See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/329759867

MIT Cheetah 3: Design and Control of a Robust, Dynamic Quadruped Robot

Conference Paper · October 2018

DOI: 10.1109/IROS.2018.8593885

CITATIONS READS
86 22,316

6 authors, including:

Gerardo Bledt Patrick Wensing

Massachusetts Institute of Technology University of Notre Dame
16 PUBLICATIONS   258 CITATIONS    72 PUBLICATIONS   1,010 CITATIONS   

SEE PROFILE SEE PROFILE

Sangbae Kim
Massachusetts Institute of Technology
79 PUBLICATIONS   6,193 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Passive Magnetic Localization View project

Industrial Automation and Robotics View project

All content following this page was uploaded by Gerardo Bledt on 18 December 2018.

The user has requested enhancement of the downloaded file.

 MIT Cheetah 3: Design and Control of a Robust, Dynamic Quadruped Robot
Gerardo Bledt1,2 , Matthew J. Powell1 , Benjamin Katz1 ,
Jared Di Carlo2 , Patrick M. Wensing3 , and Sangbae Kim1

Abstract— This paper introduces a new robust, dynamic

quadruped, the MIT Cheetah 3. Like its predecessor, the
Cheetah 3 exploits tailored mechanical design to enable simple
control strategies for dynamic locomotion and features high-
bandwidth proprioceptive actuators to manage physical inter-
action with the environment. A new leg design is presented that
includes proprioceptive actuation on the abduction/adduction
degrees of freedom in addition to an expanded range of
motion on the hips and knees. To make full use of these
new capabilities, general balance and locomotion controllers
for Cheetah 3 are presented. These controllers are embedded
into a modular control architecture that allows the robot to
handle unexpected terrain disturbances through reactive gait
modification and without the need for external sensors or
prior environment knowledge. The efficiency of the robot is
demonstrated by a low Cost of Transport (CoT) over multiple
gaits at moderate speeds, with the lowest CoT of 0.45 found
during trotting. Experiments showcase the ability to blindly
climb up stairs as a result of the full system integration.
These results collectively represent a promising step toward Fig. 1: MIT Cheetah 3. The MIT Cheetah 3 quadruped robot platform is
a platform capable of generalized dynamic legged locomotion. an electrically powered robust robot capable of untethered 3D locomotion
in uncertain terrains.
I. I NTRODUCTION
Legged robots have a disruptive potential for application
in challenging man-made and natural terrains. In contrast to intelligence, they still lag behind even basic animals in terms
wheeled or tracked vehicles, legs provide a high degree of of the physical skills of mobility that makes them attractive
versatility in how they are deployed, offering marked ad- in the first place. Many animals naturally learn to walk and
vantages for operation in irregular environments. Designing even run within hours of being born, yet the ability to do
and controlling machines to realize these potentials has long both remains a challenge for legged robots.
motivated work across the legged robotics community, with Quadruped robots have recently shown impressive ad-
increasing intensity since the onset of the recent DARPA vancements in dynamic locomotion capabilities using var-
Robotics Challenge (DRC) [1]. In the aftermath of this event, ious actuation methods and control strategies. The use of
highly-capable quadrupeds (e.g., [2], [3]) and humanoids hydraulic actuators has proven to be successful in legged
(e.g., [4], [5]) in research labs have begun to lend credibility locomotion with Big Dog by Boston Dynamics [6] and IIT’s
to vision of legged robots making their way into our human HyQ quadruped [7]. These robots take advantage of the
environments to operate by our side or in our stead. hydraulic actuation system’s ability to output large forces
Yet, despite the broad range of capabilities displayed in at the joints. ANYmal at ETH [2] features an extensive
modern machines, many demonstrations remain instances of range of motion for completing autonomous tasks in real-
success without the broad versatility or reliability necessary world situations with the use of Series Elastic Actuators
for real-world deployments. It is a conundrum that while (SEAs). SEAs characteristically have good impact mitigation
legged machines may have the capacity to outperform many properties that are crucial for high speed locomotion, as well
humans and animals in cognitive tasks of reasoning and as excellent force control capacity. The MIT Cheetah 2 made
use of a custom proprioceptive actuator design that possesses
1 Department of Mechanical Engineering, MIT, Cambridge, MA high impact mitigation, force control, and position control
02139, [email protected],[email protected],
USA:
[email protected], [email protected] capabilities. This design enabled it to autonomously jump
2 Department of Electrical Engineering and Computer Science, MIT, over obstacles [3] and bound at high speeds of 6m/s [8],
Cambridge, MA 02139, USA: [email protected] but had a limited range of motion constraining it to sagittal
3 Department of Aerospace and Mechanical Engineering, University
of Notre Dame, Notre Dame, IN 46556, USA: [email protected] plane locomotion.
1,2,3 Authors are associated with the MIT Biomimetic Robotics Lab Expanding on the results from the MIT Cheetah 2, we
and contributed equally to the success of the robot present its successor, the MIT Cheetah 3, pictured in Fig-
This work was supported by National Science Foundation [NSF-IIS-
1350879] and the Air Force Office of Scientific Research [AFOSR Grant ure 1. This new platform employs similar philosophies as its
FA2386-17-1-4661] predecessor, such as the use of proprioceptive electric motors
 TABLE I: Physical Robot Parameters
Parameter Symbol Value Units
Mass m 45 kg
Body Inertia Ixx 0.35 kg · m2
Iyy 2.1 kg · m2
Izz 2.1 kg · m2
Body Length lbody 0.600 m
Body Width wbody 0.256 m
Body Height hbody 0.200 m
Leg Link Lengths l1 , l 2 0.34 m

and simple control models for planning movement. The new Fig. 2: Cheetah 3 Leg Design. One leg showing the 3 actuators, with a
robot possesses many critical improvements, including an cutaway view of the knee actuator.
expanded range of motion, higher force production capa-
bilities, and an ability to control these forces in full 3D
generality. A new general control architecture is presented B. Leg Design and Actuation
to make full use of these capabilities. Covering both design The proprioceptive actuation strategy used on Cheetah 3
and control, this paper provides a system level overview of results in exceptional force bandwidth, force capability, and
the important components that result in the MIT Cheetah 3’s speed, at the cost of force accuracy compared to systems
robust, dynamic locomotion capabilities. Section II presents with closed-loop force control. The static torque error of
the main hardware platform design considerations. The con- Cheetah 3’s actuators is roughly ±10% primarily due to gear
trol software architecture and components are described in friction. By comparison, the series elastic actuators used in
Section III. Results from selected hardware experiments are ANYmal, which measure the deflection of a series spring
shown in Section IV. Finally, Section V discusses implica- to control output torque, have far greater torque accuracy,
tions of the robotic platform and future research directions but much lower peak torque, speed, and torque bandwidth,
that it will allow the team to explore. especially at high torques [2]. However, we believe high
force accuracy is not a requirement for dynamic legged
II. D ESIGN locomotion, and our experiments have supported this claim.
Each of Cheetah 3’s actuators consists of a custom high
A. Overview torque density electric motor coupled to a single-stage 7.67:1
planetary gear reduction. The slightly higher gear ratio
The hardware design of Cheetah 3 builds on the actuation compared to Cheetah 2 (5.8:1) was chosen to improve the
paradigm of the MIT Cheetah 1 and 2 robots [9]. By load-carrying ability and low-speed efficiency of the robot.
using high torque density electric motors with backdriveable The legs are serially actuated, but to keep leg inertia low,
single-stage planetary gear reductions, and low-inertia legs, the hip and knee actuators are co-axially located at the hip
the Cheetah 3 robot can control ground reaction forces of each leg as shown in Figure 2.
through proprioception, without the use of any force sensors, The lower link is driven by a roller chain which passes
torque sensors, or series compliance at the joints or feet. The through the upper link, providing an additional 1.15x gear
Cheetah 2 robot was designed primarily for fast locomotion reduction. By using a chain rather than a linkage, the lower
in the sagittal plane, and used high performance actuators link can rotate 330◦ , letting the robot arbitrarily change
at the hip and knee joints, but not for abduction/adduction the orientation of its knees to point forwards or backwards.
(ab/ad). Cheetah 3 has nearly identical actuators on all three While the roller chain does introduce slight torque and speed
degrees of freedom on each leg, enabling fully 3D control ripple at the link, and adds some compliance between the
of ground reaction forces. knee actuator and the lower link, in practice these effects
The Cheetah 3 legs feature a greatly expanded range of have not caused high-level performance to suffer.
motion compared to Cheetah 2. The new ab/ad actuators The hip joint can rotate continuously, limited only by the
have a range of motion of more than ±45◦ , while new length of the wires to the knee actuator, allowing the robot
hip and knee designs allow the robot operate identically to potentially operate upside-down, climb up tall obstacles,
forwards, backwards, and flipped upside-down, and poten- or use its feet for manipulation above its body. Two ab/ad
tially use its legs for simple manipulation tasks as well as actuators are located between each pair of legs, and are
locomotion. coupled to the legs by linkages. See Figure 2 for a cutaway
The robot is powered by 450 Watt-hours of on-board view of leg actuation. The leg links are machined aluminum,
Lithium Polymer batteries providing approximately 2 hours and in total have a mass of 2.7 kg for all four legs, or
of run time, although the robot has internal space for twice only 6% of the mass of the robot. The end of each leg
the battery capacity, should longer run times be necessary. has a cover made from 10mm thick 60A urethane rubber,
 TABLE II: Actuator Parameters high-level commands by giving a desired translational veloc-
Parameter Value Units ity, ṗd , and turning rate, ψ̇d . These commands are used to
Gear Ratio 7.67 plan a smooth and controllable CoM reference trajectory that
Max Torque 230 Nm is relayed to the body and leg controllers. Various controllers
Max Joint Speed 21 Rad/s
and planners use the user-input commands and the estimated
robot state to generate force commands if the leg is in
stance or position commands if the leg is in swing. The
following sections describe the implementation of the major
providing cushioning and traction regardless of the contact
components of the system on the MIT Cheetah 3.
angle between the leg and the ground. Cheetah 3’s actuation
capabilities are summarized in Table II. A. Gait Planning
Cheetah 3’s actuation system was designed to give the
The Cheetah’s gait is defined by an event-based finite
robot performance headroom for tasks including high speed
state machine that uses a leg-independent phase variable
locomotion, jumping, carrying loads, and recovering from
to schedule nominal contact and swing phases. This allows
extreme disturbances. As a simple performance metric, with
for flexible gait definitions and fluid transitions between
the leg minimally extended, the robot is capable of producing
them. Trotting, bounding, and pacing gaits are presented
a purely vertical ground reaction force of over 700 N, about
in this paper, but the framework allows implementation of
1.6 times the weight of the robot, per leg. And at 70%
any gait definition with ease. The gaits were designed to
extension, a typical configuration during operation, vertical
mimic natural animal gaits by controlling the independent
force capability exceeds 1000 N per leg. By contrast, other
phases of each leg. This nominal gait plan is modified
quadruped robots for which data are available have far lower
during unexpected contact events on the legs. Since the robot
force-to-bodyweight capabilities: At minimal leg extension,
does not use any external environment sensors, a contact
ANYmal can produce around 0.54 bodyweights per leg [2],
detection algorithm probabilistically fuses encoder measure-
and HyQ2Max 0.84 bodyweights [7].
ments, estimated force, and expected gait phase to estimate
C. Computing and Low-Level Control Architecture the likelihood that each leg is in contact with an object
Cheetah 3 has a tiered computing architecture which [11]. Scheduledcontacts are defined by independent
boolean
enables low-level leg and motor control to be run at higher variables sφ ∈ 0 = swing,  1 = contact , while estimated

loop frequencies than locomotion control, and allows easy contacts are given by ŝ ∈ 0 = swing, 1 = contact . With
expansion of computing resources as needed for future this information, the robot can differentiate between normal
sensing, planning, and navitgation tasks. Locomotion con- operation, unexpected early contacts, and late missed con-
trol and state estimation are handled by the locomotion tacts to adjust its control actions appropriately [11].
computer: an embedded PC with a 2nd Gen Core i7 lap- B. Controller Model
top CPU, running Ubuntu Linux (kernel 4.1.33) with the
CONFIG PREEMPT RT patch. The locomotion computer Due to the irregularities of the hybrid nature arising from
receives user commands and logs data using Lightweight the repeated contact mode switches as legs enter or leave
Communication and Marshalling (LCM) [10]. LCM will stance or swing, it is often difficult to use traditional control
allow additional computers for vision, planning, and other methods for balancing the robots. Rapidly transitioning
tasks to easily communicate with the locomotion computer between periods of underactuation make the problem even
in the future. Leg-level control is performed by an ARM more complex. Therefore, the controllers used on the robot
Cortex-A8-based processor. The leg controller communi- make use of a simplified control model templates to optimize
cates with the locomotion computer at 1 kHz over EtherCAT, ground reaction forces at the footstep locations.
and performs leg-level control tasks like cartesian impedance By design, the robot has light limbs with low inertia
control or joint PD control at 4.5 kHz. The higher sample as compared to the overall body. For this reason, we can
frequency at the leg level allows for high-rate joint velocity reasonably simplify the control model to ignore the effects of
filtering as well as higher bandwidth tracking during swing the legs for planning ground reaction forces from the stance
phase and other position-controlled actions. Finally, each leg feet. In particular, the Cheetah 3 controller model employs
controller sends torque commands and receives joint encoder a commonly used linear relationship [12], [13] between the
measurements from the brushless motor drives for the three robot’s COM translational p̈c and body angular acceleration
motors in each leg, which perform the current control on ω̇ b and the forces F = (F1T , F2T , F3T , F4T )T acting on each
the motors at 20 kHz sample rate with 400 Hz closed-loop of the robot’s four feet. The controller model is given by
   
bandwidth. I3 ... I3 m(p̈c + g)
F = ,
III. C ONTROL A RCHITECTURE [p1 − pc ]× . . . [p4 − pc ]× IG ω̇ b
| {z } | {z }
The overall architecture of the system is depicted in the A b
(1)
block diagram in Figure 3. Each block is designed to be
modular so it can be easily substituted without any modifi- where m and IG are the robot’s total mass and centroidal ro-
cation to the other parts of the system. The operator provides tational inertia, g is the gravity vector and pi , i ∈ {1, 2, 3, 4}
 Desired CoM
State Force Control
Joint PD
Control
Gait Swing Leg Control
Scheduler

Operator GUI CoM State KF

Leg Contact Detection

Fig. 3: System Architecture Block Diagram. The user sends velocity and turning commands as well as general tunable parameters to the main computer.
The main Cheetah control computer is composed of three main parts: higher-level planning (green), leg and body control (red), and state estimation (blue).
The force and position commands are sent to the microcontrollers for each leg that relay the motor command to the robot.

are the positions of the feet. The term [pi − pc ]× is lie in the friction pyramid and that the normal forces lie
the skew-symmetric matrix representing the cross product within feasible bounds; these switch between support-leg
(pi − pc ) × Fi . and swing-leg bounds according to the scheduled contact,
sφ , as described in Section III-A.
C. Force Control - Balance Controller
D. Force Control - Model Predictive Control
One of the Cheetah 3 support leg control modes is a
Balance Controller, an implementation (with slight modi- As an alternative to the Balance Controller, the ground
fication) of the controller described in [12]. The Balance force control block can be replaced with a model-predictive
Controller enforces PD control on the center of mass and controller (MPC) that reasons about trajectory outcomes
body orientation, while also ensuring that foot forces satisfy over a longer temporal horizon. A model-predictive force-
friction constraints. The PD control law is given by planning approach can be formulated to plan ground reaction
    forces u to minimize a cost function
p̈c,d Kp,p (pc,d − pc ) + Kd,p (ṗc,d − ṗc ) k
= (2) X
ω̇ b,d Kp,ω log(Rd RT ) + Kd,ω (ω b,d − ω) J= (xi − xref ⊤ ref ⊤
i ) Qi (xi − xi ) + ui Ri ui (5)
i=0
The desired angular acceleration reflects PD control on
SO(3) wherein the desired and actual body orientations are where Qi and Ri are positive semidefinite matrices of
described using rotation matrices Rd and R, respectively, weights, xi ∈ R12 is the robot’s state at timestep i,
and the orientation error is obtained using the exponential consisting of position, velocity, orientation, and angular
map representation of rotations [14], [15]. velocity, xref
i is the desired state of the robot, and k is the
The goal of the Balance Controller is to resolve an optimal horizon length. Solving this MPC is challenging, as orien-
distribution of leg forces F that drive the approximate COM tation dynamics have nonlinear coupled effects. In recent
dynamics to the corresponding desired dynamics given by work [17] approximate dynamics are enforced with linear
  equality constraints based on a time-varying approximation
m(p̈c,d + g) that enables formulating the MPC problem as a quadratic
bd = . (3)
IG ω̇ b,d program. The resulting controller is able to anticipate and
Since the model (1) is linear, the controller can naturally be plan around periods of flight and underactuation. Optimal
expressed as the solution of a quadratic program (QP) [16] forces are updated at 30 Hz and are sent directly to the leg
controllers [17].
F ∗ = min12 (AF − bd )T S(AF − bd )
F ∈R E. Swing Leg Control
2 ∗
+ αkF k + βkF − Fprev k2 (4) Each of the footstep locations are calculated from the
s.t. CF ≤ d corresponding hip location using a linear combination of
the Raibert heuristic [18] and a velocity-based feedback term
The cost function in (4) reflects a tradeoff between three from the capture point formulation [19]. Since the robot does
goals: driving the COM dynamics to the desired values, min- not have external environment sensors, the footstep locations
imizing the forces used, and penalizing deviations between are projected onto an assumed ground plane. Therefore, the
the current QP solution and the solution at the previous time- step location on the 2D ground plane for foot i is calculated
∗
step, Fprev . The matrix S determines the relative priority from the hip of the robot by the following
in control over the rotational and translational motion, and Tcφ
r
z0
the gains α > 0 and β > 0 dictate the influence of pstep,i = ph,i + ṗc,d + (ṗc − ṗc,d ) (6)
the force normalization and solution filtering. Constraints | 2 {z } |
kgk
{z }
Raibert Heuristic
CF ≤ d are enforced to ensure that the optimized forces Capture Point
 -BL F
+L FL

where Tcφ is the nominal scheduled contact phase (stance) p BL p+FL

-FL

-BL F
+L

p BL p FL
FL

-FL

FR BR

time, z0 is the nominal height of locomotion, and ph,i CoM

BR

B
+L

CoM +
FR FR
provides the position of the corresponding hip i. -
BL
-B
+RL

B
+R
-BR

BL

FR
p BR p FR
B
+R

p BR - p FR
A PD controller with a feedforward term is used to com- FR
Time: 1.176s | Velocity: 0.75m/s Time: 1.281s | Velocity: 0.75m/s
FR

pute joint torques for tracking the Cartesian swing trajectory

for each foot. The feedforward force is computed on the p BL
BR
-BL F
+L

p FL
-BL F
+L

p BL p FL
BR

B
+L

onboard embedded PC at 1 kHz using the leg dynamics

B
+L

FL

FL

-FL

-BR

-BR

-FL

CoM CoM
τff,i = Ji⊤ Λi (B ai,ref − J˙i q˙i ) + Ci q̇i + Gi (7)
BL

BL

B
+R
+ +
-
B
+R

p BR p FR
FR
FR p BR - FR
p FR
FR
FR FR
Time: 1.428s | Velocity: 0.75m/s Time: 1.435s | Velocity: 0.75m/s
where Ji is the foot Jacobian, Λi is the operational space
inertia matrix, B ai,ref is the reference acceleration for the Fig. 4: Predictive Support Polygon. As the robot trots with velocity to the
right, the predictive support polygon (solid black) anticipates leg touchdown
swing trajectory, qi is a vector of joint configurations, Ci and guides the CoM away from the instantaneous physical support line
is the Coriolis matrix, and Gi is the torque due to gravity. (dashed black) towards the future support as contact modes switch. The
The full controller for tracking swing trajectories is vertices of the predictive support polygon are given by ξi .

τi = Ji⊤ [Kp (B pi,ref − B pi ) + Kd (B vi,ref − B vi )] + τff,i

(8) respectively. The quantity Φ represents the total weighting
where Kp and Kd are diagonal matrices of proportional and factor for the foot. Roughly, this notifies the robot that the
derivative gains. The Cartesian PD controller runs on the leg closer the leg is to the middle of its contact phase, the more
controllers at 4.5 kHz. confidently the robot can rely on that the leg as a support
To ensure the swing-leg PD controller remains stable, the point, while the closer it is to the middle of the swing phase,
gains are scaled with the apparent mass of the leg to maintain the less the leg can be used for balance in the near future.
approximately the same natural frequency Using the phase-based weighting factors, a set of virtual
points is defined for each leg. These virtual points lend
2
Kp,j = ωdes Λjj (q) (9) a method to continuously weight the predicted relative
availability of each leg as compared to its adjacent legs.
where Kp,j is the j-th diagonal entry in Kp , ωdes is the
These virtual points slide between the 2D ground projection
desired natural frequency, and Λjj is the j-th diagonal entry
line between the foot and the corresponding adjacent foot
in the operational space inertia matrix.
during the gait cycle. Each leg, i ∈ {F R, F L, BR, BL},
F. Virtual Predictive Support Polygon has two virtual points given by
 −   
In order to take advantage of the hybrid nature of legged ξi p i p i− Φi
= (13)
locomotion, a virtual support polygon is defined to provide ξi+ p i p i+ 1 − Φ i
a desired CoM location that generalizes across all gaits. where the superscript signifies a reference to the adjacent
By anticipating contact mode switches, the virtual support leg clockwise, −, or counterclockwise, +. With the current
polygon biases away from legs nearing the end of their foot position, we can now describe the predictive support
contact phase and towards legs about to touchdown. This polygon vertex for leg i to be
strategy enables the robot to maintain its forward momentum
during the gait while using selected footstep locations to Φi pi + Φi− ξi− + Φi+ ξi+
ξi = (14)
create a smooth reference trajectory that is automatically Φ i + Φ i − + Φ i+
adapted to the footholds online. The desired CoM position is then given as the average of
We define a phase-based approach that assigns each foot’s these weighted virtual support polygon vertices
contribution to the predictive support polygon, including
N
feet in swing. The goal is to design an informed nonlinear 1 X
p̂CoM,d = ξi (15)
phase-dependent weighting strategy by exploiting the robot’s N i=1
knowledge of which feet will be the next to lift off or touch
down and when these state changes are scheduled to occur. Figure 4 shows an example evolution of this virtual predic-
Weighting factors are generated based on the contact and tive support polygon during trotting, but the method extends
swing phases to all gaits. The virtual points defined in (13) and (14) are
labeled for the front right foot. In the sequence, the CoM
1 h  φ√  
1−φ
i
Kc φ = erf σc 2 + erf σc √2 (10) moves from the support line in the middle of stance, towards
2 0 1
the future support polygon as it expects the swing feet to
1h 
−φ
 
φ−1
i
touchdown and provide support.
Kc̄φ = 2 + erf σc̄ √2 + erf σc̄ √2 (11)
2 0 1
G. Posture Adjustment on Sloped Terrain
Φ = sφ Kcφ + s̄φ Kc̄φ (12)
As mentioned, part of the Cheetah 3 controller design
where Kcφ and Kc̄φ correspond to the adaptive weight- philosophy is to create a robust algorithm without reliance
ing factor during the scheduled contact and swing phase on contact sensors and vision systems. To enable Cheetah
to traverse stairs and sloped terrain without vision, we use however, error accumulation on yaw is unavoidable without
measurements of each footstep location pi = (pxi , pyi , pzi ) fusion using exteroceptive information such as vision [23].
to approximate the local slope of the walking surface and The second stage of the state estimation uses the ori-
adjust the robot’s desired posture. In particular, the walking entation estimate 0R̂ b along with kinematic measurements
surface is modeled as a plane from the legs to estimate the base position and velocity. In
contrast to previous state estimation techniques that posed
z(x, y) = a0 + a1 x + a2 y. (16) this problem as an extended Kalman Filter [20], the two-
Coefficients a = (a0 , a1 , a2 )T of (16) are obtained through stage approach allows this secondary fusion to be posed as
the solution of the least squares problem a conventional Kalman filter. This simplifies analysis and
tuning of the filter and guarantees that the filter equations
a = (W T W )† W T pz (17) will never diverge in finite time.
 
W = 1 px py 4×3 (18) In continuous time, the process equations are modeled as
0
that finds the plane that best fits the collection of measure- ṗb = 0 vb (21)
ments of the most recent contact point for each leg. Note 0 0 0
v̇b = R̂ b ab + ag + wv (22)
that px , py and pz contain data for each leg, e.g. px = 0
ṗi = wpi ∀i = {1, . . . , 4} (23)
(px1 , px2 , px3 , px4 ), and that (W T W )† is the Moore-Penrose
pseudoinverse of W T W . As described in Section IV, the where 0 pb is the position of the body, 0 vb its velocity,
estimated walking surface is used to adjust the robot’s 0
ag = [0, 0, − g]T the gravitational acceleration, and
posture to accommodate uneven terrain. 0
pi the position of foot i. The white noise terms wv and
wpi characterize the noise in the accelerometer and any
H. State Estimation
variability in the foot positions. As in [20] the variability
Cheetah 3 estimates is it body states through a two-stage in the foot position is modeled to account for effects of foot
sensor fusion algorithm that decouples estimation of body slip in stance, or can also capture foot swing by inflating the
orientation from estimation of the body position and velocity. process noise on that foot to a high level. Viewing 0R̂ b ab +
It first uses its IMU to estimate orientation. Following this 0
ag as an input to the system, these equations represent
stage, it then fuses accelerometer measurements with leg a linear time invariant process, and thus are amenable to
kinematics to estimate base position and velocity via a conventional Kalman filtering. The continuous time model
procedure inspired by [20]. A similar two-stage scheme has is converted into discrete time with a 1ms time step using
also been recently applied for state-estimation in the HRP2 an identical procedure to [20].
humanoid, as reported in [21]. Leg kinematics provide measurements of the relative
The first stage of the state estimation employs an orienta- position vector between each foot and the body to de-
tion filter [22] using both the IMU gyro and accelerometer drift estimates 0 p̂b , 0 v̂b , and 0 p̂i for each foot. Letting
readings. The main idea of the filter is that the gyro pro- 0
prel (qi , 0R̂ b ) denote the relative foot position as computed
vides an accurate reading of the high-frequency orientation by kinematics, a measurement residual is generated
dynamics, whereas the presence of a gravity bias on the

accelerometer allows it to de-drift the estimate at a com- ep,i = 0 p̂i − 0 p̂b − 0 prel (qi , 0R̂ b )
paratively lower frequency. Letting the orientation estimate Similarly, the velocity of the foot relative to the body can be
0
R̂ b as the orientation of the body relative to the i.c.s., the computed from the leg angles, velocities, and the body orien-
filter updates this estimate according to tation and angular velocity. This computation is denoted as
0˙
  0
ṗrel (qi , q̇i , 0R̂ b , b ωb ). Under the assumption that each foot
R̂ b = 0R̂ b b ωb + κωcorr × (19)
is fixed, this provides an associated measurement residual
where κ > 0 is a correction gain and ωcorr is a correction

ev,i = −0 v̂b − 0 ṗrel (qi , q̇i , 0R̂ b , b ωb )
angular velocity to align the accelerometer reading ab with
its gravity bias Finally, a contact height hi is assumed for each foot with
  associated measurement residual:
0
ab

ωcorr = × 0R̂Tb 0 ehi = [0, 0, 1]T 0 p̂i − hi
kab k
1
Random multivariate Gaussian’s are used to capture the
The time constant of the de-drifting from this term can be measurements errors for each residual. Similar to the process
approximated by κ−1 . In practice κ is heuristically decreased noise wpi on the feet, measurement covariances for foot
during highly-dynamic portions of the gait where kab k >> i are increased to a high value during swing such that
g with the swing leg measurements are effectively ignored in this
κ = κref max(min(1, 1 − ||ab − g||/g), 0) (20) fusion process. Although space permits a full development,
the residuals are used within the innovations step of a
where g is the acceleration of gravity and κref is chosen as Kalman filter to provide estimates 0 p̂b , 0 v̂b , and 0 p̂i used
κref = 0.1. This process is effective to de-drift pitch and roll, for feedback to the Balance Control or MPC components.
 Power Cost of Transport
500 1.8

1.6 Trotting
450
Bounding
1.4 Pacing
400
Power (W)

1.2

COT
350
1
300
0.8
250 0.6

200 0.4
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Speed (m/s) Speed (m/s)
(a) (b)
Fig. 5: Power consumption and Cost of Transport for Dynamic Gaits.
Scattered points correspond to 1 second of averaged data. Lines are a linear
fit to power consumption.

IV. R ESULTS
This section presents results from experimental testing
with Cheetah 3. The selected results highlight key features
of the system: the ability to control different dynamic gaits (c)
within the same framework, design-enhanced capabilities Fig. 6: Unique Design Capabilities. The robot is able to use the controllers
that leverage Cheetah’s large range of motion, and the ability described regardless of leg orientation (a), balance in extreme configurations
(b), and make use of a wide range of motion to position its limbs (c).
to traverse stairs without external terrain information. A
video of the results is included as supplementary material.
A. Steady State Locomotion powerful custom electric motor for the ab/ad DoF, the robot
With locomotion being the primary function of the robot, gains the ability to apply large lateral forces. The robot gains
a number of experimental trials were run with three different an improved range of motion that allows for exaggerated yaw
gaits at varying speeds to quantify the efficiency of the robot motions that are depicted in Figure 6b. In addition, having no
during steady state behavior. Trotting is achieved through defined forward or top, the robot is able to rotate its legs fully
the use of both the instantaneously reactive QP controller above its body. Using the chain in the leg design gives the
described in Section III-C, as well as the convex MPC robot the ability to omni-directionally rotate the knee joint
described in Section III-D. Bounding and pacing gaits are as in Figure 6c. In the future this can be taken advantage
possible only through the use of the convex MPC as they of to allow the robot to use its legs for both locomotion as
require predictive knowledge of the upcoming flight times well as to interact with its surroundings and carry out tasks.
to remain balanced throughout the gait.
C. Blind Stair Climbing
Experiments were held indoors on a flat treadmill with
adjustable velocities. Power consumption was logged by Although the robot has no preemptive knowledge of the
measuring battery current, and does not include power con- terrain, the control architecture described allows robust loco-
sumption of the computer, which is an additional 85 watts. motion over highly unstructured and abrupt terrain changes
The gaits and controllers tested used were not optimized for such as staircases. Since there are no external sensors, the
P robot assumes a flat, unobstructed world, so it must act
CoT or efficiency. CoT is calculated as mgv , where v is the
linear velocity of the robot, m is the mass of the robot, and P reactively to its current sensed state when encountering
is electrical power consumption. Experimentally determined unforeseen obstacles. With the contact detection algorithm
CoT results for each of the three gaits are shown in Figure mentioned in Section III-A, unexpected collision with ob-
5. As opposed to animals which use internal springs and jects such as stairs can be detected and used to modify
muscles that are efficient in locking, the robot’s electric the swing foot trajectories and balance strategy through the
motors are dominated by force production cost rather than event-based gait scheduler. Once the robot senses the step,
the mechanical CoT. Therefore, the most efficient gait for it adjusts its posture based on Section III-G and expects
the robot at a given speed does not necessarily correspond contacts to be along the estimated plane.
to the one used by animals as determined in [24]. A staircase was built for testing with the US standard of a
7in (17.8cm) rise and 11in (27.9cm) run. Figure 7 shows the
B. Design-Enhanced Capabilities robot during a stair climbing experiment as it successfully
The robot’s design and robust control architecture allow reaches a platform at the top of the staircase. Several tests
for a variety of unique behaviors. Important improvements were run in which the robot successfully climbed up the
to the design over the MIT Cheetah 2 greatly enhance the stairs with failures being due to operator errors attempting
possibilities for this new robot to be used in real-world to take the robot outside of its kinematic limits rather than
applications. Since the robot ground reaction force control due to failures in the control architecture itself. The attached
planning is independent of the leg configuration, locomotion video demonstrates the ability of the robot to reject large
can occur regardless of which way the knees are pointed as disturbances as it is pulled with a rope off of the steps. In
shown in Figure 6a. With the new design incorporating the all cases it regains balance and continues to try to climb the
 [4] S. Kuindersma, R. Deits, M. Fallon, A. Valenzuela, H. Dai, F. Per-
menter, T. Koolen, P. Marion, and R. Tedrake, “Optimization-based
locomotion planning, estimation, and control design for the atlas
humanoid robot,” Autonomous Robots, pp. 1–27, 2015.
[5] N. A. Radford and et al., “Valkyrie: Nasa’s first bipedal humanoid
robot,” Journal of Field Robotics, vol. 32, no. 3, pp. 397–419, 2015.
[6] M. Raibert, K. Blankespoor, G. Nelson, R. Playter, and the Big-
Dog Team, “Bigdog, the rough-terrain quaduped robot,” in Proceed-
ings of the 17th World Congress, pp. 10822–10825, 2008.
[7] C. Semini, V. Barasuol, J. Goldsmith, M. Frigerio, M. Focchi, Y. Gao,
and D. G. Caldwell, “Design of the hydraulically actuated, torque-
controlled quadruped robot hyq2max,” IEEE/ASME Transactions on
Fig. 7: Blindly Climbing Stairs. With the control framework described Mechatronics, vol. 22, pp. 635–646, April 2017.
throughout the paper, the robot was able to successfully climb up stairs [8] H.-W. Park, P. M. Wensing, and S. Kim, “High-speed bounding with
without the use of external environment sensors. the MIT Cheetah 2: Control design and experiments,” International
Journal of Robotics Research, vol. 36, no. 2, pp. 167–192, 2017.
[9] P. M. Wensing, A. Wang, S. Seok, D. Otten, J. Lang, and S. Kim,
“Proprioceptive actuator design in the MIT Cheetah: Impact mitigation
steps after being forced off of its intended objective. This and high-bandwidth physical interaction for dynamic legged robots,”
ability to reactively sense and adapt to unexpected obstacles IEEE Transactions on Robotics, vol. 33, no. 3, pp. 509–522, 2017.
[10] A. S. Huang, E. Olson, and D. C. Moore, “Lcm: Lightweight com-
indicates that the robot is capable of robustly operating munications and marshalling,” in IEEE/RSJ International Conference
and adapting to realistic, unstructured, and changing terrains on Intelligent Robots and Systems, pp. 4057–4062, Oct 2010.
without relying on precise perception and planning. [11] G. Bledt, P. M. Wensing, S. Ingersoll, and S. Kim, “Contact model
fusion for event-based locomotion in unstructured terrains,” in IEEE
Int. Conf. on Robotics and Automation (to appear), 2018.
V. C ONCLUSIONS AND F UTURE W ORK [12] M. Focchi, A. del Prete, I. Havoutis, R. Featherstone, D. G. Caldwell,
and C. Semini, “High-slope terrain locomotion for torque-controlled
This paper has introduced the MIT Cheetah 3 quadruped quadruped robots,” Autonomous Robots, vol. 41, pp. 259–272, Jan
robot as a new legged robot platform. An improved design 2017.
over MIT Cheetah 2, this robot has a larger range of [13] B. Stephens and C. Atkeson, “Push recovery by stepping for humanoid
robots with force controlled joints,” in IEEE-RAS International Con-
motion with full 3D control capabilities. The robustness ference on Humanoid Robots, pp. 52–59, Dec. 2010.
demonstrated without the use of external sensing signifies [14] F. Bullo and R. M. Murray, “Proportional derivative (PD) control on
that it is capable of successful locomotion in challenging the Euclidean group,” in Proc. of the European Control Conference,
(Rome, Italy), pp. 1091–1097, June 1995.
situations without relying on knowledge of the environment [15] R. M. Murray, Z. Li, and S. S. Sastry, A Mathematical Introduction
prior to executing its control strategy. to Robotic Manipulation. CRC Press, 1994.
The general control architecture lends itself to the use of [16] C. Gehring, S. Coros, M. Hutter, M. Bloesch, M. Hoepflinger, and
R. Siegwart, “Control of dynamic gaits for a quadrupedal robot,” in
any number of different controllers with minimal modifica- IEEE International Conference on Robotics and Automation (ICRA),
tions to the software or hardware. Promising initial results pp. 3287–3292, May 2013.
of a novel nonlinear Policy Regularized Model Predictive [17] J. DiCarlo, P. M. Wensing, B. Katz, G. Bledt, and S. Kim, “Dynamic
locomotion in the MIT Cheetah 3 through convex model predictive
Control framework (PR-MPC) further expand on the current control.” Submitted to IROS 2018.
controllers presented by optimizing both ground reaction [18] M. H. Raibert, Legged robots that balance. Cambridge, MA, USA:
forces as well as footstep locations using simple physics- MIT Press, 1986.
[19] J. Pratt, J. Carff, S. Drakunov, and A. Goswami, “Capture point: A
based heuristics to regularize the prediction [25]. The robot step toward humanoid push recovery,” in IEEE-RAS Int. Conf. on
will intelligently decide on where to best place the feet for Humanoid Robots, (Genova, Italy), pp. 200–207, Dec. 2006.
any given gait and take advantage of its dynamics for steady [20] M. Bloesch, M. Hutter, M. Hoepflinger, S. Leutenegger, C. Gehring,
C. D. Remy, and R. Siegwart, “State estimation for legged robots
state locomotion and disturbance rejection. - consistent fusion of leg kinematics and IMU,” in Proceedings of
While the robot currently operates by taking in user inputs Robotics: Science and Systems, (Sydney, Australia), July 2012.
[21] T. Flayols, A. D. Prete, P. Wensing, A. Mifsud, M. Benallegue,
for commanded velocity and turning rate, a higher level path and O. Stasse, “Experimental evaluation of simple estimators for
planner will be developed to allow autonomous operation humanoid robots,” in 2017 IEEE-RAS 17th International Conference
in the world. This higher level planner can be naturally on Humanoid Robotics (Humanoids), pp. 889–895, Nov 2017.
[22] R. Mahony, T. Hamel, and J. M. Pflimlin, “Nonlinear complemen-
integrated in the framework both as a path planner, as well tary filters on the special orthogonal group,” IEEE Transactions on
as providing information for enhancing the controllers and Automatic Control, vol. 53, no. 5, pp. 1203–1218, 2008.
state estimation. [23] M. F. Fallon, M. Antone, N. Roy, and S. Teller, “Drift-free humanoid
state estimation fusing kinematic, inertial and lidar sensing,” in 2014
IEEE-RAS International Conference on Humanoid Robots, pp. 112–
R EFERENCES 119, Nov 2014.
[24] S. Seok, A. Wang, M. Y. Chuah, D. Otten, J. Lang, and S. Kim,
[1] “DARPA Robotics Challenge.” http://www.theroboticschallenge.org/, “Design principles for highly efficient quadrupeds and implementation
Nov. 2012. on the mit cheetah robot,” in 2013 IEEE International Conference on
[2] M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V. Tsounis, Robotics and Automation, pp. 3307–3312, May 2013.
J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, R. Diethelm, [25] G. Bledt, P. M. Wensing, and S. Kim, “Policy-regularized model
S. Bachmann, A. Melzer, and M. Hoepflinger, “ANYmal - a highly predictive control to stabilize diverse quadrupedal gaits for the MIT
mobile and dynamic quadrupedal robot,” in IEEE/RSJ International Cheetah,” in IEEE/RSJ Int. Conf. on Intelligent Robots and Systems,
Conference on Intelligent Robots and Systems, pp. 38–44, Oct 2016. 2017.
[3] H.-W. Park, P. Wensing, and S. Kim, “Online planning for autonomous
running jumps over obstacles in high-speed quadrupeds,” in Proceed-
ings of Robotics: Science and Systems, (Rome, Italy), July 2015.

View publication stats