Robustness: a new SLIP model based criterion
for gait transitions in bipedal locomotion
Harold Roberto Martinez Salazar1, Juan Pablo Carbajal2, and
Yuri P. Ivanenko3
1Artificial Intelligence Laboratory, Department of Informatics,
University of Zurich, Switzerland. martinez@ifi.uzh.ch
2Department of Electronics and Information Systems, Ghent
University, Belgium. juanpablo.carbajal@ugent.be
3Laboratory of Neuromotor Physiology, Fondazione Santa
Lucia, Italy. y.ivanenko@hsantalucia.it
October 26, 2021
Abstract
Bipedal locomotion is a phenomenon that still eludes a fundamen-
tal and concise mathematical understanding. Conceptual models that
capture some relevant aspects of the process exist but their full ex-
planatory power is not yet exhausted. In the current study, we intro-
duce the robustness criterion which defines the conditions for stable
1
arXiv:1403.0879v1  [cs.RO]  4 Mar 2014
locomotion when steps are taken with imprecise angle of attack. In-
tuitively, the necessity of a higher precision indicates the difficulty to
continue moving with a given gait. We show that the spring-loaded
inverted pendulum model, under the robustness criterion, is consis-
tent with previously reported findings on attentional demand during
human locomotion. This criterion allows transitions between running
and walking, many of which conserve forward speed. Simulations of
transitions predict Froude numbers below the ones observed in hu-
mans, nevertheless the model satisfactorily reproduces several biome-
chanical indicators such as hip excursion, gait duty factor and vertical
ground reaction force profiles. Furthermore, we identify reversible ro-
bust walk-run transitions, which allow the system to execute a robust
version of the hopping gait. These findings foster the spring-loaded
inverted pendulum model as the unifying framework for the under-
standing of bipedal locomotion.
Keywords
SLIP model, gait transitions, bipedal locomotion, human loco-
motion, biomechanics
1
Introduction
The study of bipedal locomotion has motivated the development of several
models that explain the most important principles governing the dynamics
of the observed gaits. Some researchers have adopted models that include
detailed representations of different leg components or that emulate neuro-
muscular structures using physical elements such as springs, dampers and
2
multi-segmented legs.
Although these models reproduce the dynamics of
locomotion, their use as conceptual models is not widespread due to their
complexity. In contrast, simpler models have been used extensively as con-
ceptual models of bipedal locomotion [1].
Most of these simple models were developed to explain the exchange of
kinetic and potential energy of the center of mass (CoM) of biological agents.
During walking, kinetic and potential energy of the CoM are out of phase, i.e.
the maximum height of the CoM corresponds with a minimum of its speed [2].
In consequence, the inverted pendulum (IP) model [3] is frequently used to
represent walking, since in this model the exchanges of energy are also out of
phase. Detailed analyses of the passive dynamics of the IP model constituted
a conceptual cornerstone for the development of mechanical devices capable
of stable walking without any actuators or controllers [4]. Despite its concep-
tual explanatory power, the IP model does not correctly reproduce several
aspects of human walking [5], e.g. the vertical oscillations of the CoM experi-
mentally observed are smaller than the ones predicted by the model. Inspired
in this model Srinivasan and Ruina proposed a biped model with ideal ac-
tuators on the legs [6]. They determined the periodic gaits that minimized
the work cost assuming that the leg forces are unbounded if necessary. They
found that transitions from walking to running at constant Froude number
and step length are possible only when the Froude number is one. As a re-
sult, they found an optimal walking gait that resembles the conditions of the
walking gait at human walk to run transition, but at this condition they did
not found an optimal running gait. In contrast, they identified a hybrid gait
called pendular running which is not supported with the experimental data
3
of human gait transitions. Further more, in this study the double support
phase in walking was not allowed.
Running is commonly represented with another model, the spring-loaded
inverted pendulum (SLIP) [7]. The SLIP model consist of a point mass (the
body) attached to a massless spring (the leg). During the stance phase the
spring is fixed to the ground via an ideal revolute joint that is removed during
flight phase. This model has been successfully used for the control of running
machines [8]. In terms of combining multiple gaits, the explanatory power
of the SLIP model surpasses that of the IP model, since the former can be
extended to reproduce the mechanics of human walking by adding an extra
massless spring representing the second leg, therefore unifying walking and
runnig in a single model. However, the analyses carried out with the SLIP
model had not yet explained gait transitions at constant forward speed, e.g.
from walking to running at a characteristic Froude number. Previous studies
suggested that transitions were only possible if the total energy was drasti-
cally increased or decreased to induce a considerable change in the forward
speed of the system [9]. With a simulation study [10], Srinivasan explained
gait transitions for springless bipeds model as a mechanism to minimize the
energetic cost of the locomotion.
However, in the case of springy biped
systems the walk to run transition is not predicted by work minimization
because for a certain range of stiffness it is possible to find work-free running
at very low speeds.
Given that the legs in the SLIP model are massless, their swinging motion
cannot be directly described using equations derived from Newton’s laws.
Therefore, a control policy that sets the angle of attack at touchdown (the
4
angle spanned by the landing leg and the horizontal at the time the foot
collides with the ground) must be defined a priori. Generally, the angle of
attack at touchdown is kept constant. Herein, we assume a more general
control policy: the system selects a new angle of attack at each step. The
study of the system is based on a return map. With the return map, we
can understand the evolution of the dynamical system as a function of the
selection of the gait and the angle of attack. This analysis is similar to [11,
12, 13], but in our study we define the return map at midstance. With this
analysis, we can identify the initial conditions that, under this control policy,
can perform a gait indefinitely. Instead of adding perturbations to the terrain
to measure the robustness of the system as in [14], we extended the concept
of viability introduced in [15], and assume that all the initial conditions with
a valid control policy must be able to select an angle of attack inside a range
of an arbitrary minimum size. We considered the length of the range of valid
angles of attack as a qualitative measure of the robustness. The regions in
which this control policy is valid are called robust regions, and regions where
the system can change from one gait to another are called transition regions.
In this study, we propose this definition of robustness as a criterion to
explain the onset of gait transitions, complementing the classical energetic
criterion [16, 17]. Intuitively, the robustness of a gait can be understood as
inversely related to the attentional demand required to maintain it. If highly
precise inputs are needed to continue with a gait the system must spend
more resources to select an adequate action, e.g. use of detailed models,
better estimation of states from noise sensory data, more processing time;
i.e. cognitive load or attention.
5
This new perspective is accompanied with a trade-offbetween robust-
ness and energetic cost. A similar trade-offhave been observed in bees [18]:
when flying in turbulent flows, the animal extends its lower limbs reduc-
ing the chances of rolling, but increases the drag force sacrificing forward
speed. Furthermore, the transitions found under the newly included robust-
ness criterion qualitatively reproduce experimental values of the changes in
the amplitude of the oscillations of the hip, changes in the gait duty factor
and variations of ground reaction forces. Incidentally, these transitions use
a gait pattern that we identify with hopping.
This paper is organized as follows.
In section 2, we define the mod-
els used for the simulation and introduce several concepts required for the
understanding of the results.
In section 3 we show the regions of robust
locomotion and gait transition. In that section we also compare our results
with biological data. Discussions are given in section 4 and we conclude the
paper in section 5.
2
Definitions
The time evolution of a gait is segmented in several phases, each phase is de-
scribed with a sub-model. These sub-models represent the motion of a point
mass under the influence of: only gravity (flight phase), gravity and a linear
spring (single stance phase), gravity and two linear springs (double stance
phase). The point mass stands for the body of the agent and the massless
linear springs model the forces from the legs. During walking, running and
hopping the system always goes through the single stance phase, therefore all
6
gaits can be studied and compared during this phase. We denote the maps
defined by walking, running and hopping as W, R and H, respectively. Given
an initial state xi of the model, a walking step taken with angle of attack
α is denoted xi+1 = Wα(xi) and similarly for running. As explained later a
step of the hopping gait requires two angles, therefore it can be denoted with
xi+1 = Hαβ(xi).
The state of the system is observed when its continuous trajectory passes
through a section, called S. This section is defined by the support leg forming
a right angle with the ground. At this section the state of the system is
defined by the height of the hip (i.e. height of the CoM), r, and the velocity
in the vertical direction, vy (see Appendix A for more details).
All initial conditions are given in the S section and in the single stance
phase, i.e. only one leg touching the ground and oriented vertically. (r, vy)
pairs were simulated for values of the total energy E in the range [780, 900]J
at intervals of 10 J. The model was implemented is in MATLAB(2009, The
MathWorks) and simulations were run using the step variable integrator
ode45. Experimental data analysis was performed using GNU Octave.
2.1
Viability, Robustness, symmetric gaits and biome-
chanical observables
Viability, as presented in [15], defines the easiness of taking a further step
during locomotion. That is, the wider the range of angles of attack that can
be used to take a step the easier is to take that step. In a physical platform
it is required that a valid angle of attack exists for a definite interval, since
7
real sensors and actuators have a finite resolution and are affected by noise.
A viability region in the section S contains all the states for which at least
one step can be taken selecting an angle of attack from an interval of at least
∆α, i.e. states for which if at least one iteration of the gait is applied map
into states of the same gait. For example, for the running gait, this can be
expressed as,
V R (∆α) ={x| x ∈S∧
(∃α ∈Iα, ∥Iα∥≥∆α | y = Rα (x) , y ∈S)}.
(1)
Where Iα stands for the angle interval and ∥Iα∥for its size. Narrower angle
intervals, i.e. more precise angle definition, lead to bigger viability regions
and wider intervals to smaller regions. An example of the viability regions
can be found in appendix A.
The concept of robustness is defined on top of that of viability. A state in
the robust region is a viable state that can always be mapped into the robust
region by choosing the appropriate angle of attack. This angle should be
viable, i.e. it must be selected from an interval of at least ∆α. For example,
for the walking gait, this can be expressed as,
ρW (∆α) ={x| x ∈ρW (∆α) ∧
∃α ∈Iα, ∥Iα∥≥∆α | y = Wα (x) , y ∈ρW (∆α)


}.
(2)
Where Iα stands for the angle interval and ∥Iα∥for its size. This assumes
that the controller can select an angle of attack for each step. In particular,
this includes constant angle of attack policies and some of the self-stable
8
regions identified in [9] belong to a robust region. However, this does not
mean that the system remains in the self-stable region for each step, since
that would imply that the angle of attack is selected precisely.
Instead,
robustness implies that if the system was in that region at time t, it can
remain close to it, even if the angles are selected with finite resolution.
The gaits commonly used by humans are symmetric, meaning that the
dynamical behavior of the left leg mirrors the one of the right leg. In our
model this is possible when two conditions are satisfied: the velocity in the
vertical direction at S is zero and there is an angle of attack α that can bring
the system back to the same state.
In the subsequent section we will show that the discovery of robustness as
a useful criterion to induce gait transitions allows for qualitative comparisons
with experimental biomechanical data. In particular we present results in
terms of Froude number, hip excursion, gait duty factor, and vertical ground
reaction forces. The Froude number is the ratio between the weight and the
centripetal force w2lo/g, where g is the acceleration due to gravity, lo is the
natural length of the leg and w is the angular velocity of the body around
the foot in contact with the ground. Hip excursion denotes the amplitude of
vertical oscillations of the hip. The gait duty factor is the fraction of the total
duration of a gait cycle in which a given foot is on the ground. The vertical
ground reaction force is vertical component of the normal force exerted by
the ground.
9
3
Results
We report the results obtained from the study of gait transitions in the SLIP
model following the criterion of robustness detailed in Section 2.1. It turns
out that the concept of robust gaits offer an alternative explanation for the
onset of gait transitions in bipedal locomotion, comparable with arguments
based on metabolic costs.
We begin our exposition with a detailed explanation of the conditions,
in terms of decrease of robustness, that may trigger gait transitions. From
there we move on to describe the mechanism underlying robust gait transi-
tions. The results of those two sections are combined to present qualitative
comparison with biomechanical observables, followed by a short description
of robust hopping.
The definition of robust gait applies for symmetric and non-symmetric
gaits. Figure 1a shows the area of the robust regions in the section S for dif-
ferent energies and different interval lengths ∆α. With this model we identify
three different gaits: running, walking and grounded running. Grounded run-
ning has the same phases as walking but in the transition from the single
support to the double support the vertical velocity of the center of mass is
positive while in walking the velocity is negative (Appendix A). Results show
that the grounded running gait is less robust that walking and running. For
a ∆α bigger than 0.5◦, the grounded running gait covers less than 15% of
the initial conditions in the section S.
Figure 1b shows the area of the viable transitions to the robust regions
in the section S for different energies and different interval lengths ∆α. For
10
example, the viable transition to robust running considers the initial condi-
tions outside robust running that under walking or grounded running can be
brought to robust running in one step. Given that this transitions are viable
the angle of attack can be selected from an interval of length ∆α. A similar
condition is applied to calculate the viable transition to robust walking or
robust grounded running. For a ∆α bigger than 0.75◦, the viable transition
to robust grounded running gait covers less than 10% of the initial conditions
in the section S. Figure 1c shows the total area of robust regions and viable
transitions with and without grounded running. Results show that for a ∆α
bigger than 0.5◦grounded running does not cover different initial conditions
from walking and running.
Figure 1d shows the range of forward speed for robust running and walk-
ing at several energies and different interval lengths ∆α. Results show that
the length of the interval affects the maximum Froude number in the walking
gait. The bigger the ∆α, the lower the walking Froude number. In addition
considering an interval length lower than 1◦, robust walking exists only at low
locomotion energies, while running increases robustness for higher energies.
For an interval length bigger than 1◦walking walking is not possible in all
the low energy levels.
We can draw an analogy between the results of the system with an interval
length lower than 1◦and the experimental results reported in [19], where it
was shown that imposed fast walking required higher attention than running
at similar speeds.
Furthermore, normal switching between gaits did not
required high attentional demand.
11
3.1
Conditions for transitions
We studied the transitions for a robustness criterion of ∆α equal to 1◦because
this was the limit condition in which the results of attentional demand can be
qualitatively explained by the model. In addition we focused in the walking
and running gait given that grounded running does not provide new possible
states from the ones identified in robust walking and robust running (Fig. 1c).
All the possible states of the system in the section S lie in a hemispherical
region (see equations (15)-(21) of [15] and Appendix A). In Fig. 1e-g, we
marked the apex of this hemisphere with a star symbol.
The closer the
system is to the star, the higher the forward speed of the gait. Symmetric
gaits are marked with a solid line, all symmetric gaits have vy = 0. The
figure shows that symmetric robust walking moves away from the apex of the
hemisphere as energy increases, i.e. it becomes slower. At 830 J symmetric
robust walking is constrained to the rightmost side of the viability region
reducing the speed of this gait considerably. Furthermore, at this energy the
region of symmetric walking breaks down into two unconnected segments.
This is also evident in Fig. 1d where the maximum speed of symmetric robust
walking shows a strong slowdown with a sudden change of slope. The latter is
a consequence of the rupture of the symmetric gait region. This milestone in
the evolution of the gait can be used as a natural trigger for a gait transition.
The evolution of the area of robust walking, and robust running, are
shown in detail in Figure 1e-f. This figures show that, at low energy, robust
walking covers a wide region of the viable states of the system, while at high
energy robust running covers a wider area. Around 800 J both robust gaits
12
have similar area. Based on robustness alone, this will imply a transition.
However, symmetric robust walking intersects the apex of the hemisphere
producing the fastest forward speed up to energies of 810 J, favoring walking
in terms of energy efficiency. When the energy is increased further, the area of
robust walking decreases and symmetric robust walking is constrained to low
speeds. Due to these facts, at energies close to 840 J, the speed of symmetric
robust walking and running match. For higher energies the gait transition is
imminent, since the only robust gait remaining is symmetric running.
3.2
Mechanism of gait transitions
Assuming that during locomotion the fastest robust gait patterns are pre-
ferred over slower or non-robust ones, we see that for energies below 840 J
walking is the gait of choice and for energies above that value running would
be chosen. Therefore, we study viable transitions at 840 J and compare them
with results from an experiment on human gait transition. We consider tran-
sitions only when all angles of attack used in the process can be chosen from
an interval of length 1◦or greater, i.e. we define admissible transitions using
the concept of viability (sec. 2.1).
We consider two mechanisms to execute gait transitions between symmet-
ric robust gaits (symmetric gaits are known to be self-stable and therefore
a good choice for stable locomotion, see [9]). The first mechanism, which
can only be used from walking to running, consist in moving from the robust
region of walking to the viability (non-robust) region of the same gait, and
from there select an angle of attack to go to the robust region of running.
13
This mechanism can be used in robust walking between 830 J and 840 J (see
Figure 2a). The second mechanism consist in going from a robust region of
a given gait (walking or running) directly to the robust region of a different
gait. This mechanism is applicable for robust running between 830 J and
840 J while in robust walking is only applicable around 840 J.
These mechanisms can be further constrained by selecting desired prop-
erties of the final gait. One possibility is to execute a transition in such a
way that the final gait has the same (or as close as possible) Froude number
as the initial gait. Another possibility is to execute a transition that sets
the hip excursion of the new gait to a desired value (see Figure 2b for a
graphical description). These constraints are referred in this study as strate-
gies and they are used for the comparison between our simulated results and
experimental data presented in the next section.
3.3
Qualitative Prediction of Biomechanical Observ-
ables
As we mention before, the biomechanical observables used to compare our
results with experimental data are: Froude number, hip excursion, gait duty
factor and vertical ground reaction forces. In the Appendix B-C, we extended
this comparison to include angle of attack sequences and change of phase.
We compare all our simulations against the experimental data reported in
Figure 2 of [20], we will refer to this data as “experimental data” or “the
experiment”.
Figure 2a shows the transition regions at two energy levels. We painted
14
the robust regions of running and walking with a solid color, the shaded re-
gions inside these are transitions regions where the system can change the
gait.
The diagonal shading corresponds to regions where the system can
change between robust gaits (non-symmetric) in only one step. The horizon-
tal shading delimits the region where the system can go to the non-robust
transition region, as described in 3.2. The right panel shows examples of
a transition from walking to running and another from running to walking
using the two mechanisms mentioned in the previous section. For the first
transition, the system starts at symmetric robust walking (1), in the first
step it moves to the non-robust transition region (2*) and executes the tran-
sition to robust running (3*). With two further steps the system is able to
reach symmetric robust running (4-5). The transition in the other direction
starts at symmetric robust running (5). Then the system moves to the ro-
bust transition region (6*) from which, in a single step, it changes to robust
walking (7*). With two more steps the system reaches symmetric robust
walking (8-9). In both transitions, the hip excursion was kept as constant as
possible.
Figure 2b shows the Froude number and the hip excursion of all symmetric
robust gaits at 840 J. As indicated in the figure, vertical transitions keep the
hip excursion constant, while horizontal transitions produce gaits with the
same Froude number.
Figure 3 shows time series of hip excursion and duty factor for a transition
at constant hip excursion, together with a transition at constant Froude
number. In both situations we obtain a Froude number that is about 60%
smaller than the one found in human gait transitions, which is around of
15
0.5 [20].
Nevertheless the SLIP model provides the best Froude number
estimation to the date, when compared to other simple models, e.g. the IP
model.
Ground reaction forces prior to the transition from walking to running
have three main characteristics [21].
Firstly, they present an asymmetric
double bell-shaped profile. Secondly, the earlier peak becomes bigger than
the later one and, thirdly the depression between the peaks becomes more
accentuated in the last step of walking, exactly before the transition.
In
the case of the transition from running to walking, it was reported that
the vertical ground reaction forces decrease during the steps prior to the
transition.
In Figure 4 we have plotted the vertical ground reaction forces for three
different simulated examples. The first row of panels shows transitions from
walking to running, and the second row of panels shows transitions in the
other direction.
Panels (a) and (b) show transitions keeping the Froude
number constant. Panels (c) and (d) show transitions at constant hip excur-
sion. The last example, presented in the panels (e) and (f), shows transitions
that match the change in amplitude that was observed in the experiment.
All cases qualitatively match the characteristics of the ground reactions re-
ported in [21]. The decrement in the force of the last running step is due to
the support of the second foot. A reduction of the peak in more than one
step appears only on the case where we matched the hip excursion of the
experimental data.
In Table 1, we present a summary of the comparison between the simu-
lated examples and the experimental data. Each column is discussed next.
16
Strategy
# Steps
vx
∆r
Fy
∆α
∆φ
Const. Froude number






Const. hip excursion






Fitting experiment






Table 1:
Comparison between three transition strategies and experimental data.
The symbol  indicates qualitative matching between simulation and experiment,
while the symbol  indicates the opposite. vx: forward speed of the center of
mass; ∆r: relative change in hip excursion before and after transition; Fy: vertical
ground reaction forces; ∆α: change of the angle of attack during transition; ∆φ:
change in phase of the oscillations of the hip before and after transition.
Due to the variety of transitions that can be generated with the model, the
number of steps to execute them can be select in a wide range, at least from
3 to 8 steps. From Figure 2b we can see that the Froude number of all these
transitions are lower than 0.5, this reflects the fact that the simulations have
lower forward speeds (vx) than the observed in humans.
As pointed be-
fore, the many transitions that can be simulated, permit the matching of
the relative change in hip excursion (∆r) measured in the experiment. In
all simulated transitions the vertical ground reaction forces (Fy) are qualita-
tively well reproduced. The selection of the angle of attack are qualitative
similar to what we found in the experimental case: the system moves pro-
gressively from one gait to the other changing the angle of attack at each
step. However, the oscillation of the hip before and after the simulated tran-
sitions presents a change of phase (∆φ) that not always coincide with what
is observed in reality. Details for these two observables are presented in the
the Appendix B-C.
17
3.4
Robust Hopping Gait
At 840 J we identify a transition region in robust walking where the system
can go in one step to robust running. Among the states in this transition
region, there a some that are mapped directly into the transition region of
robust running.
By selecting alternatively the right angles of attack, the
system can sequentially walk and run, producing the hopping gait. Fig. 5
shows an example of this gait. By looking at the vertical ground reaction
forces in the figure, we see the different phases that compose this gait; from
single stance phase to double stance phase then to single stance phase and
finally to flight phase.
4
Discussion
Herein we have modeled bipedal locomotion using the SLIP model. This
model conserves the total mechanical energy and at first glance it may seem
inapposite for the prediction of gait transitions, since work has to be done on
the system to increase the speed of locomotion. Nevertheless, by looking at
the behavior of the model at different energies, we can emulate the situation
where work is done on the system.
We proposed robustness as a new measure of the easiness of locomotion.
Robustness measures the level of attention that needs to be dedicated to take
a step; the more robust a gait is, the less attention that is needed to take the
next step.
According to our results, the selection of the gait can be based on two
criteria: efficiency, which is the selection of the gait with the highest forward
18
speed; and robustness, which defines how easy is to maintain the given gait.
This second criterion is consistent with the experimental results of attentional
demand in locomotion reported in [19]. Based on these criteria, walking is
the best choice for energies below 840J, and running is more appropriate for
higher energies. This resembles what is observed in human locomotion.
Using robustness as the leading criterion, we identify transition regions
that allow the system to go from one gait to the other even in the case of
imprecise angle selection. These transition regions are present for energies
from 830 J to 840 J (Fig. 2a). At 840 J, symmetric robust running and walk-
ing share all the possible velocities, facilitating gait transitions. In the case
of an increment of energy, to keep robustness and move forward faster, a
walking system can execute a transition to robust running at 840 J. The
transition can be reversed when the system decreases its energy. Note that
the mechanisms of transition shown in Fig. 2a (right panel), have the fol-
lowing properties. One mechanism connects the robust region of both gaits,
while the other one connects the non-robust viability region of walking with
robust running. The latter mechanism is not reversible, meaning that the
system cannot go from running back to this region in a single step. The
transitions connecting robust regions are reversible and the system can os-
cillate between the two gaits robustly. Is in this situation where the hopping
gait emerges. This locomotion pattern is frequently used by children when
playing joyfully.
The existence of non-empty transition regions (Fig. 2b) implies that the
system has multiple alternatives to change gaits.
These alternatives will
produce different changes of forward speed and hip excursion.
We show
19
three different scenarios: constant hip excursion, hip excursion similar to
experimental data and constant Froude number.
When the transition matches the hip excursion of the experimental data,
the Froude number varies from 0.16 in walking to 0.08 in running, while in
the experiment it is almost constant (slowly varying treadmill speed, see [20]
for details on the experiment). As explained before, in all simulated cases
the absolute values of Froude number are lower than in the experiments. The
hip excursion has an amplitude of 5.2 cm in walking and 8.3 cm which also
similar to the one reported in [20] which is around 7 cm.
When the transition keeps the Froude number constant the hip excursion
decreases from 5.7 cm in walking to 3.7 cm in running. This contradicts the
behavior observed in our experimental data. The simulated Froude number
for this transition is about 0.17.
The robustness criterion induces an underestimation of the forward speed
at gait transitions. The highest Froude number achieved using the previous
strategies is around one third of the one observed in humans (0.5). However,
given the strong simplifications in the model the result is encouraging. To
reduce the gap between simulated and experimental Froude number, the
model can be extended to include the displacement of the point where the
leg is in contact with the ground during the stance phase [22].
All transitions presented here produce similar results concerning the duty
factor. Walking has a duty factor around 0.7 and running has a duty factor
around 0.4, in accordance with the experiment. Furthermore, in all tran-
sitions from walking to running the model predicts a progressive change in
the vertical component of the reaction forces, i.e. the relation between the
20
first and the second peak of the force during the transition. This also applies
to the transitions from running to walking. In particular, the ground reac-
tion forces corresponding to transitions matching the hip excursion of the
experimental data (Fig. 4) introduces a progressive reduction of the force
peak in more than one step. All these results qualitatively reproduce the
experimental results reported in [21].
5
Conclusion
The comparison between experimental data and simulations using the SLIP
model shows that the model is not able to generate accurate quantitative
predictions. Most strikingly, the forward speed in the simulations are con-
siderable slower than that observed experimentally. This difficulty can be
overcome by adding a more detailed description of the contact between leg
and ground. Nevertheless, the SLIP model can be used as a conceptual model
to explain the many aspects of bipedal locomotion such as the mechanics of
running, walking, hopping and gait transitions.
Our findings indicate that robustness can play an important role in induc-
ing gait transition, complementing the usual view focused solely in energy
expenditure. The robustness criterion is analogous to the attentional de-
mand during locomotion and may play an important role deciding the gait
transition events. To our knowledge this is the first time such a criterion is
included in a numerical model of locomotion.
21
Acknowledgements
The research leading to these results has received
funding from the European Community’s Seventh Framework Programme
FP7/2007-2013-Challenge 2-Cognitive Systems, Interaction, Robotics- under
grant agreement No 248311-AMARSi.
Authors contribution
HMS developed the computational and mathe-
matical model, run the simulations and performed data analysis. JPC col-
laborated in development of the mathematical models, the data analysis and
the interpretation of results. YI collected and contributed the experimental
data. All authors contributed to the writing of this manuscript.
References
[1] P. Holmes, R. J. Full, D. Koditschek, and J. Guckenheimer. The dy-
namics of legged locomotion: Models, analyses, and challenges. SIAM
Rev., 48(2):207–304, 2006.
[2] G. A. Cavagna, N. C. Heglund, and C. R. Taylor. Mechanical work
in terrestrial locomotion: two basic mechanisms for minimizing energy
expenditure. Am. J. Physiol.-Reg. I., 233(5):R243–R261, 1977.
[3] S. Mochon and T. A. McMahon.
Ballistic walking.
J. Biomech.,
13(1):49–57, 1980.
[4] S. H. Collins. A three-dimensional passive-dynamic walking robot with
two legs and knees. Int. J. Robot. Res., 20(7):607–615, July 2001.
22
[5] R.J. Full and D.E. Koditschek. Templates and anchors: neuromechani-
cal hypotheses of legged locomotion on land. J. Exp. Biol., 202(23):3325–
3332, 1999.
[6] Manoj Srinivasan and Andy Ruina. Computer optimization of a minimal
biped model discovers walking and running. Nature, 439(7072):72–75,
2005.
[7] R. Blickhan.
The spring-mass model for running and hopping.
J.
Biomech., 22(1112):1217 – 1227, 1989.
[8] B. Andrews, B. Miller, J. Schmitt, and J.E. Clark. Running over un-
known rough terrain with a one-legged planar robot. Bioinspir Biomim,
6(2):026009, 2011.
[9] H. Geyer, A. Seyfarth, and R. Blickhan. Compliant leg behaviour ex-
plains basic dynamics of walking and running. P. Roy. Soc. B - Biol.
Sci., 273(1603):2861–7, November 2006.
[10] Manoj Srinivasan.
Fifteen observations on the structure of energy-
minimizing gaits in many simple biped models. Journal of The Royal
Society Interface, 8(54):74–98, 2011.
[11] M Ernst, H Geyer, and R Blickhan. Extension and customization of
self-stability control in compliant legged systems.
Bioinspiration &
Biomimetics, 7(4):046002, 2012.
[12] Michael Ernst, Hartmut Geyer, and Reinhard Blickhan. Spring-legged
locomotion on uneven ground: a control approach to keep the running
23
speed constant. In International Conference on Climbing and Walking
Robots (CLAWAR), pages 639–644. World Scientific, 2009.
[13] Andre Seyfarth and Hartmut Geyer.
Natural control of spring-
like running-optimized self-stabilization.
In Proceedings of the Fifth
International Conference on Climbing and Walking Robots (CLAWAR
2002), Professional Engineering Publishing Limited, London, pages 81–
85, 2002.
[14] Katie Byl and Russ Tedrake.
Metastable walking machines.
The
International Journal of Robotics Research, 28(8):1040–1064, 2009.
[15] Harold Roberto Martinez Salazar and Juan Pablo Carbajal. Exploiting
the passive dynamics of a compliant leg to develop gait transitions. Phys.
Rev. E, 83(6):066707, Jun 2011.
[16] R. M. Alexander. Optimization and gaits in the locomotion of verte-
brates. Physiol. Rev., 69(4):1199–1227, October 1989.
[17] A. E. Minetti, L. P. Ardigo, and F. Saibene. The transition between
walking and running in humans: metabolic and mechanical aspects at
different gradients. Acta Physiol. Scand., 150(3):315–323, 1994.
[18] Stacey A Combes and Robert Dudley. Turbulence-driven instabilities
limit insect flight performance. P. Natl. Acad. Sci.-Biol., 106(22):9105–8,
June 2009.
24
[19] Bruce Abernethy, Alastair Hanna, and Annaliese Plooy. The attentional
demands of preferred and non-preferred gait patterns. Gait Posture,
15(3):256 – 265, 2002.
[20] Yuri P Ivanenko, Francesca Sylos Labini, Germana Cappellini, Velio
Macellari, Joseph McIntyre, and Francesco Lacquaniti. Gait transitions
in simulated reduced gravity. J. Appl. Physiol., 110(3):781–8, March
2011.
[21] Li Li and Joseph Hamill. Characteristics of the vertical ground reaction
force component prior to gait transition. RQES, 73(3):229–237, 2002.
[22] Peter G. Adamczyk, Steven H. Collins, and Arthur D. Kuo. The advan-
tages of a rolling foot in human walking. J. Exp. Biol., 209(20):3953–
3963, 2006.
[23] J. W. Eaton.
GNU Octave Manual.
Network Theory Limited,
http://www.octave.org, 2002.
A
Equations of motion
We define a running gait as a trajectory that switches from the single stance
phase to the flight phase and back to the single stance phase. A walking gait
is defined as a trajectory that switches from the singles stance phase to the
double stance phase and back again to the single stance phase.
The state in the flight phase is represented in Cartesian coordinates of
25
the position of the point mass and its velocity ⃗Xff = (x, y, vx, vy)T,
˙⃗Xff =









vx
vy
0
−g









,
(3)
where g is the acceleration due to gravity.
The state in the single stance phase is represented in polar coordinates
⃗Xs =

r, θ, ˙r, ˙θ
T
, where r is the length of the spring and θ is the angle
spanned by the leg and the horizontal, growing in clockwise direction. Thus,
the equations of motion are:
˙⃗Xs =









˙r
˙θ
k
m (r0 −r) + r ˙θ2 −g sin θ
−1
r

2 ˙r ˙θ + g cos θ










.
(4)
It is important to note that θ(tTD) = α, i.e. the angular state at the time
of touchdown is equal to the angle of attack. The parameter r0 defines the
natural length of the spring.
In the double stance phase the state is also represented in polar coordi-
nates ⃗Xd =

r, θ, ˙r, ˙θ
T
, with the origin of coordinates in the new touchdown
26
point. The motion is described by:
˙⃗Xd =






















˙r
˙θ
k
m[(r0 −r) +

1 −r0
r♂

. . .
(x♂cos θ −r)] + r ˙θ2 . . .
−g sin θ
−1
r[ k
m

1 −r0
r♂

x♂sin θ . . .
+ 2 ˙r ˙θ + g cos θ]






















(5)
r♂=
p
r2 + x2
♂−2rx♂cos θ,
(6)
where x♂is the horizontal distance between the two contact points and r♂is
the length of the back leg.
The event functions are parameterized with the angle of attack and the
natural length of the springs.
Switches from the flight phase to the single stance phase are defined by:
Fff→s

⃗Xff, α, r0

:







y −r0 cos α = 0
vy < 0
,
(7)
which means that the mass is falling and the leg can be placed at its natural
length with angle of attack α. Therefore, the motion is now defined in the
single stance phase. The switch in the other directions is simply:
Fs→ff

⃗Xs, r0

: r −r0 = 0.
(8)
27
These are the only two event functions involved in the running gait. The
map from one phase to the other is defined by:
x = −r cos θ
y = r sin θ.
(9)
It is important to have in mind that the origin of the single stance phase is
always at the touchdown point.
For the walking gait, we have to consider switches between single and
double stance phases:
Fs→d

⃗Xs, α, r0

:







r sin θ −r0 cos α = 0
θ > π
2
,
(10)
which is similar to (7) with the additional condition that the mass is tilted
forward. Additionally, if we consider the sign of the vertical speed, we dif-
ferentiate between walking gait with vy ¡ 0 and Grounded Running gait with
vy ¿ 0.
The switch from the double stance phase to the single stance phase is
defined by:
Fd→s

⃗Xd, r0

: r♂−r0 = 0,
(11)
with r♂as defined in (6). The map from the double stance phase to the single
28
stance phase is the identity. In the other direction we have:
rd = r0
θd = α,
(12)
x♂= r0 cos α −rs cos θs,
(13)
where the subscripts indicate the corresponding phase.
If the system falls to the ground (y ≤0), attempts a forbidden transition
(e.g. double stance phase to flight phase), or renders vx < 0 (motion to the
left,“backwards”), we consider that the system fails.
The state of the model is observed when the trajectory of the system
intersects the section S defined in the single stance phase, i.e. only one leg
touching the ground and oriented vertically (Figure 6). The results are visu-
alized using the values of the length of the spring r and the radial component
of the velocity which, in S, equals the vertical speed vy (vx is obtained from
these values and the equation of constant energy). It is important to note
that all possible values of r, vy , and vx , for a given value of the total energy
E, lie on an ellipsoid.
E = 1
2k (r0 −r)2 + 1
2m
v2
x + v2
y


+ mgr
(14)
This intermittent observation of the system renders the continuous evo-
lution of the model into a mapping that transforms states in the section at
a time t, to states in the section at t + ∆t. The interval ∆t is the time the
system takes to reach a new vertical posture, only during periodic gaits it is
equivalent to the period of the gait.
29
Using the maps we calculated the viability regions in the section S. The
viability regions are the initial conditions that can perform an step selecting
an angle of attack from a continuous interval of length ∆α the biggest interval
size found with the system is 23◦. Figures 7-8 show different viable regions
as a function of the interval length.
B
Angle of attack estimation from empirical
data
In the experimental data of reference [20] the angle of the right limb is mea-
sure against the vertical. We use this information to estimate the angle of
the leg at landing based in two facts. First, the angle of the leg changes more
its velocity in the swing phase (the foot is not in contact with ground) than
in the support phase (the foot is in contact with the ground), and second,
as soon as the leg changes from the swing phase to the support phase there
is a big change of the angular velocity due to the impact of the food against
the ground when it lands.
The angle of attack identified using this conditions allow the comparison
of the strategy in human locomotion and the proposed model. The model
qualitatively develops a similar strategy. The difference of the angle of attack
between the steady state gait (e.g. walking or running) from the experiment
and the model is around five degrees. To facilitate the qualitative comparison
of the angle of attack, we evaluate the change of the angle of attack against
the angle of attack of walking. Using this measurement, we can avoid the
30
difference of five degrees and focus in the strategy for gait transition.
Fig. 9 and Fig. 10 show that the strategy developed with the model has
similar steps and matches the change of the angle of attack in the transition.
Fig. 9 shows a more drastic change of the angles of attack compare with the
experiment result, however the data of the experiment is from one leg which
allow the identification of the angle of attack every two steps. This can be
emulated with the model selecting only the even or the odd steps. In any of
these cases, the change of the angles of attack is going to look less drastic
and qualitatively more similar to the ones from the experiment.
C
Change of phase of hip excursion before
and after transition
Strategy
W →R
R →W
Const. Froude number
36.3◦
35.3◦
Const. hip excursion
55.3◦
51.5◦
Fitting experiment
109.0◦
110.9◦
Experiment
−35.0◦
86.8◦
Table 2: Change of phases for three strategies and experimental data. None of the
transitions shows a phase change in full accordance with the experimental data.
The absolute value of the phase change for the transition from walking to running
at constant Froude number is very close to the experimental value, however the
direction of the change is opposite.
As shown in Figure 11 (left axis), during walking and running the hip
follows and oscillatory trajectory over time. We compare the phase of these
oscillations with respect to the moment of transition. The moment of tran-
sition was identified as follows:
31
1. Calculate the analytic signal of the hip trajectory by means of the
Hilbert transform, e.g. hilbert function in GNU Octave’s signal pack-
age [23].
2. Obtain the phase of the signal from the angle of the analytic signal.
3. Take the time derivative of the phase, this is an approximation of the
frequency of the oscillations as a function of time.
4. Search for the highest peak in the frequency signal. This point separates
the regions of walking from the regions of running.
Figure 11 shows the frequency signal superimposed to the experimental data.
The transition point is indicated with a vertical arrow. Taking this point as
the origin of time, we calculate the initial phase of walking and the initial
phase of running, by means of fitting a first order polynomial to the phase
signal of each gait. This is shown in Figure 12 when applied to the exper-
imental data. The change of phase is calculated as the difference of these
initial phases normalized to the interval (−π, π]. The exact same analysis
was applied to all the signals, simulated and experimental.
The changes of phase for the three transition strategies presented in the
paper are summarized in Table 2. All the simulated examples are able to
match the direction of the change of phase in the running to walking tran-
sition. However, none of the transitions shows a phase change in full accor-
dance with the experimental data. The absolute value of the phase change
for the transition from walking to running at constant Froude number is
very close to the experimental value, however the direction of the change is
opposite.
32
Figure 1:
(Color online) Robust regions. For panels (a) - (d) the (copper) gray
color scale represents the interval size used to calculate the robust region. (a)
shows the robust region area in the section S for running (dashed line), walking
(continuous line), and grounded running (dash-dotted line). (b) shows the area of
viable transitions that brings the system to robust running (dashed line), robust
walking (continuous line), and robust grounded running (dash-dotted line) in the
section S. (c) shows the total area in the section S cover by the robust gaits
and the viable transitions. The dash-dotted line represents all the gaits, and the
continuous line represents walking and running.
(d) shows the maximum and
minimum Froude number for a robust gait at the section S for different energies.
Robust walking is depicted with the dashed line, and robust running is depicted
with the continuous line.
In panels (e) - (g) filled patches represents robust
running ((blue) light gray) and robust walking ((magenta) dark gray) in the section
S. The dashed region represents viable transition to robust running using walking
((blue) light gray), and to robust walking using running ((magenta) dark gray).
The solid black line depicts the symmetric gaits.
33
Figure 2:
(Color online) Viable transitions. In all panels (blue) light gray color
represents running and (magenta) dark gray color represents walking. (a) shows
viable transitions at two energy levels. Filled patches corresponds to robust re-
gions. Shaded regions inside these are viable transitions regions. Diagonal shading
corresponds to regions where the system can change between robust gaits (non-
symmetric) in only one step. The horizontal shading delimits the region where
the system can go to the non-robust transition region. The right panel shows two
transition using both mechanisms. See text for details. (b) shows the Froude
number versus hip excursion for symmetric robust running and walking at 840 J.
Arrows indicate: (1) constant hip excursion, (2) constant Froude number and (3)
relative change of the amplitude of the hip excursion fitted to experimental data.
0
2
4
6
8
10
12
14
0.35
0.45
0.55
0.65
0.75
-0.08
-0.04
0
0.04
0.08
Center of mass excursion [m]
Duty factor
X [m]
5
10
X [m]
15
-0.05
0
Center of mass excursion [m]
0.05
0.1
-0.1
0
0.5
0.6
Duty factor
0.7
0.8
0.4
a)
b)
Figure 3:
(Color online) Hip excursion and gait duty factor for transition at
constant hip excursion (a); and constant Froude number (b). The (blue) light
gray color represents the hip excursion and the black line represents the duty
factor. The plots show several steps before and after each transition.
34
Figure 4: Vertical ground reaction forces during transitions. The six panels show
a transition from symmetric robust walking to symmetric robust running with
three different strategies, (a)-(b) constant Froude number, (c)-(d) constant hip ex-
cursion, (e)-(f) hip excursion similar to the experimental data. The forces present
an asymmetric double bell-shaped profile. In the walking to running transition,
(a)-(c) and (e), the earlier peak becomes bigger than the later one, exactly before
the transition. The transitions in the other direction, running to walking (b)-(d)
and (f) show vertical ground reaction forces that decrease considerably in the last
running step due to the support of the second foot. The selection of a hip ex-
cursion similar to the experimental data introduces a progressive reduction of the
force peak in more than one step (f). All forces are normalized with respect to the
weight of the system.
35
Figure 5:
(Color online) Vertical ground reaction forces during hopping. Panel
(a) shows the transition regions in section S for E = 840 J; the arrows show the
states in the robust transition region that are used alternately. Panel (b) shows the
ground reaction forces for each leg. The (pink) gray rectangles show the different
flight phases. The forces from the legs are indicated with solid lines with different
colors.
Figure 6: (Color online) Illustration of the evolution of the SLIP model for running
and walking. The different phases are indicated as well as the section S where the
system is observed.
36
Figure 7: (Color online) Viability regions for running and walking. The (cooper)
gray scale color represents the viability regions for energies between [780J-810J].
The first column shows the viability region for running and the second column for
walking
37
Figure 8:
(Color online) Viability regions for walking and running.
The
(cooper) gray scale color represents the viability regions for energies between [820J-
880J].The first column shows the viability region for running and the second col-
umn for walking
38
Figure 9:
(Color online) Change of the angle of attack in the running to walk-
ing transition. The solid line represent the change of the angle of attack in the
model and the doted line represent the change of the angle of attack in a human
experiment. In both case there is a transition from running to walking.
Figure 10: (Color online) Change of the angle of attack against in the walking to
running transition. The solid line represent the change of the angle of attack in
the model and the doted line represent the angle of attack in a human experiment.
In both case there is a transition from walking to running.
39
Figure 11: (Color online) Transition point determination. Plot of the experimen-
tal data (left axis) and the the derivative of the phase signal (right axis). this
derivative gives a frequency signal that presents a peak during the transition that
is used to determine the transition point (vertical arrow).
Figure 12:
(Color online) Phase difference calculation. Taking the point of tran-
sition as the origin of time, the phase difference is calculate from the intercept of
linear fits applied to the two parts of the phase signal. Solid lines show the phase
signal for walking and running. Dashed lines show the linear fits.
40