The MIT Super Mini Cheetah: A small, low-cost

quadrupedal robot for dynamic locomotion

Will Bosworth1 , Sangbae Kim1 , and Neville Hogan1,2

Abstract— We present the MIT Super Mini Cheetah, a

small (sub-10kg) and low-cost (sub-10k$) quadrupedal robot
for dynamic locomotion. The robot can control vertical and
horizontal force and impedance at each foot and performs
dynamically stable walking, jumping, pronking, turning and
braking using simple force and impedance trajectories. We
present design specifications for dynamic legs which were used
to guide the selection of leg components (e.g., motor, gearbox,
and limb geometry). We demonstrate experimental evidence of
accurate force control during foot-ground impact and present
video of locomotion gaits performed by the robot. To the best
of our knowledge, this is the first small (sub-10kg), power
autonomous quadrupedal robot to perform such a wide range
of ballistic locomotion behaviors and demonstrate foot force Fig. 1: The Super Mini Cheetah (SMC) robot: a small
control during impact. quadrupedal robot capable of power-autonomous dynamic
behaviors such as walking, jumping, bounding, turning and
I. I NTRODUCTION
stopping.
As highlighted by the recent DARPA Robotics Challenge
[1], exploration, clean-up and rescue over dangerous terrain
are important applications for robotic legged locomotion A. Related work
which remain unsolved. Physical robots play a critical role
in controller development for legged locomotion, particularly Motivated by the high force density requirements of
for running and jumping where hard-to-model impact and locomotion, many legged robots use actuators with high
friction dynamics significantly affect system performance. intrinsic inertia and friction. Model-based inverse dynamics,
Because of the challenge in accurately modeling ground often incorporating force sensors at the end of the limb,
properties, the gap between present-day simulation and real- is one approach to control limb force with high intrinsic
robot performance increases as terrain becomes more uneven inertia actuators [2][3]. While inverse dynamic models can
and unpredictable—which is the environment that legged be accurate at low speeds, they are unable to cancel inertial
robots should provide the most value. If they have sufficient effects during large accelerations due to practical bandwidth
dynamic capability, smaller and lower cost robots enable limitations of the actuator and power electronics. Non-
more robot testing over wider terrain types. Such empirical collocated force control also exacerbates the possibility of
study is needed to bridge the gap between simulation and contact instability [4]. Furthermore, these methods require
real robot capability. time-consuming system identification, which reduces the
We introduce the MIT Super Mini Cheetah (SMC), a speed of iterative design of the hardware.
small (sub-10kg) and low-cost (sub-10k$) quadrupedal robot. Series elastic elements are often used in robot legs that are
For its size and cost the robot is unique in its ability to powered by actuators with high intrinsic inertia [5]-[9]. Some
accurately control force and impedance between each foot modern quadrupeds such as Starleth [8] use the Series Elastic
and the ground, as well as generate sufficient force and Actuator paradigm [10] to control foot forces. Other modern
motion for many dynamic behaviors. For example, the robot robots such as the Cheetah-Cub [6] use elastic elements
has jumped higher than its own leg length and performed to store and release energy throughout the gait cycle. In
gaits such as pronking, turning, and walking. The purpose these cases, an important role of series elasticity is to shape
of this paper is to describe the SMC robot design, show its the dynamic response of ground impact. However, selecting
ability to control leg forces, and demonstrate initial dynamic the best leg impedance for different behaviors remains an
behaviors achieved by the robot. These capabilities show that open research problem and modulating the impedance of
the SMC robot is suitable for locomotion testing in a variety a series-elastic leg often requires iterations of mechanical
of environments. hardware which can be prohibitively time consuming. On-
going research in variable impedance actuators is attempting
Corresponding author [email protected] to address this limitation [11].
1 Author is with the Department of Mechanical Engineering, Mas-
sachusetts Institute of Technology, Cambridge, MA 02139. Robot arms such as the Phantom haptic interface have
2 Author is with the Department of Brain and Cognitive Science, MIT. used rigid links powered by current-controlled electric mo-
tors [12]. This design paradigm can result in limbs with low where q, q̇ ∈ Rn are the generalized coordinates/velocities,
intrinsic inertia and friction which is necessary to perform D(q) is the inertia matrix, C(q, q̇) is the Coriolis and
accurate force and impedance control during mechanical centrifugal term, G(q) is the gravity vector, fc (q, q̇, τ ) is the
impact, though achieving sufficient torque density for dy- friction, τ is the input control torques, J(q) is the kinematic
namic legged robots is challenging. The MIT Cheetah is the Jacobian matrix, and Fext ∈ R2 is the horizontal and vertical
only well known robot that extends this design paradigm ground reaction force vector.
to a power-autonomous legged robot that can perform fast A low-inertia leg that acts as a perfect force source gen-
running and dynamic maneuvers [13]-[15], though its custom erates forces through interaction with the ground by directly
motors and components make it an expensive platform to relating motor torque and the leg Jacobian, as described by
replicate at present. Some legged robots use rigid limb links Eq 2.
powered by hydraulic actuators. These systems have in-
creased mechanical system complexity and are quite large— −τ ≈ J T (q)Fext . (2)
in excess of 70 kg for quadrupeds1 [16][17].
Motivated by the success of the MIT Cheetah, we devel- In real robots, this dynamic approximation can be accu-
oped the SMC robot to achieve smaller size, lower cost and rate during relatively low accelerations and when inertia
faster iteration time by emphasizing the use of commercial- and friction are minimized. Therefore, the approach cannot
off-the-shelf components and rapid prototyping methods. To control forces upon impact against the ground but may
describe our design specification and initial experimental approximate desired force during the ground contact phase
results, this paper proceeds as follows: when acceleration is relatively small. The Jacobian relation-
- Section II presents the design requirements of a leg for ship is only accurate if the connecting links between the
accurate force and impedance control and force density torque source and the foot are rigid. If there is significant
required for jumping and landing. intrinsic compliance (damping or stiffness), these energy
- Section III presents an overview of the SMC robot and storage elements must also be considered to describe the
the controller used to achieve the open-loop behaviors leg dynamics. Gear transmission ratios must be minimized,
(jumping, bounding, turning, walking and stopping). as inertia and viscous friction in the motor are reflected onto
- Section IV presents experimental data of the SMC the limb by the square of the gear ratio and dry fiction—i.e.,
robot: control of both vertical and horizontal ground modeled by Coulomb friction—is reflected by the first power
force during static standing; control of vertical ground of the gear ratio.
force during jumping; force and body motion during
steady hopping in place. A video of a variety of loco- B. Limb force and motion during ground contact
motion behaviors is included. The limb forces required for locomotion can be estimated
- Section V presents conclusions and future work. using observations from experimental biology as well as sim-
ple momentum analysis. A wide variety of legged locomotors
II. T HE DESIGN REQUIREMENTS OF A DYNAMIC LEG
exhibit patterned vertical ground reaction force profile shapes
A leg design for dynamic locomotion must balance re- during locomotion; the profile resembles a half sine wave, of-
quirements such as control fidelity, force density, range of ten with additional high frequency dynamics at the initiation
motion, and avoidance of failure modes that arise from foot- of impact. The amplitude of the profile typically increases
ground impact and the thermal losses of actuation. This with forward speed, but the vertical limb force during high
section draws on experimental biology and simple dynamic speed steady running has been measured between two to
models to develop a specification for the mass-specific force, three body weights for a variety of animals [18][19][20].
impedance, and motion needed for a leg that can perform The force profile required for a given system mass, stride
dynamic locomotion. time, and stance time can be estimated using the principle
A. Accurate force control with forward kinematics of momentum balance using a simplified shape of the force
profile [20]. In steady state running, the integral of the
Robots that perform accurate force control without inverse vertical ground reaction force that acts on a body mass must
dynamics can greatly simplify control system design and equal the integral of the gravity force through the duration
implementation. In order to use a kinematic approximation of the stride:
of a limb to accurately relate motor torque to limb force, the
actuators must be able to generate accurate torques and the  T
non-kinematic terms that are present in the dynamics of the fy (t) dt = m g T (3)
limb must be minimized. 0

Consider the dynamics of a planar n-DoF rigid body where T is the stride time, fy is the vertical ground reaction
model: force acting on the body, m is the body mass, g is gravity
and t is time.
D(q)q̈ + C(q, q̇)q̇ + G(q) + fc (q, q̇, τ ) = τ + J T (q)Fext (1)
If the ground reaction force profile is approximated as
1 Design specifications and performance of Boston Dynamics’ robots are a half sine wave with peak amplitude A and stance time
not widely available in academic literature. Their size is publicly available. tstance , as in Eq 4,
 The maximum achievable stiffness and damping through
πt
fy (t) = A sin f or 0 < t < tstance (4a) active control depends on the bandwidth of the actuator’s
tstance torque (or force) control, the precision and accuracy of the
fy (t) = 0 f or t ≥ tstance (4b) measurement of foot kinematics (Δy and Δvy ), the band-
width of the control computer, and the intrinsic impedance
then the peak amplitude of the force profile can be
of the leg structure.
estimated by calculating the integral analytically:
While the best range of impedances for high performance
mgT π locomotion remains an open research problem, the peak
A= . (5)
2 tstance limb stiffness required for locomotion can be estimated
Eq 5 describes how to estimate the peak forces that must from studies of experimental biology [18]. The role of limb
be generated by the legs (peak amplitude A) for a given damping is less documented; the damping ratio of limbs in
system mass, stride time and stance time. The difference large mammals has been observed to approach zero, while
between stride time T and stance time tstance is the time damping ratios of 0.2-0.3 have been measured in smaller
that the body would spend in flight with no legs on the insects [24].
ground. The value A can be divided between multiple legs We estimate that a useful lower limit of damping in a
with simple arithmetic. hopping machine would be one in which the forces generated
Intrinsic to the nature of mechanical work, locomotion during stance through the limb are negligible compared to the
requires limbs to exert both force and motion to move a desired forces that should be exerted on the body to move.
torso [21]. A survey of locomotors shows that the vertical For brevity, we acknowledge that a sufficiently transparent
travel of a limb during locomotion is typically about 20% limb designed to the requirements in Sec II-A will have
of the total limb length [22]. The vertical limb motion l(t) adequately low intrinsic damping.
during stance can be approximated as a half sine wave with
an amplitude Δlmax and stance time tstance . D. Limb motion during swing
πt
l(t) = Δlmax sin (6) The speed of swing-leg motion is dependent on a limb’s
tstance force capability2 , inertia and intrinsic friction. The amount
With this approximation, the peak vertical velocity of the of time available for swing leg return actions is related to
limb, vy,max , is estimated as, the stride frequency and duty cycle of a gait. We draw on
observations of biology to estimate the typical times and
Δlmax π
vy,max = . (7) motions available for this action [18][19][20].
tstance
The required torque and operating speeds for the swing
The rotational speed of the hip joint during stance can be leg motion can be estimated assuming a simple bang-bang
estimated as the quotient of the forward running speed and controller. The magnitude of the torque required to swing the
the length of the leg. leg forward, τs , is given in Eq 9. This value is dependent on
C. Limb stiffness and damping before and during ground limb rotational inertia Ileg , swing time tswing , and leg angle
contact to be swept Δθs .
Mechanical impedance control is used widely on robots
4 Δθs Ileg
involved in intermittent mechanical contact, of which legged τs = (9)
locomotion is a flagship example. While there is no con- tswing 2
sensus that pure impedance control is the most appropriate
From Eq 9, the peak joint speed to return the leg after
method to approach locomotion, many studies have shown
stance, ωs,max , is shown in Eq 10.
that tuned limb impedance can aid robust, stable, and en-
ergetically efficient locomotion [2,5-9,13-15]. In particular, τs tswing
there appears to be significant opportunity to combine direct ωs,max = (10)
2 Ileg
force profile control with tuned limb impedance [15].
If a limb can adequately approximate a force source, as E. The workspace of the leg
discussed in Sec II-A, then impedance of the limb can be
commanded as a function of the position and velocity of The functional workspace of a leg includes the range of
the limb. For example, the vertical foot force that results motion of the limb and the positions over which the forces
from impedance control, fy,imp [N], can be commanded as required for locomotion can be generated. A wide range of
a function of stiffness gain Ky [N/m] and damping gain By animals use hip rotations of +/- 30◦ from vertical during
[Ns/m] as a function of deviation from a nominal leg length ground contact [18]. Vertical travel of the limb during ground
Δy [m], and deviation from nominal leg velocity Δvy [m/s] contact has been observed to be 20% of the limb length [22].
[23].
2 Stored elastic energy could be released during recirculation, which could
fy,imp = Ky Δy + By Δvy (8) be considered part of the available force/torque for the limb action.
F. Mechanical failure the stride interval T is shown in Eq 13. The value for A and
The primary mechanical failure modes in a robot leg are the RMS torque may be dependent on the number of legs
structural failure due to high impact forces, and thermal and the number of motors used per leg.
failure of the motors due to heat dissipation. All elements    
in the drivetrain including gear teeth, bearings, structural A tstance
im,RM S = (13)
connections and rigid links must be able to withstand the km Gm Jy† 2T
cyclic loads generated during ground contact. The forces
III. T HE S UPER M INI C HEETAH ROBOT
required for locomotion discussed in Sec II-B provide a
baseline for the load requirements of the limb structure. A. Leg design: motor and structure
The design requirements introduced in Sec II-A also help The SMC robot leg draws on the design of the MIT
to minimize the impact forces upon contact with the ground. Cheetah robot which uses low-gear ratio electric motors
Introduction of foot compliance mitigates impact forces but connected by rigid limb links. A solid model of a single
compromises the force control bandwidth of the leg. SMC robot leg is shown in Fig 2. Two DC motors are
In an electric motor, thermal failure is caused by heat that mounted in a single housing, which is then mounted to the
is generated while creating large torques. The heat power torso of the quadruped. The rigid links of the leg couple the
generated as a function of motor current and motor resistance outputs of the two motor shafts to the foot, creating a five-bar
is given in Eq 11, where Pjoule is the Joule heating power closed chain linkage which has been seen in robots such as
[W], im is motor current, and Rm is motor resistance. ATRIAS [7]. This parallel motor layout allows both motors
to share vertical and horizontal foot forces, decreasing the
Pjoule = im 2 Rm . (11) motor torques needed to support high vertical loads, which
decreases the motor shaft radial load requirements. The links
Using thermal characteristics provided by the manufacturer, are manufactured using 3d printed ABS plastic; the rotational
the designer should write a safety algorithm that keeps the joints of the links are made with plain bearings and shoulder
estimated temperature of the motor below the operating screws. Each foot contains a contact sensor made using a
temperature. This algorithm typically requires estimating the pressure sensor embedded in rubber [25].
RMS current through the motors. The design of the leg proceeded iteratively following
To estimate the RMS motor current that would be used the process described in Fig 3, which refers to the design
to achieve a jumping gait of desired stride time and stance requirements in Sec II. First, a candidate motor and gearbox
time, we combine the ground-force estimates of Eqs 3-5, was selected which appeared to have adequate torque, inertia,
the relationship of motor torque and foot force in Eq 2, a friction, and load capacity. The mass of the motor determined
motor’s torque constant km and gear transmission ratio Gm , an estimate of the total mass of a quadruped. This estimated
and the Jacobian of the leg. The required torque for swing mass was used to calculate the required force and motion
leg return (Eq 9) could also be included, but we note that capabilities of the limb, which enabled a search for adequate
a leg designed with intrinsically low inertia will not require limb geometry. Detailed design of limb links was guided
significant swing-leg torque for many behaviors. by structural loading requirements (Sec II-F). Verification of
Recalling Eqs 3-5, the amplitude of vertical force A can the thermal loading was performed using the approximation
be calculated from the stance time tstance , the body mass m described in Eq 13 as well as dynamic simulation of single
and the desired stride time T . Having calculated amplitude leg hopping and quadrupedal bounding. Table I lists the
A in Eq 5, the required vertical force profile fy (t) would commercial-off-the-shelf components used in the SMC robot.
follow the half-sine wave shape of Eq 4.
The motor current im required to generate a vertical force B. Leg design: Power electronics and computation
is a function of the motor’s torque constant km [Nm/A], the Implementing endpoint force and impedance control on
gear transmission ratio Gm connected to the motor and the the SMC robot leg requires accurate measurement of the
leg Jacobian3 Jy which acts as an additional transmission motor shaft positions and a high-bandwidth control computer
from motor torques to foot force. The instantaneous vertical to calculate the configuration dependent leg Jacobian in real-
force fy as a function of motor current and leg properties is: time and calculate feedback motor torques. Each motor is
driven by a high-bandwidth current controller and fitted with
fy = km im Gm Jy (q) (12) a precise motor encoder. A microcontroller is used to read
The RMS current im,RM S is found by equating the sensors, perform computation, and send torque commands to
vertical force in Eq 12 to the required vertical force profile each motor; the total loop time during quadrupedal locomo-
that follows from Eqs 4 and 5. Using the half sine wave tion is approximately 350 μs. The current controllers provide
approximation of Eq 4 and a constant, conservative estimate accurate current tracking at better than 1 kHz.
of the vertical Jacobian, Jy† , the required current profile can C. Quadrupedal robot design
be solved analytically. The resulting RMS motor current over
The SMC robot (Fig 1) was built after first testing a single
a planar leg, the Jacobian J has horizontal and vertical components. In
3 In leg and a planar-constrained biped robot (with one front
Eqs 12 and 13, Jy corresponds to the vertical components of the Jacobian. and rear limb). The quadruped contains four identical leg
 D. Control system design
Motor
Rubber
The gaits presented in this paper are controlled using force
L1
stop and impedance commands to individual legs; the commands
L0 are organized by sequential state machines which are trig-
gered by timers or ground contact sensing, as in Fig 4. Prior
work has demonstrated stable locomotion with simple con-
L2 trollers that primarily rely on impedance (e.g., [6][14][26]).


Other hardware studies have shown that active force and
Foot pad
impedance control concepts can be combined to exploit
advantages of both approaches (e.g., [15]). The experiments
Fig. 2: A rendering of the SMC leg. Two motors are mounted performed in this work are used primarily to demonstrate the
on the “shoulder” which attaches to the torso (not shown). dynamic capability of the new SMC robot, though the ability
The leg structure consists of four rigid plastic links, resulting to extend a simple force and impedance controller to a variety
in a 5-bar linkage leg. The dimensions of the leg are: L0 = of behaviors—both forward motion and turning—provides
60, L1 = 60, L2 = 145 mm. further evidence for the potential to control locomotion using
both force and impedance commands.
A bound gait is characterized by synchronized ground
(YDOXDWHVWUXFWXUDODQGWKHUPDO
6HOHFWPRWRUDQGJHDUER[FDQGLGDWHV LQWHJULW\RIOHJ6HF,,I
contact of the front limbs and rear limbs respectively. The
forward and turning bounding gaits use single-leg state
9HULI\VZLQJPRWLRQRIOHJ6HF,,G machines which define swing and stance phases, shown
(VWLPDWHWRWDOV\VWHPPDVV
DSSUR[;PRWRUPDVV in Fig 4. The front and rear limb pairs are synchronized
(VWLPDWHLQWULQVLFIULFWLRQDQGLQHUWLDRI
OHJ6HF,,F by coupling the ground contact state of both limbs – for
example, if either front limb enters the ground state the other
(VWLPDWHIRUFHDQGPRWLRQ 'HVLJQOLQNDJHJHRPHWU\WRDFKLHYH
UHTXLUHPHQWVJLYHQV\VWHPPDVV6HF IRUFHDQGZRUNVSDFHUHTXLUHPHQWV limb must also enter the ground state.
,,E 6HF,,HVHH)LJ
A simple design methodology is used to select the force
and impedance commands to be applied during the ground
Fig. 3: A block diagram of the design process used to guide contact states:
the selection and evaluation of a motor, gearbox, and limb 1) Open-loop force profiles are designed to achieve de-
geometry for the SMC robot leg. sired open-loop behavior (as in Eqs 4 and 5).
2) Joint impedance, applied in parallel to the force profile,
TABLE I: Off-the-shelf electromechanical parts used in the provides a spring-damper “suspension” to stabilize the
robot open-loop behavior.
Vertical force profiles were designed using the concepts
Component name Manufacturer part
described in Sec II-B for hopping and running. Forward
Motor Faulhaber 3272, 32mm φ
motion was generated by applying positive horizontal force
Gearbox Faulhaber 32-3S, 23:1
profiles to the legs, while turning was achieved by treating
Shaft encoder CUI AMT103
the horizontal force profiles like a “tank-drive” system: if
Current controller Maxon Escon Module 50/5
limbs on the left side push forward, and limbs on the right
Microcontroller Cypress PSoC 5LP
side push backwards, a net turning moment will be applied
Battery Thunderpower 70C 325 mAh 3S; (8x)
to the machine. Impedance was selected empirically using a
Inertial measurement unit Vectornav VN-100 Rugged
simple simulation model and the SMC robot; prior research
[27] and high-speed video analysis provided useful insight
to select the impedance.
Some important locomotion behaviors, such as performing
braking from a drop test or forward running, may be better
modules attached to a rigid torso frame built from aluminium described using only limb impedance control. A three-
extrusion, and a PCB designed to hold the eight current legged walking gait, which could be implemented if one leg
controllers and the microcontroller. Two 12-cell (44.4 V failed, was quickly designed by moving sequentially through
nominal) battery packs were assembled from smaller COTS statically stable poses that were achieved using only limb
3-cell LiPo batteries; the drive electronics of the front and impedance commands.
rear legs operate on independent battery packs in order to
limit the voltage drop across a single battery pack. IV. H ARDWARE EXPERIMENTS
The complete SMC robot, including batteries, weighs This section presents experimental data of the SMC robot
approximately 9 kg. The limbs are 30 cm apart in length performing force control during static standing and during
and 18 cm apart in width. The leg length can range from 10 hopping. We describe tests to characterize the impedance
to 20 cm. range of the limb. Additionally, experimental data of body
 Fig. 5: Time-lapse of the SMC robot performing a pronking gait, taken at intervals of 0.167 s.

30
Fz Fx Fy Fz cmd Fx cmd

20

Force [N]
10

Fig. 4: Schematic of the single-leg state machine used to -10

organize swing and stance phases for the bounding and 0 5 10 15 20
turning behaviors described in this paper. Each state contains Time [s]
a single position command and set of impedance gains; the
stance phase also contains an open-loop force profile for Fig. 6: The results of a static force testing experiment
vertical and horizontal forces. The sequential progression measuring the performance of one leg while the SMC robot
between the states is shown by the arrows; the arrow text was standing upright. A constant vertical force (Fz ) was
describes each exit condition. Individual leg state machines commanded and the horizontal force (Fx ) followed a simple
can be combined by coupling state transitions: in a bound, if pattern of step commands. The dotted lines describe the
one front foot touches the ground (initiating the stance state), commanded foot forces and the solid lines show measured
the other front foot will also begin its stance state. force. The results show that vertical and horizontal force can
be commanded independently, though there is hysteresis in
the horizontal force.
rotation, leg kinematics, and motor torque are shown during
a hopping gait. A video attachment of the the SMC robot
performing a variety of gaits is included with this paper measured by an external load cell as well as the commanded
submission. A time-lapse of one such running gait is shown forces. The data shows that vertical force and horizontal
in Fig 5. force can be commanded independently of each other and
that the forces generally track the commanded value, though
A. Force control during standing hysteresis is visible in the horizontal force output.
Static force tests of the limb were performed to test the
ability of the leg to independently command vertical and B. Leg impedance range
horizontal ground forces. In these tests, the robot was stand- Swing-leg motion profiles were used to characterize the
ing on four legs unconstrained by any additional supports, leg impedance. In these tests, the body was fully constrained
with one leg in contact with a multi-axis load cell (ATI and the limb was not touching the ground. A lumped
SI-660-60). The robot could stand upright on three legs by parameter model of the leg was estimated by measuring the
commanding static position and impedance setpoints to these natural frequency and damping ratio of the step response
limbs. The fourth limb, which was standing over the load of the motor joint for a variety of stiffness gains (Eq 8).
cell, was then free to command forces. A constant vertical The reflected damping and inertia of the vertical travel of
force was commanded to push the foot into the ground and the foot with respect to the body is estimated as 13.3 Ns/m
a simple pattern of horizontal forces was commanded. The and 2.3 kg respectively. The intrinsic limb impedance is
magnitude of the horizontal force was selected so that the configuration dependent; these impedance values represent
foot would not slip; the magnitude of the vertical force was the leg at a length of approximately 15 cm, halfway between
selected to be approximately a quarter of the machine body its maximum length of 20 cm and minimum length of 10
weight (20 N) so that the robot would stand still. cm. Though experiments verifying the accuracy of the limb
Fig 6 shows vertical and horizontal ground reaction forces impedance at the foot have not been performed, the leg can
 80 10

Body angle [deg]

Fz roll
Fx 0 pitch
60 Fy yaw
Fz cmd
-10
Force [N]

40
-20
0 1 2 3 4
20 (a) Time [s]
0

Motor angle [deg]

1
0 -20 2
3
0 200 400 600 800 4
-40
Time [ms]

Fig. 7: The results of a vertical hopping test, showing the per- -60
0 1 2 3 4
formance of one leg. The data shows four consecutive stance (b) Time [s]
events. The dotted line shows the vertical force command;
6
commanded horizontal force was zero. The results show that

Motor current [A]

4 1
measured ground reaction force (Fz ) follows the commanded 2
force with some additional high frequency content, and the 2 3
horizontal force (Fx ) and out-of-plane force (Fy ) remain near 4
0
gs
the desired zero magnitude.
-2
0 1 2 3 4
(c) Time [s]
generate stable impedance commands corresponding to 4
kN/m stiffness and 4 kNs/m damping.
Fig. 8: Data acquired on board the SMC robot during
C. Force control during hopping hopping. (a) shows the angular orientation of the body and
Fig 7 shows the commanded and measured ground forces shows that yaw angle (heading) is not well regulated. (b)
during a vertical hopping experiment. In this trial the two shows motor angles for the four motors of the front limbs
back legs were commanded to stand upright using constant (motors 1 and 2 are part of one leg; 3 and 4 another). (c)
position and impedance setpoints. The two front limbs used shows the commanded motor current during jumping; the
controllers described in Fig 4: when the foot sensor measured dashed line shows the result of one ground contact sensor
ground contact (which requires approximately 10N) each (the ground contact sensor is a boolean value and scaled to
leg commanded a feed-forward ground force trajectory. No fit the other values on the graph). The data shows that the
additional impedance command was used during the stance robot can hop persistently using RMS motor currents below
state. The force trajectory only contained vertical force; the 2 A, which is within the ratings of the motors and drive
horizontal force command was zero. components.
The data in Fig 7 shows four consecutive steps of the
two-leg hopping behavior. The measured ground reaction
force resembles the commanded force profile in both duration Fig 8 shows body orientation (roll, pitch and yaw). The
and magnitude, with some high-frequency content in the data shows that pitch and roll angle were regulated to within
measured ground force. +/-5 degrees of horizontal. The yaw angle (heading) exhibits
drift as it is not regulated in the present controller. Angle
D. Body motion, leg motion, and motor current during and torque of four motors (two for each of the front legs)
hopping shows approximately 40 degrees of travel of each limb and
a peak current of 5 A. During this experiment, each leg was
Fig 8 shows data that was acquired on-board the SMC in contact with the ground for about 0.1 s for each 0.18 s of
robot while performing hopping in place on all four legs. flight time.
The data shows 11 strides of data over approximately three
seconds—data acquisition is limited by memory on the E. Bounding, turning, braking and walking
microcontroller. Data was acquired at 250 Hz. This gait was Numerous gaits have been performed using the controller
controlled using a controller as described in Fig 4, though described in Sec III-D; a time lapse of one forward pronking
the two front legs and two rear legs were coupled together— stride is shown in Fig 5. Forward gaits were achieved
i.e., if one front leg detected ground contact, both legs would using stance times of between 100 and 200 ms with stride
enter the stance control state. Though the front and rear legs frequencies on the order of 3 Hz. Amplitude of the vertical
were not synchronized in the controller, the resulting gait force profile in each leg (as in Eq 4) was between 60 to
resembled a pronking gait. 80 Nm and satisfactory gaits were found using vertical limb
impedances from 1 to 1.5 kN/m and up to 60 Ns/m. Forward R EFERENCES
speeds of nearly 1 m/s were achieved. Though rigorous [1] ”What Is the DARPA Robotics Challenge (DRC)?”,
characterization of stability has not been performed, the DARPA Robotics Challenge., n.d. Web. 10 July 2015,
robot exhibits dynamic stability in roll and pitch directions http://www.theroboticschallenge.org/overview.
[2] Sreenath, K., et al. ”A compliant hybrid zero dynamics controller for
which can be observed in the video attachment. The heading stable, efficient and fast bipedal walking on MABEL.” IJRR (2011).
orientation of the quadruped is not yet controlled with [3] Grebenstein, M, et al. ”The DLR hand arm system.” IEEE ICRA 2011
feedback, though intentional turning can be controlled by [4] Eppinger, S, and Seering W. ”Understanding bandwidth limitations in
robot force control.” ICRA 1987.
modifying the horizontal forces of the forward bounding gait [5] Saranli U., Buehler M., and Koditschek D., ”RHex: A simple and
to apply a net turning motion on the torso. Turning rates of highly mobile hexapod robot.” IJRR 20.7 (2001): 616-631.
π [6] Sprwitz, Alexander, et al. ”Towards dynamic trot gait locomotion:
3 rad/s were achieved. Design, control, and experiments with Cheetah-cub, a compliant
The machine can perform repeatable braking maneuvers quadruped robot.” IJRR 32.8 (2013): 932-950.
by switching to an impedance-only control for each limb’s [7] Grimes, Jesse A., and Jonathan W. Hurst. ”The design of ATRIAS 1.0
a unique monoped, hopping robot.” CLAWAR 2012.
stance phase. Applying a limb impedance command of 1.2 [8] M. Hutter, C. Gehring, M. Bloesch, M. Hoepflinger, C. D. Remy, R.
kN/m and 60 Ns/m brought turning and bounding gaits to Siegwart, StarlETH: A Compliant Quadrupedal Robot for fast, effi-
rest within a single step. Additionally, walking gaits have cient, and versatile Locomotion, Proc. of the International Conference
on Climbing and Walking Robots (CLAWAR), 2012.
been developed using impedance trajectories only. [9] Raibert, Marc, et al. ”Bigdog, the rough-terrain quadruped robot.”
Proceedings of the 17th World Congress. Vol. 17. No. 1. 2008.
V. C ONCLUSION & F UTURE W ORK [10] Pratt, Gill, and Matthew M. Williamson. ”Series elastic actuators.”
IEEE/RSJ IROS 1995.
[11] Vanderborght, Bram, et al. ”Variable impedance actuators: A review.”
We developed the SMC quadrupedal robot to address the Robotics and Autonomous Systems 61.12 (2013): 1601-1614.
opportunity that small, low-cost, dynamic robots can play [12] Salisbury, K., and Srinivasan M. ”Phantom-based haptic interaction
in understanding locomotion: empirical robot study is of with virtual objects.” Computers Graphics and Applications, IEEE 17.5
(1997): 6-10.
particular use for hard-to-model environments such as highly [13] Seok, S, et al. ”Design principles for highly efficient quadrupeds and
cluttered and variable terrain. The SMC robot was made with implementation on the mit cheetah robot.” IEEE ICRA 2013.
commercial-off-the-shelf electromechanical components and [14] Hyun, Dong Jin, et al. ”High speed trot-running: Implementation of a
hierarchical...”, IJRR 33.11 (2014): 1417-1445.
modern rapid prototyping methods to decrease cost and [15] Park, H, Chuah M, and Kim S. ”Quadruped Bounding Control with
increase replicability. We demonstrated the robot leg’s ability Variable Duty Cycle via Vertical Impulse Scaling.” (IEEE IROS 2014).
to perform force and impedance control during locomotion [16] Semini, C, et al. ”Design of HyQa hydraulically and electrically
actuated quadruped robot.” J of Sys & Cont Eng (2011).
and showed the dynamic capability of the robot with data [17] Ackerman E., (2015 Feb 10). “Spot Is Boston Dynamics’
acquired by the robot while performing hopping as well Nimble New Quadruped Robot”. Retrieved September 2015 from
as with video demonstration. The benefit of limbs that are http://spectrum.ieee.org/automaton/robotics/robotics-hardware/spot-is-
boston-dynamics-nimble-new-quadruped-robot.
capable of force and impedance control is made apparent by [18] Farley, C. et al. ”Running springs: speed and animal size.” Journal of
the variety of locomotion behaviors that were demonstrated experimental Biology 185.1 (1993): 71-86.
with a very simple design methodology. The ability to [19] Weyand, Peter G., et al. ”Faster top running speeds are achieved
with greater ground forces not more rapid leg movements.” Journal
generate such a variety of gaits bodes well for using the of applied physiology 89.5 (2000): 1991-1999.
SMC robot to study gaits on numerous terrain types and [20] Hudson P et al., ”High speed galloping in the cheetah (Acinonyx
environments. jubatus) and the racing greyhound (Canis familiaris): spatio-temporal
and kinetic characteristics.” J exp bio 215.14 (2012): 2425-2434.
Immediate continued work includes characterizing the [21] Hogan, Neville. ”Mechanical impedance of single-and multi-articular
stability of open-loop gaits such as those described in systems.” Multiple muscle systems. Springer NY, 1990. 149-164.
this paper, as well as further integrating state estimation [22] Blickhan R., and Full R. ”Similarity in multilegged locomotion:
bouncing like a monopode.” J of Comp Phys A 173.5 (1993): 509-517.
to enable complete model-based feedback control. Further [23] Hogan, Neville. ”Impedance control: An approach to manipulation.”
characterization of the force and impedance control of the American Control Conference, 1984. IEEE, 1984.
leg on different ground surface types is ongoing. We intend [24] Garcia, Mariano, et al. ”Damping and size: Insights and biological
inspiration.” ISAMAM 2000
for the SMC robot to be a platform for a broad range of [25] Chuah, M, and Kim S. ”Enabling force sensing during ground locomo-
legged applications; as such, we documented the mechanical tion: A bio-inspired, multi-axis, composite force sensor using discrete
requirements of the limb to provide a guide for further pressure mapping.” Sensors Journal, IEEE 14.5 (2014): 1693-1703.
[26] Pratt, J et al. ”Virtual model control: An intuitive approach for bipedal
iteration of the leg platform for different dynamic locomotion locomotion.” IJRR 20.2 (2001): 129-143.
goals or size scales. [27] Bosworth W, Kim S, and Hogan N. ”The effect of leg impedance on
stability and efficiency in quadrupedal trotting.” IEEE IROS 2014

VI. ACKNOWLEDGEMENTS

The authors thank the DARPA Maximum Mobility and

Manipulation (M3) Program and the Louis G. Siegle Fel-
lowship for supporting this work.
The authors thank Jonas Whitney, Debbie Ajilo, Michael
Farid, and Michael Chuah for significant contributions to the
SMC robot development.