Академический Документы
Профессиональный Документы
Культура Документы
To formulate disturbance-rejection measures for limit cycle largest Floquet multiplier to the unit circle [19], as we will
walkers, two concepts from general nonlinear dynamics anal- show in Section V.
ysis [17] are in use. The basin of Attraction (BoA) involves a 3) Largest allowable deterministic disturbance: Some re-
global assessment of the system states that result in convergence searchers have measured the disturbance rejection of their
to the limit cycle. The Floquet multipliers indicate the rate of biped by measuring the maximum size or range of a spe-
convergence on a step-to-step basis for small deviation from the cific deterministic disturbance that the biped can manage
limit cycle (linearized stability). Disturbance rejection is also without falling. Examples of such disturbances are an im-
occasionally quantified by the largest disturbance a limit cycle pact at the hip [5], a stepdown in the floor [20], and a
walker can handle. As shown in Section II, none of these mea- slope [21].
sures are satisfactory for the evaluation and optimization of limit The benefit of this measure is the good correlation with
cycle walkers. The BoA and the largest allowable disturbance actual disturbance rejection, in case it is obtained for a
take too much computation time or experimentation time (if at representative set of real world disturbances. Drawbacks
all possible), while the faster Floquet multipliers have limited of this measure are the long computation or experimental
correlation with actual disturbance rejection. time and the fact that the boundary of what a biped can
In this paper, we present a new measure for limit cycle walk- handle needs to be exceeded to establish this measure,
ers that has a short calculation time and good correlation with meaning in practice that the biped will have to fall.
actual disturbance rejection. We call the new measure the Gait
Sensitivity Norm (g/e2 ). The definition of the new mea- III. THE GAIT SENSITIVITY NORM
sure will be given in Section III. Then, in Section IV and V,
the Gait Sensitivity Norm will be subjected to a comparative A. General Concept
study with other existing measures. In Section VI, we apply the We introduce a new measure, which we call the Gait Sen-
Gait Sensitivity Norm to a real prototype. The discussion and sitivity Norm g/e2 . It quantifies the effect of a set of
conclusion follow in Section VII and VIII. disturbances on a walking gait. To establish the Gait Sensi-
tivity Norm, we define a generalized system description for a
limit cycle walkers gait having a set of disturbances e as the
II. EXISTING MEASURES FOR LIMIT CYCLE WALKERS
system input and a set of gait indicators g as system output
For limit cycle walkers, three measures to quantify distur- (Section III-B). The Gait Sensitivity Norm g/e2 measures
bance rejection are currently in use. All three have specific the size of the dynamic response of this system (Section III-C).
properties that make them unsuitable in practice when evaluat- The selection of disturbances e and gait indicators g are open
ing or optimizing performance. The measures and their respec- to the designer, which adds expert knowledge to this measure.
tive properties are listed later. The first measure, the BoA, will be A meaningful selection of e and g is crucial for the success of
used for a validity check of our new measure in Section IV. The the Gait Sensitivity Norm; an example is given in Section VI.
other two measures are subjects of a quantitative comparison 1) Disturbances e. The set of disturbances consists of those
with our new measure in Section V. disturbances that are of interest to the designer. The set
1) Basin of Attraction: The BoA is the total set of system should instigate the important failure modes (ways to fall)
states of a limit cycle walker for which its gait converges of the limit cycle walker. This can include disturbances
to its nominal limit cycle [18], [19]. This measure involves that effect the motion once per step, such as floor irregu-
a full evaluation of the nonlinear system behavior starting larities, and continuous disturbances such as sensor noise
at all possible system states. or torque ripple. The designer has to weigh the distur-
The benefit of the BoA is the good correlation between bances to account for the different magnitudes in which
the actual disturbance rejection and the distance from the they occur in practice.
limit cycle to the borders of the BoA. The drawback is the 2) Gait indicators g. The gait indicators should quantify the
large amount of computational or experimental time (if at characteristics of the walking gait that are directly re-
all possible) that is required to obtain the BoA as well as lated to the failure modes. Various gait indicators can be
the arbitrary interpretation of the BoA (e.g., dependent on used, such as step width, step time or ground clearance at
the choice of system states, involving multiple units, hard midswing. For instance, step time indicates how close a
to quantify due to complex shape). 2-D walker is to falling forward and falling backward (see
2) Largest Floquet multiplier: Floquet multipliers indicate Section IV-C), and ground clearance indicates how close a
how fast small deviations from the limit cycle converge walker is to the failure of toe stubbing. The gait indicators
on a step-to-step basis [4], [5], [17]. For a stable limit are weighted by dividing them over the absolute value of
cycle, the Floquet multipliers have to be within the unit the gait indicator for which falling is expected (e.g., the
circle; the closer to zero, the faster the convergence rate. nominal minimal ground clearance).
The benefit of the Floquet multipliers is that they require a 3) Size of Dynamic Response g/e2 . The dynamic re-
short calculation or experimental time as they involve only sponse of the system is the variability of the gait indica-
small deviations from the limit cycle (a linearization ap- tors as a result of the input disturbances. This dynamic
proach). The drawback is the limited correlation between response is an essential factor in measuring disturbance
the actual disturbance rejection and the distance from the rejection. This claim is supported by several human gait
HOBBELEN AND WISSE: A DISTURBANCE REJECTION MEASURE FOR LIMIT CYCLE WALKERS: THE GAIT SENSITIVITY NORM 1215
To obtain the Gait Sensitivity Norm, we define a system de- The linearization of the stride function S is performed at the
scription of a limit cycle walkers gait with disturbances e as fixed point v of S, which is the intersection of the gait limit
input and gait indicators g as output. The system describes cycle with the Poincare section. The value of g at the fixed point,
how the gait indicators vary on a step-to-step basis as a re- g , is the nominal value of the gait indicators:
sult of small disturbances that occur when the walker is in its v = Sv (v , 0)
nominal limit cycle motion. The system response can be ob-
tained in two ways: 1) direct identification of the inputoutput g = Sg (v , 0). (2)
relation of the system and 2) through a statespace system To perform the linearization, the fixed point needs to be known.
description. It is found by performing a NewtonRaphson search as done by
1) Direct InputOutput Identification: The first option, di- McGeer [5]. The linearized model will consider small deviations
rectly identifying the inputoutput relation, is especially suitable from the fixed point v and small disturbances e, resulting in
for application on real prototypes. One can apply a single im- small deviations of the gait indicators from their nominal values
pulse disturbance to the system (e.g., a stepdown in the floor) g.
and measure the fluctuation of the gait indicators g as a result of The statespace representation of the complete linearized dis-
that. The disturbance needs to be large enough to have a distin- crete system model that describes the limit cycle walker, as
guishable effect, but small enough to make sure that the walker depicted in Fig. 2, is
does not fall. To incorporate multiple disturbances, separate ex-
periments for every disturbance have to be performed. vn +1 = A vn + B en
2) StateSpace System Description: The second option, the gn = C vn + D en (3)
statespace system description, is better suited for simulation
as it needs less simulation time to obtain the system response to where:
multiple disturbances.
r A = S v (v , 0) , which is the linearized version of the stan-
vn
A limit cycle walker can be described as a linear discrete dard stride function, the mapping from initial conditions
statespace system by performing a linearization of its gait on v of step n to the initial conditions of step n + 1, also
a step-to-step basis. This process of linearization has first been known as the Jacobian J;
described by Hurmuzlu [4], and later, by McGeer [5]. Subse-
r B = S v (v , 0) , which is a sensitivity matrix that describes
en
quently, it has frequently been used by passive dynamic walk- the effect of the disturbances e in step n on the initial
ing researchers. It is an essential part of the calculation of the conditions v of the next step;
Floquet multipliers of a limit cycle walker, as mentioned in r C = S g (v , 0) , which is a sensitivity matrix that gives the
vn
Section II. deviation of the gait indicators g in a step that starts out
The linearization process is based on the discrete step-to-step with initial conditions v;
analysis called Poincare mapping. A walking step is interpreted r D = S g (v , 0) , which is a sensitivity matrix that gives
en
as a Poincare map, the mapping of the initial conditions of the deviation of the gait indicators g in a step in which
the step (vn ) to the initial conditions of the next step (vn +1 ). disturbances e are present.
This nonlinear mapping S was termed the stride function by All the sensitivity matrices A, B, C, and D in (3) are numer-
McGeer. It is found by performing a full dynamic simulation ically determined through a perturbation analysis of the stride
of one stride. For this paper, the stride function S has been function. This means that, for all initial conditions and all dis-
extended to incorporate the gait indicators gn and the effect of turbances, a full dynamic simulation of one stride needs to be
1216 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 6, DECEMBER 2007
Fig. 4. Typical walking step. Just after footstrike, the swing leg (heavy line)
swings forward past the stance leg (thin line) until the swing leg hits the ground
and a new step begins. is the angle between the stance leg and the slope
Fig. 3. Discrete system responses to an impulse and a white noise disturbance
normal, is the angle between the two legs, l is the leg length, rf is the foot
input and their relation to the H 2 -norm of the system (E[.] is the expected value
radius, M is the hip mass, m is the foot mass, k is the hip spring stiffness, g
operator). Each point represents a new walking step (k-axis). These responses
is the gravitational acceleration, and is the slope angle. Adapted from Garcia
are given here as examples, but are in fact real responses of the model described
et al. [26].
in Section IV in which e is a floor height difference and g is step time.
performed in which one initial condition or disturbance is per- IV. COMPARATIVE STUDY
turbed by a small amount. After this has been done, the complete
To demonstrate the value of the Gait Sensitivity Norm, we
dynamic response can be calculated through a series of matrix
subject it to several test cases. The tests are performed on
multiplications.
an extended version of the simplest walking model by Garcia
et al. [26], (see Fig. 4). We vary the foot radius, ground slope, and
C. Calculating the H2 Norm hip spring stiffness. For each of the three parameter variations,
To quantify the size of the dynamic response of our limit we record the effect on three disturbance rejection measures:
cycle walker system, the Gait Sensitivity Norm calculates the 1) the reciprocal of the Gait Sensitivity Norm (1/ ge 2 ) for
H2 norm (.2 ) of this system. The H2 norm measures system which the choice of disturbance e and gait indicator g is
responses to both impulse inputs and white noise inputs, as explained in Section IV-C;
shown in Fig. 3. 2) the distance between the largest Floquet multiplier and
The relation of the H2 norm to the impulse response of a the unit circle (1 max(||));
system is especially useful in determining the Gait Sensitivity 3) the largest allowable deterministic disturbance
Norm in real prototypes. In case of direct inputoutput identifi- (max(|e|)), using the same disturbance e as for the
cation as described in Section III-B, the Gait Sensitivity Norm Gait Sensitivity Norm.
ge 2 can directly be obtained from the discrete response of For all three of these measures, a higher value should qualify
the gait indicators g to a single disturbance e0 better disturbance rejection; that is why we take the reciprocal
value of the Gait Sensitivity Norm. The BoA measure is not
q
g 1 included in this study as it is not a quantitative measure (it is a
= (gk (i) g (i))2 (4)
e |e0 | i=1 shape).
2 k =0
in which gk (i) is the value of the ith gait indicator k steps A. Comparison Criteria
after the disturbance has occurred and q is the number of gait
indicators. The three disturbance rejection measures 1/ ge 2 , 1
In case the dynamic response of the walker is obtained through max(||) and max(|e|) will be compared based on two criteria.
simulation using the statespace description, the Gait Sensitivity 1) Correlation With Actual Disturbance Rejection: A mea-
Norm involves solving (5) sures prediction is good when the measure is highly cor-
related to the actual disturbance rejection; the correlation
coefficient r2 gives the part of the change in the actual dis-
g
= trace(DT D) + trace(BT (AT )k CT CAk B) turbance rejection that is explained by respective measures
e
2 k =0 (in percentages). The number r2 does not tell whether
(5) the correlation is positive or negative. In this comparative
in which trace (X) is the sum of the elements on the main study, only a positive correlation is good, and thus, we will
diagonal of the matrix X. give the sign of the correlation with the value of r2 .
HOBBELEN AND WISSE: A DISTURBANCE REJECTION MEASURE FOR LIMIT CYCLE WALKERS: THE GAIT SENSITIVITY NORM 1217
The actual disturbance rejection is the walkers ability to One can imagine that the walker will not manage to overcome
prevent a fall in the presence of disturbances. We will a high step up, and consequently, fall backward. A large step
establish this ability given the presence of Gaussian white down in the floor will result in falling forward.
noise disturbances as this gives a good approximation of The choice of gait indicator g is the step time T , as we
common real-world disturbances. We define the actual hypothesize that there is a strong relation between step time
disturbance rejection as the magnitude (95% confidence and both of the failure modes. The larger the variations of step
interval) of a Gaussian white noise disturbance for which time T , the closer the model is to a fall. We reason as follows.
the walker is able to prevent falling for 95% of the time in Falling forward occurs when a limit cycle walker is not able to
an 80-step trial. This 80-step trial is long enough to capture put its swing leg in front of its stance leg fast enough to prevent
the effect of slow convergence after a disturbance and short the fall [20]. This implies that a limit cycle walker is closer
enough to result in reasonable computational times. The to falling forward when its step time is smaller, suggesting an
choice of disturbance is floor height variations, the same expected relation between step time and the failure mode of
as will be used in establishing the Gait Sensitivity Norm. falling forward. Falling backward of a dynamic walker is due
The rationale for this choice will be given in Section VI-C. to the fact that the walker has insufficient energy content to
2) Relative Calculation Time. The three measures as well as pass the apex (point of highest potential energy content, i.e.,
the actual disturbance rejection are calculated with a sim- midstance) during walking. When a walker is close to falling
ulation in a MATLAB environment on a 1.9 GHz Intel backward, it will have a low level of kinetic energy around the
Pentium M processor with 1 GB of RAM. The computa- apex, causing a slow motion, and thus, long step time. This
tional time needed to calculate the measures is compared suggests an expected relation between step time and the failure
to the time it takes to calculate the actual disturbance re- mode of falling backward.
jection and given in percentages. To test the hypothesis and establish the relation between the
step time and falling, we performed a comparison between the
B. Model Description step time variability and the BoA of our model. For all points
(sets of initial conditions) in the BoA of this model, we simulate
Fig. 4 shows the extended version of the simplest walking
the fully nonlinear dynamic response to those specific initial
model by Garcia et al. [26]. Their original model is a 2-D model
conditions and calculate the 2-norm of the step time response
consisting of two rigid links with unit length l, connected at the
T 2
hip. There are three point masses in the model, one in the hip
with unit mass M and two infinitesimally small masses m in the
feet. The model walks down a slope in a gravity field with unit T 2 (v0 ) = (Tk T )2 (6)
magnitude g. The extended model adds a torsional spring in the k =0
hip with stiffness k and arced feet with arc radius rf . Due to
where
the normalization of the model (M = 1, l = 1, g = 1), the only
free model parameters are the slope angle , the spring stiffness 0
v0 = .
k, and the arc radius rf . Similar compass gait models have been 0
used by other researchers [27][30].
The dynamics of the model consists of two parts. The first If the expected relation between step time and falling exists,
part is the continuous dynamics that describes the motion of the the BoA and the 2-norm of the step time response should be
stance and swing leg in between footstrikes. The footscuffing related. Fig. 5 shows that indeed this relation exists; the contour
during midstance that inevitably occurs with 2-D straight-legged map of T 2 and the BoA that are depicted show a relation in
(knee-less) models is ignored. The second part of the dynamics shape. For increasing T 2 , the contour map monotonically in-
is the discrete impact that describes the footstrike, as this is mod- creases and takes the shape of the boundary of the BoA, where
eled as a fully inelastic instantaneous collision. The equations falling backward or falling forward occur. In the direction in
of motion of both parts are given in the Appendix. which the edges of the BoA are further away, T 2 increases
more slowly, which shows that the largest allowable disturbance
C. Choice of Disturbance and Gait Indicator and T 2 are inversely related. This finding confirms the hy-
pothesis that step time is a good indicator for the chance of
We will use the model to test our new disturbance rejection
either falling forward or falling backward.
measure. As mentioned in Section III-A, the Gait Sensitivity
Norm requires a meaningful selection of disturbances e and gait
indicators g. This selection depends a bit on the particular model V. PROOF OF CONCEPT: RESULTS COMPARISON
under consideration. In this case, the simple 2-D walking model This section presents the results of the comparative study. We
has only two failure modes: it can fall forward or backward. vary three parameters in the model of Section IV-B:
So, a selection of e and g is meaningful if both failure modes 1) the foot radius rf ;
are instigated by the disturbances e and measured by the gait 2) the hip spring stiffness k;
indicators g. 3) the slope angle .
The choice of disturbance e for this study is floor irregulari- For each parameter variation, the limit cycle is found. Then,
ties, as they induce both falling forward and falling backward. the three disturbance rejection measures are calculated and
1218 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 6, DECEMBER 2007
Fig. 5. The shape of the BoA of the simplest walking model on a slope of Fig. 6. Actual disturbance rejection and three disturbance rejection measures
= 0.004 rd (rf = 0 and k = 0) [19], accompanied by a contour map of the for increasing foot radius of the extended simplest walking model, slope angle
root-mean-square of the step time response, T 2 . The contour map monotoni- = 0.004 rd and hip spring stiffness k = 0 are left unchanged. For rf > 0.90,
cally increases from the location of the fixed point (white star) toward the edges the limit cycle is unstable and the disturbance rejection reduces to zero. The four
of the BoA. Outside the BoA, the walker falls forward or backward [19]. quantities are scaled for visibility. For the comparison, we are only interested
in relative changes in the quantities due to changing foot radius, the absolute
values are not important.
Fig. 9. Range of allowed single floor height differences for changing slope
angle. For > 0.007rd, the shape of this range is highly irregular. The definition
Fig. 11. (a) The physical 2-D limit cycle walker called Meta. (b) This
of max(|e|) in this irregular shape is indicated by the arrow.
prototype is subjected to a parameter study establishing the Gait Sensitivity
Norm through measurement trials with a single step-down disturbance. (c) For
each parameter setting in the study Meta also performed a 100-step trial over
a floor with randomly varying height to evaluate the correlation between the
measured Gait Sensitivity Norm and actual disturbance rejection.
Fig. 10. Actual disturbance rejection and three disturbance rejection measures
for increasing slope angle of the extended simplest walking model.
TABLE IV
QUANTITATIVE COMPARISON MEASURES FOR CHANGING SLOPE ANGLE
Fig. 12. (a) Gait Sensitivity Norm measured on the physical prototype Meta.
The box plots show the result of 20 measurement trials for each of the four-
parameter settings. (b) Results of four 100 step trials walking over a floor
with randomly varying height. For each parameter setting, the results show the
amount of steps in the trial that were successful (100 steps minus the steps in
which a fall occurred). The results in (a) and (b) show the same trend in the
effect of the parameter setting on the disturbance rejection of the prototype.
the confidence level of our measurement by canceling out the The fact that the Gait Sensitivity Norm performs much better
variability due to noise that is secondary to our purposefully than the Floquet multipliers in terms of predicting a walkers
applied floor disturbance (e.g., sensor noise). actual disturbance rejection is mainly determined by two factors.
A full validation of the measured Gait Sensitivity Norm with 1) The Gait Sensitivity Norm incorporates the effect of real-
actual disturbance rejection, as was done for the simple model world disturbances instead of looking at perturbations in
in Section V, is unfortunately not practically feasible in case of the walkers state space. By doing this, the Gait Sensi-
a real prototype. In practice, establishing the actual disturbance tivity Norm incorporates the likelihood of certain state
rejection as defined in Section IV-A would require an excessive perturbations occurring in reality in contrast to assuming
laboratory setup (large diversity in floor properties), extremely all eigenmodes are excited equally.
long experimental times, and many, possibly destructive, falls. 2) The use of gait indicators weighs the relevance of the
Only an approximate validation of the Gait Sensitivity Norm walkers eigenmodes with respect to actual failure modes.
is possible by obtaining a rough indication of the actual distur- Instead of observing the eigenmodes that have the slowest
bance rejection. convergence as is done with the largest Floquet multiplier,
This indication is gained by letting the prototype walk over the Gait Sensitivity Norm observes a specific combination
a fixed floor with randomly varying floor height and measuring of eigenmodes that brings the walker closest to a fall.
the amount of successful steps and falls (unsuccessful steps) that
occur during this trial [Fig. 11(c)]. The percentage of successful
steps in the total set of steps is an indicator of the prototypes B. Choice of Disturbances and Gait Indicators
actual disturbance rejection. In this study, we have built such a The choice of disturbances e and gait indicators g is crucial
randomly varying floor by the use of a set of wooden boards for the success of the Gait Sensitivity Norm. The choice to
with varying thicknesses (multiples of 4 mm). The order in use floor height variations as disturbance and step time as gait
which these boards were placed was based on a random num- indicator in this study turns out to be successful for the simple
ber generator with a uniform distribution. The total floor length 2-D walking model.
allowed us to have the prototype walk 100 steps (for each pa- In general, the choice of disturbances e is free to the designer.
rameter setting). Due to material and space limitations, we built One can choose either to study a walkers ability to cope with
the floor in six separate pieces. Fig. 12(b) shows the results of one specific disturbance that seems essential to the designer
the experiment. These results show the same trend in the effect or try to build a complete set of disturbances that represents
of the parameter setting on the disturbance rejection of the pro- the real world. We think that, in general, for various walkers,
totype. Parameter setting 1 has the lowest disturbance rejection using floor height variations is a good choice. It is a disturbance
and parameter setting 3 the highest, parameter settings 2 and 4 that is constantly present in real-world situations, and it has a
perform in between these extremes. As the measurements on the large influence on walkers behavior. This is shown by the fact
long random floor take a long time to perform, we did not repeat that most existing bipedal robots have a hard time coping with
it as often as the Gait Sensitivity Norm measurements to get the unexpected floor irregularities and only perform successfully on
same confidence level. Nonetheless, the experiment indicates well-conditioned surfaces.
that the measured Gait Sensitivity Norm on the real prototype The choice of gait indicators g is not as open as the choice of
gives a good prediction of the actual disturbance rejection. disturbances e. For every application, the designer needs to sum-
marize the possible failure modes and check whether the chosen
VII. DISCUSSION set of gait indicators is related to all of those. In Section VI-C, we
showed how this was done for our simple 2-D walking model. In
The results in Section V show that the Gait Sensitivity Norm is
general, the choice of gait indicators involves understanding the
the only disturbance-rejection measure in this study that is able
main failure modes that underlie a specific walker. It is our belief
to give a good prediction of the disturbance rejection of a simple
that such specific understanding is crucial and necessary in the
2-D walking model in a short calculation time. Section VI shows
construction of a disturbance-rejection measure, as the highly
that the Gait Sensitivity Norm also gives a good prediction of
nonlinear nature of limit cycle walking prevents a more gen-
the disturbance rejection of a more complex, real walking robot.
eralized definition. We intend to study models and prototypes
In this section, we will give an explanation of this performance
of incrementally increasing complexity to compose a complete
and discuss the general use of the Gait Sensitivity Norm in case
understanding of the failure modes underlying bipedal walking,
other walkers are involved.
and consequently, a complete set of gait indicators. It should
be noted that, although this study only involves a single gait
A. Explanation of the Gait Sensitivity Norms Performance
indicator, the Gait Sensitivity Norm allows the use of a vector
For all disturbance-rejection measures studied, the Gait Sen- of indicators when necessary.
sitivity Norm as well as the Floquet multipliers excel in achiev- At this stage of our research, we expect that using step time
ing fast calculation times. This is because both are based on as the gait indicator will be effective for most bipedal walkers.
the concept of linearizing a walkers step-to-step behavior us- The results in Sections V and VI show this is true for both a
ing the Poincare analysis method. This linearization drastically simple and a more complex, realistic prototype walker. Also,
decreases the calculation time needed to establish a limit cycle we examined the effect of using other gait indicators on the
walkers behavior. performance of the Gait Sensitivity Norm. Table V shows that,
1222 IEEE TRANSACTIONS ON ROBOTICS, VOL. 23, NO. 6, DECEMBER 2007
g0 + g0,p ert = Sg (v0,p ert , 0) at every step. The nominal value of the gait indicators g
is estimated by taking the mean value of the gait indicators
A(:, j) = v1,p ert /
over multiple steps. As the experiment(s) will usually not
C(:, j) = g0,p ert / start out with the fixed point initial conditions v , the mea-
surements on the first couple of steps out of an experiment
will have to be omitted.
where A(:, j) indicates the jth column of matrix A.
3) Set up separate disturbance experiments: For every dis-
4) Calculate sensitivity matrices B and D: The sensitiv-
turbance out of the set e, a separate experiment should
ity matrices B = Sv (v , 0)/en and D = Sg (v , 0)/en
be conducted. In every experiment, the respective distur-
are obtained in a similar fashion as matrices A and C.
bance should be applied once during a single step (e.g.,
In this case, a single stride is simulated in which one dis-
one single stepdown in the floor). The size of the distur-
turbance e(j) is present and the initial conditions are set
bance should be big enough to cause a distinguishable
to be the fixed-point values v . This is repeated for all
effect, but small enough to ensure that falling does not
elements of e.
occur. Before this disturbance is applied, the prototype
5) Calculate the H2 -norm: To complete the calculation of the
needs to be able to settle down into its nominal limit cy-
Gait Sensitivity Norm, (5) from Section III-C is solved, using
cle motion. The best way of doing this is creating a large
the matrices A, B, C, and D that have been determined.
enough disturbance-free section before the disturbance.
Whether the convergence to the limit cycle motion was
Physical Prototypes
successful can be checked by observing the value of the
1) Choose disturbances e and gait indicators g : This choice gait indicators g in the step just before the disturbance and
should be based on the knowledge on the failure modes of comparing it with g .
the model, as exemplified in Section VI-C. 4) Measure variability of gait indicators: In the steps follow-
2) Estimate the fixed point and g : To estimate the fixed point ing the purposely applied disturbance (e.g., k from 1 to
and the nominal value of the gait indicators g , one (or 20), the value of the gait indicators gk should be measured.
multiple) walking experiment needs to be performed in 5) Calculate the H2 -norm: The practical estimation of the
which the presence of disturbances is minimized. During Gait Sensitivity Norm is completed by calculating (4) from
this experiment, the gait indicators g have to be measured Section III-C.
(1rf )((1rf )+rf cos ())((1rf ) sin ()rf sin ()) 2 rf (1 rf ) sin ()2
= +
(1 rf ) + rf + 2rf (1 rf ) cos ()
2 2 (1 rf ) + rf2 + 2rf (1 rf ) cos ()
2
((1rf ) cos ()+rf cos ())((1rf ) sin ()rf sin ()) (1rf ) sin ()rf sin ()
+ + k
(1rf )2 +rf2 +2rf (1rf ) cos () (1 rf )2 + rf2 + 2rf (1 rf ) cos ()
+ =
+ =
(1 rf )2 cos ( ) + rf (1 rf )(cos ( ) + cos ( )) + rf2
+ =
(1 rf )2 + rf2 + 2rf (1 rf ) cos ( )
rf (1 rf ) + (1 rf ) cos ( )(1 (1 rf ) cos ( ))
+ = (1 rf )
(1 rf )2 + rf2 + 2rf (1 rf ) cos ( )
ACKNOWLEDGMENT [20] M. Wisse, A. L. Schwab, R. Q. van der Linde, and F. C. T. van. Der. Helm,
How to keep from falling forward: Elementary swing leg action for
The authors would like to thank Frans van der Helm, Arend passive dynamic walkers, IEEE Trans. Robot., vol. 21, no. 3, pp. 393
Schwab, and Richard van der Linde for helpful comments. 401, Jun. 2005.
[21] J. Pratt, C.-M. Chew, A. Torres, P. Dilworth, and G. Pratt, Virtual model
control: An intuitive approach for bipedal locomotion, Int. J. Robot.
REFERENCES Res., vol. 20, no. 2, pp. 129143, Feb. 2001.
[22] R. M. Guimaraes and B. Isaacs, Characteristics of the gait in old people
[1] J. Pratt and R. Tedrake, Velocity-based stability margins for fast bipedal who fall, Int. Rehab.. Med., vol. 2, no. 4, pp. 177180, 1980.
walking, presented at the 1st Ruperto Carola Symp. Fast Motions [23] J. M. Hausdorff, H. K. Edelberg, S. L. Mitchell, A. L. Goldberger, and
Biomech. Robot.: Optim. Feedback Control, Heidelberg, Germany, 2005. J. Y. Wei, Increased gait unsteadiness in community-dwelling elderly
[2] P. B. Wieber, On the stability of walking systems, presented at the Int. fallers, Arch. Phys. Med. Rehabil., vol. 78, no. 3, pp. 278283, 1997.
Workshop Humanoid Hum. Friendly Robot., 2002. [24] B. E. Maki, Gait changes in older adults: Predictors of falls or indicators
[3] M. Vukobratovic, A. Frank, and D. Juricic, On the stability of biped of fear?, J. Am. Geriatr. Soc., vol. 45, no. 3, pp. 313320, 1997.
locomotion, IEEE Trans. Biomed. Eng., vol. 17, no. 1, pp. 2536, Jan. [25] J. M. Hausdorff, Gait variability: Methods, modeling and meaning, J.
1970. Neuroeng. Rehab., vol. 2, no. 19, 2005.
[4] Y. Hurmuzlu and G. D. Moskowitz, Role of impact in the stability of [26] M. S. Garcia, A. Chatterjee, A. Ruina, and M. J. Coleman, The simplest
bipedal locomotion, Int. J. Dyn. Stab. Syst., vol. 1, no. 3, pp. 217234, walking model: Stability, complexity, and scaling, ASME J. Biomech.
1986. Eng., vol. 120, no. 2, pp. 281288, 1998.
[5] T. McGeer, Passive dynamic walking, Int. J. Rob. Res., vol. 9, no. 2, [27] H. Miura and I. Shimoyama, Dynamic walk of a biped, Int. J. Robot.
pp. 6282, 1990. Res., vol. 3, no. 2, pp. 6074, 1984.
[6] E. Garcia and P. A. G. de Santos, New dynamic energy stability margin [28] S. Kajita and K. Tani, Study of dynamic biped locomotion on rugged
for walking machines, in Proc. Int. Conf. Adv. Robot., Coimbra, Portugal, terrainDerivation and application of the linear inverted pendulum
2003, pp. 10141019. mode, in Proc. Int. Conf. Robot. Autom., 1991, vol. 2, pp. 14051410.
[7] C. Chevallereau, G. Abba, Y. Aoustin, F. Plestan, E. R. Westervelt, [29] A. Goswami, B. Thuilot, and B. Espiau, A study of the passive gait of
C. C.-D. Wit, and J. W. Grizzle, Rabbit: A testbed for advanced con- a compass-like biped robot: Symmetry and chaos, Int. J. Robot. Res.,
trol theory, IEEE Control Syst. Mag., vol. 23, no. 5, pp. 5779, Oct. vol. 17, no. 12, pp. 12821301, 1998.
2003. [30] A. D. Kuo, Energetics of actively powered locomotion using the simplest
[8] M. Vukobratovic and B. Borovac, Zero-moment pointThirty years of walking model, ASME J. Biomech. Eng., vol. 124, no. 2, pp. 113120,
its life, Int. J. Human. Robot., vol. 1, no. 1, pp. 157174, 2004. 2002.
[9] E. B. Lee and L. Markus, Foundations of Optimal Control Theory. New [31] S. H. Collins, M. Wisse, and A. Ruina, A 3-D passive-dynamic walking
York: Wiley, 1967. robot with two legs and knees, Int. J. Robot. Res., vol. 20, no. 7, pp. 607
[10] A. Goswami, Foot rotation indicator (FRI) point: A new gait planning 615, 2001.
tool to evaluate postural stability of biped robots, in Proc. Int. Conf.
Robot. Autom., 1999, pp. 4752.
[11] S. H. Collins, A. Ruina, R. Tedrake, and M. Wisse, Efficient bipedal Daan G. E. Hobbelen received the M.Sc. degree
robots based on passive-dynamic walkers, Science, vol. 307, no. 5712, in mechanical engineering from Delft University of
pp. 10821085, 2005. Technology, Delft, The Netherlands, in 2003, where
[12] S. Kajita and K. Tani, Experimental study of biped dynamic walking, he is currently working toward the Ph.D. degree.
IEEE Control Syst. Mag., vol. 16, no. 1, pp. 1319, Feb. 1996. His current research interests include bio-
[13] K. Ono, R. Takahashi, and T. Shimada, Self-excited walking of a biped inspired robotics, legged locomotion, mechatronics,
mechanism, Int. J. Robot. Res., vol. 20, no. 12, pp. 953966, 2001. (multibody) dynamics and (complaint) control.
[14] T. Takuma, K. Hosoda, and M. Asada, Walking stabilization of biped
with pneumatic actuators against terrain changes, in Proc. Int. Conf.
Intell. Rob. Syst., 2005, pp. 27752780.
[15] R. Q. Linde, Passive bipedal walking with phasic muscle contraction,
Biol. Cybern., vol. 81, pp. 227237, 1999.
[16] M. Wisse and J. v. Frankenhuyzen, Design and construction of Mike:
A 2D autonomous biped based on passive dynamic walking, in Proc. Martijn Wisse (M02) received the M.S. and Ph.D.
2nd Int. Symp. Adapt. Motion Animals Mach., Kyoto, Japan, 2003, degrees in mechanical engineering from Delft Uni-
pp. 143154 versity of Technology, The Netherlands, in 2000 and
[17] S. H. Strogatz, Nonlinear Dynamics and Chaos. Cambridge, MA: 2004, respectively.
Westview, 2000. He is currently an Assistant Professor of Hu-
[18] R. Q. van der Linde, Bipedal walking with active springs: Gait synthesis manoid Robotics, Delft University of Technology.
and prototype design, Ph.D. dissertation, Delft Univ. Technol., Delft, The
Netherlands, 2001.
[19] A. L. Schwab and M. Wisse, Basin of attraction of the simplest walking
model, in Proc. ASME Des. Eng. Tech. Conf., Pittsburgh, PA, 2001, p. 9.