You are on page 1of 5

2011 IEEE/RSJ International Conference on

Intelligent Robots and Systems


September 25-30, 2011. San Francisco, CA, USA

A Nonlinear Series Elastic Actuator for Highly Dynamic Motions


Ivar Thorson, Darwin Caldwell

Abstract A novel revolute nonlinear series elastic actuator


called the HypoSEA is presented. The actuator uses a hypocycloid mechanism to stretch a linear spring in a nonlinear way.
The actuator is optimized for highly dynamic tasks such as
running and jumping, as it features a 120Nm torque capability
and more than 30J of passive energy storage. When combined
with a suitable controller, using the spring as an energy buffer
can greatly reduce the work done by the rotor during periodic
motions. The design has exceptionally low reflected mechanical
impedance, making it robust against repeated impact loads. The
nonlinear stiffening spring is optimized for the nonlinearities
typically found in revolute-jointed hopping robots, and may
be adjusted offline using a pretensioning mechanism. Finally,
the low effective stiffness around the zero-torque equilibrium
allows for extremely sensitive force control.
Fig. 1.
The HypoSEA: a nonlinear series elastic actuator with large
energy storage capabilities and nonlinear stiffness optimized for running
and jumping. The motor can be seen on the left, and the long output link
containing a spring on the right.

I. INTRODUCTION
In nature elasticity is found everywhere. Muscle fibers
and animal tissues have significant elasticity that plays an
important role during locomotion [1]. Elastic models of
walking [2] and running [3] concisely describe the motion of
an animals center of mass by computing forces generated
by a virtual elastic element. During the last two decades
researchers have begun to more seriously study the effects of
purposely introducing elastic elements into actuation systems
to improve force control fidelity, safety, and efficiency.
These so-called series elastic actuators have a variety of
forms and purposes: [4] used linear ballscrews to precisely
control linear forces; [5] used a revolute mechanism to create
an adjustable nonlinear spring; [6] and [7] use variable geometry linkages to adjust mechanical stiffness; other researchers
focus on how compliance relates to safety concerns [8] [9].
In this paper, we will consider an electric series elastic actuator for jumping and running robots. Jumping and running
are highly dynamic tasks that require good power-to-weight
ratios as well as high torque actuation. Despite the excellent
controllability of electric actuators, many hopping robots use
instead pneumatics [10] [11] or hydraulics [12] [13] [14] to
achieve the necessary power-to-weight ratios for sustained
hopping motions. Relatively few hopping robots incorporate
both electric actuation and series elasticity [15] [16] [17]
[18] and further research is needed.
The monopod robot Thumper [19] weighs approximately
30kg and appears to store approximately 75J of energy
internally using high performance fiberglass springs [16].
The actuator presented in this paper is in a similar class of
energy storage as it weighs 8kg and stores >30J of energy in
a compact space, and is nearly comparable to the excellent
design in [5] that weighs 2kg and stores up to 16J of energy.

Although extensive literature exists on nonlinearity in


variable stiffness actuators, few designs are designed towards
a particular tasks nonlinearity. Generally, a linear stiffness
is the objective. This is unfortunate, as much of the elasticity
found in nature is nonlinear. Besides being biomimetic,
nonlinearly stiffening springs have desirable characteristics
including superior dynamic range and resolution during force
control.
Previously, [20] created an adjustable nonlinear mechanism that is suitable for lightweight hopping robots. However
unlike the work presented in this paper, [20]s stiffness curves
do not appear to be optimized for the torques produced
during hopping. Also, although the actuator presented in
[21] possesses nonlinear characteristics, these nonlinearities
are purposely non-smooth and contain internal collisions
beneficial for stability but not for energetic efficiency or
control.
This paper is organized as follows. In section II, the highlevel design of the actuator is presented and the mechanism
by which the desirable nonlinearity is created is shown.
Section III presents the mathematical model describing the
actuator. Section IV details a simple torque controller which
controls the robot. Section V lists various mechanical parameters, force bandwidth spectra, and step responses of the
actuator.
II. ACTUATOR DESIGN
The HypoSEA pronounced hi-po-see and short for
Hypocycloid-based Series Elastic Actuator is shown in
Figure 1. Concisely, the actuator uses a hypocycloid mechanism to stretch a single tension spring a relatively large

Istituto Italiano di Tecnologia, Department of Advanced Robotics.

ivar.thorson@iit.it
978-1-61284-456-5/11/$26.00 2011 IEEE

390

Differential
Rotor
r

G1

Hypocycloid
mechanism
e

Output Link
l

Gear Reduction
e

G2

Fig. 3.

Block diagram of the inertial connections of the HypoSEA.

Thanks to the differential, the mechanism works for


continuous rotation at the output.
In addition to the advantages presented by this hypocycloid
mechanism, the HypoSEA also features improved torquecontrol performance by minimizing reflected inertia with
a low-reduction compound planetary gear system, a hightorque brushless DC motor, and 2:1 gear reduction stage G2
to reduce the 120 degree maximum elastic element windup
to a more useful 60 degrees, as will be described in the
sequel.

Fig. 2. A schematic diagram of hypocycloid mechanism in the HypoSEA


actuator showing the mechanism by which the tension spring is stretched
by a distance d.

distance along a centerline. Because the amount of energy


stored in a spring is proportional to the square of the
deflection distance, a large deflection is desirable for storing
large amounts of energy. The HypoSEA was designed with
this fact in mind.
Figure 2 shows a schematic representation of the mechanism by which the HypoSEA stretches a linear spring in a
nonlinear way. The interface to the mechanism is the straight
line passing though the center of both circles, which defines
a deflection angle e . The inner circle is assumed to roll
without slipping inside the larger circle. One end of a tension
spring is attached to point on the circumference of this
inner circle. As the mechanism is deflected by an angle e ,
point traces a perfectly straight vertical line, stretching
the spring by a distance d. A pretension distance p may be
added to change the nonlinearity of the system. l0 is simply
listed to show the natural length of the tension spring.
There are several advantages to using a hypocycloid mechanism in this way:

III. ACTUATOR MODEL


The torque and angle relationship of the actuator are
shown in Figure 3. The blocks shown in the figure are the
rotor position r , gear reduction ratio G1 , differential point,
gear reduction ratio G2 , hypocycloid mechanism torquedeflection relationship e (e ), and the output link position l .
The rotor/transmission inertia Ir , the link inertia Il , damping
terms br and bl on the rotor and link masses respectively
are also considered. The center of mass of the output link
ml is off-axis by a distance ll . The differential element and
gearing mean that the displacement angle of the nonlinear
element e and transmission ratios G1 and G2 satisfy

A single spring stores energy regardless of the direction


of elastic deflection relative to the rotor.
The spring is stretched in a perfectly straight line and
does not pull the spring off-axis during stretching as a
cam mechanism would.
For delicate manipulation tasks the inherent nonlinearity
of the compliant element provides improved torque
control resolution at low deflections without sacrificing
high-torque capability under larger deflections.
The torque-deflection relationship of the elastic mechanism may be varied from approximately quadratic to
approximately linear by increasing spring pretension.
The long stroke of the mechanism enables tremendous
energy storage relative to its size.
The torque-deflection relationship of the elastic mechanism is well matched for revolute-jointed running
robots.

e = G2 (l G1
1 r )

(1)

l = G1 r = G2 e

(2)

The equation of motion of the system is thus




 


Ir
0
r
br 0
r
+
+
0 bl
0 ll2 ml + Il
l
l




m
G2 G1
1
e (e ) =
ext
G2

(3)

where m is the electrically generated torque on the rotor, and


ext is any external torques to the system. The relationship
between elastic deflection angle e and elastic torque e (e )
may be derived from the three geometric relations shown in
Figure 2.
d(e ) = 2r(1 cos e )

391

(4)

e (e ) = Fe 2r sin e

(5)

Fe (e ) = Ke (d + p)

(6)

Torque vs Deflection for Hypocycloid Mechanism

Transmission Torque e [Nm]

40

20

p=0.00
p=0.01
p=0.02
p=0.03
p=0.04
Linear
Quadratic

-20

-40
-3

-2

-1
0
1
Deflection Angle e [rad]

h
k , k
Fgrf
l

Fig. 4. The torque-deflection curve for the HypoSEA transmission for


various pretension distances p. By adjusting the pretension, quadratic or
linear stiffness relationships may be easily approximated.

Fig. 5. Simplified two-link leg for estimating the knee torques in a worstcase jumping configuration.

Knee Torque vs Interior Angle for Various k0

We can write the torque explicitly as


e (e ) = Ke (2r)(sin e )(2r(1 cos e ) + p)

200

k0 = 3.0
k0 = 2.5
k0 = 2.0
k0 = 1.5
k0 = 1.0

(7)
150
Torque k [Nm]

where Ke is the stiffness of the physical spring, r is the radius


of the inner circle, d(e ) is the distance from the intersection
of the inner circle and the vertical axis with the bottom of
the outer circle, and p is the pretension displacement of the
spring at e = 0.
A graph of the torque-deflection relationship for various
pretension distances p is shown in Figure 4, using the values
of r = 0.025m, Ke = 10kN/m.
From Figure 4 we see that the useful range of the
mechanism appears to be about 120 degrees. Above this
deflection, the incremental stiffness begins to decrease, which
could create instability. Therefore, the mechanism is mechanically limited to keep the derivative of the stiffness curve
monotonic, and a gear reduction G2 = 2 is used to reduce
the maximum deflection to 60 degrees. We can also see
that the torque-deflection relation is approximately quadratic
for low pretensions and becomes more linear as pretension
is increased.

100

50

0
0

0.5

1.5
2
Knee Interior Angle k [rad]

2.5

Fig. 6. The torque-deflection relation required for a jumping robot with


various knee angles at touchdown to create a Fgr f given by equation 8. The
areas under each curve are equal and equivalent ground reaction forces are
created.

this vertical force may be written as a function of the knees


interior angle k :

A. SLIP Torques for Revolute Joints


(k ) = Fgr f l cos

Let us show that the HypoSEAs elastic nonlinearity is


particularly suited for the kind of torques that a hopping
robot with a revolute jointed knee would require. If we
assume a simplified model of a large mass directly above
the foot, that the thigh and shank are the same length l, and
that the knee joint must do all the work to support the body,
we arrive at the configuration shown in Figure 5.
The robot collides with the ground at starting knee angle
k0 and hip height h0 . We wish to produce a SLIP model
ground reaction force Fgr f to bounce the robot up again.
Fgr f = Kleg (h0 h)

k
2

(9)

A graphs of the torques for various starting knee angles


k0 which produce a bouncing motion for mass m = 10kg
on a SLIP style virtual leg are shown in Figure 6. We can
see that to produce identical ground reaction forces, landing
with a straighter leg results in more nonlinear knee torques
with lower peak torques, as well as a larger deflection angle
k0 k . Conversely, landing with a bent knee results in more
linear knee torques with less deflection. Choosing the best
landing posture will depend on other mechanical considerations and cannot be made at this time. The important thing
to note is that the HypoSEA can approximate all of these
curves by adjusting the pretension.

(8)

where Kleg is the modeled whole-leg stiffness according


to the SLIP [3] model. The torque at the knee that produces
392

TABLE I
M ECHANICAL PARAMETERS

IV. CONTROLLER
Since the actuator is designed from the ground up for
excellent force control and energy storage, we will not study
an output link position controller in this paper. Instead,
we will explain the means by which an output link torque
controller was implemented.
First, a desired torque on the output link is specified to
the controller. The desired torque on the output link will
of course depend on the task at hand. For demonstration
purposes, let us follow the signal path of a simple gravity
compensation algorithm as it works its way to the motor.
The torque that will compensate for gravitational forces on
the output link of the actuator are:
ld = ml gll sin(l )

Description
Total mass
Actuator diameter
Maximum link length
Maximum output torque*
Minimum Resolvable Torque*
Maximum controllable output velocity*
Rotor-Link Gear Reduction
Link-Elasticity Gear Reduction
Link mass
Link CoM distance
Encoder counts per revolution
Motor poles
Motor torque sensitivity
Motor back-emf constant
Motor resistance
Motor inductance
Motor electrical time constant
Linear Spring Constant
Maximum spring pretension
Maximum spring deflection
Maximum spring potential energy*

l
l
G1
G2
ml
ll
Kt
Kem f
Rm
Lm
Tm
Ke
p
d

(10)

where l = 0 corresponds to the orientation where the output


link arm hangs straight down, ml is the mass of the output
link, and ll is the output link centroids distance from the
center of rotation.
This desired link torque ld must be generated by a
deflection by the nonlinear spring, so we invert equation 7 to
get the desired elastic deflection. This was done numerically,
and an approximate polynomial model was then used in the
realtime computations.

Value
8.9
14.0
60
126
<0.02
10.2
18.33
2.0
5.3
11.5
8192
12
0.30
0.30
2.64
2.76
1.04
10.09
40
72
42.3

Unit
kg
cm
cm
Nm
Nm
rad/s
kg
cm
poles
Nm/A
V/rad/s
ohms
mH
ms
N/mm
mm
mm
J

Normalized Response to a Step Voltage


1

Current
Rotor Position
Elastic Deflection
Velocity
Link Position

0.8
0.6
0.4

(11)

0.2
Value

ed (e ) = K1 e + K2 e2 + ...

From this desired elastic deflection ed , we can compute the


an instantaneous desired rotor position rd
rd = G1 G1
2 ed

0
-0.2
-0.4

(12)

-0.6

A desired rotor torque rd is then computed from a PD


controller which moves the actuator towards desired position
rd :
rd = KD r KP (rd r )
(13)

-0.8
-1
21.5

21.6

21.7

21.8

21.9

22

22.1

22.2

22.3

22.4

22.5

22.6

Time

where KP and KD are the proportional and derivative gains.


Finally, the desired motor current imd is found from the
desired motor torque since imd = Kt1 rd , and the voltage to
create this torque is found by modeling the motor as an RL
circuit with a back-emf proportional to the angular velocity:

Fig. 7. A measured open loop voltage step response with no output load.
The figure is normalized for visualization purposes. Note how the effect of
the elastic deflection on all other quantities can clearly be seen.

dim
+ Kem f r
(14)
v pwm = Rm im + Lm
dt
This pulse-width-modulation voltage v pwm was then sent
to the motor driver board, which amplified the signal and
handled commutation.

Figures 7, 8. Figures 9 and 10 show the frequency response


of the simple PD control system described in this paper.
Figure captions are self-explanatory. Remember that since
the actuator is nonlinear, these frequency spectrum plots
are only valid for the particular torque offset and torque
amplitude at which the signal was taken.

V. RESULTS

VI. CONCLUSION

The measured mechanical parameters of the actuator are


listen in Table I. The motor controller being used to test the
actuator can presently only safely supply 10A of current to
the board, so quantities in the table marked with a star are
based on extrapolation from the actuator operation at slightly
less than half of the intended design current.
The complete description of the system identification
method of the actuator is beyond the scope of the paper, but
we can present open loop data taken from the actuator in

The actuator works well and in laboratory tests shows


promising torque control capability at force control levels
four orders of magnitude apart. The system is able to
resolve torques less than 0.02Nm exactly how much less
depends on the minimum spring pretension and stiction
yet is still able to produce torques as large as 120Nm.
Future work includes developing more sophisticated gainscheduling controllers for HypoSEA to more fully utilize the
nonlinearity of the mechanical system, particularly during the
393

Open Loop Frequency Response


1000

Torque Control Frequency Response with Load


10

Applied Voltage
Current
Elastic Deflection
Rotor Work
Elastic Work

100

Desired Torque
Actual Torque

Torque Magnitude

Magnitude

10

0.1

0.1

0.01

0.01
0.001
0.001

0.0001

0.0001
0.1

10

100

0.1

Frequency

10

100

Frequency

Fig. 8. A measured open loop voltage frequency response. Magnitudes are


not to scale. Note the mechanical resonant frequency is approximately 2Hz,
the frequency at which most of the work in human locomotion occurs, and
that rotor work is reduced at this resonant frequency.

Fig. 10. The measured torque control frequency spectrum of the actuator
when loaded with a 2kg mass hanging from the link and creating a
gravitational torque. Note that the actuator can now track low frequency
torques more accurately due to increased link inertia and therefore lower
link velocities. Because of the hypocycloid mechanisms increased effective
stiffness farther from torque equilibrium, the small torque signal frequency
response has also increased to about 6Hz. For high torque levels required
in a hopping robot, the actuators bandwidth will easily exceed 10Hz.

Torque Control Frequency Response, No Load


10

Desired Torque
Actual Torque

Torque Magnitude

[5] S. Wolf and G. Hirzinger, A new variable stiffness design: Matching


requirements of the next robot generation, in ICRA 2008, May 19-23,
2008, Pasadena, California, USA, 2008, pp. 17411746.
[6] R. Schiavi, G. Grioli, S. Sen, and A. Bicchi, VSI-II: a novel prototype
of variable stiffness actuator for safe and performing robots interacting
with humans. in ICRA. IEEE, 2008, pp. 21712176.
[7] A. Jafari, N. Tsagarakis, B. Vanderborght, and D. Caldwell, A novel
actuator with adjustable stiffness (AwAS), in IROS 2010. IEEE,
2010, pp. 42014206.
[8] M. Zinn, B. Roth, O. Khatib, and J. Salisbury, A new actuation
approach for human friendly robot design, The international journal
of robotics research, vol. 23, no. 4-5, p. 379, 2004.
[9] A. Bicchi and G. Tonietti, Fast and soft-arm tactics, IEEE Robotics
and Automation Magazine, vol. 11, no. 2, pp. 2233, 2004.
[10] K. Hosoda, Y. Sakaguchi, H. Takayama, and T. Takuma, Pneumaticdriven jumping robot with anthropomorphic muscular skeleton structure, Autonomous Robots, vol. 28, no. 3, pp. 307316, 2010.
[11] R. Niiyama, A. Nagakubo, and Y. Kuniyoshi, Mowgli: A bipedal
jumping and landing robot with an artificial musculoskeletal system,
in Robotics and Automation, 2007 IEEE International Conference on.
IEEE, 2007, pp. 25462551.
[12] M. H. Raibert, Legged robots that balance. Cambridge, MA, USA:
Massachusetts Institute of Technology, 1986.
[13] G. Zeglin, Unirooa one legged dynamic hopping robot, PhD Thesis,
MIT. 1991.
[14] S. Hyon and T. Mita, Development of a biologically inspired hopping
robot-Kenken, in ICRA 2002, vol. 4. IEEE, 2002, pp. 39843991.
[15] M. Ahmadi and M. Buehler, The ARL Monopod II running robot:
Control and energetics, in ICRA 1999, pp. 16891694.
[16] J. Hurst, The role and implementation of compliance in legged
locomotion, Ph.D. dissertation, Carnegie Mellon University, 2009.
[17] J. Rummel, F. Iida, J. Smith, and A. Seyfarth, Enlarging regions of
stable running with segmented legs, in ICRA 2008. IEEE, 2008, pp.
367372.
[18] S. Curran and D. Orin, Evolution of a jump in an articulated leg with
series-elastic actuation, in ICRA 2008. IEEE, 2008, pp. 352358.
[19] I. Poulakakis and J. Grizzle, Modeling and control of the monopedal
robot thumper, in ICRA 2009. IEEE, 2009, pp. 33273334.
[20] B. Vanderborght, N. Tsagarakis, C. Semini, R. Van Ham, and D. Caldwell, MACCEPA 2.0: Adjustable compliant actuator with stiffening
characteristic for energy efficient hopping, in ICRA 2009. IEEE,
2009, pp. 544549.
[21] M. Hutter, C. D. Remy, and R. Siegwart, Design of an articulated
robot leg with nonlinear series elastic actuation, in Mobile Robotics:
Solutions and Challenges, 2009, pp. 645652.

0.1

0.01

0.001
0.1

10

100

Frequency

Fig. 9. The measured torque control frequency spectrum of the actuator


when unloaded. Note the ability to accurately track torques is relatively
limited. Below 1Hz, the maximum rotor velocity limits the torque tracking
ability, and above 3Hz the low effective stiffness of the hypocycloid
mechanism prevents torques from being developed quickly enough.

motion planning phase. Although the existing prototype has


a disappointingly low power-to-weight ratio, it is perhaps
overly robust, and with some effort perhaps a significant
fraction of the mass may be cut from the design.
ACKNOWLEDGMENTS
Thanks especially to Gianluca Pane for help with the
mechanical engineering. Thanks also to Irene Sardellitti and
Bram Vanderborght for their support and discussion.
R EFERENCES
[1] R. Alexander, Elastic mechanisms in animal movement. Cambridge
University Press Cambridge:, 1988.
[2] H. Geyer, A. Seyfarth, and R. Blickhan, Compliant leg behaviour
explains basic dynamics of walking and running, Proceedings of the
Royal Society B: Biological Sciences, vol. 273, no. 1603, p. 2861,
2006.
[3] R. Blickhan, The spring-mass model for running and hopping,
Journal of Biomechanics, vol. 22, no. 11-12, pp. 12171227, 1989.
[4] G. Pratt and M. Williamson, Series elastic actuators, IROS 1995.
Published by the IEEE Computer Society, 1995, p. 399.

394

You might also like