Mit Cheetah Actuator PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Author’s proof: Please see official final version here http://dx.doi.org/10.1109/TRO.2016.

2640183

Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth
Physical Interaction for Dynamic Legged Robots
Patrick M. Wensing1 , Albert Wang1 , Sangok Seok2 , David Otten3 , Jeffrey Lang3 , and Sangbae Kim1

Geared*Motor*with*
Abstract—Designing an actuator system for highly-dynamic Geared
Geared
Geared MotorMotor
Motor with with Torque(Force)
Force/Torque*Sensor*
withTorque(Force)
Torque(Force) Series*Elas(c*Actuator*
SeriesSeries
Series ElasticElastic Actuator
Actuator
legged robots has been one of the grand challenges in robotics Sensor
Sensor
Sensor
Elastic Actuator
research. Conventional actuators for manufacturing applications High Gear Ratio High Gear Ratio
High
HighGear
GearRatio
Ratio High
HighGear
GearRatio
Ratio Spring,Spring,
Spring,
have difficulty satisfying design requirements for high-speed Transmission
Transmission Stiff Stiff
Transmission Stiff TransmissionEncoder
Transmission
Transmission EncoderEncoder
locomotion, such as the need for high torque density and the SensorSensor
Sensor
Leg
Leg Leg Leg
Leg Leg
ability to manage dynamic physical interactions. To address motor
motor motor motor motor
motor
this challenge, this paper suggests a proprioceptive actuation
paradigm which enables highly-dynamic performance in legged
machines. Proprioceptive actuation uses collocated force control
at the joints to effectively control contact interactions at the (a)(a)%%(a)
(a) High torque (b) (b)
(b)(b)%%
High
Hightorque
torque
feet under dynamic conditions. Modal analysis of a reduced density density
motor motor
densitymotor
leg model and dimensional analysis of DC motors address the
main principles for implementation of this paradigm. In the
realm of legged machines, this paradigm provides a unique Propriocep(ve**
Proprioceptive
Proprioceptive
Proprioceptive
combination of high torque density, high-bandwidth force control, actuator
Actuator*
actuator
actuator Low Lowleg
Lowinertia
inertia inertia
leg leg

and the ability to mitigate impacts through backdrivability. We


introduce a new metric named the ‘impact mitigation factor’
(IMF) to quantify backdrivability at impact, which enables design (c)%% Low Low
LowGear
Gear Gear Ratio
Ratio
Ratio
Transmission
Transmission
Transmission
comparison across a wide class of robots. The MIT Cheetah leg (c) (c)
(c)
is presented, and is shown to have an IMF that is comparable
Fig. 1. Three different electromagnetic (EM) actuator concepts. (a) High-ratio
to other quadrupeds with series springs to handle impact. The geared motor with torque sensor, (b) Series elastic actuator, (c) Proprioceptive
design enables the Cheetah to control contact forces during force control actuator.
dynamic bounding, with contact times down to 85 ms and peak
forces over 450 N. The unique capabilities of the MIT Cheetah,
achieving impact-robust force-controlled operation in high-speed
3D running and jumping, suggest wider implementation of this often conflicting requirements on the design of EM actua-
holistic actuation approach. tors. Design processes for EM components are traditionally
governed by performance metrics such as the motor con-
I. INTRODUCTION stant, torque constant, peak torque, continuous torque, thermal
capacity, and others [1]. Beyond the EM components, the
Physical interactions with the environment play a crucial transmission (speed reducer) is a critical additional component
role in many emerging applications of robotics. Whether in which affects the control bandwidth and ability to handle
legged locomotion or disaster response, the need to control collisions with the environment [2]. Multi-objective design
and exploit interaction with the environment introduces unique trade offs thus need to be weighed in the selections of both
challenges to actuator design. As we focus on these applica- the EM components and transmissions. Three common EM
tions, it is imperative to shift our actuator paradigms away actuator concepts that manage these trade offs for robotics are
from technologies that have been optimized for the assembly shown in Fig. 1 and are discussed throughout the paper.
line floor. Most manufacturing robots have been designed to We propose to address this intricate actuator design problem
perform rapid and accurate pick-and-place position control through a new paradigm for EM actuators in legged machines
sequences in well-known environments. Future mobile robots, called proprioceptive actuation. As a key feature, this allows
however, must be capable to manage broader dynamic and ‘proprioceptive’ foot force control through internal torque
force-critical interactions in unstructured environments. The control. This can be accomplished without exteroceptive sen-
actuator paradigm presented in this paper offers an approach sory feedback that is known to cause non-collocated sensing
for electromagnetic (EM) actuators to address these needs. problems upon collision which result in contact instability [3].
The design of actuators to manage physical interactions The paradigm of proprioceptive actuation has had success
is particularly challenging for application to dynamic legged in low-force haptic devices such as the PHANTOM [4], but
locomotion. Legged locomotion involves repeating dynamic previously has not been employed for high-force applications
events such as impact, high-force interaction with uncertain such as legged locomotion. Overall, the paradigm eliminates
terrain, and rapid leg swing. These events place unique and the need for physical springs, stiffness modulating mecha-
Authors are with the 1 Department of Mechanical Engineering at the nisms, and force/torque sensors, and enables high-bandwidth
Massachusetts Institute of Technology, Cambridge, MA, 02139, USA, 2 Naver force control without contact force feedback.
Corporation, Seongnam, South Korea, and the 3 Department of Electrical Proprioceptive actuation is a unique paradigm to address
Engineering and Computer Science at the Massachusetts Institute of Tech-
nology, Cambridge, MA, 02139, USA. corresponding email: sangbae at many of the design challenges facing legged locomotion.
mit.edu These challenges are presented in Section II with a summary
of the previous approaches developed to handle them. Through The power of the motor often becomes a major metric in
modal analysis of a simple prototype of a legged system, designing motors for traditional applications. However, the
and dimensional analysis of motor performance across scale, high power achieved by high-speed continuous operation does
Section III shows that the proprioceptive paradigm is uniquely not apply to legged machines, where operation is mostly
situated to manage high-force, high-bandwidth contact interac- intermittent and varying directions continuously. In fact, the
tions imperative to high-speed locomotion. We view handling power density of EM actuators themselves (continuous up
impact as the most difficult of these contact interactions, to 7 kW/kg [10], 3-5 kW/kg [11]) can significantly exceed
and propose a new metric in Section III-B called the Impact that of biological muscle (Max. 0.3 kW/kg) [12], and of
Mitigation Factor (IMF). The IMF quantifies the normalized most manufacturer-recommended operating specifications for
inertial impedance of a floating-body robot, capturing the commercial motors.
effects of actuator design to reduce impulsive forces at impact.
The MIT Cheetah design, presented in Section IV, embodies B. Energy Efficiency
the principles of proprioceptive actuation. By minimizing the
Energy efficiency is a critical metric for legged machines.
overall impedance of the actuator, the design is able to mitigate
Locomotion is a typically energy dissipative process and the
impacts comparably to existing quadrupeds with series elastic
loss mechanisms are categorized into three major modes:
actuators (SEAs), while maintaining high-bandwidth open-
Joule heating, transmission, and interaction losses [13]. We
loop force control (Section V). A discussion in Section VI de-
previously described design principles to minimize these losses
tails the high-level benefits on this paradigm which embraces
for highly-efficient legged robots. Their implementation, in the
a notion of ‘less is more’ in mechanical design. We believe
design of the MIT Cheetah, has led to a machine that rivals
this approach provides a robust and practical new paradigm
the efficiency of animals [13]. However, this design was not
for legged machines.
explicitly optimized for efficiency.
Portions of this article (in Section III-C) have appeared
Transmission selections in particular have coupled implica-
previously in a conference paper [5]. Significant additional
tions for Joule heating and interaction losses. Low gear ratios
theoretical analysis and experiments are reported here beyond
will reduce the reflected inertia but can cost higher energy
the previous work.
in generating torque via Joule heating. On the other hand, if
the gear ratios used in a robot are too high for its application,
II. C HALLENGES IN ACTUATOR DESIGN FOR HIGH - SPEED they may improve energetics in theory but can break a leg upon
LEGGED ROBOTS contact or even could prevent a desired dynamic motion due to
excessive mechanical impedance. This tradeoff should further
One of the main challenges in actuator design is to identify
be considered based on the robot’s functional requirements
critical metrics for the target application and tradeoffs between
such as payload, required travel distance per charge, and
the metrics across design parameters. For dynamic legged
travel speed. For instance, a load-carrying robot walking at
locomotion, we identified the following key actuator char-
slow speeds could energetically benefit from a higher gear
acteristics: 1) high mass-specific torque (torque density), 2)
ratio in comparison to an agile robot running at high speeds.
efficiency, and 3) impact mitigation capability. The following
Gear efficiencies and frictional properties play a large role
subsections describe these three aspects in more detail.
in balancing energetic tradeoffs. Yet, while there exist many
methods to identify gear energetics of a fabricated design
A. Torque density [14], accurately modeling gear energetics a priori for design
optimization is complicated by a variety of factors [15]. Still,
Achieving high torque with a minimum actuator mass is
whereas energy efficiency is a metric we should seek to
critical in running robots. This is due to in part to the high
maximize, impact mitigation is an essential requirement.
ground reaction forces (GRFs) required to propel the body.
The maximum normal GRF on each leg is about three times
the bodyweight in a human running at 4.5 m/s [6]. In data for a C. Impact mitigation
dog galloping at 9 m/s [7], it reaches 2.6 times the bodyweight While satisfying the torque and power requirements, the leg
during a ground phase of approximately 70 ms (20% of gait should be able to mitigate impact forces upon collision. There
period). While long legs provide a large workspace, they are a number of approaches in the literature which may use
intensify the torque requirements to produce these forces. active control or passive mechanical dynamics to minimize
To meet GRF requirements at the foot, large moment arms collisional forces surrounding an impact event.
dictate high joint torques. For example, the maximum torque Researchers have developed several approaches that employ
at the knee required for a 75 kg human to jump is around impedance control with high-gear-ratio actuators to achieve
200 Nm [8]. This is remarkable considering that the torque force reduction after impact. Apparent impedance can be
from a 1270 kg Toyota Corolla is 170 Nm at 4000 rpm [9]. modulated by implementing impedance control [16] with
Simply employing large motors to meet torque requirements torque sensors at each joint [17], [18]. However, this control
is not viable in mobile robot design. Higher actuator masses approach used with high-gear-ratio actuators has not demon-
contribute to GRF requirements, and thus increase torque strated robustness to impacts or high-bandwidth force control
requirements. This relationship highlights the importance of capabilities for dynamic legged locomotion. Previous work
maximizing torque density for actuators in legged machines. has addressed human-robot impacts in industrial settings with
any residual structural stiffness of the leg. We assume that this
mb mb interface spring has no preload. This assumption is captured
in the model by placing the origin in the vertical direction at a
height which corresponds to the spring’s rest length at impact.
✓ m` m` Letting q = [yb , ✓]T , the dynamics of the leg model follow
ks
ŷ ŷ
I I H q̈ + K q + g = ST ⌧ (1)
ki ki from standard Lagrangian methods, with
h i h i
mb + m` rm` (mb + m` )g
H= rm` 2
m` r + I g = m gr (2)
`
f f h i
(a) (b) ki rki
K = rk r2 k S = [ 0 1] . (3)
i i
Fig. 2. Prototype systems to understand the effects of actuation and leg
design on mechanical bandwidth and impact force magnitude. (a) Prototype A modal analysis of the system is tractable, providing insight
with rigid transmission and (b) prototype with series elastic actuation (SEA). into both the impact dynamics and the mechanical bandwidth
The SEA rotational actuator is assumed connected to mass b, with the SEA
spring between the actuator output and the pinion. of the leg. This simple linear model captures many features
pertinent to the design of multi-DoF legs, such as the effects
of reflected actuator inertia on the GRFs following impact.
the DLR Lightweight Robot III [19]. Initial impact events The ground reaction force f (t) onto the leg is provided by
are detected, and effective control post impact is shown the interface spring
to reduce overall collision severity. As a main difference,
f (t) = ki y` (t) = [ki rki ] q(t) . (4)
proprioceptive actuation provides a mechanical approach to | {z }
minimize the severity of initial and post-impact forces, which :=C

can otherwise be detrimental in high-speed collisions without Letting s be the Laplace variable, the Laplace transform
sufficient compliance. of the ground reaction force F(s) is related to the Laplace
The series elastic actuation (SEA) paradigm [20], [21], transform of the torque T (s) and initial conditions through:
[22] has been employed in legged machines to reduce ini- ⇥ ⇤
F(s) = C(s2 H + K) 1 ST T (s) g/s + H(sq0 + q̇0 )
tial rigid-body impacts by purposefully adding mechanical
(5)
elasticity in series with an actuator. More recently, numerous
To analyze properties of the GRF, drop-test conditions are
designs have been presented for variable stiffness actuators
assumed. The leg is assumed stationary with respect to
(VSAs) to mitigate impacts. Stiffness modulation in VSAs
the body at an impact with initial velocity ẏb,0 . As such,
has been achieved through a variety of mechanisms. Designs T
may strategically load nonlinear springs [23] to modulate q0 = [0, 0] , and q̇0 = [ẏb,0 , 0] . The Laplace transform of
stiffness or may use linear springs by stretching them in the ground reaction force F(s) can then be found through
a nonlinear way [24]. Vanderborght et al. [25] provide a algebraic expansion of (5):
thorough review and classification of the many ways to provide rki mb
F(s) = T (s)+
and modulate stiffness in VSAs. Other SEA designs, such d(s)
as in the quadruped StarlETH, have demonstrated successful ki (mb m` r2 + Imb + Im` ) ⇣ g⌘
execution of controlled variable leg impedance using fixed ẏb,0 + (6)
d(s) s
stiffness joint SEAs [26]. While SEAs offer a great potential
where
actuator solution for legged robots, their force bandwidth can
suffer in comparison to designs without added compliance. d(s) = (mb I + m` I + r2 mb m` ) s2 + ki (mb r2 + I) . (7)
The next section will quantify this effect in a simple leg model.
This
p second-order system has a single natural frequency !n =
ki /me where me is the effective mass felt by the spring
III. P ROPRIOCEPTIVE F ORCE C ONTROL ACTUATION
A. Impact force analysis in a simplified leg model mb I + m` I + r2 mb m`
me = . (8)
mb r 2 + I
This section studies how the design parameters of an
actuation system affect its impact dynamics and force control. It is interesting to note that the reflected actuator inertia
A simple model, shown in Fig. 2, was developed to capture directly modulates the effective mass
essential design parameters for a legged system. The model lim me = m` , and lim me = m` + mb , (9)
consists of a body mass mb at height yb . Forces delivered to I!0 I!1

the mass are modulated by an actuated rack and pinion. A where any increase in I leads to an increase in me
pinion of radius r is driven by an ideal actuator with output dme r2 m2b
torque ⌧ . An inertia I represents the total rotational inertia of = > 0. (10)
dI (mb r2 + I)2
the actuator including gear transmission. The rack, with mass
m` , abstracts the entire leg and interfaces with the ground Further, letting

through a Hookean spring with stiffness ki . This stiffness @y` @y`
J := = [1, r] (11)
represents a lumped stiffness of the ground, foot covering, and @yb @✓
Bandwidth (Hz) Max. Impact Force (N) !1
10 3
10 4
4
10 4
104 !2 q
10 Ir 2 mb m`
!n !
˜2 = ks/
r 2 mb m` +I(m` +mb )
2 2
!
~1
10 10
103:5 !
~2
103:5

Frequency (Hz)
2
10

0 0
10 10 3
q
Inertia (kg"m2 )

Inertia (kg"m2 )
10
103 Ir 2 mb
!
˜1 = ks/
I+r 2 mb
-2 -2 2:5
10 10 10 10 1
102:5

10 -4 10 -4 102
102
10 0 1 2 3 4 5
10 -6 10 -6 101:5 10 10 10 10 10
101:5 Series Spring Constant ks (Nm/rad)

10 -8
10 -4 10 -2 10 0 10 2
10 -8
10 -4 10 -2 10 0 10 2
Fig. 4. Influence of series spring stiffness ks on the natural frequencies
Leg Mass (kg) Leg Mass (kg) of the system, !1 and !2 . The mechanical bandwidth, as determined by the
lowest natural frequency, increases with ks . Parameters r = 0.1 m, I =
Fig. 3. Influence of leg mass m` and actuator inertia I on open-loop force 0.010 kg · m2 , ki = 106 N/m, m` = 3 kg, and mb = 32 kg, roughly match
control bandwidth !n and maximum impact force assuming an initial velocity that of the MIT Cheetah running on rigid ground with a 3 mm rubber footpad.
of ẏb,0 = 1 m/s. Parameters ki = 106 N/m, r = 0.1 m, and mb = 32 kg
were used to roughly match the MIT Cheetah on rigid ground with a 3 mm
rubber footpad.
The system kinetic T and potential V energies take the form:
1 1 2 1
T = mb ẏb2 + I ✓˙m + m` ẏ`2 , and (14)
2 2 2
as the Jacobian for leg mass, straightforward algebra verifies 1 1
V = g mb yb + g m` yl + ki y`2 + ks (✓m ✓)2 . (15)
2 2
1 T 1
me = JH J . (12) A Lagrangian development can be followed to derive the
dynamics, again resulting in a linear system (1).
Thus, me is precisely a task-space (operational-space) inertia Following this development, the ground reaction force for
[27] measured at the foot prior to the interface spring. With this the extended system again matches (5). The torque-to-force
connection, Section III-B generalizes the analysis to higher- transfer function H⌧ (s) satisfies
DoF articulated mechanisms. ks rki mb
H⌧ (s) = C(s2 H + K) 1 T
S = (16)
Returning to the system (6) and examining the impact force ↵s4 + s2 +
in more detail, we assume ⌧ = 0 in order to isolate the passive
where
mechanical properties of the mechanism. The inverse Laplace
transform of (6) then provides ↵ = Ir2 mb m` , (17)
p = ks (Imb + Im` + r2 mb m` ) + Ir2 ki mb , and (18)
f (t) = ẏb,0 ki me sin(!n t) + me g(1 cos(!n t)) . (13) = ks ki (mb r2 + I) . (19)
Letting j represent the imaginary variable, the four poles s =
Figure 3 shows the dependency of the maximum impact force
±j !1,2 of (16) provide the two natural frequencies of this
and mechanical bandwidth !n on the leg design parameters I
system as s
and m` . Due to the dependence of impact force and bandwidth p
⌥ 2 4↵
on me , a combination of low leg mass and low actuator inertia !1,2 = . (20)
simultaneously maximizes open-loop force control bandwidth 2↵
and minimizes impact force magnitudes. For high leg mass, the Graphs of these two frequencies are shown in Fig. 4. As
actuator inertia has comparatively less effect on these metrics. intuition may suggest, !1 ! !n as ks ! 1, where !n is
This analysis was extended to the case of including series the natural frequency for the previous prototype with rigid
compliance into the pinion actuation, as shown in Fig. 2(b). transmission. At low transmission stiffnesses ks , the interface
The addition of compliance helps to soften impact, but is spring ki and leg mass m` do not affect the mechanical
known to have a detrimental effect on closed-loop force bandwidth q
control bandwidth. In terms of design, the requirements of ! Ir 2 mb
˜ 1 := ks/ I+r 2m ⇡ !1 (21)
a closed-loop controller to fight the natural dynamics of b

the system in high-frequency regimes can be minimized by At high stiffnesses ks , a similar, yet less physically meaningful
designing to maximize the open-loop mechanical bandwidth result holds for !2
q
of the mechanism. For this extended model, we approximate Ir 2 mb m`
!
˜ 2 := ks/ r2 m m ⇡ !2 . (22)
its mechanical bandwidth with its lowest natural frequency. +I(m +m ) b ` ` b

In this extended case, the system dynamics can again be Although it is possible to achieve a closed-loop force control
placed in the form of (1). A series spring of stiffness ks is bandwidth beyond the lowest natural frequency of the system,
modeled between a motor angle ✓m and a spring output angle this process is sensitive to model-based information and re-
✓. The extended system thus has configuration q = [yb , ✓, ✓m ]. quires additional actuation effort to fight the natural dynamics
of the system. This is further complicated in applications of such a comparison. The new development in Section III-B2
non-collocated force control which must compensate for the illuminates the effects of actuator design on previous impact
detailed dynamics of potentially many transmission elements analyses, while Section III-B3 formulates the new IMF metric
between the sensor and the actuator. to quantify inertial backdrivability across different robots.
Thus, we argue that the open-loop mechanical bandwidth 1) Rigid-Body Impact Dynamics: Given a floating-body
of a mechanism ultimately limits the practical closed-loop system, with base coordinates qb 2 R6 and internal (joint)
bandwidth. Indeed, with the numbers used in the figure above coordinates qj 2 Rn its dynamics can be compactly described
!n = 79.7 Hz, a value similar to that reported with our frame- through
work in Section V. The addition of a stiffness ks = 70 Nm/rad,  
similar to that in StarlETH [26], provides !1 = 13.5 Hz, Hbb Hbj q̈b
+ h(q, q̇) = ST ⌧ + JT f (23)
similar to the 9 Hz closed-loop bandwidth reported in their Hjb Hjj q̈j
LQR-based SEA torque control.
Through this analysis, it follows that to minimize impact where q = [qb , qj ]T 2 Rn+6 , h(q, q̇) 2 Rn+6 includes the
forces and maximize mechanical bandwidth, a design should Coriolis, gravity, and spring-dependent terms, J 2 Rm⇥(n+6)
be sought with minimal reflected actuator inertia, minimal leg is a contact Jacobian, and f 2 Rm represents the contact force.
mass, and minimal actuator compliance. Reflected inertia in For point-foot contacts considered here, m = 3. In a case of
particular has been shown to play an important role in the series compliance at the joints, qj may include both joint and
effective mass that governs collisional dynamics. The next motor angles. In contrast to the previous section, we do not
section will address how to generalize and quantify these model any ground compliance, and assume that impact can be
effects in more complete leg models. considered as an impulsive event. As a result, the analysis is
idealized in comparison to that in the previous section.
Given a state q, q̇ just before a foot impact, the system hits
B. The Impact Mitigation Factor (IMF) the ground with a velocity v = J q̇ and experiences a contact
In more complex mechanisms, factors such as actuator impulse ⇢ 2 R3 as given by [29]:
placement and the structure of the leg articulation ultimately
⇢= ⇤v . (24)
determine how reflected actuator inertias affect the back-
drivability of the robot. While backdrivability includes both ⇤ 2 R3⇥3 is the operational-space inertia matrix (OSIM) [27],
velocity-dependent and inertia-dependent effects [2], inertia- [30], [31] of the system felt at the contact and is given by
dependent effects are much more difficult to shape through
1
closed-loop impedance control [16], [28]. Thus, these passive ⇤= JH 1
JT . (25)
inertial effects are purely dependent on the inherent design
characteristics of the robot. Of these inherent characteristics, 2) Actuator Effects on the OSIM: It is important to note
reflected actuator inertias play an important role. They directly how the reflected inertias of EM actuators affect the OSIM.
determine the degree to which the body inertia contributes to Generally, the mass matrix H, as partitioned in (23), provides
impacts, as observed in the simple leg model in (9). As a a kinetic energy metric
result, the highest loads felt in the legs, gearboxes, and other
1 T
transmission components are governed, in large part, by design T = q̇ H q̇ (26)
decisions centered on reflected actuator inertia. 2
Previous work has addressed the role of the effective contact where T includes the kinetic energy of rigid-body links as well
inertia in modeling impact events [29], which is used in as rotor inertias and gears of EM actuators. Thus, the mass
Section III-B1. This previous work modeled the values of matrix can always be partitioned into a matrix that accounts
impact impulses under specific impact conditions, which may for the kinetic energy of all rigid-body links Hrb and a matrix
be of use in comparing designs for a specific robot. For un- Hmot that accounts for the kinetic energy of the moving parts
structured environments where impact conditions may not be within link-mounted actuators (rotors, gears, etc.)
known, however, maximizing overall inertial backdrivability is
imperative to mitigate impact forces. H = Hrb + Hmot . (27)
With these impact-related motivations, this section quanti- Hmot is often approximated as [32]
fies how effectively the free dynamics of the mechanism are
at reducing impact impulses in a floating-body robot. To em- Hmot = diag(01⇥6 , I1 , . . . , In ) (28)
phasize the importance of actuator design on impact dynamics,
impulse reduction is evaluated through comparison to a design where Ii represents the reflected rotational inertia on DoF
with worst-case reflected inertia, one where all joints are i and scales as the square of its associated gear ratio. This
rigidly locked. Through this approach, a new metric, called the approximation is only valid for actuators with large gear ratios,
impact mitigation factor (IMF), is introduced to quantify the where the only significant kinetic energy of the motor elements
normalized inertial backdrivability of the mechanism. Despite is from rotational kinetic energy along their rotational axes.
the importance of backdrivability to physical interaction, there Regardless of this approximation, it can be seen from
are not yet meaningful metrics to compare backdrivability ⇣ ⌘ 1
1
across machines and to include EM actuator properties in ⇤ = J (Hrb + Hmot ) JT (29)
that any increase to the actuator inertia Hmot (in a positive impacts. Given a direction x, we define the directional IMF
semidefinite sense) necessarily gives rise to an increase in (DIMF) ⇠x as
1 T
contact inertia ⇤, generalizing the simplified result in (10). x ⇤x
⇠x = 1 12 T (36)
In the worst-case scenario, if reflected inertias become 2 x ⇤L x
arbitrarily large on each joint, the system will roughly behave
as if all the joints are locked. This worst-case design can where ⇠x similarly inherits the property that 0  ⇠x  1
be used as a benchmark to evaluate the effects of reflected from the fact that ⇤L ⇤. This DIMF has a clear physical
inertias in designs with the joints free to move. Partitioning interpretation as well. The kinetic energy lost due to an impact
the Jacobian as with velocity v is
⇥ ⇤ 1
J = Jb Jj , (30) T = v⇤v . (37)
2
an impact velocity v would give rise to a contact impulse in Thus, ⇠x quantifies the percent decrease in energetic losses
a locked system ⇢L as afforded through the free dynamics of the mechanism when
impacting in direction x.
1
⇢L = Jb Hbb1 JTb v. (31) 4) Series Elasticity and the IMF: When compliance is
added to a stiff transmission model, portions of the actuator
With this insight, we define a locked contact inertia ⇤L as
inertia are no longer inertially coupled to the foot. This inertial
⇤L = Jb Hbb1 JTb
1
. (32) decoupling affects ⇤ and the associated impact impulse. The
compliance, roughly, redistributes some of the impact impulse
Intuitively, the apparent inertia of the system with joints locked to a window of time following impact. Changes in the values
should be greater that the inertia felt with the joints free to of the stiffness modify how spread out the forces become, but
rotate. This can be shown more formally, that ⇤L ⇤ in a do not change the initial impulse, which is governed by inertial
positive semidefinite sense. Proof is given in the appendix. It properties alone. In this sense, adding compliance to a rigid
is important to note that ⇤L is invariant with changes in gear transmission model will affect the IMF, while changes to the
ratio (assuming negligible change in overall mass). values of stiffness will not. Considering non-impulsive impact
events over a finite window of time, which should always be
3) Impact Mitigation: Development and Metric: Given an the case for collisions in reality, represents an important area
impulse that is experienced in the locked system, it is natural for future extension of the IMF. Further details on the IMF for
to consider: How effectively are the free dynamics of the a system with SEAs are provided in comparison to the design
mechanism at mitigating this impulse? That is, assuming the of the MIT Cheetah leg in Section V.
same impact velocity, how would the impact in the free system
compare to its non-backdrivable equivalent? Given the locked
impulse ⇢L , equal impact velocities v are achieved when the C. Geometric considerations for electromagnetic motors
free system experiences an impulse From the analysis of the previous section, we seek to deter-
mine principles for actuator selection that minimize actuator
⇢ = ⇤⇤L 1 ⇢L (33) inertia while providing the high torques necessary for high-
Roughly, the term ⇤⇤L 1 characterizes how much inertia is speed locomotion. Many design parameters of EM actuators
felt at impact in comparison to the overall (locked) system contribute to overall performance. However, we consider gap
inertia. The reduction in impulse is then given by radius to be a critical parameter for the purposes of maximiz-
ing torque density and providing transmission transparency.
⇢L ⇢= I ⇤⇤L 1 ⇢L . (34) The gap radius of an EM actuator is the radius of the magnetic
| {z } interface between the rotor and the stator. While gap radius
:=⌅
is one of many possible condensed indicators of performance
where ⌅ is introduced as an Impact Mitigation Matrix (IMM), [33], it is directly related to torque density and torque per total
and its determinant inertia. Torque per total inertia, in turn, is directly related to
⇠ := det(⌅) (35) the available bandwidth of the actuator.
Ignoring edge effects, the axial length of the motor does not
as the Impact Mitigation Factor (IMF). This new factor has affect torque density and torque per inertia because increasing
interesting properties that pertain to backdrivability. It is shown axial length is equivalent to adding identical motors on the
in the appendix that 0  ⇠  1. An IMF ⇠ = 1 corresponds to same axle. In Figure 5, Emoteq HT series motor characteristics
a system with perfect inertial backdrivability that eliminates all are plotted against gap radius. Motors of the same gap radius
impact, whereas ⇠ ! 0 as ⇤ ! ⇤L . Due to the normalization and various lengths have overlapping values of torque density
provided from the locked case (with roughly infinite reflected and torque per inertia, whereas gap radius directly affects those
actuator inertia), the IMF is a nondimensional quantity which characteristics. For extreme geometries beyond these designs,
enables comparison across machines of different scale. edge effects degrade the metrics as length decreases.
Further, one may be interested in the capability of a design 1) Effects of Geometry on Motor Performance Metrics:
to mitigate impact in a particular direction, for instance in To understand the main effects of motor geometry on torque
the case of a hopping robot subject to predominantly vertical density and torque per inertia, we consider a class of designs
10 6 Stator l Assumptions:
tr , ts , M = constant
4 Rotor rg
10
2
10
rg : gap radius
0 l: motor length ts
10 range of motor geometries
tr : rotor thickness
ts : stator thickness tr
-2
10 M: mass

10 -4 = =J / rg !1:6
Fig. 6. Motor performance tradeoffs are considered for a constant mass bud-
-6
= =M / rg 0:8 get under assumptions of constant rotor/stator thickness and electromagnetic
10 Kt2 =R / rg 4:1 similarity.

10 0.5 10 1 10 1.5 10 2
Gap Radius rg (mm) of wires in the cross section n / rg . Thus, the relationship
becomes
Fig. 5. Motor performance metrics ⌧ /J, ⌧ /M , and Kt2 /R extracted from ⌧2
Emoteq frameless HT series motors.
/ rg3 l . (41)
I 2R
2) Comparison to Catalog Data: Figure 5 shows values
for these metrics across data collected from Emoteq HT
under the following simplifying conditions. First, the radial series motors. The max continuous torque as reported on data
thickness of the rotor and stator are assumed fixed. Second, the sheets was used to approximate conditions of electromagnetic
cylindrical geometries of the rotor and stator are approximated similarity. Within this dataset, the torque density was found
to first order as thin walls. As a result, the actuator mass proportional to rg0.8 ; torque per rotor inertia proportional to
is given by M = 2⇡rg l(ts ⇢s + tr ⇢r ) and rotor inertia by rg 1.6 ; and torque production efficiency proportional to rg4.1 .
J = 2⇡ltr ⇢r rg3 , where rg is the gap radius, l is the axial These factors of proportionality, rg0.8 , rg 1.6 and rg4.1 are
length, and t⇤ and ⇢⇤ are the radial thickness and density, relatively consistent with the proportions in our dimensional
respectively, of the stator and rotor. Third, current density is analysis rg1 , rg 1 , and rg3 from (38), (39), and (41) respectively.
assumed constant [34]. Neglecting edge effects, this results The similarity is to be taken considering that the dimensions
in conditions of thermal and electromagnetic similarity, where of these motors used in Fig. 5 do not exactly match our
the steady-state stator temperature and average shear stress s assumptions, and in particular the assumption of constant
on the rotor are constant across designs. The resulting torque stator and rotor thickness. In the Emoteq HT series motors, the
from this stress ⌧ = 2⇡rg2 l s does depend on motor geometry. stator and rotor thicknesses scale by ts / rg0.8 and tr / rg0.4
Therefore, within this class of designs, we predict that respectively. The incorporation of these trends into the pre-
torque density and acceleration capability (torque per rotor vious analysis would lead to a lower ⌧ /J, lower ⌧ /M , and
inertia) scale by the following relationships with rg , and we higher Kt2 /R as observed in the motor catalog data. Changing
observe no effect from axial length. rotor/stator thickness, however, has a nonlinear effect on the
magnitude of the magnetic field at the rotor-stator air gap,
⌧ /M / rg (38)
and heavily influences stator winding design. As a result, the
⌧ /J / 1/rg (39) effects of these changes are difficult to model accurately in
general. The remainder of the section returns to considerations
Another important characteristic is torque production effi- of fixed rotor and stator thickness across designs.
ciency, which is related to the motor constant KM . KM 2
is 3) Optimizing Geometry Under Design Constraints: The
equivalent to the torque squared per unit ohmic power loss. theoretical motor performance metrics (38) and (39) can be
Torque is generated from sheer stress s / IBn/A where I helpful in making initial decisions related to actuator geometry
is the motor current, B is the field strength of the magnets, n while considering design constraints. Such considerations are
is the number of wires in a cross section perpendicular to the unique relative to previous isometric scaling analyses that
axis, and A / rg is the area of the cross section. Using this fixed geometry [34], [35]. Suppose a desired output torque
relationship, torque production efficiency is given as ⌧out and a fixed mass budget M for an actuator in the class
Kt2 ⌧2 rg2 l2 B 2 n2 Aw considered previously. Under this mass budget, motor length
2
KM = = 2 / (40) is constrained to scale by l / 1/rg as shown in Fig. 6. To
R I R ⇢lw
meet the output torque requirements across different designs,
where Kt is the torque constant, R is the terminal resistance, suppose that a gear train with ratio N is used.
Aw is the cross section area of each wire, lw the total length As an initial approximation, assume that the gear box has
of wire, and ⇢ is the the resistivity of the wire material. negligible added mass, inertia, and friction torques from the
An additional set of assumptions is used to simplify this gears themselves. The required gear ratio as a function of gap
relationship. First, a wire length lw / nl is used, which radius is then
assumes all stator wire contributes to winding coils. Second, ⌧out
N= / 1/rg (42)
a fixed wire gauge is assumed, which implies that the number 2⇡rg2 l s
with a reflected rotor inertia through the gearbox of τ"
Jref = Jr N 2 = 2⇡rg3 ltr ⇢r N 2 / rg0 (43) Demag.
Demagne'za'on*
torque
Torque
when the scalings for l and N are considered. While the total
reflected rotor inertia and the total torque at the output shaft Sat.
Satura'on
Torque
stay the same, increasing the gap radius lowers the required Torque
Intermittent Operation
gear ratio. In summary, in a design space where the motor
region
mass and the output torque requirement are held constant, and
gearboxes are idealized, (43) shows that total reflected inertia Cont.
Con'nuous ~0.3kW/kg
torque
remains constant across variations in gap radius. Torque
Continuous operation
Considering the non-idealized effects of gear transmissions, region
however, shows the benefits of high-gap radius designs when 0
the actuator mass budget is fixed. The results of constant operation range recommended by ω"
output torque and constant reflected inertia only hold across motor manufacturer
different design geometries if the mass, friction, and inertia Fig. 7. Different operational limits of a motor in torque-speed space.
of the gears are ignored. If we assume instead that mass, The green area represents the manufacturer-recommended operation that is
friction, and gear inertia increase monotonically with gear ratio bounded by the voltage limit and continuous torque determined by the thermal
limit. The blue area shows the extended operating range of the motor by
N , this implies in turn that ⌧out will decrease and Jr will raising voltage and speed but maintaining current under the continuous torque
increase monotonically with N . Thus, these considerations limit. The motor may intermittently exceed the continuous torque limit and
favor a larger gap radius motor when mass budget is fixed. operate in the transient operation space shown in peach color. Beyond the
saturation torque, parts of iron in the stator become magnetically saturated
Such a design will have a smaller gear ratio, fewer gear stages, and the torque/current relationship becomes highly nonlinear as the torque
and less gear mass, resulting in less friction loss, higher torque constant decreases. This nonlinearity can be compensated for in control to
density, higher bandwidth, and higher IMF. provide accurate torque delivery. At the demagnetization torque, magnetic
fields from the stator begin to demagnetize the rotor magnets, damaging the
Therefore, in this design space, there is no tradeoff, and in motor.
its purest form, this analysis advocates for high-gap-radius
direct-drive robots [36]. However, the geometry required for
Planetary gear
direct drive to have enough torque is typically infeasible. For train (1:5.8)
Base

example, in order to design a direct-drive motor for the MIT Knee Hip
CoM
Stator Stator
Cheetah, one such motor would have been 76.2 cm in diameter
with a 5 mm axial length. Considering these limitations, the Knee Hip
Rotor
optimal actuator for a given mass will thus consist of a motor Rotor

with the largest gap radius as allowed by space and the smallest Output to hip
gear ratio as required by torque specifications. Encoder
4) Other Considerations: For extreme geometries, edge mount

effects begin to degrade the benefits of high-gap-radius de- Encoder


signs. As a result, it is important to view this analysis as mount
Output to
providing a guiding principle to focus design pressure for linkage
actuators in legged machines. Outside the class of motors
considered, or at edges of the design space, detailed EM, Fig. 8. Shoulder module section view for a front leg of the MIT Cheetah.
winding, and manufacturing considerations should limit the
degree to which gap-radius effects are regarded as dominant
without further modeling. short ground contact times. For example, the ground contact
In addition to motor geometry optimization, we can increase time of each leg during 6 m/s bounding is around 60 ms.
torque density by utilizing the intermittent torque of the In order to control the ground reaction forces during such a
motor, which is much higher than the continuous torque limit. short period of time, the system must have high-bandwidth
These operating regimes are depicted in Fig. 7. Many robotic force control and survive frequent high impact forces. The
actuators operate under the continuous torque limit, although MIT Cheetah is the first embodiment of the proprioceptive
the duty factor of high torque usage in most legged robots is actuation concept in legged locomotion.
small. In legged locomotion, the continuous torque limit of the Figure 8 shows the detailed design of the leg module.
motor does not limit torque capability. Unlike most EM motor The four hip modules of the robot are identical with minor
applications, the joint torque profiles in legged locomotion differences in segment length ratios among the three links and
constantly fluctuate. between the front and rear legs. As discussed in Section III-C,
it is critical to maximize the ratio of torque to inertia. We
IV. P ROPRIOCEPTIVE ACTUATION IN THE MECHANICAL selected design parameters for motor with the largest radius
DESIGN OF THE MIT C HEETAH LEG that could fit within a 12.5 cm diameter space, which is
The MIT Cheetah is designed to emulate various locomotion constrained by the size of the MIT Cheetah. The first version
capabilities of quadrupeds such as walking, running, and jump- of the MIT Cheetah, which trotted up to 6 m/s [13], used
ing. Such behaviors involve repeated high impacts followed by an Emoteq HT-5001 frameless motor. The second version
of the MIT Cheetah, capable of bounding outdoors, uses Femur
custom motors designed by the MIT team [37]. The motor 18 mm m1 = 1.55 kg
was specifically designed to maximize the saturation torque 245 mm I1 = 80.9 kg cm2

for a given mass, where the major design tradeoff is between


2

1.5
Tibia
ohmic loss and the saturation torque. The saturation torque 1
220 mm
m2 = 0.157 kg
82 mm
density of the custom designed motor is around 27 Nm/kg 0.5
2
I2 = 13.1 kg cm2

(standalone without module or gearing), significantly higher 0


1.5

Foot
than 9 Nm/kg in the Emoteq HT-5001. Given this high torque
1
-0.5 84 mm 182 mm m3 = 0.116 kg
0.5

density, a 3 cm motor length is able to be used on each axis.


-1

ẑ I3 = 7.4 kg cm2
-1.5
ŷ 0

The rotor inertia of the custom motor is three times that of the -2
-0.5
-2 -1.5 -1 -0.5
x̂ 0 0.5 1 1.5 2

Emoteq HT-5001 and the attached leg inertia is slightly larger -1

than the previous MIT cheetah. -1.5

Fig. 9. Visual description of the configuration dependence of the impact


Given the large torque capacity of the motor, we chose
-2

mitigation factor (IMF), and main dynamics parameters for a leg of the MIT
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

the gear ratio (1:5.8)1 to meet the normal GRF generation Cheetah. The dashed line represents an ideal directional IMF (DIMF) of 1,
while the blue represents the computed DIMF.
requirements for running at a range of speeds up to 13.5 m/s.
We chose a one-stage planetary gear train with four planet 1 1
gears. Such a low gear ratio provides a higher IMF, beneficial
for highly-dynamic locomotion. Unlike traditional serial-link 0.8 0.8
robots in which actuators are present at every joint, two
0.6 0.6
actuators and the gear train are coaxially located at the hip

IMF 9

9x^
of the leg to minimize the total moment of inertia, which 0.4 0.4
allows for a compact and robust design. One of these actuators
directly actuates the hip, while the other actuates the knee 0.2 0.2

through a parallel linkage. The structure of the leg is also


0 0
designed to minimize mass and leg inertia, and thus maximize 0 1 2 3 0 1 2 3
Knee Angle (rad) Knee Angle (rad)
the impact mitigation factor. Bending stress is minimized in the
leg structure by distributing tensile forces to the tendons. This
method allows significantly lower inertia of the leg without 1 1

compromising leg strength [38]. The shoulder module contains 0.8 0.8
the motors (2.00 kg), heat-treated steel gear trains (0.42 kg),
and framing (1.23 kg) for a total mass of 3.65 kg. A leg with 0.6
Cheetah
0.6

three ABS plastic links is attached to each module, with a


9y^

9z^
0.4
SEA Cheetah 0.4
combined mass of 850 g. The lightweight leg allows the center StarlETH
of mass to be located inside of the shoulder module. As a 0.2
Hubo 0.2
result, a small moment of inertia of the leg allows for rapid SEA Hubo
movement and high-bandwidth force control. 0
0 1 2 3
0
0 1 2 3
Knee Angle (rad) Knee Angle (rad)

V. A SSESSMENT AND R ESULTS Fig. 10. Impact mitigation factor and directional impact mitigation factors
A. Impact Mitigation Factor versus knee angle. The coordinate system for these numbers is shown in Fig. 9
for the Cheetah with similar coordinate systems relative to the sagittal plane
In order to assess the backdrivability of the MIT Cheetah for the other morphologies.
leg design, the impact mitigation factor (IMF) was computed
and compared to other legged systems. As noted in Section
III-B, the IMF is configuration dependent. This configuration planetary gears, as estimated from CAD models. The rotor
dependency is shown visually in Fig. 9. The figure shows the has inertia 3.0279 ⇥ 10 4 kg m2 along its rotational axis,
directional IMF (DIMF) ⇠x for directions x in the sagittal with a 5.8:1 gear reduction, resulting in a reflected inertia of
plane. For the Cheetah, the IMF correlates most strongly with 0.0102 kg m2 . Other main parameters of the leg are given in
knee angle, as shown in Fig. 10. Due to the large mass and Fig. 9 for reference.
inertia of the components within the Cheetah’s motors, (28) The IMF of the Cheetah design was also compared with
was not used for simplification in the calculation of the IMF. approximate StarlETH and HUBO Plus models. Rigid-body
The motor rotors, for instance, have rotational inertias on out- masses, CoM locations, and kinematics for StarlETH were
of-plane principal axes that are the same order of magnitude taken from an available publication [26]. Previous work pro-
as for the leg links. These inertial contributions would be vided estimated rotational inertias in the sagittal plane [39]. In-
otherwise ignored through the use of (28). The IMF numbers ertias in other directions were computed from an equi-density
account for the mass and inertias of the motor rotors and assumption on a bounding box estimated from graphical mod-
1 This gear ratio was also driven by practical considerations, as it is the
els. No reflected actuator inertia was modeled for StarlETH, as
largest single-stage ratio available for a standard planetary configuration with its series elastic actuators decouple reflected inertia from the
commodity English geartooth options. endpoint. The HUBO model was taken from a URDF of the
Step Input Test (100N, Radial Direction)

160
HUBO Plus robot [40]. The DC actuators assumed at the joints
had a rotor inertia of 3.33 ⇥ 10 6 kg m2 with a 160:1 gear 140
From 𝑓 × 𝑡 = 0.35
reduction, producing a total reflected inertia of 0.0852 kg m2 . 120 Driver & Motor Current Bandw

Force(N)
Leg System Bandwidth: 103.7H
No reflected inertia effects of the gear reduction were modeled.

(N)
100 Stabilization Time: 55 ms
IMFs were computed assuming impact in the center of the 80

Force
bottom of the foot.
Figure 10 shows that the SEA-actuated StarlETH has the 60

highest impact mitigation factor over a range of knee an- 40 To fully control 400Hz Bandwid
gles. For these computations, the virtual leg in each system
Force SensorSensor
Force
Commanded Force
control frequency should be at
20 Commanded
Estimated fromForce
Motor Usually use x5 to x10 for good
remained upright, while the knee angle was used to modulate Estimated from Motor Target: 4kHz control frequen
0
virtual leg length. Low IMFs for small knee angles are in (250us loop time)
0 10 20 30 40 50 60 70 80 90 100
most part due to reduced backdrivability in the ẑ direction, as Time(ms)
Time (ms)
indicated by ⇠ẑ . A similar degradation in ⇠x̂ is observed in
StarlETH and Cheetah when the leg becomes fully collapsed Fig. 11. Step-input test on the Cheetah robot leg. The force estimated from
at a knee angle of ⇡ rad. The IMF of HUBO experiences the motor is determined from the measured current in the stator windings
less peak-to-peak variability in IMF across knee angle. This using an assumption of static force transfer from the actuator to the foot.
difference can be attributed to the additional articulation in System Rise time Tr (ms) Bandwidth (Hz)
the HUBO hip and ankle which further prevent its upper-body motor driver to winding current 0.875 400
mass from being felt at impact even for small knee angles. motor driver to foot force 3.38 103.7
To assess the relative contribution of reflected actuator TABLE I
inertia to the IMF of HUBO and Cheetah, hypothetical modifi- S YSTEM BANDWIDTH ESTIMATION FROM STEP - INPUT TEST.
cations of these machines were considered that included SEAs
at the joints. Averaged over the configurations considered, the
IMF of unmodified Cheetah was 90% that of the hypothetical
work demonstrated the capabilities of proprioceptive force
SEA Cheetah. In comparison, the IMF of unmodified HUBO
control to realize a virtual stiffness at the foot. The next
was 52% that of a hypothetical SEA HUBO. HUBO’s design
subsection goes beyond impedance control and demonstrates
does include significant actuator mass distally in the leg, which
open-loop force control for more general force profiles.
in part causes this difference.
C. Proprioceptive Force Control Test
B. Step-input test on the MIT Cheetah leg To investigate the performance of proprioceptive force
To initially evaluate proprioceptive force control, the Chee- control for dynamic locomotion, we performed in situ force
tah robot leg was mounted inside of an axial material testing control tests during unconstrained 3D bounding. The robot
device (Zwick Roell BX1-EZ005.A4K-000). Its stock force bounded in free space using a control law from previous
sensor was replaced with a six-axis force-torque sensor (ATI work [41]. The bounding controller would output desired
delta, SI-660-60 calibration) which can measure up to 1980 N forces fpro 2 R3 for each stance leg, to be realized through
in the z-axis with 0.25 N resolution. proprioceptive actuation. These desired forces were mapped
To identify the open-loop force control bandwidth of the to joint torques through
leg, a 100 N step-input test was executed. A pure radial
⌧ d = S` JT fpro . (44)
force was commanded and mapped to desired leg torques
assuming static loading conditions. Desired joint torques were Here J 2 R is a foot Jacobian and S` 2 R
3⇥18
is a 3⇥18

then realized using closed-loop motor current control with selector matrix for the joints in the leg. The desired torque
custom motor drivers and current control taking place at a ⌧ d 2 R3 contains torques for the two proprioceptive actuators
rate of 20 KHz. The currents from these drivers were used to nominally in the sagittal x̂-ẑ plane, as well as for a more
estimate the force at the foot, again under an assumption of traditional hip ab/ad actuator which nominally affects the
static loading conditions. Fig. 11 shows the results of this test. lateral ŷ forces. Coordinate systems follow those in Fig. 10.
In this figure, the black line is the commanded force, the red The ATI delta force-torque sensor was embedded into the
line is the estimated force from measured motor currents, and contact surface such that the front left leg of the robot could
the blue line is the force measured by the external sensor. The run on the sensor without interference to the gait. Figure 12
discrepancy between the estimated force and the exteroceptive shows a figure of this experimental setup.
sensed force highlights the presence of structural compliance Figures 13-15 show the proprioceptive force control track-
in the leg. This structural compliance effectively decreases the ing in the sagittal plane where the proprioceptive actuators
bandwidth of the proprioceptive force control below that of have force control authority. The force control figures show
the electrical dynamics. Assuming a second-order response, high duty ratio bounding (Fig. 13) in comparison to medium
the response bandwidth of the leg system can be estimated and low duty ratio bounding (Figs. 14 and 15). Lower duty
based on the measured rising time, as seen in Table I. ratios require shorter contact times down to 85 ms in the most
Further details on similar experiments with the MIT Cheetah extreme case considered here. Again, we emphasize that aside
leg can be found in our previous publication [5]. This previous from statics (Eq. 44), no model-based information was used to
50

Force fx (N)
0

-50

-100

-150
0 100 200 300 400 500 600
Time (ms)
Proprioceptive
500 Measured
Evaluation Period

Force fz (N)
400

300

200

100
Fig. 12. Experimental setup for proprioceptive force control measurements. 0
The ATI delta force/torque sensor is embedded into the contact surface so 0 100 200 300 400 500 600
Time (ms)
that the front left leg could step on the sensor without interference to the gait.

50 Fig. 14. Force tracking for duty ratio D = 0.410 bounding.


Force fx (N)

0 50

-50

Force fx (N)
0

-100
-50

-150
0 100 200 300 400 500 600 700 800 900 1000 -100
Time (ms)
-150
Proprioceptive 0 100 200 300 400 500 600
500
Measured Time (ms)
Proprioceptive
Force fz (N)

400 Evaluation Period


Measured
300 500 Evaluation Period
Force fz (N)

200
400
100
300
0
0 100 200 300 400 500 600 700 800 900 1000 200
Time (ms)
100

Fig. 13. Force tracking for duty ratio D = 0.579 bounding. 0


0 100 200 300 400 500 600
Time (ms)

Fig. 15. Force tracking for duty ratio D = 0.298 bounding.


translate desired force into commanded motor current across
any of these results.
Table II quantifies the average tracking accuracy of the to reduce impact forces, high IMF designs indicate favorable
proprioceptive force control approach. Force control accuracy actuator characteristics more broadly related to dynamic in-
was evaluated following each impact transient, during the teractions with the environment. High IMF designs capture
periods shown in gray in Figures 13-15. These periods start an ability to transparently control the interaction forces, and
when the measured vertical force crosses the proprioceptive to protect both the environment and the robot in instances of
force, and end when the proprioceptive force goes to zero. In collision. These benefits have practical importance, improving
general, higher duty ratio bounding results in lower average the lifetime of fragile transmission components, reducing the
errors. need for bulky structural designs, and reducing leg dynamics
High duty ratio bounding in Fig. 13 shows one of the which complicate swing-leg control.
reasons to pursue proprioceptive force control. The vertical As shown in the previous section, the IMF is also tightly
force of the first hop in particular, shows a low-frequency related to improved bandwidth in proprioceptive force control
(⇡20 Hz) unforced oscillation in the vertical direction. This (closed-loop torque control with open-loop ground reaction
additional vibration mode is caused by unmodeled dynamics force control). In dynamic locomotion, foot force control
(compliance) of the leg structure and mechanisms. Future in short stance times is critical for balance control. Non-
design efforts may be directed towards further increasing collocated force control lacks robustness to unmodelled high-
the stiffness-per-mass ratio of the legs in order to raise
the natural frequency of the structural dynamics. It is these
unmodeled transmission dynamics that limit the bandwidth of Duty Ratio (D) Avg. Err. fx (%) Avg. Err. fz (%)
0.579 24.35 18.20
non-collocated force control schemes. 0.410 27.54 14.19
0.298 33.52 22.91
VI. D ISCUSSION TABLE II
P ROPRIOCEPTIVE FORCE TRACKING ACCURACY VERSUS DUTY RATIO
The proprioceptive actuation paradigm offers a number of (D). G ENERALLY, TRACKING IMPROVES WITH HIGHER D.
qualitative benefits which extend beyond the analysis pre-
sented. While designing for high IMF has been motivated
the promising performance of the proposed actuator design
paradigm by evaluating the force production capabilities of a
leg from the MIT Cheetah.
These results encourage broader adoption of proprioceptive
actuators to manage physical interaction for emerging appli-
cations in robotics. From disaster response to assistance in
the home, proprioceptive actuators may endow next generation
robots with the ability to stably make and break contact, while
providing the high-bandwidth and force-critical capabilities to
react in unstructured environments. Through further refinement
Fig. 16. The ability to mitigate impacts and manage high-bandwidth contact directed toward these applications, we believe the proposed
interactions has been an important capability in results beyond those presented approach will enable the development of actuators that yet
here, such as the execution of a running jump over large obstacles [43].
exceed the capabilities of biological muscles in every aspect.

frequency dynamics incited around impact, causing contact ACKNOWLEDGMENT


instability [3]. The ability to stably make and break contact
though proprioceptive force control motivates broader appli- This research was sponsored by the DARPA M3 program
cation of this paradigm for high-force interaction. through contract number W31P4Q-13-1-0014.
Increasing the torque density of a EM actuator itself is criti-
cal to achieve conflicting specifications for high IMF and high
A PPENDIX
efficiency. Since the gear ratio is a critical design parameter
I MPACT M ITIGATION FACTOR
in the tradeoff between the torque generation efficiency and
the IMF, increasing torque density of the motor itself enables In the main body of the text, the impact mitigation factor
a designer to the reduce actuator gear ratios without compro- for a leg was introduced as
mising either of the aspects. Through exploiting this strategy,
the MIT Cheetah has already achieved a high efficiency (total ⇠ = det I ⇤⇤L 1 (45)
cost of transport of 0.5) rivaling animals in a similar scale
[13], [42]. Further research to understand and quantify the where ⇤ 2 Rm⇥m is the contact-space inertia matrix of the
exact energetic tradeoffs due to gear transmission selections leg, and ⇤L 2 Rm⇥m is the contact-space inertia matrix if all
may yet enable operation beyond the efficiency envelope of of the joints were locked. In this appendix, we will show that
legged animals in nature. ⇤  ⇤L and 0  ⇠  1.
While space and scope have prevented their description, Using notation from the main body of the text:
these advances in actuator design for impact mitigation and  1
" #
⇥ ⇤ Hbb Hbj JT
high-bandwidth physical interaction have been paramount to- ⇤ 1
= Jb Jj
Hjb Hjj
b
JT
. (46)
j
wards this unique platform achieving recent feats such as
running at up to 6 m/s and successfully jumping over obstacles Then, defining A := (Hjj Hjb Hbb1 Hbj ) 1 it follows from
up to 80% of the Cheetah’s nominal leg length [43]. Figure the matrix inversion lemma that
16 shows snapshots from an example jumping motion. ⇣ ⌘ 1
Hbb Hbj Hjj 1 Hjb Hbb1 Hbj A
⇤ 1 =J JT (47)
VII. C ONCLUSIONS 1
AHjb Hbb A

This paper proposed a new actuator paradigm for high-speed Hbb1 + Hbb1 Hbj AHjb Hbb1 Hbb1 Hbj A
running robots, provided analyses to demonstrate the central =J AHjb Hbb1 A
JT . (48)
tenants of this paradigm, and presented experimental results
of force tracking tests. While proprioceptive actuation was Factorizing this matrix,
effectively implemented in early haptic devices, such as the h i h iT
PHANTOM, a prototype leg model illuminates its benefits Hbb1 Hbj Hbb1 Hbj

1
= ⇤L 1 + J I
A I
JT , (49)
to manage impact and open-loop force control for legged
machines. Geometric scaling analysis indicates that increasing where the second term on the right hand side is positive
the gap radius benefits torque density, which plays a critical semidefinite since A is a Schur complement of H. Thus,
role in system energetics for locomotion. In order to quantify
1
the backdrivability afforded through this paradigm, we have ⇤ ⇤L 1 and ⇤  ⇤L . (50)
introduced the Impact Mitigation Factor (IMF) which is appli- 1 1
cable to robots driven by EM actuators with or without series From this fact, it follows that 0  ⇤L 2 ⇤⇤L 2  I. Thus
elasticity at the joints. A successful implementation of the ⇣ ⌘
1 1
actuator design principles was shown to allow for high-force 0  det I ⇤L 2 ⇤⇤L 2  1, and (51)
proprioception to deliver desired forces through contact with
only motor current sensing. The experimental results show 0  det I ⇤⇤L 1  1 . (52)
R EFERENCES Damme, R. V. Ham, L. Visser, and S. Wolf, “Variable impedance
actuators: A review,” Robotics and Autonomous Systems, vol. 61, no. 12,
[1] D. C. Hanselman, Brushless permanent magnet motor design. The pp. 1601–1614, 2013.
Writers’ Collective, 2003. [26] M. Hutter, C. Gehring, M. Hopflinger, M. Blosch, and R. Siegwart,
[2] W. Townsend and J. Salisbury, “Mechanical design for whole-arm ma- “Toward combining speed, efficiency, versatility, and robustness in an
nipulation,” in Robots and Biological Systems: Towards a New Bionics?, autonomous quadruped,” IEEE Transactions on Robotics, vol. 30, no. 6,
ser. NATO ASI Series, P. Dario, G. Sandini, and P. Aebischer, Eds. pp. 1427–1440, Dec 2014.
Springer Berlin Heidelberg, 1993, vol. 102, pp. 153–164. [27] O. Khatib, “A unified approach for motion and force control of robot
[3] S. D. Eppinger and W. P. Seering, “Three dynamic problems in robot manipulators: The operational space formulation,” IEEE Journal of
force control,” in IEEE International Conference on Robotics and Robotics and Automation, vol. 3, no. 1, pp. 43–53, Feb. 1987.
Automation, 1989, pp. 392–397. [28] A. Albu-Schaffer, C. Ott, U. Frese, and G. Hirzinger, “Cartesian
[4] T. H. Massie and J. K. Salisbury, “The PHANTOM haptic interface: A impedance control of redundant robots: recent results with the DLR-
device for probing virtual objects,” in Proceedings of the ASME winter light-weight-arms,” in IEEE International Conference on Robotics and
annual meeting, symposium on haptic interfaces for virtual environment Automation, vol. 3, Sept. 2003, pp. 3704 – 3709.
and teleoperator systems, vol. 55, no. 1. IOS Press, 1994, pp. 295–300. [29] Y.-F. Zheng and H. Hemami, “Mathematical modeling of a robot
[5] S. Seok, A. Wang, D. Otten, and S. Kim, “Actuator design for high collision with its environment,” Journal of Robotic Systems, vol. 2, no. 3,
force proprioceptive control in fast legged locomotion,” in IEEE/RSJ pp. 289–307, 1985.
International Conference on Intelligent Robots and Systems, Oct 2012, [30] H. Asada, “Dynamic analysis and design of robot manipulators using
pp. 1970–1975. inertia ellipsoids,” in IEEE Int. Conf. on Robotics and Automation, vol. 1,
[6] M. F. Bobbert, M. R. Yeadon, and B. M. Nigg, “Mechanical analysis of Mar 1984, pp. 94–102.
the landing phase in heel-toe running,” Journal of biomechanics, vol. 25, [31] P. Wensing, R. Featherstone, and D. E. Orin, “A reduced-order recursive
no. 3, pp. 223–234, 1992. algorithm for the computation of the operational-space inertia matrix,”
[7] R. M. Walter and D. R. Carrier, “Ground forces applied by galloping in IEEE Int. Conf. on Rob. and Automation, 2012, pp. 4911–4917.
dogs,” J Exp Biol, vol. 210, no. 2, pp. 208–216, 2007. [32] R. Featherstone, Rigid Body Dynamics Algorithms. New York, NY:
[8] A. Tsiokanosa, E. Kellisb, A. Jamurtasa, and S. Kellisc, “The rela- Springer, 2008.
tionship between jumping performance and isokinetic strength of hip [33] J. Sensinger, “Selecting motors for robots using biomimetic trajectories:
and knee extensors and ankle plantar flexors,” Isokinetics and Exercise Optimum benchmarks, windings, and other considerations,” in IEEE
Science, vol. 10, no. 2, pp. 107–115, 2002. International Conference on Robotics and Automation (ICRA), 2010,
[9] “Toyota Corolla 2016 model brochure,” 2016. [Online]. Available: pp. 4175–4181.
http://www.toyota.com/corolla/ebrochure/ [34] B. Hannaford, P.-H. Marbot, P. Buttolo, M. Moreyra, and S. Venema,
[10] [Online]. Available: http://www.teslamotors.com “Scaling of direct drive robot arms,” The International Journal of
[11] [Online]. Available: http://www.scorpionsystem.com Robotics Research, vol. 15, no. 5, pp. 459–472, 1996.
[12] I. W. Hunter and S. Lafontaine, “A comparison of muscle with artificial [35] R. S. Wallace and J. M. Selig, “Scaling direct drive robots,” in IEEE
actuators,” in IEEE Solid-State Sensor and Actuator Workshop, 1992, International Conference on Robotics and Automation, vol. 3, May 1995,
pp. 178–185. pp. 2947–2954 vol.3.
[13] S. Seok, A. Wang, M. Chuah, D. J. Hyun, J. Lee, D. Otten, J. Lang, and [36] H. Asada and K. Youcef-Toumi, Direct-drive robots: theory and prac-
S. Kim, “Design principles for energy-efficient legged locomotion and tice. MIT press, 1987.
implementation on the MIT Cheetah robot,” IEEE/ASME Transactions [37] N. Farve, “Design of a low-mass high-torque brushless motor for appli-
on Mechatronics, vol. 20, no. 3, pp. 1117–1129, June 2015. cation in quadruped robotics,” Master’s thesis, Massachusetts Institute
[14] G. Abba and P. Sardain, “Modeling of frictions in the transmission of Technology, 2012.
elements of a robot axis for its identification,” IFAC Proceedings [38] A. Ananthanarayanan, M. Azadi, and S. Kim, “Towards a bio-inspired
Volumes, vol. 38, no. 1, pp. 7–12, 2005. leg design for high-speed running,” Bioinspiration & Biomimetics, vol. 7,
[15] A. Wang and S. Kim, “Directional efficiency in geared transmissions: no. 4, p. 046005, 2012.
Characterization of backdrivability towards improved proprioceptive [39] M. Hutter, C. Remy, M. Hoepflinger, and R. Siegwart, “Efficient and
control,” in IEEE International Conference on Robotics and Automation, versatile locomotion with highly compliant legs,” IEEE/ASME Transac-
May 2015, pp. 1055–1062. tions on Mechatronics, vol. 18, no. 2, pp. 449–458, April 2013.
[16] N. Hogan, “Impedance control: An approach to manipulation: Part [40] “OpenRAVE model of the KHR-4 “Jaemi” Hubo,” 2013. [Online].
II–implementation,” Journal of Dynamic Systems, Measurement, and Available: https://github.com/daslrobotics/openHubo
Control, vol. 107, no. 1, pp. 8–16, 1985. [41] H.-W. Park, M. Y. Chuah, and S. Kim, “Quadruped bounding control
[17] G. Hirzinger, N. Sporer, A. Albu-Schaffer, M. Hahnle, R. Krenn, with variable duty cycle via vertical impulse scaling,” in IEEE/RSJ
A. Pascucci, and M. Schedl, “DLR’s torque-controlled light weight robot International Conference on Intelligent Robots and Systems (IROS
III-are we reaching the technological limits now?” in IEEE International 2014), Sept 2014, pp. 3245–3252.
Conference on Robotics and Automation, vol. 2, 2002, pp. 1710–1716. [42] H.-W. Park, S. Park, and S. Kim, “Variable-speed quadrupedal bounding
[18] C. Ott, A. Albu-Schaffer, A. Kugi, and G. Hirzinger, “On the passivity- using impulse planning: Untethered high-speed 3D running of MIT
based impedance control of flexible joint robots,” IEEE Transactions on Cheetah 2,” in IEEE International Conference on Robotics and Automa-
Robotics, vol. 24, no. 2, pp. 416–429, 2008. tion, 2015.
[19] S. Haddadin, A. Albu-Schäffer, and G. Hirzinger, “Requirements for [43] H.-W. Park, P. Wensing, and S. Kim, “Online planning for autonomous
safe robots: Measurements, analysis and new insights,” The International running jumps over obstacles in high-speed quadrupeds,” in Proceedings
Journal of Robotics Research, vol. 28, no. 11-12, pp. 1507–1527, 2009. of Robotics: Science and Systems, Rome, Italy, July 2015.
[20] G. A. Pratt and M. M. Williamson, “Series elastic actuators,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems,
vol. 1, 1995, pp. 399–406.
[21] P. Gregorio, M. Ahmadi, and M. Buehler, “Experiments with an elec-
trically actuated planar hopping robot,” in Experimental Robotics III.
Springer Berlin Heidelberg, 1994, pp. 267–281.
[22] M. H. Raibert, “Legged robots,” Commun. ACM, vol. 29, no. 6, pp.
499–514, Jun. 1986.
[23] J. Hurst, J. Chestnutt, and A. Rizzi, “The actuator with mechanically
adjustable series compliance,” IEEE Transactions on Robotics, vol. 26,
no. 4, Aug 2010.
[24] I. Thorson and D. Caldwell, “A nonlinear series elastic actuator for
highly dynamic motions,” in 2011 IEEE/RSJ International Conference
on Intelligent Robots and Systems, Sept 2011.
[25] B. Vanderborght, A. Albu-Schaeffer, A. Bicchi, E. Burdet, D. Caldwell,
R. Carloni, M. Catalano, O. Eiberger, W. Friedl, G. Ganesh, M. Gara-
bini, M. Grebenstein, G. Grioli, S. Haddadin, H. Hoppner, A. Jafari,
M. Laffranchi, D. Lefeber, F. Petit, S. Stramigioli, N. Tsagarakis, M. V.
Patrick M. Wensing (S’09, M’15) received the Jeffrey Lang (F’98) received his SB (1975), SM
B.S., M.S., and Ph.D. degrees in Electrical and (1977) and PhD (1980) degrees from the Department
Computer Engineering from The Ohio State Uni- of Electrical Engineering and Computer Science
versity, Columbus, OH, USA, in 2009, 2013, and at the Massachusetts Institute of Technology. He
2014 respectively. He is currently a Postdoctoral joined the faculty of MIT in 1980 where he is now
Associate in the Biomimetic Robotics Laboratory the Vitesse Professor of Electrical Engineering. He
within the Department of Mechanical Engineering served as the Associate Director of the MIT Labo-
at the Massachusetts Institute of Technology. ratory for Electromagnetic and Electronic Systems
His research interests lie at the intersection of from 1991 to 2003 and now serves as an Associate
dynamics, optimization, and control as applied to Director of the MIT Microsystems Technology Lab-
agile, intelligent, and physically interactive robotic oratories.
systems. He was awarded an NSF Graduate Research Fellowship for his Professor Lang’s research and teaching interests focus on the analysis,
dissertation research on balance control strategies for humanoid robots. At design and control of electromechanical systems with an emphasis on: rotating
MIT, the results of his postdoctoral work on the MIT Cheetah 2 robot have machinery; micro/nano-scale (MEMS/NEMS) sensors, actuators and energy
received considerable publicity worldwide, with features from TIME, WIRED, converters; flexible structures; and the dual use of electromechanical actuators
and the Wall Street Journal. He served as Co-Chair for the Robotics and as motion and force sensors. He has written over 280 papers and holds
Automation Society Student Activities Committee from 2012 to 2014. 26 patents in the areas of electromechanics, MEMS, power electronics and
applied control. He has been awarded 4 best-paper prizes from IEEE societies,
and has received two teaching awards from MIT. He is a coauthor of
”Foundations of Analog and Digital Electronic Circuits” published by Morgan
Kaufman, and the editor of, and a contributor to, ”Multi-Wafer Rotating
MEMS Machines: Turbines Generators and Engines” published by Springer.
Professor Lang is a Fellow of the IEEE, and a former Hertz Foundation Fellow.
He served as an Associate Editor of ”Sensors and Actuators” between 1991
Albert Wang (S’14) received the B.S. and M.S. and 1994. He has also served as the General Co-Chair and Technical Co-Chair
degrees from the Department of Mechanical En- of the 1992 and 1993 IEEE MEMS Workshops, respectively, and the General
gineering, Massachusetts Institute of Technology, Co-Chair of the 2013 PowerMEMS Conference.
Cambridge, MA, USA, in 2010 and 2012 respec-
tively. He is currently a Ph.D. Candidate in the
MIT Biomimetic Robotics Lab, working on the
HERMES humanoid project. His research interests
include dynamic legged locomotion, actuation sys-
tems, and system architecture, and shared autonomy
for teleoperation.

Sangok Seok (S’10, M’14) received the B.S. and


M.S. degrees from the School of Mechanical and
Aerospace Engineering, Seoul National University, Sangbae Kim (M’08) received the B.S. degree in
Seoul, Korea, in 2002 and 2004, respectively, and the mechanical engineering from Yonsei University in
Ph.D. degree in Mechanical Engineering from the Korea, in 2001 and earned the M.S and Ph.D. de-
Massachusetts Institute of Technology, Cambridge, grees in the Biomimetics Laboratory, Stanford Uni-
MA, USA, in 2014. versity in 2004 and 2008 respectively. Prof. Sangbae
He was with the Korean branch of National In- Kim is currently the director of the Biomimetic
struments as an applications engineer and a mar- Robotics Laboratory and an Associate Professor of
keting engineer from 2004 to 2009, and with the Mechanical Engineering at MIT.
Mechatronics R&D center, Samsung Electronics as His research focuses on bio-inspired robotic plat-
a principal engineer from 2014 to 2015. He is currently a leader of the robotics form design by extracting principles from complex
group at Naver Labs. His research focus is on service robot platforms, which biological systems. Kim’s achievements on bio-
enable ambient intelligence, utilizing space and mobility. Dr. Seok received inspired robot development include the world’s first directional adhesive
the IEEE/ASME Transactions on Mechatronics best paper award for 2016. inspired from gecko lizards, and a climbing robot, Stickybot, that utilizes
the directional adhesives to climb smooth surfaces, featured in TIME’s best
inventions of 2006. The MIT Cheetah achieves stable outdoor running at an
efficiency of animals, employing biomechanical principles from studies of
the best runners in nature. This achievement was covered by more than 200
articles. He was a recipient of the King-Sun Fu Memorial Best Transactions
on Robotics Paper Award (2008), DARPA YFA (2013), and NSF CAREER
David Otten received the B.S. and S.M. degrees (2014) award.
from the Massachusetts Institute of Technology
(MIT), Cambridge in 1973 and 1974 respectively.
In 1974 he joined the MIT Electric Power Systems
Engineering Laboratory (EPSEL) as a staff engineer.
Since 1984 he has been a Principal Research Engi-
neer in the renamed Laboratory for Electromagnetic
and Electronic System (LEES) at MIT. His research
interests include instrumentation, power electronics,
and the micromouse robot contest.

You might also like