the impedance of the ground on the fly and
combine it with a model-based approach to adjust the
desired impedance as the preliminary parameters of the
contact model are updated.
7. Acknowledgements
This work was partially funded by the Spanish National
Plan for Research, Development and Innovation through
grant DPI2013-40504-R. Mr. Juan Carlos Arevalo and Mr.
Manuel Cestari would like to thank the Spanish National
Research Council and the Spanish Ministry of Economy
and Competitiveness for funding their PhD research.
8. References
[1] Alexander, R. M. (1988). Elastic mechanisms in
animal movement. Cambridge University Press,
Great Britain.
[2] Altendorfer, R., Moore, N., Komsuoglu, H., Bueh
ler, M., Brown, H. B., McMordie, D., Saranli, U., Full,
R., and Koditschek, D. E. (2001). Rhex: A biological
ly inspired hexapod runner. Autonomous Robots,
11(3):207213.
[3] Arampatzis, A., Braggemann, G. P., and Metzler, V.
(1999). The effect of speed on leg stiffness and joint
kinetics in human running. Journal of Biomechan
ics, 32(12):13491353.
[4] Arevalo, J. C., Carrillo, X., Cestari, M., SanzMerodio, D., and Garcia, E. (2013a). System identi
fication applied to contact modeling: An
experimental investigation. In Robotics and Auto
mation, 2013. Proceedings. ICRA13. IEEE Interna
tional Conference on, volume 1:3699-3706.
(Accepted).

[5] Arevalo, J. C., Cestari, M., Sanz-Merodio, D., and

Garcia, E. (2012). Impedance control for a bioins
pired underactuated leg. In 15th International
Conference for Climbing and Walking Robots,
CLAWAR12. Baltimore.
[6] Arevalo, J. C., Cestari, M., Sanz-Merodio, D., and
Garcia, E. (2013b). Event-driven ground-impedance
identification for legged robots. In 16th Internation
al Conference for Climbing and Walking Robots,
CLAWAR13.
[7] Arevalo, J. C. and Garcia, E. (2012). Impedance
control for legged robots: An insight into the
concepts involved. Systems, Man, and Cybernetics,
Part C: Applications and Reviews, IEEE Transac
tions on, 42(6):1400 1411.
[8] Blickhan, R. (1989). The spring-mass model for
running and hopping. Journal of Biomechanics,
22(11):12171227.
[9] Blickhan, R. and Full, R. J. (1993). Similarity in
multilegged locomotion: bouncing like a monop
ode. Journal of Comparative Physiology A: Neuro
ethology, Sensory, Neural, and Behavioral
Physiology, 173(5):509517.
[10] Box, G. E. P., Jenkins, G. M., and Reinsel, G. C.
(1976). Time series analysis. Holden-day San
Francisco.
[11] Cavagna, G. A., Franzetti, P., Heglund, N. C., and
Willems, P. (1988). The determinants of the step
frequency in running, trotting and hopping in man
and other vertebrates. The Journal of Physiology,
399(1):8192.
[12] Clark, B. D. (1988). Mechanics and control on the
hind limb of bobwhite quail running and landing
on substrates of unpredictable mechanical stiffness.
PhD thesis, Chicago University.
[13] Dupont, P. E., Schulteis, C. T. M., Millman, P. A.,
and Howe, R. D. (1999). Automatic identification of
environment haptic properties. Presence, 8(4):394
411.
[14] Farley, C. T., Glasheen, J., and McMahon, T. A.
(1993). Running springs: speed and animal size.
Journal of Eperimental Biology, 185(1):7186.
[15] Farley, C. T. and Gonzalez, O. (1996). Leg stiffness
and stride frequency in human running. Journal of
Biomechanics, 29(2):181186.
[16] Ferris, D. P. and Farley, C. T. (1997). Interaction of
leg stiffness and surface stiffness during human
hopping. Journal of Applied Physiology, 82(1):15
22.
[17] Ferris, D. P., Liang, K., and Farley, C. T. (1999).
Runners adjust leg stiffness for their first step on a
new running surface. Journal of Biomechanics,
32(8):787794.
[18] Ferris, D. P., Louie, M., Farley, C. T., Ferris, D. P.,
Louie, M., and Farley, C. T. (1998). Running in the

Juan C. Arevalo, Daniel Sanz-Merodio, Manuel Cestari and Elena Garcia:

Identifying Ground-Robot Impedance to Improve Terrain Adaptability in Running Robots

11

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

[29]

[30]

12

real world: adjusting leg stiffness for different

surfaces. Proceedings of the Royal Society of
London. Series B: Biological Sciences, 265(1400):
989994.
Full, R. J. and Koditschek, D. E. (1999). Templates
and anchors: neuromechanical hypotheses of
legged locomotion on land. Journal of Experimental
Biology, 202(23):33253332.
Garcia, E., Arevalo, J., Munoz, G., and Gonzalez-deSantos, P. (2011). Combining series-elastic actuation
and magneto-rheological damping for the control of
agile locomotion. Robotics and Autonomous
Systems, 50(10):827839.
Giraldi, G. and Dharf, I. (2002). Literature survey of
contact dynamics modelling. Mechanism and
Machine Theory, (37):12131239.
Grizzle, J. W., Hurst, J., Morris, B., Park, H. W., and
Sreenath, K. (2009). MABEL, a new robotic bipedal
walker and runner. In American Control Confer
ence, 2009. ACC09., pages 20302036.
Gustafsson, F. and Gustafsson, F. (2000). Adaptive
filtering and change detection, volume 5. Wiley
Online Library.
Ham, R., Sugar, T., Vanderborght, B., Hollander, K.,
and Lefeber, D. (2009). Compliant actuator designs.
Robotics & Automation Magazine, IEEE, 16(3):81
94.
Ham, R., Vanderborght, B., Damme, M., Verrelst, B.,
and Lefeber, D. (2006). MACCEPA: the actuator
with adaptable compliance for dynamic walking
bipeds. Climbing and Walking Robots, pages 759
766.
Hurst, J. and Rizzi, A. (2008). Series compliance for
an efficient running gait. Robotics & Automation
Magazine, IEEE, 15(3):4251.
Hurst, J. W., Chestnutt, J. E., and Rizzi, A. A. (2004).
An actuator with physically variable stiffness for
highly dynamic legged locomotion. In Robotics and
Automation, 2004. Proceedings. ICRA04. 2004
IEEE International Conference on, volume 5, pages
46624667.
Kerdok, A. E., Biewener, A. A., McMahon, T. A.,
Weyand, P. G., and Herr, H. M. (2002). Energetics
and mechanics of human running on surfaces of
different stiffnesses. Journal of Applied Physiology,
92(2):469478.
Kikuuwe, R. and Yoshikawa, T. (2002). Robot
perception of environment impedance. In Robotics
and Automation, 2002. Proceedings. ICRA02. IEEE
International Conference on, volume 2, pages 1661
1666.
Kikuuwe, R. and Yoshikawa, T. (2003). Recognizing
surface properties using impedance perception. In

Int J Adv Robot Syst, 2015, 12:1 | doi: 10.5772/59888

Robotics and Automation, 2003. Proceedings.

ICRA03. IEEE International Conference on, volume
1, pages 15391544.
[31] Lejeune, T. M., Willems, P. A., and Heglund, N. C.
(1998). Mechanics and energetics of human locomo
tion on sand. Journal of Experimental Biology,
201(13):20712080.
[32] Ljung, L. and Soderstrom, T.(1987). Theory and
practice of recursive identification. MIT press series
in signal processing, optimization, and control (4),
Massachussets.
[33] Ljung, L. (1987). System identification: theory for
the user. Preniice Hall Inf and System Sciencess
Series, New Jersey, 7632.
[34] MacLean, K. (1996). The haptic camera: A techni
que for characterizing and playing back haptic
properties of real environments. In Proceedings of
ASME Dynamic Systems and Control Division,
volume 58, pages 459467.
[35] McMahon, T. A. (1985). The role of compliance in
mammalian running gaits. Journal of Experimental
Biology, 115(1):263282.
[36] McMahon, T. A. and Cheng, G. C. (1990). The
mechanics of running: How does stiffness couple
with speed? Journal of Biomechanics, 23:6578.
[37] McMahon, T. A. and Greene, P. R. (1979). The
influence of track compliance on running. Journal
of Biomechanics, 12(12):893904.
[38] Nichol, J. G., Singh, S. P. N., Waldron, K. J., Palmer,
L. R., and Orin, D. E. (2004). System design of a
quadrupedal galloping machine. The International
Journal of Robotics Research, 23(10-11):10131027.
[39] Raibert, M., Blankespoor, K., Nelson, G., and
Playter, R. (2008). BigDog, the rough-terrain
quadruped robot. In Proceedings of the 17th World
Congress IFAC. Seoul, South Korea.
[40] Raibert, M. H. (1986). Legged robots that balance.
MIT Press.
[41] Raibert, M. H. (1990). Trotting, pacing and bound
ing by a quadruped robot. Journal of Biomechanics,
23:7998.
[42] Semini, C. (2010). HyQ - Design and Development
of a Hydraulically Actuated Quadruped Robot.
PhD thesis, Italian Institute of Technology and
University of Genoa.
[43] Zamparo, P., Perini, R., Orizio, C., Sacher, M., and
Ferretti, G. (1992). The energy cost of walking or
running on sand. European Journal of Applied
Physiology and Occupational Physiology, 65(2):
183187.