Control of A High Performance Bipedal Robot using Viscoelastic

Liquid Cooled Actuators

Junhyeok Ahn1 , Donghyun Kim2 , Seunghyeon Bang3 , Nick Paine4 , Luis Sentis3†

Abstract— This paper describes the control, and evaluation of

a new human-scaled biped robot with liquid cooled viscoelastic
actuators (VLCA). Based on the lessons learned from previous
work from our team on VLCA, we present a new system design
embodying a Reaction Force Sensing Series Elastic Actuator
and a Force Sensing Series Elastic Actuator. These designs
are aimed at reducing the size and weight of the robot’s
actuation system while inheriting the advantages of our designs
such as energy efficiency, torque density, impact resistance
and position/force controllability. The robot design takes into
consideration human-inspired kinematics and range-of-motion,
while relying on foot placement to balance. In terms of actuator
control, we perform a stability analysis on a Disturbance
Observer designed for force control. We then evaluate various
position control algorithms both in the time and frequency
domains for our VLCA actuators. Having the low level baseline
established, we first perform a controller evaluation on the
legs using Operational Space Control. Finally, we move on to Fig. 1. DRACO is a high performance biped equipped with liquid cooled
VLCAs. It is designed to nimbly maneuver in cluttered environment.
evaluating the full bipedal robot by accomplishing unsupported
dynamic walking.
have small motors for the ankles. This choice is intended to
I. INTRODUCTION significantly reduce leg distal mass and allow for faster swing
During the DARPA Robotics Challenges (DRC), several cycles during dynamic locomotion. Faster swing cycles can
international teams explored the use of humanoid robots in have benefits for collision recovery.
emergency response tasks. Some of the robots employed ATRIAS was built to mimic a spring-mass model for fa-
in the DRC were humanoid bipeds. Biped robots could cilitating unsupported passive-ankle dynamic walking. Since
have benefits over other embodiments in tasks such as then it has achieved walking in various terrains and running
maneuvering in tight spaces. [9], [10]. However, its controller assumes a mechanical
In terms of design, the DRC humanoid robots SCHAFT approximation of the robot design to an idealized pendulum
[1] and JAXON [2] were built to improve heat dissipation model with most of its mass located at the Center-of-Mass
and thermal management capabilities through the use of (CoM). This can limit the distribution of masses and topol-
liquid-cooled electric actuators. Another two humanoids, ogy of unsupported passive-ankle robots. Besides ATRIAS
ESCHER [3] and VALKYRIE [4], [5] were built with force there is not much research published on the design of robots
control and ground impact resistance via the use of Series capable of unsupported passive-ankle walking.
Elastic Actuators (SEAs). These robots possess actuated Motivated by agile locomotion, our paper introduces new
ankles for locomotion which is good for standing manip- bipedal robot technology with emphasis on power-dense
ulation but at the same time result in a larger leg distal electric actuation and the achievement of unsupported dy-
mass. In turn, it slows down the stepping strides and makes namic balancing. DRACO, illustrated in Fig. 1, is an adult-
the robot heavier. Biped robots like ATRIAS, CASSIE, and size lower-body biped robot designed to maneuver nimbly
MERCURY [6]–[8] are either void of ankle actuation or in cluttered environments. The lower-body mechanical ar-
chitecture was designed to reduce volume and weight for
1 J. Ahn is with the Department of Mechanical Engineer- the rated height and payload. Three electric actuators drive
ing, The University of Texas at Austin, TX, 78712, USA
[email protected] each leg’s hip motions. One located near the pelvis providing
2 D. Kim is with the Department of Mechanical Engineering, hip rotation. Another one located on the outer lateral hip
Massachusetts Institute of Technology, MA, 02139, USA providing hip abduction/adduction. And another one located
[email protected] on the upper thigh providing hip flexion/extension. Knee
3 S. Bang and L. Sentis are with the Department of Aerospace
Engineering and Engineering Mechanics, The University of Texas flexion/extension is provided by an actuator located on the
at Austin, TX, 78712, USA [email protected], lower part of the thigh. Finally, ankle pitch is provided by a
[email protected] small actuator located on the leg’s calf with the purpose of
4 N. Paine is with Apptronik, Austin, USA
[email protected] statically balancing the robot. Thus during walking, DRACO
† L. Sentis is the corresponding authors. is designed to perform passive-ankle dynamic locomotion
Fig. 2. DRACO’s joints approximate the range of motion of the adult human body with 1.5, 1.19, 2.27, 2.14, 1.55 rad for hip rotation, abduction, flexion,
knee flexion and ankle flexion respectively. (a), (b), (c-1), (c-2), (d), and (e) shows ankle flexion/extension, hip flexion/extension, hip external rotation, hip
internal rotation, knee flexion/extension and hip abduction/adduction. Color maps in (e) and (f) represent the density score for the workspace of the foot.

without relying on active ankle torques. evaluation of a high performance biped robot with VLCAs
DRACO is actuated by Viscoelastic Liquid Cooled Ac- that can achieve unsupported passive-ankle balancing. An-
tuators (VLCAs) which include viscoelastic elements in the other contribution is the study of stability and performance
drivetrain in order to improve joint position controllability analysis for the proposed joint and multi-DOF controllers.
as reported in [11]. Liquid cooled Reaction Force Sensing The remainder of the paper is organized as follows: Section II
Series Elastic Actuators (RFSEA) are used for the hip presents the mechatronic design of DRACO. Stability and
and knee joints, reducing actuation weight while increasing performance of VLCA are studied in Section III. OSC and
energy efficiency, torque density, impact resistance and posi- unsupported dynamic balancing via WBC are described in
tion/force controllability. Liquid-cooled Force Sensing Series Section IV. Finally, Section V concludes the paper.
Elastic Actuators (FSEA) drive the pitch ankle joints for
II. M ECHATRONIC D ESIGN
ankle flexion/extension control. To increase heat dissipation
on the electric motors, these VLCAs use liquid-cooling A. Mechatronics Overview
motor jackets [12] enabling 2.5x higher continuous torques The DRACO biped is 1.30 m tall, weighs 37 kg, and
on all joints compared to conventional electric actuators. achieves a similar range of motion than an adult human
For actuator control, we use a decoupled control strategy regarding leg and hip motions. DRACO has ten actuated
as described in [5], which relies on a rigid joint model. Each DOFs including three for the hip structure, one for the knee
low level actuator controller acts as an idealized force or joint and one for the ankle. The range of motion of each DOF and
position source to facilitate the use of control architectures workspace of the robot’s foot are shown in Fig. 2. Gener-
for multi Degree Of Freedoms (DOFs), such as Whole Body alized coordinate of DRACO, and mechatronic components
Control (WBC) [13]. Previously, we showed high fidelity including actuators, auxiliary sensors and electronic boards
control of SEAs via Disturbance Observer (DOB) controllers are illustrated in Fig. 3.
designed with the assumption of a time invariant nominal Unlike many humanoid robots, it does not have ankle roll
SEA model [14] and the improvement of joint position con- actuation. This allows to significantly reduce distal mass and
trollability by placing viscoelastic materials on the actuator’s therefore enhance swing speed motion. Based on the lessons
drivetrain [11]. In this paper, we prove the robust stabilization learned from previous works [14], [16], [17], the hardware
capabilities of our DOB controllers. We evaluate various joint has been designed with performance and mechanical safety
position feedback controllers depending upon 1) using either consideration while reducing it’s overall weight and the risk
motor and spring encoders versus linear potentiometers for of overheating motors.
feedback, and 2) whether to include force inner feedback
loop to decrease mechanical friction and stiction. Finally, B. Viscoelastic Liquid Cooled Actuators
we implement and test Operational Space Control (OSC) To achieve the design objectives and protection against
[15] and dynamic balance control using WBC [8]. We external impacts, VLCAs are employed to actuate the robot
demonstrate accurate and stable actuator position tracking joints. VLCA is a family of prismatic SEA with a viscoelastic
in the operational space and unsupported dynamic balancing material instead of metal springs and active liquid cooling,
with well-regulated motor temperatures thanks to the liquid first introduced in our previous work [11]. There we in-
cooling system. vestigated its power density, energy efficiency, high-fidelity
The main contribution of this paper is on the control and force control and joint position control. In this section, we
(a) (b) RFSEA

IMU

Force sensor

Medullua board FSEA

Axon control board

Load Part

Motor Part

Rubber Part
Liquid reservoir and fan

Fig. 3. (a) shows generalized coordinate of the robot. Floating base frame and actuated DoFs are represented as black and blue axis, respectively. (b)
Mechatronic Parts and Schematic diagrams of the VLCAs.

introduce RFSEA type VLCA actuators and FSEA type to incorporate in the calf of the leg. For more detailed
VLCA actuators used in DRACO. information, readers are referred to our previous work [11].
The schematic diagrams of the VLCAs and the nomen-
clatures are shown in Fig. 3. In the diagram, Jm (kg m2 ), III. ACTUATOR C ONTROL
Bm (N s/m), Mr (kg), Br (N s/m), Ml (kg) and Bl (N s/m)
are motor inertia, motor damping coefficient, elastomer mass, In Fig. 4(a), we outline our overall robot control struc-
elastomer damping coefficient, load mass and load damping ture which contains a multi-joint control block coordinating
coefficient. θm (rad), xr (m) and xl (m) are displacement multiple decoupled joint controllers. In order to control the
of the motor, elastomer and the output load which are robot’s dynamic locomotion behavior effectively, the low
the actuator states. τm (N m) is motor torque which is level actuator controllers are designed to deliver certain high
the actuator input, and τl (N m) and q(rad) are the joint performance specifications. In our recent work, [11], we
torque and position which are the joint output. Nm (rad/m) studied joint position controllability and torque controllabil-
(Nl (rad/m), respectively) is the speed reduction ratio of the ity for liquid cooled viscoelastic actuators. In this section, we
motor (output joint, respectively) provided by the ball screw extend this analysis by studying the effects of different types
(the actuator position, respectively). To measure the actuator of sensors for joint position feedback as well as performing
states, we place a quadrature encoder at motor and elastomer a stability analysis of DOBs used for force control.
side to measure motor angle, θm , and elastomer deflection,
xr . In addition, we adopt a linear potentiometer to measure
absolute position, xl , of the actuator. A. Model & Identification
The RFSEA transmits mechanical power when the BLDC Let us consider the transfer functions for RFSEA actuators
motor turns a ball nut via a low-loss timing belt and pulley, [18] using the nomenclatures shown in Fig. 3,
which causes a ball screw to exert a force to the actuators
output. Rigid assembly consists of the electric motor, the
 
kτ Pm (s) Pr (s) + Pl (s)
ball screw, and the ball nut connected in series to the Pi→θm (s) = 2
compliant viscoelastic element, which in turn connects to Nm Pm (s) + Pl (s) + Pr (s)
the mechanical ground of the actuator. When the actuator kτ Pm (s)Pr (s)
Pi→xr (s) = −   (1)
exerts a force, it causes the viscoelastic element to contract Nm Nm 2 P (s) + P (s) + P (s)
m l r
and extend. The liquid cooling system allows to increase
the maximum continuous current by a factor of 2.5 without Pi→F (s) = Kr Pi→xr (s)
thermal failure. PF →q (s) = Nl Pl ,
The FSEA liquid cooled actuator transmits mechanical
via a BLDC motor in series to the ball screw. The FSEA where i(A), F (N) and kτ (N m/A), are motor current, actu-
type VLCA includes a compliant element between the ball ator force and motor constant. P◦→4 (s) represents transfer
screw and the actuator output. As a result of the drivetrain, functions with input signal ◦ and output signal 4. In addition
it provides a long, thin and lightweight design that is ideal Pm , Pr and Pl correspond to the motor, elastomer, and load
(a) TABLE I
Section. IV Motor
Joint ACTUATOR PARAMETERS
current
Motor
Robot command Actuator
current VLCA
Control
Single-joint Robot
Control controller Sensed Actuator Jm Bm Ml Bl Mr Br Kr
Joint
Section. V state Joint 1
state 1.6e−5 2.0e−4 2953 ∼ ∞ 2.0e4 1.3 1.6e4 9.5e6
Joint 10
d
(b)
G(s) high fidelity force control of SEAs measuring force via com-
pression of a compliant element and [5], [20] studied PID,
model-based and DOB structures to achieve high fidelity
force tracking. Since our actuator model considers a variable
load, unknown a priori, we first design a nominal plant
DOB
based on an infinite load mass assumption. Fig. 4(b) depicts
(c)
our force feedback controller with a DOB where QA , QB
and Cτ correspond to low pass filters and a PD controller
respectively, with expressions,
C• = P• + D• s
(i) (ii)
Inner Force Loop 1 (3)
Q• = √ .
(s/ωc )2 + 2(s/ωc ) + 1
Here, P• and D• are proportional and derivative gains, and
ωc is the cut-off frequency of the filter defined by Q• .
In addition, Pi→F represents the actual actuator plant with
∞
Fig. 4. (a) shows our decoupled control approach. (b) represents a block motor current as input and actuator force as output. Pi→F
diagram for force feedback control incorporating a DOB structure. (c) shows represents a model of the actuator plant using the infinite
a block diagram of our actuator position feedback controller.
load mass assumption.
transfer functions with expressions, In this section, we provide a formal analysis on the
robustness and stability of DOB-based controllers under
1
Pm (s) = uncertain loads. In order to study the performance of our
s2
Jm + Bm s DOB controller given the time varying output load we will
1
Pr (s) = 2
(2) apply perturbation theory analysis [21]. We derive the state-
Mr s + B r s + K r space equations of our DOB, G(s) in Fig. 4(b) by using the
1
Pl (s) = , method explained in [22], resulting in the equations,
2
Ml s + B l s
x1 = F
Note that Ml in the above equation is indefinite since it varies
with the joint configuration and the contact state of the robot ẋ1 = x2
∞
reflected onto the actuator. Pi→F in Fig. 4(c) represents the ẋ2 = ψ > z + φ> x + gC(ζ − ξ) + g(ur + d) (4)
same plant as Pi→F but with an infinite load mass Ml ' ∞, ż = Sz + Gx
i.e. an ideal rigid contact.
ż = Sz + Gx,
kτ and Nm are obtained from product sheets, Nl is derived
from a pre-computed look-up table, and Kr is approximated and
by measuring elastomer displacements given known applied
forces. To obtain other parameters, we do so via system iden- g 1 >
tification techniques. We generate a motor current following τ ξ̇ = Aξ + BC(ζ − ξ) + B − ψ z + ψ > z
g g
an exponential chirp signal, with frequencies between 0.01
g
>
+ (φ − φ) x + B(ur + d) (5)
Hz to 200 Hz and measure actuator force as an output signal g
while 1) constraining the actuator output to a fixed position
τ ζ̇ = Aζ + BC(ζ − ξ) + Bur .
(Ml ' ∞) and 2) letting the actuator to move freely. Note
that by fixing the actuator output, its open loop transfer func- Here, ur and F are the input reference and the measured
tion becomes second order such that motor and elastomer force respectively. x = [x1 , x2 ]> and z ∈ R are the internal
parameters can be identified independently. Combining the states and zero dynamics state of the plant Pi→F . The 2D
system identification tests with constrained and free moving vectors, ξ and ζ represent states corresponding to the filters
outputs, we identify the rest of the parameters as shown in QA and QB respectively. The matrices ψ, φ, g, S, G, A,
Table I and compute the bode plots shown in Fig. 5(a). B and C are unknown but bounded plant parameters. In
addition, the state and the plant parameters expressed with
B. Force Feedback Control symbol overlines represent the same vector and operators
∞
Many different methods have been proposed for control- for the nominal plant Pi→F . For example, S, G, ψ, φ, g
ling series elastic actuators using force feedback. [19] studied and z represent plant parameters and zero dynamics for the
 (a)
nominal plant. Based on the identified parameters of our

Magnitude (dB)
system shown in table I and the second order Butterworth 60 0

Phase (deg)
filter of Eq. (3) with 60Hz cutoff frequency for both QA and 40
-90
Locked Output
QB , the state-space parameters become Unlocked Output
-180
20
   
0 1 0 Frequency (Hz) Frequency (Hz)
√
 
A= , B= , C= 1 0 , (b) (c)
−1 − 2 1
1.3

Magnitude (dB)
ψ = −9.53e3, φ = 1e4[−1.49, − 0.01], 1.25

joint position (rad)

 
S = −0.0433, G = 1 0 , g = 1.31e4. 1.2

1.15
Eq. (4) and (5) are in the standard form for singular

Phase (deg)
perturbation analysis where τ represents the perturbation pa- 1.1 switch
desired step input
(i) 100 10 0 0
rameter [23]. The variables x, z, z are called slow dynamics 1.05
(ii) 30 4 0 0
(i) 70 8 3 1
while the variables ξ and ζ are called fast dynamics and the 1
2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6
following theorem holds: time (s) Frequency (Hz)

Theorem. The proposed DOB structure (Fig. 4(b)) is ro- Fig. 5. (a) shows experimental and estimated bode plots of input
bustly stable and converges to the performance of the nom- current versus output force based on two different actuator output set-ups
for system identification. (b) shows experimental step responses of three
inal plant under uncertain loads. different joint position feedback controllers. Selected gains and the switch
to measure feedback signals are shown in the table. (c) shows experimental
Proof. [22] proved that if the unknown variables φ, ψ,S, and estimated joint position control performance for each controllers.
G and g are bounded with g 6= 0, there exists a τ ∗ > 0 such
that, for all 0 < τ < τ ∗ , the DOB structure is robustly stable the close loop systems for each switch option can be derived
if 1) the zero dynamics of the actual actuator plant are stable, from inspecting the block diagram resulting in:
2) the boundary-layer subsystem, Eq. (4), is exponentially Cq PF →F d PF →q
(i)
stable. In our case, Pq→qd = (6)
1 + Dq Cq PF →F d PF →q
1) The actual plant, Pi→F , has stable zeros given the (ii) PF →q PF →F d Cq
identified actuator parameters of Table I, Pq→qd = , (7)
1 + Dq Cq PF →F d Nl Pi→θ /(Nm Pi→F )
2) The system matrix of Eq. (5) is
where PF →F d represents the transfer function of the inner
A − gg BC g
 
g BC ,
force control loop which has the expression,
−BC A + BC Pi→F (kτ Nm + Cτ )
PF →F d =
. (8)
and its characteristic polynomial is pa (s)pb (s) with 1 + Pi→F Dτ Cτ
  Based on the block diagram, we design three different
pa (s) = sI − A joint position controllers to compare their performance: 1)
 
pb (s) = sI − A + g−g

BC ,
 using motor and elastomer quadrature encoder feedback
g (i)
without inner force feedback control (Pq→qd , Cτ = 0),
where g = 2.4M 3.2e4Ml
. This system matrix is always 2) using linear potentiometer feedback without inner force
l +3.1 (ii)
Hurwitz, since pa (s) and pb (s) have negative roots feedback control (Pq→qd , Cτ = 0), and 3) using motor
for all possible Ml , which results in the exponential and elastomer quadrature encoder feedback with inner force
(i)
stability for Eq. (5). feedback control (Pq→qd , Cτ 6= 0).
To compare the closed loop systems, we empirically
choose gains such that the position control loops of con-
C. Position Feedback Control
trollers 1) and 2) from above behave as critically damped
In this subsection, we design different types of position systems. To increase feedback gains, we first increase Dq
controllers: 1) by measuring actuator position by either before the system gets unstable and then choose the highest
adding up motor quadrature encoders and elastomer quadra- stable Pq , such that the step response of the joint position
ture encoders, or directly using a linear potentiometer, and controller, Pqd →q , does not overshoot.
2) by including force feedback control within the position In the case of using position control with the inner force
control loop. Fig. 4(c) shows our joint position control control loop, i.e. controller 3) from previous, gain selection
structure using PD control, C• , and including time delays, become more complex due to the dependencies between Cτ
T• = e−T• s . The switch labeled (i) uses the option with mo- and Cq . We observed that the transfer function of a joint
tor and elastomer quadrature encoders and (ii) uses the option position control structure embedding a force control loop
with a linear potentiometer to measure actuator position. The could be represented as the multiplication of two second
force feedback control loop enclosed with a black dotted box order systems [24]. We then proposed a method to make the
where Cτ can be set to zero if we want to remove this loop combined system critically damped given a desired natural
from the joint position controller. The transfer functions of frequency. For our comparative analysis, we increase the
Fig. 6. Operational Space Control Test. (a) shows DRACO moving its left ankle following a Cartesian trajectory. (b) shows desired and measured joint
torques for tracking performance. (c) shows desired and measured Cartesian position tracking performance. Both torque and Cartesian position trackers are
tracking closely the reference trajectories.

force loop gains and decrease the joint position control gains are desired Cartesian trajectory accelerations, and Cartesian
according to stability constraints. In this way, we emphasize position and velocity errors. J is for the jacobian of the
the role of the inner force control loop to see its effect. left ankle and J is for the dynamically consistent pseudo
We now compare controllers 1), 2) and 3), as represented inverse, which is defined as J , A−1 J > (J A−1 J > )−1 .
with green, black and blue color, respectively, in Fig. 5. The robot then sends the computed joint torques command
In the figure, we choose gains to make the closed loop through the EtherCAT network using the embedded Axon
systems critically damped and match the natural frequency boards for joint control. The Axon boards implement each a
of controller 3) to controller 1). Controller 1) performs torque controller as described in Section III-B. The result is
better than 2) since the quadrature encoders give higher shown in Fig. 6 and demonstrates the performance of OSC on
quality signals than the linear potentiometer. We now analyze DRACO. Because of the use of a DOB, the joint controllers
controller 3) based on inner force feedback control. We display robustness despite load uncertainty.
notice that controller 3) only allows for smaller values of Pq
and Dq than controller 1) for stability reasons. In conclusion, B. Unsupported Dynamic Balancing
although the use of inner force feedback control reduces Here, we demonstrate the ability of DRACO to achieve
friction effects, it ends up reducing joint position gains which unsupported dynamic balancing by means of the WBC
decrease position accuracy. As a result, for our locomotion and TVR algorithms (see Fig. 8). Dynamic balancing is
tests we use controller 1) instead of controller 3). achieved via stabilizing leg contact changes (coordinated
by the state machine block) triggered by either pre-defined
IV. ROBOT C ONTROL
temporal specifications or foot contact sensors. The sequence
Building on our actuator control study above, we devise of contact phases is represented by a list of tuples specifying
and test two multi-joint controllers for DRACO. We will the phase name and its time duration, i.e. {(DS, 0.01s)
first evaluate DRACO’s performance using and instance of → (LFi , 0.03s) → (SWi , 0.33s) → (LNi , 0.01s)}1 . In
Operational Space Control [15]. After that we will evaluate addition, DRACO can detect sudden velocity changes on its
DRACO using our newest unsupported dynamic locomotion ankle movement as a trigger mechanism to detect contact.
controller, which consists of two parts, WBC and Time-To- We use ankle velocity trigger as sensors to terminate the
Reversal (TVR) planner [8]. SWi phase.
A. Operational Space Control The xk symbols represent a set of operational space tasks.
Arrows, xk → xl , express priorities in that the left set of
For this test, we first fix DRACO to a table as shown in
tasks xk has higher priority than the right set of tasks xl for
Fig. 6(a) and generate a Cartesian trajectory for the robot’s
arbitrary k and l indexes. In turn, WBC handles priorities by
left ankle to follow in the forward direction. We command a
solving a prioritized inverse kinematics problem. Below are
sinusoidal trajectory with amplitude of 0.3 m and frequency
the task and priority assignments to the phases that we use
of 1.4 Hz. The lateral and vertical Cartesian directions of the
for the robot’s dynamic balancing behavior:
ankle are controlled to stay at a fix point. Torque commands
for each robot joint is computed according to the OSC DS, LN, LF : {x1 → x3 }
control law, SW : {x1 → x2 → x3 → x4 },
τ = AJ (ẍ + Kp e + Kd ė − J˙ q̇) + b + g, (9)
where the task sets are defined as
where A, b, g, q̇ and τ are inertia, coriolis, gravity forces, 1 DS, LF , SW and LN mean Double Support, Lifting, Swing, and
i i i
joint velocities and joint torque commands written with Landing phases respectively. The subscript i represents the swing leg type,
respect to the robot’s generalized coordinates. ẍ, e and ė either the robot’s right leg or its left leg.
(a) (b)

(c) 0.2 (d) Joint Position (rad) Hip Rotation

Motor temperature ('C)
0.1 0.02 34
33
0 0
desired Hip Abduction
-0.1 0.05 sensed
stance 35
-0.2 planned -0.05 30
landing Hip Flexion
-0.3 -0.5 35
0.2 -0.6 30
Knee Flexion
0.1 1.2 60 130
1 40 w/o liquid cooled
0 30
(i th) torso desired 1.4 Ankle Flexion
-0.1 (i th) torso sensed
(i+1 th) torso desired 34
1.2 33
(i+1 th) torso sensed
-0.2 -0.05 0 0.05 0.1 0.15 0.20 23.2 23.6 24 24.4 24.8 5 10 15 20 25 30 35 40 45
y(m) time(s) time(s)

Fig. 7. The sequences (a) and (b) show DRACO dynamically self-balancing without any mechanical support. (c) shows the phase space plot of the robot’s
CoM in the lateral direction with respect to the robot’s frame for two consecutive balancing steps. (d) shows desired and measured joint positions as well
as estimated core winding motor temperatures during dynamic balancing.

• x1 ∈ R2 : Right and Left Hip Rotation Task Axon embedded controllers for real-time execution. For joint
• x2 ∈ R: Swing Foot Ankle Flexion Task position control we rely on the elastomer quadrature encoders
• x3 ∈ R3 : Torso Roll, Torso Pitch, Torso Height Task for feedback as explained before. However, we turn off the
• x4 ∈ R3 : Swing Foot Position Task inner force feedback controllers described in Section III-C to
increase the joint position accuracy. This is important to land
The Right and Left Hip Rotation Task and the Torso Roll,
the small feet near the desired foot locations with minimal
Pitch, and Height Task are set to [0◦ , 0◦ ] and [0◦ , 0◦ , 1m],
errors.
respectively, in order to make DRACO face forward and
maintain its torso upright. For the SW task assignment, The behavior resulting from integrating the new liquid-
we incorporate the Swing Foot Ankle Flexion Task with cooled viscoelastic in the DRACO biped robot with the
a desired value of 80◦ in order to detect sudden velocity WBC and TVR control algorithms is shown in Fig. 7.
changes when touching the ground. The swing Foot Position DRACO is able to achieve unsupported dynamic balancing
Task is driven by b-spline trajectory computation that steers without falling. The accompanying video demonstrates this
the swinging foot to a desired landing location given by the capability. The data is plotted in Fig. 7 for two consecutive
TVR planner. After all operational space tasks are specified, steps. In the phase space plots of that figure, we can see
the WBC controller shown in Fig. 8 provides the computation the velocity of the torso being effectively reversed based on
of sensor-based feedback control loops and motor commands the TVR planner which aims precisely at achieving such
to achieve the desired goals. As a result, the entire body outcome. Reversing velocity allows the robot to converge to
of the robot, the actual plant, will execute the commands the coordinate origin while dynamically stepping. Readers
to dynamically balance without support. For this particular are referred to Appendix in [25] for more details about
experiment we use both an IMU in combination with a the phase space plot of the linear inverted pendulum. In
motion capture system for CoM state estimation. addition, the use of liquid cooling safely regulates core
The computed motor commands are then sent out to the winding temperatures of the electric motors. In that same
 Locomotion Controller Actuator Control
[4] N. Radford et al., “Valkyrie: NASA’s First Bipedal Humanoid Robot,”
Journal of Field Robotics, vol. 32, no. 3, pp. 397–419, May 2015.
Kinematic WBC Joint Position [5] N. Paine, J. S. Mehling, J. Holley, N. A. Radford, G. Johnson, C.-
Feedback Control L. Fok, and L. Sentis, “Actuator control for the nasa-jsc valkyrie
Tasks
Dynamic WBC humanoid robot: A decoupled dynamics approach for torque control
State Machine of series elastic robots,” Journal of Field Robotics, vol. 32, no. 3, pp.
378–396, 2015.
[6] C. Hubicki et al., “Atrias: Design and validation of a tether-free
3d-capable spring-mass bipedal robot,” The International Journal of
Fig. 8. This figure describes a control architecture to achieve unsupported Robotics Research, vol. 35, no. 12, pp. 1497–1521, 2016.
dynamic balancing. [7] Z. Xie, P. Clary, J. Dao, P. Morais, J. Hurst, and M. van de Panne,
“Iterative reinforcement learning based design of dynamic locomotion
figure, we can see that when turning off the liquid cooling skills for cassie,” in arXiv:1903.09537, 2019.
system, the knee motors increases temperature beyond 130 ◦ [8] D. Kim, S. J. Jorgensen, J. Lee, J. Ahn, J. Luo, and L. Sentis, “Dy-
namic locomotion for passive-ankle biped robots and humanoids using
which can damage the motor windings. In contrast, when whole-body locomotion control,” arXiv preprint arXiv:1901.08100,
turning on liquid cooling, the motor temperature remains 2019.
below 60 ◦ all the time during balancing which is considered [9] A. Ramezani, J. W. Hurst, K. A. Hamed, and J. W. Grizzle, “Per-
formance analysis and feedback control of atrias, a three-dimensional
a very safe temperature. Overall, without the cooling system, bipedal robot,” Journal of Dynamic Systems, Measurement, and Con-
we could not achieve agile locomotion for our lightweight trol, vol. 136, no. 2, p. 021012, 2014.
system due to overheating. [10] K. Sreenath, H.-W. Park, I. Poulakakis, and J. W. Grizzle, “Embedding
active force control within the compliant hybrid zero dynamics to
V. C ONCLUDING REMARKS achieve stable, fast running on mabel,” The International Journal of
Robotics Research, vol. 32, no. 3, pp. 324–345, 2013.
Overall, our main contribution has been on the control, [11] D. Kim, J. Ahn, O. Campbell, N. Paine, and L. Sentis, “Investiga-
and evaluation of a new adult-size humanoid bipedal robot, tions of a robotic test bed with viscoelastic liquid cooled actuators,”
IEEE/ASME Transactions on Mechatronics, vol. 23, no. 6, pp. 2704–
dubbed DRACO with control considerations on the VLCA 2714, Dec 2018.
actuators. DRACO is able to achieve unsupported dynamic [12] N. Paine and L. Sentis, “Design and comparative analysis of a
balancing with only ten actuators, and despite the ankle retrofitted liquid cooling system for high-power actuators,” in Actu-
ators, vol. 4, no. 3. Multidisciplinary Digital Publishing Institute,
actuators being much weaker than human ankles. This per- 2015, pp. 182–202.
formance is possible due to a combination of mechanics that [13] L. Sentis and O. Khatib, “Synthesis of whole-body behaviors through
reduce distal mass, the use of high power dense actuators, hierarchical control of behavioral primitives,” International Journal of
Humanoid Robotics, vol. 2, no. 04, pp. 505–518, 2005.
high quality sensing, the integration of a robust planner, and [14] N. Paine, S. Oh, and L. Sentis, “Design and control considerations for
stiff controllers that control the robot’s body, foot, and joint high-performance series elastic actuators,” IEEE/ASME Transactions
positions with high accuracy. on Mechatronics, vol. 19, no. 3, pp. 1080–1091, June 2014.
[15] O. Khatib, “A unified approach for motion and force control of robot
In the future, we plan to remove the use of the motion manipulators: The operational space formulation,” IEEE Journal on
capture system for CoM state estimation and rely on IMU Robotics and Automation, vol. 3, no. 1, pp. 43–53, February 1987.
and vision. For localization we plan to explore the integration [16] M. Slovich, N. Paine, K. Kemper, A. Metger, A. Edinger, J. Weber,
and L. Sentis, “Building hume: A bipedal robot for human-centered
of stereo RGB cameras for dense range data and high hyper-agility,” in Dynamic Walking Meeting, vol. 4, 2012, p. 2.
frame-rates. Another part of this project will consists on the [17] T. Jung, J. Lim, H. Bae, K. K. Lee, H. Joe, and J. Oh, “Develop-
addition of an upper body with two robotic arms for loco- ment of the humanoid disaster response platform drc-hubo+,” IEEE
Transactions on Robotics, vol. 34, no. 1, pp. 1–17, Feb 2018.
manipulation behaviors. [18] Y. Park, S. Oh, and H. Zoe, “Dynamic analysis of reaction force
sensing series elastic actuator as unlumped two mass system,” in
ACKNOWLEDGMENT Industrial Electronics Society, IECON 2016-42nd Annual Conference
The authors would like to thank the members of the of the IEEE. IEEE, 2016, pp. 5784–5789.
[19] J. Pratt, B. Krupp, and C. Morse, “Series elastic actuators for high
Human Centered Robotics Laboratory at The University of fidelity force control,” Industrial Robot: An International Journal,
Texas at Austin and the company Apptronik for their great vol. 29, no. 3, pp. 234–241, 2002.
help and support. This work was supported by the Office [20] K. Kong, J. Bae, and M. Tomizuka, “A compact rotary series elastic
actuator for human assistive systems,” IEEE/ASME transactions on
of Naval Research, ONR Grant #N000141512507 and the mechatronics, vol. 17, no. 2, pp. 288–297, 2012.
National Science Foundation, NSF Grant #1724360. [21] H. Shim, G. Park, Y. Joo, J. Back, and N. H. Jo, “Yet another tutorial
of disturbance observer: robust stabilization and recovery of nominal
R EFERENCES performance,” Control Theory and Technology, vol. 14, no. 3, pp. 237–
249, 2016.
[1] Y. Ito, S. Nozawa, J. Urata, T. Nakaoka, K. Kobayashi, Y. Nakanishi, [22] H. Shim and Y. Joo, “State space analysis of disturbance observer
K. Okada, and M. Inaba, “Development and verification of life- and a robust stability condition,” in 2007 46th IEEE Conference on
size humanoid with high-output actuation system,” in 2014 IEEE Decision and Control, Dec 2007, pp. 2193–2198.
International Conference on Robotics and Automation (ICRA), May [23] H. K. Khalil, Nonlinear systems. Prentice hall Upper Saddle River,
2014, pp. 3433–3438. NJ, 2002, vol. 3.
[2] K. Kojima et al., “Development of life-sized high-power humanoid [24] Y. Zhao, N. Paine, S. J. Jorgensen, and L. Sentis, “Impedance
robot jaxon for real-world use,” in 2015 IEEE-RAS 15th International control and performance measure of series elastic actuators,” IEEE
Conference on Humanoid Robots (Humanoids), Nov 2015, pp. 838– Transactions on Industrial Electronics, vol. 65, no. 3, pp. 2817–2827,
843. 2018.
[3] C. Knabe, J. Seminatore, J. Webb, M. Hopkins, T. Furukawa, [25] J. Ahn, O. Campbell, D. Kim, and L. Sentis, “Fast kinody-
A. Leonessa, and B. Lattimer, “Design of a series elastic humanoid for namic bipedal locomotion planning with moving obstacles,” in 2018
the darpa robotics challenge,” in 2015 IEEE-RAS 15th International IEEE/RSJ International Conference on Intelligent Robots and Systems
Conference on Humanoid Robots (Humanoids), Nov 2015, pp. 738– (IROS), Oct 2018, pp. 177–184.
743.