Professional Documents
Culture Documents
1 Introduction
Animals have segmented legs with various geometric configurations depending on
their scale and the habitat in which they live. Large-sized mammalian animals have
upright posture to reduce the bending moments acting about their leg joints. Small
animals have more bent legs (crouched legs) because acceleration is crucial for their
survival even though this type of leg produces reduced effective mechanical advan-
tage [2]. Despite the rich biological precedent available, most analytical studies on
running have used the prismatic legs assumed in the Spring-Loaded Inverted Pen-
dulum (SLIP) model of running [1]. Only a few studies have been published that
represent running behavior by means of segmented-leg models [11]. Blickhan et al.
showed that the geometric configuration of a segmented leg influences the demand
for energy production, structural stability and velocity transmission from muscle
groups to the leg tip [2]. Rummel and Seyfarth showed that the two-segment-leg
model offers a much larger stability region (i.e., the range of the leg impact angles
that guarantee stable running) than the standard SLIP model could [11].
Jae Yun Jun · Duncan Haldane · Jonathan E. Clark
Dept. Mechanical Engineering, Florida State University, 2525 Pottsdamer Street Rm A229,
Tallahassee, Florida 32310
e-mail: {jaeyun,haldadu,clarkj}@eng.fsu.edu
m
l2
K −θ
β
l
ψ γ
y l1 F
leg
x α
(a) (b)
Fig. 1 (a) Half-circle leg employed on EduBot[9], a RHex-like robot. (b) Two-segment-leg
model to approximate the dynamics of running with a half-circle leg. The parameters and the
variables of the model are explained in the text.
Each stride is comprised of one stance phase and one flight phase. The stance
phase is described by two dimensional generalized coordinates, q = [ψ , θ ]T , and
T
generalized speeds, q̇ = ψ̇ , θ̇ (see Figure 1(b)). The equation of motion for stance
phase is
M(q)q̈ + N(q, q̇) + G(q) = 0 (1)
where M is the 2×2 inertia matrix, N is the 2×1 Coriolis force vector, and G is a 2×
1 gravitational force vector. All angles are defined to be positive counterclockwisely.
The motion of the Center of Mass(CoM) during flight phase is simulated by using
a simple ballistic model. The transition from stance phase to flight phase occurs
when the condition y ≤ lo sin α is triggered, where α is some predefined leg impact
angle with respect to the horizontal line, and y is the height of the CoM from the
ground. The transition from flight phase to stance phase occurs when the condition
l > lo is satisfied.
As shown by Rummel and Seyfarth in [11], we relate the leg torsional stiffness
to a reference stiffness defined as k10% = F10% /Δ l10% , where Δ l10% is a reference
leg compression at 10% of the effective leg rest length (i.e. 0.1lo), and F10% is the
leg force at the reference leg compression. Dimensionless reference stiffness (k̃10% )
are used in the rest of the presented work where k̃10% is defined as k10% lo /mg. The
relationship between the reference stiffness (k10% ) and the torsional stiffness (K) is
shown in [11].
The leg ground reaction force is directed from the leg contact point with the
ground to the center of mass (CoM), and its magnitude has the following expression
l K(βo − β )
Fleg = (2)
l1 l2 sin β
where l1 , l2 , l, K, βo , β are the length of the distal leg segment, the length of the
proximal leg segment, the effective leg length, the leg’s torsional stiffness, the leg’s
intersegmental rest angle, and the leg’s intersegmental angle, respectively. From
762 J.Y. Jun, D. Haldane, and J.E. Clark
the leg force calculation, the horizontal and vertical ground reaction forces can be
expressed as
Fx = −Fleg sin γ
(3)
Fy = Fleg cos γ
Δα ĚĞŐ Δα ĚĞŐ
βŽ ĚĞŐ
βŽ ĚĞŐ
;ĂͿ ;ďͿ
Δα ĚĞŐ Δα ĚĞŐ
βŽ ĚĞŐ
βŽ ĚĞŐ
;ĐͿ ;ĚͿ
Fig. 2 Range of the leg’s impact angles of the stable running gaits (in degrees) for various
ratios of the length of the leg’s proximal segment (l2 ) to the length of the leg’s distal segment
(l1 ) and for various leg’s intersegmental rest angles (βo ). (a) for k̃10% = 12.46, (b) for k̃10% =
21.50, (c) for k̃10% = 28.50, and (d) for k̃10% = 38.50.
From Figure 2, one can observe that the size of the range of the stable leg impact
angles is largest when l2 /l1 = 1 for all considered leg stiffness values and leg inter-
segmental rest angles. Three aspects can be observed as the leg stiffness is increased.
First, the stability region expands in both directions of the 2D-domain formed by the
ratio of the lengths of the leg segments and the leg intersegmental rest angle with the
leg stiffness increase. Second, the size of the stability region increases from about
9o for k̃10% = 12.46 to about 29o for k̃10% = 38.50. Third, the leg intersegmental rest
angle that corresponds to the maximum size of the stability region shifts from about
90o to 170o as the leg stiffness value is increased from k̃10% = 12.46 to k̃10% = 38.50.
Some preliminary conclusions could be stated from the obtained results. First, it
is desirable to design legs with the same lengths of the proximal and distal segments.
This suggests that the current curved leg design is not optimum in terms of gait
stability, suggesting the first leg design change: altering the l2 /l1 ratio from about
0.35 in the current design to l2 /l1 = 1.
Second, for a given leg stiffness, there exists a leg intersegmental rest angle βo for
which the size of the range of the stable leg impact angles is maximized. Third, the
size of the stability increases with the leg intersegmental rest angle when l2 /l1 =
1 and when an appropriate leg stiffness value is chosen. This suggests the second
design change, increasing the leg stiffness k̃10% , and increasing βo =90o to βo =150o.
764 J.Y. Jun, D. Haldane, and J.E. Clark
τ Hip
m
l2
−θ γ l
K, B β F
leg
ψ −
τ Hip
LJ l1 l
dž
(a) (b)
Fig. 3 (a) Torque-driven and damped two-segment-leg model. A torque actuator (τHip ) is
added at the hip to model the motor that actuates half-circle leg, and a torsional damper (B) is
incorporated at the intersegmental joint to model the energetic losses of the leg. (b) Bipedal
test robot with curved legs is attached to a boom located at the center of the track.
The resulting equations of motion are similar to Eq. 1 with the difference of the
addition of the torque actuator (τHip ) and the torsional damper (B). The states in the
energy-dissipative system are the same as those of the conservative system.
The new expression for the leg force is given in Eq. (4).
l K(βo − β ) − Bβ̇
Fleg = (4)
l1 l2 sin β
And the expressions for the horizontal and vertical ground reaction forces are given
in Eq. (5)
τ
Fx = −Fleg sin γ − Hip cos γ
τHipl (5)
Fy = Fleg cos γ − l sin γ
where Fleg is shown in Eq. 4, γ is the angle between the effective leg and the vertical
line, and τHip is the amount of torque applied by the torsional actuator located at the
hip.
Compliant Leg Shape, Reduced-Order Models and Dynamic Running 765
3.2 Controller
A periodic function known as Buehler Clock [12] (used by RHex-like robots) is
considered as a reference signal for controlling the leg motion. This function is a
periodic two-piecewise linear function with one steeper slope for the fast swing
phase and one flatter slope for the slow stance phase. Four parameters characterize
this periodic function: desired leg touch-down angle (γA ), desired leg lift-off angle,
(γB ), duration of the stance phase (tB ), and the stride period (T ) (see [6] for details).
The control input signal is generated by using a simple proportional-derivative (PD)
controller:
τHip = kp (γdes − γ ) + kd(γ̇des − γ̇ ) (6)
where kp and kd are the proportional and the derivative gains, γdes and γ̇des are the
desired leg angular position and speed dictated by the reference signal, and γ and γ̇
are the actual angular position and speed of the effective leg, respectively.
3.3 Optimizer
Each of the three considered legs (legs with (βo = 90o , l2 /l1 = 0.35), (βo = 90o ,
l2 /l1 = 1) and (βo = 150o , l2 /l1 = 1)) requires distinct controller parameter settings
for stable running, and these parameters need to be optimized independently, done
here by using a direct search method (Nelder-Mead algorithm), in order to compare
their respective running performances. Two optimization cost functions have been
considered: stability and efficiency. For stability, the norm-2 of the maximum eigen-
value of the Jacobian matrix of the linearized Poincaré map is used as the stability
measure (see Section 2.2). For efficiency, specific resistance is used. Specific resis-
tance is defined as P/mgv, where P is the average mechanical power expenditure
within a stride, m is the body mass, g is the gravitational acceleration, and v is the
average forward running speed over a stride.
The results of this optimization process are described and compared to the exper-
imental results in Section 4.3.
Three sets of legs are fabricated based on the simulation results and are listed in
Table 1.
Vinyl Ester resin with 2% MEKP activator was used to create the composite
legs. This styrene based resin offers more control over composite compliance than
traditional epoxy systems. By post-curing the styrene based composite, stiffness can
be controllably increased.
Two types of fiberglass were used in the reinforcing matrix in the compliant legs.
A very light 6781 weave comprised of S2 glass was applied in alternating layers
with a heavier S-glass double bias weave. A portion of the composite needed to be
rigid to modify the location of the characteristic pivot. To create a light weight, rigid
composite section, a coring material called Soric was used. Soric is comprised of
small hexagonal foam sections interspersed with a light binding matrix.
The original legs (leg A, Figure 4(a)) are comprised of three layers of S2-6781
alternated with three layers of double bias for a total of six layers. The resulting
composite was post cured at 80◦ C for thirty minutes. The leg B (Figure 4(b)) is
comprised of three layers of S2-6781 and two layers of double bias. Additionally,
stiffening Soric was applied to the layup to shift the characteristic pivot to the mid-
dle of the leg. The resulting composite was post cured at 80◦C for fifteen minutes.
Finally, the leg C (Figure 4(c)) was fabricated in multiple sections: a rigid attach-
ment to the motor mount, a flexural segment and a rigid rolling contact adapter.
These sections were then bonded together using a methacrylate adhesive. The flex-
ural segment was fabricated separately to avoid the mechanical imperfections that
can result when pieces with complex geometry are VARTM processed. The flexural
section is comprised of four layers of 3K T-300 carbon fiber applied in alternating
45◦ layers. The rigid motor mount attachment section is made by reinforcing the
flexural layup with Soric. The rolling contact adapter is comprised of ten layers of
3K T-300 carbon fiber. After the legs have been fabricated, a section of rubber tread
is applied to increase traction.
Both the original leg (leg A) and leg B have a stiffness of 1,100 N/m (k̃10% =
12.46), leg C has a stiffness of 2,516 N/m (k̃10% = 28.50). The stiffness of the legs
was verified using a customized MTS Insight material testing machine. A leg mount
is attached to the load cell; the other end of the leg is placed on a linear slide rail
to allow rolling contact. The MTS system compresses the leg by 10% of the rest
A) (b) βo = 90o , l2 /l1 = 1, k̃10% = 12.46 (leg B) (c) βo = 150o , l2 /l1 = 1, k̃10% = 28.50
(leg C).
768 J.Y. Jun, D. Haldane, and J.E. Clark
length, or 1.14 cm, and the (k̃10% )) value is calculated to verify that the leg is the
correct stiffness.
Table 3 Simulation results by running with the controller parameters optimized for gait
stability
Study Leg |λmax | vx [m/s] Mechanical Power [W] SR
1 A 0.1762 1.1603 8.0460 0.6890
2 B 0.3949 1.7947 6.6977 0.3814
3 C 0.0723 0.9899 5.4196 0.5440
Table 3 shows that the running with leg C is the most stable (lowest |λmax |), and
running with leg B is fastest (highest vx ) and most efficient (lowest specific resis-
tance). Figure 5 and Figure 6 show the simulation-to-experiment comparison re-
sults for running with leg A and leg B, respectively. Repeatable running experiments
could not be performed with leg C because the leg repeatedly fractured at the flexion
point of the leg. To overcome this, new leg design is required in the future. In addi-
tion, the PD controller used on the robot operates in voltage mode, and, therefore,
the gains employed for experimental running trials needed to be modified manually
because the gains optimized in the simulation are for torque mode (see Eq. 6).
Figure 5 and Figure 6 show six graphs respectively: (a) height of CoM (ex-
periment and simulation); (b) electric current drawn by right motor (experiment);
(c) ground reaction forces (experiment); (d) forward speed of CoM (experiment
and simulation); (e) hip torque applied for right leg motion (simulation); and (f)
ground reaction forces (simulation). The horizontal axis of all six graphs repre-
sents time normalized by the respective stride period (T = 0.2840s for Study 1 and
T = 0.2780s for Study 2).
For Figures 5(a), 5(d), 6(a) and 6(d), the normalized time 0 (or 1) corresponds
to the instant of time when leg maximum compression occurs. On the other hand,
for Figures 5(b), 5(c), 5(e), 5(f), 6(b), 6(c), 6(e) and 6(f), the normalized time 0 (or
1) corresponds to the instant of time when leg touch-down occurs. The figures are
represented in this way because the data collected from the current sensor and the
force sensor are not synchronized with the data collected from the boom encoders.
The time normalization process for the data collected from the boom encoders is
possible by detecting the instant of time when the height of CoM is the lowest. For
the data collected from the current sensor and the force sensor, this process is easier
by detecting the instant of time when the leg touches the ground.
Compliant Leg Shape, Reduced-Order Models and Dynamic Running 769
^ƚƵĚLJϭ;>ĞŐͿ
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ s'Z&
,'Z&
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ s'Z&
,'Z&
Fig. 5 Simulation-to-experiment comparison results for leg A (Study 1 shown in Table 3).
The horizontal axis of all the graphs is time normalized by the stride period (T = 0.2840s).
^ƚƵĚLJϮ;>ĞŐͿ
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ s'Z&
,'Z&
s'Z&
,'Z&
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ
Fig. 6 Simulation-to-experiment comparison results for leg B (Study 2 shown in Table 3).
The horizontal axis of all the graphs is time normalized by the stride period (T = 0.2780s).
770 J.Y. Jun, D. Haldane, and J.E. Clark
The running with leg A for Study 1 is experimentally observed to be very sta-
ble and repeatable. The height fluctuations of the CoM observed during experimen-
tal running match closely to those predicted by the simulation model (Figure 5(a)).
However, the forward velocity of the robot deviates substantially from the simula-
tion prediction (Figure 5(d)). It appears that this discrepancy occurs mainly because
the simulation model does not capture the foot slippage that occurs in actual running
experiments. Figure 5(b) shows that during about 30% of the stride period the right
motor does negative work. Afterwards, the electric current draw grows positively in
order to correct the error produced between the desired and actual leg trajectories.
As the leg moves towards the instant of the lift-off events, the drawn current gets
smaller because the error signal is reduced and reaches its minimum value during
flight phase. Figure 5(e) shows the amount of applied torque for the right leg motion
in simulation level. Mechanical power is considered in simulation, but in experi-
ments electric power is used instead. The amount of applied torque can be related to
the electric current drawn by the motor using motor constant specified by the motor
manufacturer. Figure 5(b) and Figure 5(e) show similar trends. Note that in simula-
tion no torque is applied during flight phase because the leg is considered massless
and is assumed to track the desired leg trajectory perfectly. Figure 5(c) and Figure
5(f) show that the peak measured vertical ground reaction force is much lower than
predicted from the simulation. One source of this difference is the rolling contact
motion caused by running with curved legs which the two-segment-leg model does
not capture (because the model assumes pin joint contact model). Rolling contact
motion during stance phase has been shown to reduce the peak vertical ground re-
action forces, and the larger the radius of the rolling foot curvature, the smaller the
peak vertical ground reaction force is obtained for both walking [13] and running [7].
Similar results are shown running with leg B for Study 2 (Figure 6). The average
forward speed for this study is larger than for Study 1 as predicted by the simula-
tion model (Table 3). The running motion with the optimal controller parameters
obtained for Study 2, however, was not as stable as running with Study 1 as one can
observe from Figure 5(a) and Figure 5(d).
Specific resistance values for Study 1 and 2 are estimated from the averaged
forward speeds and the estimated average electric powers shown in Table 4. These
results show that running with leg B is faster and more efficient than running with
leg A for these particular studies, and this result matches to the simulation results
shown in Table 3.
When the optimization cost function is changed from the gait stability measure
(|λmax |) to specific resistance (SR), neither leg A nor leg B could run stably with the
Compliant Leg Shape, Reduced-Order Models and Dynamic Running 771
Table 5 Simulation results by running with the controller parameters optimized for specific
resistance
Table 6 shows that running with leg B is again faster and more efficient than
running with leg A, despite the fact that both experimental runnings diverge greatly
from the simulation results. All simulation and experimental results shown in this
772 J.Y. Jun, D. Haldane, and J.E. Clark
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ
^ƚƵĚLJϰ;>ĞŐͿ
;ĂͿ ;ďͿ
^ŝŵƵůĂƚŝŽŶ
džƉĞƌŝŵĞŶƚ
^ƚƵĚLJϱ;>ĞŐͿ
;ĐͿ ;ĚͿ
Fig. 7 Simulation-to-experiment comparison results for Study 4 (leg A) and Study 5 (leg B).
The stride period for Study 4 is T = 0.3692s, and for Study 5, T = 0.3226s.
section seem to indicate that running with leg B results in faster and more stable
gait than running with leg A, although more experiments need to be performed to
confirm this.
addition, curved legs have other aspects that the described model does not capture:
leg length change, rolling motion, and the stiffness directionality. The influence of
these aspects on the running performance (such as stability, efficiency and speed of
running) remains to be analyzed. Better results may be obtained if the controller
optimization process is performed on the actual robotic running. Another interest-
ing result that may bear further investigation is that the gaits optimized for stability
appear to map better to the robot than those optimized for specific resistance.
Despite the differences between the simulations and the experiments, the results
obtained from both systems seem to indicate that running with leg B (βo = 90o ,
l2 /l1 = 1, k̃10% = 12.46) is faster and more efficient than running with leg A (βo =
90o , l2 /l1 = 0.35, k̃10% = 12.46), and these results seem to be due to the ratio of the
lengths of the leg segments chosen for leg B.
References
1. Blickhan, R.: The spring mass model for running and hopping. Journal of Biomechan-
ics 22, 1217–1227 (1989)
2. Blickhan, R., Seyfarth, A., Geyer, H., Grimmer, S., Wagner, H., Gunther, M.: Intelli-
gence by mechanics. Philosophical Transactions of the Royal Society London, Series A
(Mathematical, Physical and Engineering Sciences) 365, 199–220 (2007)
3. Galloway, K.C.: Passive variable compliance for dynamic legged robots. Ph.D. thesis
(2010)
4. Galloway, K.C., Clark, J.E., Koditschek, D.E.: Design of a tunable stiffness composite
leg for dynamic locomotion. In: Proc. ASME IDETC/CIE (2009)
5. Hyon, S.H., Mita, T.: Development of a biologically inspired hopping robot - kenken. In:
Proceedings of IEEE International Conference on Robotics and Automation (2002)
6. Jun, J.Y., Clark, J.E.: Dynamic stability of variable stiffness running. In: Proceedings of
IEEE International Conference on Robotics and Automation (2009)
7. Jun, J.Y., Clark, J.E.: Effect of rolling on running performance. In: Proceedings of IEEE
International Conference on Robotics and Automation (2011) (submitted)
8. Kim, S., Clark, J.E., Cutkosky, M.R.: isprawl: design and tuning for high-speed au-
tonomous open-loop running. International Journal of Robotics Research 25, 903–912
(2006)
9. Komsuoglu, H.: Towards a comprehensive infrastructure for construction of modular
and extensible robotic systems. Tech. rep., Dept. of Computer and Information Science,
University of Pennsylvania (2007)
10. Raibert, M.: Bigdog, the rough-terrain quadruped robot. In: The International Federation
of Automatic Control (2008)
11. Rummel, J., Seyfarth, A.: Stable running with segmented legs. International Journal of
Robotics Research 27, 919–934 (2008)
12. Saranli, U., Buehler, M., Koditschek, D.E.: Rhex: A simple and highly mobile hexapod
robot. International Journal of Robotics Research 20, 616–631 (2001)
13. Whittington, B.R., Thelen, D.G.: A simple mass-spring model with roller feet can induce
the ground reactions observed in human walking. Journal of Biomechanical Engineering-
Transactions of the Asme 131 (2009)
14. Zhang, Z.G., Kimura, H.: Rush: a simple and autonomous quadruped running robot.
Proceedings of the Institution of Mechanical Engineers, Part I (Journal of Systems and
Control Engineering) 223, 323–336 (2009)