2012 IEEE/RSJ International Conference on

Intelligent Robots and Systems

October 7-12, 2012. Vilamoura, Algarve, Portugal

Actuator Design for High Force Proprioceptive Control in Fast Legged

Locomotion
Sangok Seok, Albert Wang, David Otten and Sangbae Kim

Planetary gear
Abstract— High speed legged locomotion involves high ac- Hip motor
celeration and extensive loadings of the leg, which impose Knee motor
critical challenges in actuator design. We introduce actua-
tor dimensional analysis for maximizing torque density and
transmission ‘transparency’. A front leg prototype developed
CoM
based on insight from the analysis is evaluated for direct
proprioceptive force control without force sensors. The vertical
stiffness controlled leg was tested on a material testing device to
calibrate the mechanical impedance of the leg. By compensating Tendon
Ab/Adduction
transmission impedance from commanded torque, the leg was joint
able to estimate impact force. For the impact test, the mean Encoder mount
absolute error as a ratio of full scale sensor force is 0.041 in
the 3406 N/m stiffness experiment and is 0.049 in the 5038 N/m
experiment. The results indicate that prescribed force profile
control is possible during high speed locomotion. Fig. 1. The solid model of the front leg design of the MIT cheetah
robot. The leg is designed to maximize backdrivability and transparent force
I. INTRODUCTION production. The rotational inertia of the leg is minimized by locating all
drive components at the shoulder. The center of mass of the leg is located
High speed locomotion entails several critical challenges 3cm below the center of the rotation of the shoulder joint.
in legged robot design. One of the major apparent challenges
is the high force requirement. As observed in biology, ground
reaction force increases with speed and is directly related to
the duty factor 1 . In steady state running, the total vertical Although the power mass density of EM (electromagnetic)
impulse during one period (T) of cyclic locomotion must be actuators (continuous up to 7 kW/kg [35], 3-5 kW/kg[36])
equal to the total gravitational impulse to satisfy momentum significantly exceeds biological muscle (Max. 0.3 kW/kg)
conservation. [1], the high power is available only at high speed with
relatively low torque compared to muscles. Higher gear
ZT reduction can increase torque density, a critical requirement
Fz dt = mgT in legged locomotion, but this increases actuators’ passive
0 impedance (reflected inertia, friction, damping) which limits
the bandwidth and significantly compromises transmission
The equation implies that a smaller duty factor entails
‘transparency’[29], critical for high speed force control.
higher ground reaction forces. Typically, faster running re-
Gear friction is usually highly nonlinear [23][24][22]and
quires a higher stride frequency and lower duty factor [4].
significantly compromises force control performance [10]
This leads to high effective ground reaction forces during
because the mechanical impedance of the gear train can cause
running. The maximum normal ground reaction force on
non-desirable force in fast dynamics. The critical trade off in
each leg is around three times the bodyweight in human
EM actuators seems to be the trade off between high torque
running at 4.5 m/s [2] and 2.6 times the bodyweight in dog
density and low actuator impedance.
galloping at 9 m/s [3].
In addition to the high force requirement, variable There are several approaches that enable impedance con-
impedance seems highly desirable. Hurst [16] emphasized trol without compromising torque density of the motors.
the importance of the variable stiffness in series elastic Many researchers modulate the apparent impedance by em-
actuators. In human running, significant changes in effective ploying torque feedback and full state feedback, but the
leg stiffness are observed in the variation of speed [34] and performance is limited to relatively slow speed dynamics
ground stiffness [33]. [31][30] and is not suitable for high speed locomotion.
Employment of series elastic actuators[12] allows variable
This work was supported by Defence Advanced Research Program mechanical impedance to achieve by tuning the stiffness
Agency M3 program of the elastic elements. A dual actuator is used in a
Authors are with the Department of Mechanical Engineering at Mas- manipulator to improve stiffness modulation [13]. Tunable
sachusetts Institute of Technology, Cambridge, MA, 02139, USA , corre-
sponding email: sangok at mit.edu stiffness of series elasticity is incorporated to geared motors
1 the portion of time that leg stays on the ground for manipulation [14][15], and for running robots [16][21].

978-1-4673-1736-8/12/S31.00 ©2012 IEEE 1970

The hybrid approach on macro-mini actuators incorporates Torque

advantages of both high force pneumatic actuators with

Demag. Torque Recommended
minimal impedance and small EM actuators [17] for human- (~10X of cont. Torque)
operation space
(from manufacturer)
safe robotic manipulators. Most of these approaches require
Extended continuous
additional actuators and increase actuator complexity and operation space
inertia of manipulators. Pneumatic actuators are implemented
V= Vrecommended Transient operation
in robotic arm design due to its intrinsically low passive Continuous space
torque
impedance and relatively high force density but they suffer
from limited bandwidth [9][20]. Direct drive actuators are
employed to minimize friction caused by the gear train Angular speed

[19][7]. The major disadvantage of direct drive arms is

significantly lower torque density. Fig. 2. The EM motor performance diagram in torque-speed space.

High bandwidth force control can be achieved by em-

ploying open loop impedance control 2 with low mechanical
impedance (low inertia linkages and actuators) system ar- The power of EM motors is usually determined by the
chitecture. This approach is well-executed in haptic display recommended voltage provided by the manufacturer. The
devices [11], which employ low impedance transmission sys- recommended voltage is set concerning the thermal failure
tems such as cable transmissions and linkages that maximize in the case that users command voltage without the knowl-
the ‘transparency’[29] closely related to ’backdrivability’ of edge of the coil temperature. This is usually not useful in
the actuator. Utilizing low mechanical impedance, direct a current-controlled system where current can be limited
torque commands can produce forces at the end effector and under the continuous torque or coil. Within the continuous
remove the need for force sensor feedback. The dynamic current limit, the motor can generate much higher power
stability issue [32] is significantly improved by collocating than the rating provided by the manufacturer with a fixed
the sensing 3 and the system remains stable through high voltage limit. Figure 2 shows operation limits of motors in
speed user input. Although haptic display devices success- torque-speed space. Considering that electric Joule heating is
fully reproduce desired virtual mechanical impedance in a proportional to current squared times resistance and current is
small force range, it is not clear how to realize this approach directly related to torque, not speed, the power can be greatly
in the systems that require high-force high speed operations. increased by increasing speed under the continuous current
For highly dynamic legged locomotion, force control limit. This is represented in the blue area in Figure 2. Speed
should be realized in a simple and robust system. The can be limited by increased viscous damping of bearings and
mechanical system should be able to withstand high ground air, electric commutation frequency limit, and so on. In most
reaction forces followed by high impacts on the ground. In robotics applications the duty factor of the motor operation
this paper, we introduce the dimensional analysis of EM is low and the operation trajectory in torque-speed space
actuators focusing on maximizing torque density and the changes depending on the situation. In the applications where
mechanical impedance of a high force ‘transparent’ actuation there are frequent physical interactions with environments
system for high speed legged locomotion. We present a leg including legged locomotion, the duty factor of the robot is
prototype designed for the MIT Cheetah robot, a running significantly low. If the duty factor of the actuator operation
quadruped, realizing high speed and high torque. The low is known, we can utilize the transient operation space shown
impedance of the leg allows direct torque control of the in pink in Figure 2. The robots often need to produce torques
actuators and true proprioceptive force control without a higher than the continuous operating torque limit for a short
force sensor. Section II discusses dimensional analysis of EM period of time. One can always achieve higher power in
actuators regarding a critical design trade-off in EM motors. high-speed by providing higher voltage but this requires
In section III, we describe the mechanical design aspect of higher gear ratio that leads to a higher impedance of the
the front leg of MIT cheetah leg based on the dimensional actuator, reducing the efficiency and increasing the total
analysis of actuators. Finally, through the stiffness control weight. Achieving a high torque actuator allows lower gear
test of the leg, we demonstrate how the leg can be used to ratio and lower mechanical impedance. Therefore, power
generate accurate forces during running. density of EM actuators is not a well defined metric and does
not indicate the actual delivered power if combined with a
II. D IMENSIONAL ANALYSIS ON E LECTROMAGNETIC gear train. A clearer metric to maximize in actuator design
MOTORS for robotics is torque density (mass specific torque).
A. Power density vs. Torque density
B. Gap radius: a critical metric in EM motor design
Power density (mass specific power) is often used as There are a number of EM actuator design parameters
a major metric to evaluate the capacity of the actuators. that affect the overall performance of the motors. One of the
2 Technically, there is a current feedback loop to control torque of the critical design parameter among them is the gap radius. The
motor instead of using force sensor feedback. gap radius is the distance from the rotating axis to the center
3 Here, sensing is current sensing at the motor to realize desirable torques of the gap between permanent magnets and the stator. Figure

1971
 6 y = -1.5734x + 6.7254 01='

%!"

Continuous torque density (Nm/kg)

4 01;' !"#$%&'()'*+,"%-%..''
rgap log(torque/inertia)
/012'3"%.'#*'",..4' Maxon RE
01<'
2 log(torque/mass)
5'6'7170089':'717;<='

Torque/inertia (mNm/ gcm2)

tstator log(Kt^2/R) y = 0.8023x - 1.1813 01>' $#" Emoteq HT
0 0'
0.5 1 1.5 2
Mass ~ rgap trotor -2 71=' $!"
Torque ~ rgap2
Rotor inertia ~ rgap3 71;'
-4
Torque density ~ rgap y = 4.1107x - 8.1444 71<' ?,9#@'A!' #"
Torque/inertia ~ 1/ rgap -6
log(gap diameter) 71>' 5'6'717789':'717=8>'

(a) (b) 7' !"

7' >7' <7' ;7' =7' 077' 0>7' 0<7'
!" #!" $!!" $#!"
Outside diameter (mm) Outside diameter(mm)
Fig. 3. The specifications of the motor are scaled with the gap radius
assuming electromagnetic similarity.
Fig. 4. (left) A comparison between Maxon RE series and Emoteq
frameless motors assuming the frame weight be 50% of the motor for
Emoteq motors. (right) Acceleration capability(torque/rotor inertia).

3(a) shows hypothesized brushless motor specifications as

functions of rgap (gap radius). The gap radius is a strong pa-
rameter in motor characteristics. In contrast, the axial length reflected rotor inertia and the total torque at the output shaft
of the motor does not affect torque density and torque per stay the same. In general, higher gear ratios add mass, inertia
inertia because increasing axial length is equivalent to adding and friction loss. Thus the increased motor radius would
identical motors on the same axle. In Figure 3(b), Emoteq improve torque per inertia by reducing the gear requirement.
HT series motor characteristics are plotted against gap radius. Therefore, it is optimal to choose the motor with largest gap
Motors of the same gap radius and various lengths have radius and add the least amount of gearing to meet the torque
overlapping values of torque density and torque per inertia. requirement.
As we vary the gap radius, assuming that the electromagnetic Figure 4 shows the data collected from Maxon motors and
characteristics of a section (marked as dotted line in Figure Emoteq motors to evaluate the scaling predictions. We use
3) remains same with same radial thicknesses of stators and the continuous torque as a representative metric because it
rotors, we can estimate how motor specifications scale with includes the thermal characteristics whereas demagnetization
the gap radius. We call this electromagnetic similarity. As the torque (sometimes denoted as peak torque) does not represent
gap radius becomes larger, the moment arm and the surface how long the motor can sustain that torque 5 . The thermal
area inside of the gap increase proportionally. The mass of limitation of the motor should be analysed depending on
the motor is proportional to rgap ; the torque is proportional the duty factor of the application. The plot in Figure 3
2 ; the rotor inertia is proportional to r 3 . Therefore,
to rgap 0.8 ; torque
gap shows that the torque density is proportional to rgap
we can predict that torque density is proportional to rgap and −1.5
per rotor inertia is rgap ; the torque production efficiency is
the acceleration capability (torque/inertia) is proportional to 4 . This shows a consistent trend consid-
proportional to rgap
1/rgap . Another important characteristic is torque production ering these specification of these motors do not necessarily
efficiency which is related to the motor constant KM . KM 2 is
follow our assumptions in modelling due to manufacturing
equivalent to the torque squared per unit ohmic power loss. considerations.
The following equation describes the how it is related to
2 = Kt = τ ∝ 0
2
gap 2 nl B2 r2 A III. M ECHANICAL DESIGN OF THE LEG
motor parameters. KM R I2 R ρ where Kt is
the torque constant, I is the motor current, R is the terminal Using the torque density analysis introduced in section II,
resistance, n is the number of wires in the cross section, l0 is we selected the largest radius motor that could fit within the
the length of the motor along the axis, B is the field strength 5 inch space constraint for the MIT Cheetah leg. Given the
of the magnets, rgap is the radius of the motor gap, A is the large torque of the motor, only a minimal gear ratio (1:5.8)
cross section area of each wire and ρ is the the resistivity of is required, whereas most EM actuators used in legged robot
the wire material. Given a particular wire gauge, the number tend to have large gear ratios (more than 1:100) which
of wires in the cross section n scales linearly with the radius make the actuator non-backdrivable and inefficient. Unlike
2 3 .
of the motor. Thus the relationship becomes Iτ2 R ∝ rgap traditional serial link robots in which actuators are present
These models can be very helpful to design an actuator at every joint, two actuators and gear trains are coaxially
for a given set of requirements (e.g. torque density, inertial, located at the hip of the leg in Figure 1. The knee joint is
viscous, and frictional mechanical impedance). Suppose we driven by a steel linkage at the output of the knee motor.
want to choose a motor and gearset with the largest torque This design minimizes the rotational inertia of the leg and
per inertia while keeping the output torque and motor mass helps reduce the mass in the motor frames. The desired peak
constant. If we double the radius, the motor length is halved, torque is roughly 100 Nm considering the maximum ground
the motor torque doubles and the gear ratio is halved. The reaction forces estimated from data in [3]. Each motor is
motor rotor inertia increases by 4 times but the reflected rotor connected to one stage of planetary gearing with 4 planet
inertia through the gears4 scales by one fourth. The total
5 Usually, demagnetization torques are 10 times larger than the continuous
4 the reflected inertia is proportional to gear ratio squared torque.

1972
 Supply Voltage 88.8 V Supply Supply Feedforward
gears to distribute the stress. The peak torque of the motor Voltage

is 21 Nm.
Speed2 x Gain1
The structures of the leg are also designed to minimize Speed
_

mass and leg inertia. The humerus and radius are made Current
Speed x
Command x Gain2
+
Σ
Command
of foam-core composite plastic and the foot is molded
_
with an embedded webbing tendon that provides compliance iA
ABC
id
_ Σ d-Axis Gain
+
Σ x
vd
d-q
vA

and minimizes the stress on the radius. Bending stress is iB to

d-q
to
ABC
vB

iC Transform iq vq Transform vC
minimized in the leg structure by distributing tensile forces _ Σ q-Axis Gain
+ Σ x
+ +
to the tendons. This method allows significantly lower inertia Current
Command x Gain3
Encoder
Command + Σ
of the leg without compromising leg strength [37]. The Encoder +
Speed Supply Feedforward
shoulder module that contains the motors, gear trains and Speed x Gain4

frame weighs 4.2 kg. The humerus weighs 160 g and the Calculate Speed Speed
from Change in
lower limb including foot mass is 300 g. The moment of Encoder Position

inertia of the whole leg at the shoulder joint 0.058 kgm2 in

a straight configuration. Fig. 5. A block diagram of the control architecture of the current driver.

IV. P ROPRIOCEPTIVE F ORCE C ONTROL

A. Controller
paper, the controller uses pure proportional gain only. The
The leg dynamics are described by the following general control equation is given by τ = J T (−K p x) + fc (q, τ) where
equation in joint space τ the desired output torque command, J is the Jacobian, K p
H(q)q̈ +C(q, q̇)q̇ + g(q) + fd (q̇) + fc (q, τ) = τ − J T Fext is the virtual spring stiffness, x is the Cartesian coordinate of
the foot and fc (q, τ) is the compensator for torque dependant
where H(q) is the mass matrix, C(q, q̇) is Coriolis force friction.
g(q) is gravity, fd (q̇) is damping friction, fc (q, τ) is torque
dependent Coulomb friction, τ is the torque applied at the B. Experiment Setup
motors, J is the Jacobian matrix and Fext is the external To compare the actual force to calculated force from the
reaction force exerted on the foot in Cartesian space. The motor current, the Cheetah robot leg was mounted inside
impedance control algorithms in this system are similar to of an axial material testing device (Zwick Roell BX1-
those used in haptic devices. As in haptics, the passive EZ005.A4K-000) which can generate ±5 kN loads and
impedance caused by inertia and damping can hamper the speeds of 500 mm/s with 0.2 m stroke length. The standard
accurate reproduction of forces [29]. Therefore, the prevail- force sensor was replaced with a six axis force torque sensor
ing design has been to increase ‘transparency’ between the (ATI delta, SI-660-60 calibration) which can measure up to
end effector and actuator by reducing dynamic effects from 1980 N in the z-axis with 0.25 N resolution. Pure vertical
mass and friction. Transparency allows the force exerted by force was obtained by tuning the leg configuration and
the motors to match the end effector forces thus eliminating controller so that all other forces and torques were close
the need for a non-collocated sensor. Instead of a force sensor to zero (less than 1% of the maximum force during the test).
at the foot, force can be determined from the applied actuator The actual force data, the estimated force from motor current
torque and from displacements in joint space measured by the and z-axis displacement were simultaneously measured at
motor encoder. This approach avoids the possible instability 500 Hz. Force commands from position feedback described
caused by unmodeled modes between the force actuator and in Section IV-A were executed every measurement cycle. The
the non-collocated sensor [32]. force sensor data was then compared to the proprioceptive
The MIT Cheetah robot leg is specifically designed to measurement. Tests were conducted by creating a virtual
minimize inertia and friction. In the dynamics equations, linear spring in the controller using only P control without I
Coriolis terms do not contribute much and can be neglected and D gain. Cyclic sawtooth position profiles were exerted on
[19]. Also the motor force is large compared to gravity and the leg and the speed was varied over experiments from 6.67
damping, so the low speed force relationship becomes τ ≈ mm/s to 426.7 mm/s. An impact test was also conducted to
J T Fext + fc (q, τ). The approximation greatly simplifies the show high bandwidth control in conditions closer to running.
control because double differentiation of the encoder signal The experiment was performed with an impact velocity of
to calculate acceleration is not robust on digital systems. As 500 mm/s.
a first approximation it is also possible to ignore torque de-
pendent friction. It can be compensated later as feedforward C. Controller Hardware
in control after being characterized in experiment. An FPGA based real-time controller (NI sbRIO-9642) was
To enable proprioceptive control, the robot first converts used as the main controller. This controller integrates a 2
joint space displacements to Cartesian space using forward mega gate FPGA (Xilinx Spartan II) and a 400 MHz real-
kinematics. A PID loop on position is then applied around time processor(Freescale MPC5200). To reduce the load on
a zero set point and a new torque command is sent to the the real-time processor, all low level communication and
motor. In the case of the virtual linear spring shown in this interfacing with the hardware is programmed in FPGA. In the

1973
 FPGA, 12 parallel UART lines are emulated for communica-
NI sbRIO-9642 Motor Drivers
tion with eight custom brushless DC motor drivers and four
Real Time Controller FPGA
i i (2B) Dynamixel EX-106+ smart motors in the abduction joints.
Parallel UART d-q
PD Control (500Hz) PCI Bus
Emulator
UART(500kbps) Feedback Additionally, data from the analog six axis force sensor is
Jacobian Transpose
(4 kHz)
θ, i θ, i, status (4B) converted in FPGA for logging. This force sensor is used for
Cheetah leg
data comparison with the converted forces from the motor
F F τ i current. Figures 5 and 6 show the motor driver architecture
End Jacobian Brushless
Environment x x θ and overall control architecture respectively.
Effector (transmission) Motors

θ In this experiment, the left front leg’s shoulder and knee

Absolute θ motors were activated and the Jacobian transformation, PD
Encoders
control, and FT sensor data acquisition are processed every
Fig. 6. A condensed block diagram of the control architecture of the
2 ms in the real-time controller. The experimental data are
hardware. streamed from the main controller to PC via UDP (User
Datagram Protocol) communication to prevent delay in the
main loop time.

D. Results and Discussion

Data processing on the cyclic tests is used to find the
torque dependent friction. Figure 8(a) shows the hysteresis
curve as the leg is loaded and unloaded. The behavior is
consistent over a range of cycling speeds. Figure 8(c) shows
the discrepancy between the measured force profile from
the force sensor and the estimated force from the motor
current during the 7 mm/s cyclic test. In order to make
a perfect spring in force-displacement space, the estimated
force must be compensated. Figure 8(b) shows the cyclic
measured force vs displacement plot at 7 mm/s showing the
Fig. 7. (left)Experimental setup of the leg on a material testing device. two distinct behaviors corresponding to leg compression and
(right) MIT cheetah prototype extension. The hysteresis curve corresponds to the energy
consumed by friction in each loading and unloading cycle.
The friction force changes direction depending on whether
the leg is compressing or extending. To compensate for the
280 260
behavior, two linear fits are used to determine the actual
240 a 240 b
200 200
force vs displacement relation in both movement directions.
During the compression, the actual force vs displacement
Force(N)

160
Force(N)

160

120 120
slope is 0.886 of the estimated force vs displacement slope.
80
80
40
The value is 0.696 for extension of the leg. In order to
427 mm/s 40 Force Displacement Curve
0 213 mm/s
7 mm/s 0
Linear Fit of Compression
Linear Fit of Extension
apply compensation, the estimated force is multiplied by
-40
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 the appropriate ratio value depending on the direction of
Displacement (mm) Displacement (mm)
300
leg movement. Figure 8(c) shows the estimated force before
320 Compensated (3406 N/m)
260 c 280 d
Force Sensor (3406 N/m)
Compensated (5083N/m)
and after compensation as the measurement from the force
220 Force Sensor (5083N/m)
240 sensor. The compensated curve and actual force match well.
Force(N)

180
Force(N)

200
140 160
After the correct values for friction compensation are found
100 120 using data from the cyclic displacement test, the compen-
60 80
Force from FT sensor 40
sation can be applied to other arbitrary movements. Figure
20 Force from Motor Current
0
-200
Compensated Force 0 8(d) shows the force profiles from impact tests conducted
5000 10000 15000 20000 25000 30000 35000 0 50 100 150 200 250 300 350 400 450 500
Time (ms) Time (ms) with two virtual spring settings. The compensated force
matches the force sensor measurement in both cases. The
Fig. 8. (a) Cyclic force vs displacement plot for various speeds and total mean absolute error as a ratio of full scale sensor force
displacements, (b) Linear fit on the cyclic force vs displacement plot at 7
mm/s showing the two distinct regions corresponding to leg compression and is 0.041 in the 3406 N/m experiment and is 0.049 in the
extension. The two linear fits are used to determine force compensation on 5038 N/m experiment. At higher virtual stiffness, there is
the force estimated using motor current, (c) Plot of force over time showing some discrepancy when the direction of movement changes.
the force estimated from motor current, the compensated force and the true
measured force from the force torque sensor, (d) Force vs time plot during This suggests that compensation may have to be adjusted
several impact test performed at with controller virtual stiffness of 3406 depending on the algorithm of the force controller. The
N/m and 5083 N/m. Both the compensated force from the motor current leg configuration used for the experiment produces only
and the force sensor data are shown.
force in the vertical direction. Therefore the loading is on
each joint is equal to the exerted vertical force. Thus the

1974
torque dependent friction can be taken as a function of [14] Koganezawa, K., “Mechanical stiffness control for antagonistically
vertical force. This lumped approach works well and gives driven joints”, in Proc. of the IEEE/RSJ International Conference on
Intelligent Robots and Systems. IEEE/RSJ, August 2005, pp. 2512-
sufficiently good results during impact and should therefore 2519.
deliver high performance during robot running. [15] Tonietti, G., Schiavi, R., and Bicchi, A., “Design and control of
a variable stiffness actuator for safe and fast physical human/robot
V. CONCLUSIONS interaction”, in Proc. 2005 IEEE International Conference on Robotics
and Automation, April 2005, pp. 582-533.
This paper presents the force production capabilities of [16] Hurst, J. W., Chestnutt. J. E., and Rizzi, A. A., “An actuator with
the leg on the MIT Robotic Cheetah. We have shown that physically variable stiffness for highly dynamic legged locomotion”, in
Proc. of IEEE International Conference on Robotics and Automation.
the mechanical design principles for maximum transparency New Orleans, LA, USA: IEEE, April 2004, pp. 4662-4667.
allows for high force proprioception even with simple con- [17] Shin, D.,Sardellitti, I., Park, Y.L., Khatib, O., Cutkosky M. “Design
trollers. These legs will be used on the robot to perform high and Control of a Bio-Inspired Human-Friendly Robot”, The Interna-
tional Journal of Robotics Research April 2010 vol. 29 no. 5 571-584.
speed running and will be able to execute the prescribed [18] Migliore, S.A., Brown, E.A., DeWeerth, S.P., “Novel nonlinear elastic
force commands. The supplement video to this paper shows actuators for passively controlling robotic joint compliance”, Journal
balanced standing and jumping using proprioceptive force of Mechanical Design 129(4):406-412, 2007.
[19] H. Asada, K.Y. Toumi, Direct-Drive Robots, Theory and Practice, MIT
control. Future research includes the characterization of gear Press, Cambridge, 1987.
impedance to complete the model of the actuator as a full [20] Caldwell, D.G., Medrano-Cerda, G.A., Goodwin, M., “Control of
package including the motive device and the transmission pneumatic muscle actuators”, Control Systems, IEEE, Issue Date: Feb
1995, Vol.15, 1, page(s): 40-48.
whose model will enable precise force control. Energy ef- [21] K. Galloway, J. Clark, and D. Koditschek, “Design of a tunable stiff-
ficiency can also be improved by developing high torque ness composite leg for dynamic locomotion”, in ASME IDETC/CIE,
motors suited for the joint velocities and forces in legged 2009
[22] He S., Gunda, R., Singh, R., “Effect of sliding friction on the dynamics
locomotion. of spur gear pair with realistic time-varying stiffness”, Journal of
Sound and Vibration Vol. 301, Issues 3-5, 3 April 2007, Pages 927-
ACKNOWLEDGMENT 949.
[23] Grotjahn, M., Daemi, M., Heimann, M., “Friction and rigid body
This program is sponsored by DARPA M3 program. identification of robot dynamics”, International Journal of Solids and
Structures Volume 38, Issues 10-13, March 2001, Pages 1889-1902.
R EFERENCES [24] REBBECHI, F. OSWALD and D. TOWNSEND, “Measurement of
gear tooth dynamic friction”, ASME, DE-Vol. 88, Proceedings of the
[1] I. Hunter and S. Lafontaine, “A comparison of muscle with artificial 7th Power Transmission and Gearing Conference, (1996), pp. 355-363.
actuators”, in Tech. Dig. IEEE Solid State Sensors Actuators Work- [25] M. Raibert, “Trotting, pacing, and bounding by a quadruped robot”,
shop, pp. 178-185, 1992. J. Biomechanics, 23, Suppl.1, 79–98,1990.
[2] M. F. Bobbert and M. R. Yeadon and B. M. Nigg, “Mechanical analy- [26] H. Kimura, Y. Fukuoka and A.H. Cohen, “Adaptive Dynamic Walking
sis of the landing phase in heel-toe running”, Journal of Biomechanics, of a Quadruped Robot on Natural Ground based on Biological
Vol. 25, no. 3, page, 223-234, 1992. Concepts”, Int. Journal of Robotics Research, Vol.26, No.5, pp.475-
[3] R. M. Walter and D. R. Carrier, “Ground forces applied by galloping 490, 2007.
dogs”, J Exp Biol, vol. 210, no. 2, pp. 208-216, 2007. [27] M. Raibert, K. Blankespoor, G. Nelson, R. Playter and BigDog Team,
[4] L. D. Maes, M. Herbin, R. Hackert, V. L. Bels, and A. Abourachid, “BigDog, the rough-terrain quadruped robot”, 17th World Congress
“Steady locomotion in dogs: temporal and associated spatial coordi- The International Federation of Automatic Control Seoul, Korea, July
nation patterns and the effect of speed”, J Exp Biol, vol. 211, no. 1, 6-11, 2008
pp. 138-149, 2008. [28] I. Poulakakis and M. Buehler, “Modeling and experiments of unteth-
[5] Lee, Y-T, H-R Choi, W-K Chung and Y. Youm, “Stiffness Control ered quadrupedal running with a bounding gait: the scout ii robot”,
of a Coupled Tendon-Driven Robot Hand”, IEEE Control Systems, International Journal of Robotics Research, 24(4):239-256, 2005.
pp.10-19, 1994. [29] C. R. Carignan and K. R. Cleary., “Closed-loop force control for
[6] K. Koganezawa, “Back-drivable and Inherently Safe Mechanism for haptic simulation of virtual environments” Haptics-e [Online], vol (2).
Artificial Finger”, Proceedings of Robotics: Science and Systems, Available: http://www.haptics-e.org.
June, 2010, Zaragoza, Spain [30] C. Ott, A. Albu-Schaeffer, A. Kugi, and G. Hirzinger, “On the passivity
[7] Puddy, B., Hunter, I., “Design and optimization strategies for muscle- based impedance control of flexible joint robots”, IEEE Trans. Robot.
like direct-drive linear permanent-magnet motors”, The International Automat., vol. 24, no. 2, pp. 416-429, 2008.
Journal of Robotics Research June 2011 30: 834-845. [31] G. Hirzinger, N. Sporer, A. Albu-Schaeffer, M. Haehnle, R. Krenn, A.
[8] Ishida, T., Takanishi, A., “A Robot Actuator Development With High Pascucci, and M. Schedl, “DLR’s torque-controlled light weight robot:
Backdrivability”, Robotics, IEEE Conference on Automation and III. Are we reaching the technological limits now?”, in Proc. Int. Conf.
Mechatronics, p 1-6, June 2006 Robot. Automat. ICRA, Washington D.C., 2002, pp. 1710-1716.
[9] Goldfarb, M., Barth, E., Gogola, M. Wehrmeyer, J., “Design and [32] S. D. Eppinger and W. P. Seering, “Three dynamic problems in robot
Energetic Characterization of a Liquid-Propellant-Powered Actuator force control”, IEEE Trans. Robot. Automat., vol. 8, pp. 751-758,
for Self-Powered Robots”, IEEE / ASME Trans Mechatronics, Vol. 8, Dec.1992.
pp. 254-262. [33] D. P. Ferris, M. Louie, and C. T. Farley, “Running in the real world:
[10] S. P. Buerger and N. Hogan., “Novel Actuation Methods for High adjusting leg stiffness for different surfaces”, Proceedings of the Royal
Force Haptics”, Advances in Haptics 2010, Mehrdad Hosseini Zadeh Society B: Biological Sciences, vol. 265(1400), 1998.
(Ed.), ISBN: 978-953-307-093-3, InTech. [34] C. T. Farley and O. Gonzalez, “Leg stiffness and stride frequency in
[11] Massie, T., Salisbury, K., “The PHANTOM Haptic Interface: A human running”, Journal of Biomechanics, vol. 29, no. 2, pp. 181-186,
Device for Probing Virtual Objects”, Proc. ASME Winter Annual 1995.
Meeting, Symposium on Haptic Interfaces for Virtual Environment [35] http://www.teslamotors.com/
and Teleoperator Systems. [36] http://www.scorpionsystem.com/
[12] Pratt, G., Williamson, M., “Series Elastic Actuators”, Proc. IEEE/RSJ [37] Ananthanarayanan, A., Azadi, M., Kim, S.,“Towards the Bio-inspired
Int Conf on Human Robot Interaction and Cooperative Robots, 1995, Legs Design for High Speed Running”, Accepted for publication in
pp. 399-406. Journal of Bioinspiration and Biomimetics, July 2012 (accepted)
[13] Kim, B.S., Song, J.B., Park, J.J., “Serial-Type Dual Actuator Unit With
Planetary Gear Train: Basic Design and Applications”, IEEE/ASME
Transactions on Mechatronics, Vol. 15, No. 1, Feb, 2010

1975