ArticlePDF Available

Actuator Control for the NASA-JSC Valkyrie Humanoid Robot: A Decoupled Dynamics Approach for Torque Control of Series Elastic Robots

Authors:
  • Houston Mechatronics Inc

Abstract and Figures

This paper discusses the actuator-level control of Valkyrie, a new humanoid robot designed by NASA's Johnson Space Center in collaboration with several external partners. Several topics pertaining to Valkyrie's series elastic actuators are presented including control architecture, controller design, and implementation in hardware. A decentralized approach is taken in controlling Valkyrie's many series elastic degrees of freedom. By conceptually decoupling actuator dynamics from robot limb dynamics, the problem of controlling a highly complex system is simplified and the controller development process is streamlined compared to other approaches. This hierarchical control abstraction is realized by leveraging disturbance observers in the robot's joint-level torque controllers. A novel analysis technique is applied to understand the ability of a disturbance observer to attenuate the effects of unmodeled dynamics. The performance of this control approach is demonstrated in two ways. First, torque tracking performance of a single Valkyrie actuator is characterized in terms of controllable torque resolution, tracking error, bandwidth, and power consumption. Second, tests are performed on Valkyrie's arm, a serial chain of actuators, to demonstrate the robot's ability to accurately track torques with the presented decentralized control approach.
Content may be subject to copyright.
Actuator Control for the
NASA-JSC Valkyrie Humanoid Robot:
A Decoupled Dynamics Approach for Torque Control
of Series Elastic Robots
Nicholas Paine
Electrical and Computer Engineering Department
The University of Texas
Austin, TX 78712 USA
npaine@utexas.edu
Joshua S. Mehling
NASA Johnson Space Center
Houston, Texas, TX 77058 USA
joshua.s.mehling@nasa.gov
James Holley
NASA Johnson Space Center
Houston, Texas, TX 77058 USA
james.j.holley@nasa.gov
Nicolaus A. Radford
NASA Johnson Space Center
Houston, Texas, TX 77058 USA
nicolaus.a.radford@nasa.gov
Gwendolyn Johnson
Mechanical Engineering Department
The University of Texas
Austin, TX 78712 USA
gwendolynbrook@gmail.com
Chien-Liang Fok
Mechanical Engineering Department
The University of Texas
Austin, TX 78712 USA
liangfok@utexas.edu
Luis Sentis
Mechanical Engineering Department
The University of Texas
Austin, TX 78712 USA
lsentis@austin.utexas.edu
Abstract
This paper discusses the actuator-level control of Valkyrie, a new humanoid robot designed
by NASA’s Johnson Space Center in collaboration with several external partners. We focus
on several topics pertaining to Valkyrie’s series elastic actuators including control architec-
ture, controller design, and implementation in hardware. A decentralized approach is taken
in controlling Valkyrie’s many series elastic degrees of freedom. By conceptually decou-
pling actuator dynamics from robot limb dynamics, we simplify the problem of controlling
a highly complex system and streamline the controller development process compared to
other approaches. This hierarchical control abstraction is realized by leveraging disturbance
observers in the robot’s joint-level torque controllers. We apply a novel analysis technique
to understand the ability of a disturbance observer to attenuate the effects of unmodeled
dynamics. The performance of our control approach is demonstrated in two ways. First,
we characterize torque tracking performance of a single Valkyrie actuator in terms of con-
trollable torque resolution, tracking error, bandwidth, and power consumption. Second, we
perform tests on Valkyrie’s arm, a serial chain of actuators, and demonstrate its ability to
accurately track torques with our decentralized control approach.
1 Introduction
As an entry into the 2013 DRC Trials, NASA-JSC formed a team with several external partners and led the
development of Valkyrie (see Figure 1). Valkyrie was designed to perform tasks required both for responding
to disasters, such as the Fukushima Daiichi nuclear disaster in Japan (Nagatani et al., 2013), and for
advancing human spaceflight by one day assisting human explorers in extraterrestrial settings such as Mars.
While many areas of expertise were required to produce a new humanoid robot in less than 12 months, this
paper focuses on one area, namely the methods used to control Valkyrie’s series elastic actuators (SEAs). In
the following discussion, we present our overall control approach, an outline of Valkyrie’s multi-joint control
architecture, the actuator controller design and its implementation in hardware.
Figure 1: a) NASA-JSC’s Valkyrie Humanoid Robot. Valkyrie has 44 actuated degrees of freedom and
a suite of sensors including stereo vision, laser range fingers, sonar depth perception, and tactile feedback.
b) Safe Human Interaction. Series elastic actuators and compliant torque control enable safe interaction
with humans and with unexpected environmental collisions. c) DRC Competition. Valkyrie was one of
16 entries in the 2013 DARPA Robotics Challenge Trials. In this image Valkyrie is performing a valve turn
task.
NASA-JSC has a long history of developing robotic actuators and their controllers, beginning with the
Robonaut project and extending to a number of other systems (Ambrose et al., 2000; Bluethmann et al.,
2003; Bridgwater et al., 2012; Reiland et al., 2013). Robonaut 2, for example, is the first humanoid robot in
space and is currently aboard the International Space Station (Diftler et al., 2011).
For Valkyrie, the actuation control requirements were driven by the needs of the holistic robot control
approach and by the mechanical design of its actuators. Because the DRC tasks required movement and ma-
nipulation in uncertain environments, compliant control approaches were favored over high-gain rigid control
approaches. Consequently, series elastic actuators were chosen as a means of achieving compliant control
and protecting both Valkyrie and external objects from unexpected collisions. The naturally low output
impedance provided by integrated passive compliance makes series elastic actuators particularly effective at
stable interactions with the environment. This stability as well as improvements in shock tolerance, energy
storage capability, power output, and force sensing are among the many benefits of the SEA architecture
widely cited in the literature (Pratt and Williamson, 1995; Pratt et al., 2002; Paluska and Herr, 2006).
1.1 Multi-Joint Series Elastic Control Architecture Background
How to best address compliance within the context of a full body coordinated control architecture is an
open question. Early investigations into elastic joint robots treated each motor as an ideal torque source and
controlled both actuator and limb dynamics with a centralized algorithm (Spong, 1987). Signal latencies and
the complexities of modeling the higher order dynamics inherent in a multi-Degree-Of-Freedom (DOF), series
elastic robot are difficult to overcome. Nevertheless, approaches based on this method are widely used today
and prove quite effective (Albu-Schaffer et al., 2007; Ott et al., 2008; Li et al., 2012). It has been suggested,
however, that robots, particularly those designed to mimic the broad versatility of humans and animals,
could benefit from a more embedded, collocated control of their actuators (Pratt et al., 2004). Physically,
or at least conceptually, decentralizing joint-level control allows for the assumption of a rigid body model
at the high level, actuated by ideal joint torque sources at the low level. This approach is taken in (Kim
et al., 2012; Sentis et al., 2013; Hutter et al., 2013). The success of such decentralized control architectures
directly depends on the performance capabilities of joint-level torque controllers, which must closely match
an ideal torque source model.
While effective series elastic robots can be designed using either a centralized or decentralized control archi-
tecture, one less obvious advantage perhaps tips the scale in favor of the decentralized approach. Hierarchical
controller tuning and validation, that is, the ability to test each joint in a multi-DOF system individually,
before integration into the whole robot, could speed development time and ease the process of testing higher
level functionality once the full robot is assembled. To achieve this result, torque control of each actuator
must truly be decoupled from the rest of the system.
1.2 Joint-Level SEA Control Background
Many different joint-level control architectures exist for torque control of series elastic actuators. Some
measure spring force and control motor force using some variant of PID control structures (P, PD, etc.)
(Pratt and Williamson, 1995; Sensinger and Weir, 2006; Hurst et al., 2010; Ragonesi et al., 2011; Garcia
et al., 2011). If friction and backlash are large, a single-loop PID force controller may become unstable before
the desired force tracking is achieved. To remedy this issue, an inner position or velocity control loop may
be used with an outer force control loop as proposed by (Robinson, 2000) and (Pratt et al., 2004). This idea
has been adopted and carried on by many others, translating force control into a position or velocity tracking
problem (Wyeth, 2006; Vallery et al., 2007; Lagoda et al., 2010; Thorson and Caldwell, 2011; Taylor, 2011).
Other work shows how steady state tracking and disturbance rejection can be significantly improved through
the use of disturbance observers (Kong et al., 2009; Kong et al., 2012; Paine et al., 2014). Building on these
approaches, the Valkyrie SEA force control architecture presented in this paper is closely related to previous
work on high-performance SEAs (Paine et al., 2014). It differs, however, in that PD feedback is used to
shape the dynamic response of the actuator, eliminating the need for an inverse-dynamics feedforward term
and improving the controller’s phase margin.
1.3 Summary of Our Approach and Paper Outline
This paper offers new results in two areas. First, we document the actuator control approach used in the field
by the Valkyrie DRC robot. We provide details on actuator control performance using quantitative metrics
and experimental results. We also demonstrate how these actuator controllers are used in tasks performed by
Valkyrie prior to and during the DRC Trials. Second, we present generic methods for controlling the torque
output of robots driven by series elastic actuators. These methods include both the conceptual decoupling
of multi-joint versus single-joint dynamics and the joint-level series elastic torque controller that makes this
conceptual decoupling feasible. In the process of characterizing this torque controller, we introduce a new
method for analyzing the disturbance rejection capability of disturbance observers, which we apply to the
case of a series elastic actuator with variable load inertia.
We begin by outlining Valkyrie’s multi-joint control approach and the requirements it places on the single-
joint torque controllers. We then take a detailed look at the torque control implementation applied to
Valkyrie’s various series elastic actuators in Section 3, providing both helpful tuning techniques and quan-
titative torque control metrics based on single-joint experimental results. Section 4 then uses a SEA plant
with finite load inertia to gain insight into the disturbance rejection capabilities of disturbance observer
control structures. Given this knowledge, in Section 5 we apply our proposed distributed torque controllers
to Valkyrie’s arm, a multi-link chain of SEAs, demonstrating the achieved joint torque tracking performance
in a real-world scenario.
2 The Role of Actuator-Level Control in Valkyrie
The selection and design of actuator-level control in a series elastic robot depends on the holistic robot
control strategy. In this section, we give a brief overview of our holistic decentralized control approach and
provide examples of the broad utility provided by joint-level torque controllers.
2.1 A Decentralized Control Approach
Valkyrie uses a decentralized control approach as shown in Figure 2b. This approach was chosen over a
centralized approach (Figure 2a) for several reasons. First, using hierarchical control abstractions reduced
the complexity of the robot model from a flexible-joint to a rigid-joint model. This simplification allowed our
team to leverage existing whole-body control techniques which assumes a rigid-joint model actuated by ideal
torque sources (Sentis et al., 2013). Second, a powerful, distributed embedded control element, using NASA-
JSCs proprietary ‘Robonet’ high speed serial bus interface and ‘Turbodriver’ motor controllers, is co-located
at each joint on Valkyrie. Because of this embedded processing capability, our single-joint controllers are
able to utilize highly effective dynamic-model-based control schemes. Third, reducing the coupling between
central and peripheral systems decreases overall communication latency due to the reduced number of signals
required by the multi-joint controller, meaning they may be updated at a faster rate. Fourth, a decentralized
control approach naturally lends itself to an incremental testing methodology, which aides development and
debugging.
In the decentralized control approach used on Valkyrie, actuator-level dynamics are abstracted away from
the central multi-joint controller. The multi-joint controller models the robot as rigid bodies actuated by
joint torques. The outputs of the multi-joint controller are desired joint torques which are then passed to
a subordinate set of single-joint controllers. The single-joint controllers model the actuator dynamics and
enforce the received torque commands. This approach differs from centralized control approaches, which do
not abstract actuator dynamics from the multi-joint model (Spong, 1987; Albu-Schaffer et al., 2007; Ott
et al., 2008).
Figure 2: Two different approaches for controlling SEA-driven robots. a) Centralized Approach. In this
approach, a multi-joint controller models both rigid-body dynamics and actuator dynamics. The multi-joint
controller takes joint states as an input and outputs desired motor currents or positions. b) Decentralized
Approach (used on Valkyrie). In this approach, actuator-level dynamics are abstracted from the multi-
joint controller. The multi-joint controller models the robot as rigid bodies actuated by joint torques. The
multi-joint controller generates desired joint torques which are then passed to a subordinate set of single-
joint controllers. The single-joint controllers model the actuator dynamics and enforce the received torque
commands.
τ
+
-
e
+
+
Torque
Control
q
B
K
s
d
q
+
-
e
d
q
..
Robot
di
τg+
Single-joint controller
Multi-joint
controller
q
τk
τTorque
Control
q
Robot
d
i
Single-joint controller
Multi-joint
controller q
τk
b) Joint-level impedance control modea) Joint-level torque control mode
s
q
.
Figure 3: Two actuator control modes supported by Valkyrie. a) Torque Control Mode. This joint-level
control mode matches the decentralized torque abstraction model shown in Figure 2b where the multi-joint
controller sends desired torques to the single-joint controller. b) Impedance Control Mode. Desired joint
position (qd), velocity ( ˙qd), stiffness (K), damping (B) and gravity compensation torque (τg) are sent to the
single-joint controller. An impedance control law is enforced at the single-joint level to minimize latency in
the control loop, resulting in higher possible stiffness and damping gains. Note: Both control modes rely on
an accurate torque feedback controller.
2.2 Primary Actuator Control Modes
During the early stages of the project, the control system design of Valkyrie closely matched the decentralized
model shown in Figure 2b. The desired joint torque signal created a natural interface between multi-joint and
single-joint control domains. Figure 2b represents the robot control system from the multi-joint controller’s
perspective. If we instead consider a single-joint controller’s perspective, the same control system could be
represented by Figure 3a. The single-joint controller receives a desired joint torque (τd) and uses a torque
feedback controller (discussed in Section 3) to enforce this command.
Later in the project, a higher importance was placed on rejecting disturbances caused by multi-joint model
uncertainty. As a result, high joint impedance became a new control target. When high impedance controllers
were implemented at the multi-joint level, latencies from single- to multi-joint communication would limit the
degree to which impedance could be increased before incurring control loop instability. As a solution to this
issue and following work by (Pratt et al., 2004), we created another control paradigm at the single-joint level
that we dubbed “impedance control mode”, which allowed the position and velocity feedback to occur locally
on the single-joint controller (see Figure 3b). This control paradigm increases achievable joint impedance
due to a significant reduction of control loop latency compared to those incurred in “torque control mode”.
In both control modes, a control loop that provides accurate torque control is required. In the next section, we
present the design of the torque feedback controller and demonstrate its performance on Valkyrie’s actuation
hardware.
3 Implementation of Torque Feedback Control Using Series
Elastic Actuators
As discussed in Section 2, torque control plays an important role in the control of Valkyrie. In this section,
we introduce Valkyrie’s series elastic actuators and give details on how they are controlled to accurately
track torques.
Valkyrie’s torso, legs, and arms house a combination of rotary and linear SEAs (Figure 4) driven by brushless
DC motors1. The rotary SEAs use harmonic drives, while custom designed torsion springs2act as the
compliant element. The spring’s deflection is measured to sense joint torques. The linear SEAs use roller
screw drivetrains and commercial-off-the-shelf die springs as the compliant element. The linear actuators
have redundant force feedback in the form of sensed deflection of the die springs and load cells in the actuator
output linkage. All of these actuators are driven by NASA’s ‘Turbodriver’ motor controllers which perform
the joint-level control discussed below.
Figure 5 shows the torque controller implemented in Valkyrie. The control plant is an SEA with a locked
output, as shown in Figure 5b. The inner PD compensator is tuned to produce the desired frequency response
based on this locked-output assumption. A disturbance observer (DOB) is then used to reject deviations
from this nominal locked-output model and maintain torque tracking accuracy. Further discussions on DOB
disturbance rejection for non-locked-output scenarios will be discussed in Section 4.
From Figure 5b the relation between torque applied to the spring (τm) and spring deflection (θ) is a second
order dynamic system with jmrepresenting effective motor inertia felt by the spring, bmrepresenting effective
motor-side damping felt by the spring, and krepresenting spring stiffness:
θ(s)
τm(s)=1
jms2+bms+k.(1)
Defining the following variables to be: τk: spring torque, i: motor current, N: motor speed reduction,
kτ: motor torque constant, η: drivetrain efficiency, we can apply the following relations: τk=(Hooke’s
law) and τm=where β=Nkτη. The control plant Pfrom motor current to spring torque is then found
to be
1During the DRC Trials 2013, the compliant element of the leg actuators was removed and replaced with rigid material.
This change was made several weeks before the DRC in an attempt to improve Valkyrie’s locomotion performance.
2Spring stiffness for each joint was chosen based on a fixed desired spring deflection at peak joint torque.
Figure 4: Valkyrie’s series elastic actuators. For the rotary actuators, spring deflection corresponding to
joint torque is sensed using a Renishaw optical sensor. For the linear actuators, load cells are included in
addition to the spring deflection sensor to provide redundant force feedback. The loadcells are placed closer
to the joint output giving them better dynamic sensing performance but suffer from a higher noise floor than
the spring deflection sensor due to their analog signal properties.
Figure 5: a) Schematic representation of a series elastic actuator. b) Plant model (P) used for control
design. Note that a locked-output assumption is used. c) Diagram of the torque feedback controller used
on Valkyrie’s series elastic actuators. The PD compensator is used to shape the dynamics of the torque
response while the DOB is used to improve disturbance rejection, especially at low frequencies. This control
structure is based on the controller presented in (Paine et al., 2014).
P(s) = τk(s)
i(s)=(s)
τm(s)
τm(s)
i(s)=Nkτηk
jms2+bms+k=βk
jms2+bms+k.(2)
Assuming kis calibrated beforehand, all of the parameters in (2) can be found using system identification
techniques with the actuator output locked.
As depicted in Figure 5c, the closed-loop torque-tracking transfer function3(Pc) is composed of a feedforward
term (N1η1kτ
1) and a feedback term (P D). The feedforward term is used to scale desired actuator
torques into approximate actuator output torques to minimize control effort from the feedback term. The
feedback term is represented by the following transfer function:
P D(s) = i(s)
e(s)=kds+kp.(3)
For simplicity, we do not model the effects of a low-pass filter that is applied to the derivative term. In
practice, the cutoff frequency for the kdsterm in (3) is chosen to be sufficiently higher than the actuator’s
maximum closed-loop system bandwidth (200 Hz versus 70 Hz).
We can then solve for Pcas follows:
Pc(s) = τk(s)
τr(s)=P·B1+P·P D
1 + P·P D =(kd)s+k(1 + βkp)
jms2+ (bm+kd)s+k(1 + βkp).(4)
3.1 Simplification of feedback gain selection
When faced with highly parameterized feedback controllers, it is often unclear how each parameter should
be chosen. Clearly, a trade-off between poor performance (low gains) and poor stability margin (high gains)
must be found. In practice, gain tuning is often performed manually to quickly locate parameters that
balance these two trade-offs. From (3) we see that two parameters must be found (kpand kd). Instead of
choosing kpand kddirectly, it can be more insightful to consider an equivalent pair of feedback parameters:
bandwidth and damping ratio. This is because a desired “shape” of the frequency response can be chosen by
selecting a desired damping ratio, and then bandwidth may be chosen independently to satisfy the trade-off
between performance and stability. Here, we derive the equations mapping damping ratio to kpand kd.
Notice that the characteristic polynomial of (4) is a second order system that can be represented in terms
of an effective mass ( ˆ
M), spring, ( ˆ
K) and damper ( ˆ
B):
ˆ
M=jm(5)
ˆ
B= (bm+kd) (6)
ˆ
K=k(1 + βkp).(7)
The damping ratio (ζ) for such a second order system is
3Note that in this paper we label the open-loop plant as Pwhile closed-loop transfer functions use the format Px,xbeing
a unique subscript.
ζ=ˆ
B
2pˆ
Mˆ
K
(8)
Combining (5), (6), (7) and (8) forms a relation between desired closed-loop damping ratio (ζd) and kp,kd:
ζd=(bm+kd)
2pjmk(1 + βkp).(9)
Solving (9) for kdyields
kd=2ζdpjmk(1 + βkp)bm
.(10)
Using (10), a desired damping ratio can be chosen and kdcan then be automatically calculated for a given kp.
In tuning Valkyrie’s torque controller, kpwas used to represent bandwidth, therefore simplifying the tuning
of the PD compensator to a single degree of freedom. Using this one parameter, the trade-offs between
performance and stability could easily be changed on-the-fly while ensuring a dynamic response with the
desired damping ratio. On Valkyrie, we chose a desired damping ratio of 0.9 to produce a flat torque response
with little overshoot (ζd= 1.0 is critically damped, ζd= 0.7 is underdamped with minimum settling time).
3.2 Disturbance Observer
The use of a disturbance observer (DOB) applied to an inner PID/PD control loop has been shown to
significantly enhance the torque tracking capability of SEAs (Kong et al., 2012; Paine et al., 2014). DOBs
have several useful properties for our specific application. First, they preserve and enforce a dynamic plant
model through the use of a nominal model. This means that the shape of the closed-loop frequency response
of Pcwill not be altered by adding a DOB. The DOB will try to maintain this characteristic response in
the presence of either 1) external disturbances or 2) plant model variations. The latter characteristic is of
central importance to our approach and is discussed in detail in Section 4. Secondly, DOBs excel at removing
steady state error, and therefore effectively serve as an integral feedback term. This characteristic is useful
in minimizing controlled torque resolution as discussed in Section 3.3.
A DOB applied to Pcis shown in Figure 5c. Pn
1represents the inverse nominal closed-loop model (using
the locked-output constraint). Qis a low-pass filter which is used both to make Pn
1proper and to tune
the frequency (fq) up to which disturbances are rejected. In our implementation, the Qfilter takes the form
of a second order butterworth filter
Q(s) = 1
(s/ωq)2+ 1.4142(s/ωq)+1 (11)
where ωq= 2πfq.
In the tuning of Valkyrie’s disturbance observers, fqvalues in the range of 20 Hz to 70 Hz were found to
adequately reject unmodeled disturbances while maintaining high control loop stability.
Figure 6: Torque tracking and disturbance rejection of Valkyrie’s torque control implementation. A 10 Nm
sinusoidal reference torque is tracked with intentional disturbances (right half) and without intentional
disturbances (left half). Tracking error peaks around 1 Nm.
3.3 Quantifying Torque Control Performance
In the development of Valkyrie, it was essential to establish useful performance metrics so that the effects of
controller modifications or topological changes could be quantified. In this section we discuss the performance
metrics used to measure torque control performance.
Torque resolution determines an actuator’s minimum controllable torque magnitude. An important distinc-
tion must be made between sensed resolution and controlled resolution. Sensed torque resolution is the
minimum torque magnitude an actuator is able to observe and is purely determined by the type of sensor
used. Controlled torque resolution is the minimum torque magnitude that is both observed and acted upon.
Controlled torque resolution depends on sensed torque resolution as well as an actuator’s mechanical prop-
erties (friction, for example) and the properties of the torque controller. A controller with small steady state
error is required to minimize controlled torque resolution.
The controlled torque resolution of Valkyrie was measured by placing an actuator in torque control mode
and placing a series of loads on the actuator output. The torque from the minimum load that caused motion
was determined to be the actuator’s controlled torque resolution. For the elbow actuator of Valkyrie, the
controlled torque resolution was measured to be 0.002 Nm.
10−1 100101102
−70
−60
−50
−40
−30
−20
−10
0
10
20
Magnitude (dB)
Frequency Response from Desired Torque to Torque Error
Frequency (Hz)
Open loop
Closed loop (PD)
Closed loop (PD+DOB)
a)
PD
b)
Open loop PD+DOB (closed-loop)
OL
PD+DOB
Reduction of low frequency error
Figure 7: a) Frequency responses of a Valkyrie elbow actuator with a fixed output. The dashed line repre-
sents the open-loop plant response while the solid blue line represents torque tracking with the closed-loop
controller shown in Figure 5c. The closed-loop response significantly increases torque tracking bandwidth
and removes the resonant peak seen in the open-loop actuator response. b) Torque tracking error versus
signal frequency. Adding PD feedback reduces tracking error at frequencies within the actuator bandwidth.
The full controller (PD+DOB) further reduces error in low frequencies compared to the PD controller.
As a second measure of torque tracking performance, we performed a test where the actuator was commanded
to track a sinusoidal reference torque. We then created disturbances by manually applying loads to the
actuator output. By measuring the maximum torque tracking error, we obtained a rough estimate of nominal
torque tracking performance. Figure 6 shows these results. As can be seen, torque tracking error remains
relatively constant despite the added disturbances. Maximum torque tracking error both with and without
disturbances was found to be approximately 1.0 Nm.
Torque bandwidth is an important metric that establishes the maximum signal frequency an actuator is able
to accurately track. Figure 7a shows the torque tracking bandwidth of Valkyrie’s elbow actuator using a
fixed-output constraint (matching Figure 5b). Using PD feedback, we extend the bandwidth4of the SEA
by a factor of 5.3 compared to the passive bandwidth of the SEA (70 Hz compared to 13 Hz). The torque
error plot (Figure 7b) illustrates the effect of two different controllers on the torque tracking accuracy. The
“PD feedback” line demonstrates the maximum performance a proportional controller can achieve before
becoming unstable. The “PD+DOB” line, representing the full torque controller implemented in Valkyrie
(that of Figure 5c), clearly illustrates torque tracking improvements in the low frequency range. This
improved low-frequency torque tracking benefits the controlled torque resolution metric.
As a final metric, we consider the power consumption of a Valkyrie SEA in torque control mode. Based
on the previous discussion of feedback gain selection in Section 3.1, at a first glance, the primary factors in
determining feedback gains appear to be performance and stability. A study of actuator energetics quickly
demonstrates that a third factor, power consumption, is of critical importance as well. Efficiency is especially
important for Valkyrie due to its reliance on battery power.
Figure 8 shows the power consumption of Valkyrie’s elbow actuator for three different scenarios. In each
scenario, a chirp signal is generated as a torque reference. The three scenarios are differentiated by the
4Bandwidth (BW) of a second order system is defined to be the point in the magnitude response where the value reaches
-3 dB. This value can be calculated given the system’s natural frequency (ωn) and damping ratio (ζ):
BW =ωn·q(1 2ζ2) + p4ζ44ζ2+ 2.
Figure 8: Power consumption of a Valkyrie knee actuator. A desired torque chirp signal was tracked using
a PD controller (Pcin Figure 5c) for different values of kp, thus altering the closed-loop system bandwidth
(BW). Power was measured using knowledge of motor current and motor velocity. As is shown, while the
actuator is physically capable of tracking torque signals with frequency content above the bandwidth of
the passive system (11 Hz), doing so requires large power consumption and thus increases motor heating.
Therefore, a balance must be established between torque tracking performance and power consumption to
avoid overheating the actuator.
Table 1: Valkyrie Torque Control Performance Metrics
Metric Value Units
Controllable torque resolution 0.002 Nm
Nominal torque tracking error 1.0 Nm
Maximum torque bandwidth 70 Hz
selection of kpin the torque feedback loop. As is shown in Figure 8, torque tracking bandwidth plays a
significant role in determining the efficiency of a torque controlled SEA. Increasing torque tracking bandwidth
from 30 Hz to 82 Hz results in an average increase of power consumption at the actuator’s passive bandwidth
frequency by a factor of 3.6 (97W versus 27W).
These power consumption considerations must be balanced with the needs of the multi-joint controller
upstream of the torque controlled SEA. If this centralized controller requires high bandwidth torque tracking,
the passive actuator cutoff frequency must be large enough to reduce the discrepancy between passive and
active torque bandwidth, as this discrepancy determines peak power consumption (B`elanger, 1995). Because
the spring of an SEA heavily influences the passive actuator frequency, its selection should consider these
energy-related issues. A large amount of work remains to better understand these relations, but the trends
observed here are useful in directing future study.
Table 1 summarizes the torque control performance metrics of Valkyrie’s elbow actuator discussed in this
section. Valkyrie’s other series elastic joints share a similar performance profile.
4 Decoupling Dynamics of Multi-Joint SEA-Driven Robots
The goal of our torque control approach is to make each SEA appear to the multi-joint controller as an
ideal torque source, or at least a low-pass filtered torque source. That is, we want to avoid modeling internal
actuator dynamics at the multi-joint level, and instead only model the effects of the rigid body system. Prior
work in this area has demonstrated that effective decoupling of the fast actuator-level dynamics from the
slower multi-body dynamics is indeed possible (Ott et al., 2003). However, in this work torque errors remain
large (around 50Nm) and thus do not adequately abstract an actuator as an ideal torque source.
One option is to employ a decentralized joint-level control approach, which has been shown to work well
for rigid-joint position controlled robots (Nakao et al., 1987; Godler et al., 1999). In these approaches
subordinate joint-level controllers are co-located at each joint and use feedback to compensate for multi-
body dynamics. Interestingly, the authors found that the largest disturbances to affect their joint controllers
were caused by variation in the apparent load inertia due to changing robot pose. (Nakao et al., 1987) found
that a Disturbance Observer (DOB) could fully reject the undesirable behavior resulting from these model
variations that were unknown to their low level controllers. In this section, we study the effectiveness of a
similar approach applied to the control of Valkyrie’s actuators. In contrast to the rigid actuator position
controllers of (Nakao et al., 1987; Godler et al., 1999), our work focuses instead on the torque-tracking of
a SEA in a multi-joint environment. As a result, the control plant models, their variation due to changing
load inertia, and the DOB’s ability to cope with this variation all differ from these previous studies.
Outside of the decentralized control community, several methods exist for controlling the torque output of
a SEA with no knowledge of the load inertia (Vallery et al., 2007; Kong et al., 2009; Kong et al., 2012;
A. Schepelmann and Geyer, 2012). However, their disregard for the magnitude of the load inertia also
limits their use; no guidance or condition is provided for the case where load inertia becomes too small and
inevitably deteriorates tracking performance or causes instability, as it is known to do (Pratt and Williamson,
1995). In fact, few studies have considered the effects of load inertia on the torque-control performance of
SEAs (Wyeth, 2006) and instead usually qualify that their controllers assume “sufficiently large” load inertia.
How large is “sufficiently large?” The methods presented in this section can be used to identify this minimum
load inertia, or more generally, the maximum tolerable deviation from the control plant’s nominal dynamics,
given the DOB-based controller presented previously in Section 3.
We begin the study of a joint-level DOB’s ability to reject the effects of multi-body dynamics by modeling
the effects of variable actuator load inertia on the controller presented in Section 3. As previously discussed,
the controller in Section 3 is tuned for the case where the load inertia is locked and is not altered thereafter.
That is, no adaptive control techniques or gain scheduling based on robot pose is used. Instead, a DOB is
used to attenuate disturbances due to control plant variation and maintain the desired closed-loop dynamics.
This feature greatly benefits the tuning procedure for the many actuators used in Valkyrie. Each actuator
can be separately tuned on a bench with minimal parameter tuning required once the actuator is assembled
into a multi-joint system.
4.1 Augmented Control Plant Model
To carry out our analysis, we must augment the locked-output plant model (2) by including the effects of
load inertia. The modeling approach we take follows work presented in (Kong et al., 2009) and is extended
to map motor torques to spring torques.
A multi-input model for an SEA is shown in Figure 9, with τmrepresenting motor torque, θmmotor angle,
jmmotor inertia, bmmotor damping, kspring stiffness, τLexternal torque applied at the load, θLload angle,
jLload inertia, and bLload damping. The state equations for the system are
Figure 9: An example of a variable load inertia scenario using the abduction/adduction shoulder joint of
Valkyrie. Approximate actuator-level models are shown for each case, a large load inertia corresponding to
the out-stretched case and a smaller load inertia corresponding to the bent-elbow case.
jm¨
θm+bm˙
θm+k(θmθL) = τm(12)
jL¨
θL+bL˙
θL+k(θLθm) = τL.(13)
Assuming no external torque input, combining (12) and (13), and representing spring deflection as
θd=θmθL, yields a relation (α) between motor angle and spring deflection
α(s) = θd(s)
θm(s)=jLs2+bLs
jLs2+bLs+k.(14)
Using Hooke’s law and (14), spring torque (τk) is then represented by
τk(s) = d(s) = (s)θm(s).(15)
Combining (12) and (15) yields a transfer function from motor torque to motor angle
θm(s)
τm(s)=1
jms2+bms+α(s)k(16)
which, combined with (14), provides the transfer function from motor current to spring torque for any load
inertia
P(s) = τk(s)
i(s)=τm(s)
i(s)
θm(s)
τm(s)
θd(s)
θm(s)k=βα(s)k
jms2+bms+α(s)k.(17)
Figure 10: a) Series of SEA control plants (17) for varying load inertia. Reducing load inertia is
shown to increase resonant frequency and alter effective level of damping in the response. b) Closed-loop
(Pcin Figure 5c) torque response for a series of SEA plants with decreasing jL.For larger values
of jL, the bandwidth of the system remains fairly constant while only low frequency behavior changes.
Equation (17) represents a version of the fixed-output control plant model (2) augmented with the effects of
finite load inertia.
4.2 Sensitivity Analysis to Variations of Load Inertia
Plotting (17) for a wide range of load inertias provides a visualization and intuition towards understanding
how an SEA will respond to different load inertias (see Figure 10a). When jLis infinite (bold black line), the
system exhibits the familiar second order underdamped response, assuming underdamped system parame-
ters. As jLdecreases, low frequency behavior changes noticeably and the resonant peak moves to higher
frequencies. In the extreme case, when jL= 0, the load inertia no longer carries energy and the system
becomes the first order mass-damper system defined by jmand bm.
Similarly, we may simulate the dynamics of the closed-loop response (Pcfrom Figure 5c) for varying plant
load inertias (see Figure 10b). In this case, we see that despite significant variation in load inertia, the high
frequency dynamics remain relatively constant. The main variation in the closed-loop response occurs at
low frequencies, except for very small values of jL.
4.3 DOB Disturbance Rejection
Our objective is to leverage the disturbance rejection capability of a DOB to create a controller that is
insensitive to variation in load inertia. The result shown in Figure 10b demonstrates the qualitative behavior
of finite load inertia on the inner PD compensator, Pc. In this section we seek to understand a DOB’s
disturbance rejection capability and compare this capability with the disturbances shown in Figure 10b. If
the DOB can successfully remove the disturbances of Figure 10b, we will have achieved our goals of creating
Figure 11: Disturbance observer Region Of Convergence (DROC) for a DOB with a given Qfilter. A control
plant Pcwill converge to an error less than or equal to δif Pcis contained within the shaded area between
Pnand Pb.
an inertia-independent torque controller and thus decentralize torque control of Valkyrie’s SEAs.
The structure of a DOB can be seen referring back to Figure 5. Given a nominal model (Pn) of a control
plant (Pc), the measured plant output (τk) is passed through the nominal model inverse to produce an
estimate (τe) of the control plant input (τr). A lumped disturbance value (d) is calculated by subtracting τr
from τeand is compensated for by subtracting dfrom the desired value (τd). A low pass filter (Q) is used
both for tuning purposes and to make the inverse nominal plant model proper.
The transfer function for the DOB shown in Figure 5 is
PDOB =τk
τd
=Pc
1 + Q(PcPn
11) (18)
The Qfilter in (18) affects the range of disturbances that are rejected by the DOB. Consider the case where
Q= 1. In such a case, deviations from Pnare rejected at all frequencies, resulting in PDOB =Pn. In
practice, actuator limitations and sensor noise place a limit on the upper bound of the cutoff frequency for
Q. Nonetheless, this basic example demonstrates that large disturbances, or large deviations of the control
plant, may be compensated for by using sufficiently high Qfilter cutoff frequencies.
In the remainder of this section we analyze the relation between the disturbance rejection capability of a
DOB and the variations in closed-loop plant behavior of Pccaused by varying load inertia (Figure 10b). The
result of this analysis is a criteria that guarantees bounded tracking error of the DOB. (Kong and Tomizuka,
2013) provides a similar criteria for determining stability bounds of DOB-based systems with multiplicative
uncertainties, but does not give the same error bound guarantee presented here.
The underlying idea behind our method of characterizing DOB disturbance rejection is as follows (see
Figures 5 and 11). A DOB attenuates deviations of the control plant, Pc, from some nominal model, Pn.
The discrepancy between Pnand the DOB-compensated system (PDOB ) is the tracking error of the full
closed-loop system. Because the DOB only attenuates deviations, tracking error can never be reduced to
zero if finite deviations of Pcversus Pnexist. However, if a minimum acceptable tracking error is specified
(δ), the worst-case deviation of the control plant from the nominal plant that satisfies the tracking error can
be found (Pb). Therefore, any Pccontained within the area bounded by the best case scenario (Pn, zero
deviation) and the worst case scenario (Pb,δdeviation) satisfies the minimum tracking error requirement.
We refer to the area between Pnand Pbas the Disturbance observer Region Of Convergence (DROC).
Figure 12: These figures demonstrate how a disturbance observer can eliminate the effects of plant deviations
caused by changes to actuator load inertia. a) Closed-loop frequency response of Pcfor different load inertias.
The dashed line represents 20% error (<2dB) compared to the nominal high load inertia model while the
shaded area represents the DOB region of convergence. Both the maximum and minimum arm inertia lines
are contained within the DROC, meaning a DOB can be used to attenuate the disturbance due to these
effects. b) Closed-loop frequency response from desired torque (τd) to measured spring torque (τk) for the
full torque controller as shown in Figure 5c with a finite actuator load inertia. A series of responses to
plants with varying inertias is shown representing the full inertial operating range of the Valkyrie shoulder
abduction/adduction actuator (jL= 1.26 0.69 Kg ·m2). Note the insensitivity of the system to load
inertia variation.
A more formal description of this idea is presented here. Let us define an error tolerance (δ) relative to the
nominal plant model
Pd=Pn(1 δ) (19)
where Pdis the desired maximum allowable deviation from the nominal plant. Setting PDOB =Pdand
solving (18) for Pcyields
Pb=Pc=Pd(1 Q)
1QPdPn
1.(20)
Pbrepresents the boundary plant transfer function that satisfies the error tolerance, δ. In other words, Pb
determines the maximum deviation from the nominal plant model for which the DOB is able to compensate.
4.4 DOB Rejection of Disturbances Due to Variable Load Inertia
Analyzing the results shown in Figure 10b using our DROC method gauges how a DOB will reject dis-
turbances due to load inertia variation and thus maintain the desired ideal torque source abstraction. In
this analysis, we model a Valkyrie shoulder actuator for the two load inertia cases shown in Figure 9 using
the control parameters shown in Table 2. We also simulate a smaller load inertia (0.16 Kg ·m2), which is
included to demonstrate a case which violates the DROC. Figures 12 and 13 show these results.
Figure 12a illustrates how the closed-loop responses for both values of arm inertia remain within the DROC.
Table 2: Variable Inertia Simulation Parameters
Parameter Value Units
kp1.5 A/Nm
ζd0.9
fq20 Hz
δ20 %
Figure 13: Time domain representations of Figures 12a and 12b. In a) the step response of only the inner
portion of the proposed force controller (Pc) to various load inertias is shown. Again, for these control
parameters, all values of load inertia except jL= 0.16Kg ·m2remain within the DROC. In b) step responses
are shown for the same SEA plant parameters as in a), but a DOB is now applied. The DOB is shown to
dramatically improve low-frequency performance, maintaining the full range of Valkyrie’s arm inertia within
the allowable plant deviation.
Because both responses remain within the DROC, a DOB applied to either closed-loop response is guaranteed
to bring the system response to within δ(<2dB in this case) of the nominal fixed-output plant model. We
can visualize this result in Figure 12b where the perturbations to the full system transfer functions due to
load inertia variation are imperceptible for the two arm inertia cases and only become perceptible in the
scenario where jL= 0.16 Kg ·m2. These results also show that the stability of the system is minimally
influenced over the full range of Valkyrie’s arm inertia due to the relationship between the system bode plot
and stability (gain and phase) margins. Because of this relationship, the DROC is also a useful tool for
assessing the variation of stability for systems with DOBs.
Figure 13 shows the data of Figures 12a and 12b in the form of time domain step responses. Here, it is
again shown how the DOB is able to compensate for large deviations of the control plant, provided these
deviations primarily occur at frequencies below the Qfilter cutoff frequency.
5 Experimental Results on Valkyrie
In this section, we put the disturbance rejection capability of the DOB to the test in the form of two
experiments using Valkyrie’s arm, a serial chain of four SEAs. From Section 4 we have learned that the
DOB is capable of rejecting disturbances at frequencies below its Qfilter cutoff frequency. In this section, the
complexity of disturbances experienced by the DOB are increased beyond the variable inertia case considered
in Section 4.4. Now, the controller must attenuate disturbances from the full dynamics of a serial chain of
Figure 14: Experiment 1: Human Interaction. Valkyrie arm joints J1 through J4 are placed in torque
control mode and are actively tracking a constant desired torque signal. A human applies motions to the
arm causing all four joints to move.
−4
−2
0
Torque (Nm)
Actuator Torque Tracking
j1TrqMeas
j1TrqDes
−8
−6
−4
−2
Torque (Nm)
j2TrqMeas
j2TrqDes
−2
−1
0
Torque (Nm)
j3TrqMeas
j3TrqDes
7 8 9 10 11 12
−3.5
−3
−2.5
−2
−1.5
Time (s)
Torque (Nm)
j4TrqMeas
j4TrqDes
−0.2
0
0.2
Actuator Position
Position (rad)
−0.6
−0.4
−0.2
Position (rad)
−0.2
0
0.2
Position (rad)
7 8 9 10 11 12
−2
−1.5
−1
−0.5
Position (rad)
Time (s)
J1
J2
J3
J4
Figure 15: Data from Experiment 1: Human Interaction. Joint motion is caused by the human who
is interacting with Valkyrie’s arm. The peak tracking error for each joint is: J1: 0.85 Nm, J2: 2.2 Nm, J3:
0.63 Nm, J4: 0.64 Nm.
Figure 16: Motion generation controller used to test joint torque control in Experiment 2. The parameter
Krepresents joint stiffness while Brepresents joint damping.
Table 3: Valkyrie Control Parameters used in Experiments 1 and 2
Parameter J1 J2 J3 J4 Units Notes
kp1.5 1.5 1.5 1.5 A/Nm
ζd0.9 0.9 0.9 0.9
fq50 20 50 50 Hz
K100 100 50 50 Nm/rad Experiment 2 only
B10 10 5 5 Nm ·s/rad Experiment 2 only
four SEAs along with the disturbances introduced by control action of each of their motors.
The goal of the two experiments shown here is to assess the torque tracking capability of the distributed
joint-level torque controllers, which possess no knowledge of neighboring joints. The only models used in
these tests are the nominal high-output-impedance models used by the DOB as discussed in Section 3. The
control parameters used in these experiments are listed in Table 3.
In the first experiment, joints one through four5are placed into torque control mode and are commanded
to track a constant torque value. A human then grasps Valkyrie’s forearm and applies motions such that
all four joints move (see Figure 14). Figure 15 shows the data from this experiment. Low torque tracking
errors are maintained, despite the interaction forces and motions from the human.
In the second experiment, joints one through four are again placed into torque control mode. In this
experiment, however, desired joint torques are generated by the controller shown in Figure 16. This is a
simple proportional-derivative controller which creates an apparent joint stiffness (K) and damping (B)
based on a desired joint position (qd) and velocity ( ˙qd). We use this controller to generate motion in joints
one through four; we are less concerned with the position tracking error (qe) in this particular experiment.
Figure 17 shows the basic arm motion generated during Experiment 2. Correspondingly, Figure 18 shows
the data from Experiment 2.
These two experiments demonstrate that the disturbance attenuation properties of the joint-torque controller
are able to suppress the dynamics of neighboring joints, producing accurate torque tracking. However, by
visual inspection of Figures 15 and 18 it is difficult to evaluate the achieved tracking performance based
on torque error magnitude alone. A helpful metric in this scenario is the maximum torque tracking error
relative to the rated maximum joint torque. Using this metric, we find that the torque error relative to the
rated joint torque of each joint in Experiment 1 is: J1: 0.4%, J2: 0.76%, J3: 0.97%, and J4: 0.98%. For
5Refer back to Figure 2b for joint naming conventions.
Figure 17: Experiment 2: Coordinated Motion. Valkyrie arm joints J1 through J4 are placed in torque
control mode and are supplied desired torque signals according to the controller shown in Figure 16. As a
result, each joint roughly tracks a sinusoidal position.
−10
−5
0
5
Torque (Nm)
Actuator Torque Tracking
j1TrqDes
j1TrqMeas
−30
−20
−10
0
10
Torque (Nm)
j2TrqDes
j2TrqMeas
−4
−2
0
2
Torque (Nm)
j3TrqDes
j3TrqMeas
4 6 8 10 12
−4
−2
0
2
Time (s)
Torque (Nm)
j4TrqDes
j4TrqMeas
−5
0
5
Actuator Torque Tracking Error
Error (Nm)
−5
0
5
Error (Nm)
−2
−1
0
1
2
Error (Nm)
4 6 8 10 12
−5
0
5
Error (Nm)
Time (s)
−0.2
−0.1
0
0.1
0.2
Actuator Position
Position (rad)
−1
−0.5
0
Position (rad)
−0.2
−0.1
0
0.1
0.2
Position (rad)
4 6 8 10 12
−0.8
−0.7
−0.6
−0.5
−0.4
Position (rad)
Time (s)
J1
J2
J3
J4
Figure 18: Data from Experiment 2: Coordinated Motion. The purpose of this experiment is to
demonstrate the torque tracking accuracy of a serial chain of SEAs controlled using the methods presented
in this paper. The vertical black lines in the figure represent the time where each joint begins tracking the
desired torque signal. The peak tracking error for each joint is: J1: 2.2 Nm, J2: 2.5 Nm, J3: 0.9 Nm, J4:
0.65 Nm.
Figure 19: Valve Turn at the DRC Trials. Valkyrie using the actuator control methods described in this
paper to turn a valve during the DRC Trials, December 2013 in Miami, Florida, USA.
Experiment 2 these numbers are: J1: 1.16%, J2: 0.87%, J3: 1.38%, and J4: 1%. When considering that the
accuracy of many sensors fair no better than this in terms of full-scale accuracy6, we consider our results to
be quite strong. Based on this metric, our results also fare well compared to leading research in the field of
accurate torque control of SEAs with unmodeled disturbances. To date and to the authors’ best knowlege,
the leading results in literature achieve tracking accuracies of 1.6% of full scale (Kong et al., 2009) and 15%
of full scale (Kong et al., 2012).
The proposed actuator control methodology has also been used by Valkyrie to perform useful tasks in the
field, such as the valve turn task during the DRC Trials 2013 (Figure 19).
6 Conclusions and Discussion
The challenge tasks created for the DARPA Robotics Challenge set an unprecedented bar for the application
of advanced robotics. In response to this challenge, the NASA-JSC Valkyrie team designed a unique hu-
manoid robot. The combination of human-inspired dexterity, physical compliance, and an impressive suite
of sensory modes distinguished Valkyrie from the other DRC entries.
Valkyrie’s use of series elastic actuators inspired adoption of a decentralized torque control architecture. To
realize such an architecture, excellent disturbance rejection was required from the joint-level controllers to
provide actuators that can be abstracted as near-ideal torque sources, regardless of kinematic configuration.
In addition to proposing control methods aimed at this goal, we have provided a tool, the DROC, which
may be used to indicate in which conditions a DOB-controlled SEA meets this “near-ideal torque source”
specification.
We characterized Valkyrie’s SEA torque control performance with several empirical tests. In the first set of
tests, we characterized performance of a single actuator, obtaining measures of controllable torque resolution
(0.002 Nm), nominal torque tracking error (1 Nm), and torque tracking bandwidth (up to 70 Hz). These
numbers fare well in comparison to other torque-controlled SEAs which achieve bandwidths of 5-25 Hz (Pratt
and Williamson, 1995), 16 Hz (Vallery et al., 2007), 10 Hz (Kong et al., 2009) and 19 Hz (Sensinger and
Weir, 2005). In addition, actuator power consumption was considered. We demonstrated its critical relation
to torque tracking bandwidth, where a 30 Hz increase in bandwidth produced a 3.6x increase in power
consumption. The performance of our control approach was also shown in two tests using Valkyrie’s arm, a
serial chain of SEAs. In these tests, Valkyrie’s actuators accurately tracked torques to within 1.38% of their
rated torque values, an improvement compared to leading results in literature (1.6% and 15%) (Kong et al.,
6http://www.futek.com/files/pdf/Product%20Drawings/lcm300.pdf
2009; Kong et al., 2012).
Several unanswered questions remain for future study. First, a side effect of nearly-ideal torque source
actuators is a significant reduction in effective joint friction. This low-friction environment leads to a heavy
reliance on rigid body model accuracy. Accurate gravity compensation, for instance, is crucial to employ in
such systems as the torque-controlled actuators are easily backdriven by the weight of the robot. Therefore,
methods for validating and improving the rigid body model in hardware should be adopted when applying the
near-ideal torque source approach presented here. Second, when the rigid body model is inaccurate, as was
the case during much of Valkyrie’s development, controllers must be employed that can compensate for these
model discrepancies. Such controllers must rely more on position feedback effort than on the feedforward
effort produced from the model-based portion of the multi-joint controller (such as gravity compensation).
Maximizing the output impedance of these controllers7, 8 can be beneficial when there are rigid body model
discrepancies, a topic not well understood for the case of SEAs. This area of research deserves further study.
Acknowledgments
The Valkyrie robot was the product of many people’s efforts, most of whom are not listed as authors of this
paper. The work presented in this paper represents only a small portion of the efforts undertaken by the
Valkyrie team as a whole.
References
A. Schepelmann, M. T. and Geyer, H. (2012). Development of a testbed for robotic neuromuscular controllers.
In Proceedings of Robotics: Science and Systems.
Albu-Schaffer, A., Ott, C., and Hirzinger, G. (2007). A unified passivity-based control framework for position,
torque and impedance control of flexible joint robots. Int. J. Robot. Res., 26(1):23–29.
Ambrose, R. O., Burridge, R. R., Bluethmann, W., Aldridge, H., Askew, R. S., Rehnmark, F., Diftler, M.,
Magruder, D., and Lovchik, C. (2000). Robonaut: NASA’s space humanoid. IEEE Intelligent Systems,
15(4):57–63.
B`elanger, P. R. (1995). Control Engineering: A Modern Approach. Saunders College Publishing, Orlando,
Florida.
Bluethmann, W., Ambrose, R., Diftler, M., Askew, S., Huber, E., Goza, M., Rehnmark, F., Lovchik, C., and
Magruder, D. (2003). Robonaut: A robot designed to work with humans in space. Autonomous Robots,
14(2-3):179–197.
Bridgwater, L. B., Ihrke, C., Diftler, M. A., Abdallah, M. E., Radford, N. A., Rogers, J., Yayathi, S., Askew,
R. S., and Linn, D. M. (2012). The robonaut 2 hand-designed to do work with tools. In Robotics and
Automation (ICRA), 2012 IEEE International Conference on, pages 3425–3430. IEEE.
Diftler, M. A., Mehling, J., Abdallah, M. E., Radford, N. A., Bridgwater, L. B., Sanders, A. M., Askew,
R. S., Linn, D. M., Yamokoski, J. D., Permenter, F., et al. (2011). Robonaut 2-the first humanoid
robot in space. In Robotics and Automation (ICRA), 2011 IEEE International Conference on, pages
2178–2183. IEEE.
Garcia, E., Arevalo, J., Sanchez, F., Sarria, J., and Gonzalez-de Santos, P. (2011). Design and development
of a biomimetic leg using hybrid actuators. In Intelligent Robots and Systems (IROS), IEEE/RSJ
International Conference on, pages 1507 –1512.
7Perhaps even beyond the bounds of passivity (Ott et al., 2008)
8The “impedance mode” in Figure 3 was designed for this purpose
Godler, I., Inoue, M., Ninomiya, T., and Yamashita, T. (1999). Robustness comparison of control schemes
with disturbance observer and with acceleration control loop. In Industrial Electronics, ISIE Proceedings
of the IEEE International Symposium on, volume 3, pages 1035–1040 vol.3.
Hurst, J., Chestnutt, J., and Rizzi, A. (2010). The actuator with mechanically adjustable series compliance.
Robotics, IEEE Transactions on, 26(4):597 –606.
Hutter, M., Remy, C., Hoepflinger, M., and Siegwart, R. (2013). Efficient and versatile locomotion with
highly compliant legs. Mechatronics, IEEE/ASME Transactions on, 18(2):449–458.
Kim, J., Kwak, H., Lee, H., Seo, K., Lim, B., Lee, M., Lee, J., and Roh, K. (2012). Balancing control of
biped robot. In IEEE international Conference on Systems, Man, and Cybernetics.
Kong, K., Bae, J., and Tomizuka, M. (2009). Control of rotary series elastic actuator for ideal force-
mode actuation in human-robot interaction applications. Mechatronics, IEEE/ASME Transactions on,
14(1):105 –118.
Kong, K., Bae, J., and Tomizuka, M. (2012). A compact rotary series elastic actuator for human assistive
systems. Mechatronics, IEEE/ASME Transactions on, 17(2):288 –297.
Kong, K. and Tomizuka, M. (2013). Nominal model manipulation for enhancement of stability robustness
for disturbance observer-based control systems. International Journal of Control, Automation, and
Systems, 11(1):12 – 20.
Lagoda, C., Schouten, A., Stienen, A., Hekman, E., and van der Kooij, H. (2010). Design of an electric series
elastic actuated joint for robotic gait rehabilitation training. In Biomedical Robotics and Biomechatronics
(BioRob), 3rd IEEE RAS and EMBS International Conference on, pages 21 –26.
Li, Z., Tsagarakis, N., and Caldwell, D. (2012). A passivity based admittance control for stabilizing the com-
pliant humanoid coman. In Humanoid Robots (Humanoids), 12th IEEE-RAS International Conference
on, pages 43–49.
Nagatani, K., Kiribayashi, S., Okada, Y., Otake, K., Yoshida, K., Tadokoro, S., Nishimura, T., Yoshida, T.,
Koyanagi, E., Fukushima, M., and Kawatsuma, S. (2013). Emergency response to the nuclear accident
at the Fukushima Daiichi Nuclear Power Plants using mobile rescue robots. Journal of Field Robotics,
30(1):44–63.
Nakao, M., Ohnishi, K., and Miyachi, K. (1987). A robust decentralized joint control based on interference
estimation. In Robotics and Automation. Proceedings. IEEE International Conference on, volume 4,
pages 326–331.
Ott, C., Albu-Schaffer, A., Kugi, A., and Hirzinger, G. (2003). Decoupling based cartesian impedance control
of flexible joint robots. In Robotics and Automation, Proceedings. IEEE International Conference on,
volume 3, pages 3101–3107 vol.3.
Ott, C., Kugi, A., and Hirzinger, G. (2008). On the passivity-based impedance control of flexible joint robots.
Robotics, IEEE Transactions on, 24(2):416–429.
Paine, N., Oh, S., and Sentis, L. (2014). Design and control considerations for high-performance series elastic
actuators. Mechatronics, IEEE/ASME Transactions on, 19(3):1080–1091.
Paluska, D. and Herr, H. (2006). Series elasticity and actuator power output. In Robotics and Automation,
ICRA Proceedings IEEE International Conference on, pages 1830 –1833.
Pratt, G. and Williamson, M. (1995). Series elastic actuators. In Intelligent Robots and Systems. ’Hu-
man Robot Interaction and Cooperative Robots’, Proceedings. IEEE/RSJ International Conference on,
volume 1, pages 399 –406 vol.1.
Pratt, G., Willisson, P., Bolton, C., and Hofman, A. (2004). Late motor processing in low-impedance
robots: impedance control of series-elastic actuators. In American Control Conference, Proceedings of
the, volume 4, pages 3245 –3251 vol.4.
Pratt, J., Krupp, B., and Morse, C. (2002). Series elastic actuators for high fidelity force control. Industrial
Robot: An International Journal, 29(3):234 – 241.
Ragonesi, D., Agrawal, S., Sample, W., and Rahman, T. (2011). Series elastic actuator control of a pow-
ered exoskeleton. In Engineering in Medicine and Biology Society,EMBC, 2011 Annual International
Conference of the IEEE, pages 3515 –3518.
Reiland, M. J., Hargrave, B., Platt, R., Abdallah, M. E., and Permenter, F. N. (2013). Architecture for
robust force and impedance control of series elastic actuators. US Patent 8,525,460.
Robinson, D. W. (2000). Design and Analysis of Series Elasticity in Closed-loop Actuator Force Control.
PhD thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY.
Sensinger, J. W. and Weir, R. F. (2005). Design and analysis of a non-backdrivable series elastic actuator.
In Rehabilitation Robotics, ICORR. 9th International Conference on, pages 390–393.
Sensinger, J. W. and Weir, R. F. (2006). Unconstrained impedance control using a compact series elastic
actuator. In Mechatronic and Embedded Systems and Applications, Proceedings of the 2nd IEEE/ASME
International Conference on, pages 1 –6.
Sentis, L., Petersen, J., and Philippsen, R. (2013). Implementation and stability analysis of prioritized
whole-body compliant controllers on a wheeled humanoid robot in uneven terrains. Autonomous Robots,
35(4):301–319.
Spong, M. (1987). Modeling and control of elastic joint robots. J. Dyn. Sys., Meas., Control., 109(4):310–318.
Taylor, M. D. (2011). A compact series elastic actuator for bipedal robots with human-like dynamic perfor-
mance. Masters thesis, Carnegie Mellon University.
Thorson, I. and Caldwell, D. (2011). A nonlinear series elastic actuator for highly dynamic motions. In
Intelligent Robots and Systems (IROS), IEEE/RSJ International Conference on, pages 390 –394.
Vallery, H., Ekkelenkamp, R., van der Kooij, H., and Buss, M. (2007). Passive and accurate torque control of
series elastic actuators. In Intelligent Robots and Systems, IROS IEEE/RSJ International Conference
on, pages 3534 –3538.
Wyeth, G. (2006). Control issues for velocity sourced series elastic actuators. In Australian Conference on
Robotics and Automation.
... Legged robots such as quadrupeds and humanoid robots often benefit from elastic mechanisms to protect the hardware from high force impacts spikes during walking. This approach has been featured on Valkyrie [54], [48], THOR , ESCHER [32], UXA-90 [55], [56], iCub [57], [58], and WALK-MAN [59]. As displayed in Fig. 2, these actuators are designed with either linear or rotary motion to drive parallel or in-line with the joint, respectively. ...
... This section focuses on PANDORA's control hierarchy broken into high and low levels of control and discusses PANDORA's mechatronics design including the layers of sensing and feedback between the control layers. This control and mechatronics approach is consistent with other legged robots that utilize similar planning approaches and feedback sensor types [32], [54], [16], [26], [35]. This paper focuses on presenting the joint-level control approach which is designed within the Low-Level Controller (LLC). ...
... One of the most popular approaches for controlling humanoid robots is utilizing linear actuators attached to drive the joint in parallel. An elastic component is often connected to the gear train acting as a passive spring resulting in a mechanical low-pass filter between the motor and its loads [51], [49], [48], [74], [75], [76], [54]. For a linear system, this combination is known as an LSEA and is often paired with the DOB control strategy. ...
Preprint
Full-text available
In this work, the joint-control strategy is presented for the humanoid robot, PANDORA, whose structural components are designed to be compliant. As opposed to contemporary approaches which design the elasticity internal to the actuator housing, PANDORA's structural components are designed to be compliant under load or, in other words, structurally elastic. To maintain the rapid design benefit of additive manufacturing, this joint control strategy employs a disturbance observer (DOB) modeled from an ideal elastic actuator. This robust controller treats the model variation from the structurally elastic components as a disturbance and eliminates the need for system identification of the 3D printed parts. This enables mechanical design engineers to iterate on the 3D printed linkages without requiring consistent tuning from the joint controller. Two sets of hardware results are presented for validating the controller. The first set of results are conducted on an ideal elastic actuator testbed that drives an unmodeled, 1 DoF weighted pendulum with a 10 kg mass. The results support the claim that the DOB can handle significant model variation. The second set of results is from a robust balancing experiment conducted on the 12 DoF lower body of PANDORA. The robot maintains balance while an operator applies 50 N pushes to the pelvis, where the actuator tracking results are presented for the left leg.
... Another popular technique to estimate joint torques is the use of so-called Series Elastic Actuators (SEA) [16]. A drawback of SEAs is that, while the additional compliance is convenient for torque measurements, it may be a limitation for higher level controllers not explicitly designed to deal with a compliant system. ...
Preprint
In this paper, we present algorithms to estimate external contact forces and joint torques using only skin, i.e. distributed tactile sensors. To deal with gaps between the tactile sensors (taxels), we use interpolation techniques. The application of these interpolation techniques allows us to estimate contact forces and joint torques without the need for expensive force-torque sensors. Validation was performed using the iCub humanoid robot.
... Researchers were able to build useful dynamic models for position and force control applications by considering all external disturbances as fictional disturbance variables. Series elastic actuators (SEAs) have attracted more attention recently because of their potential advantages in the physical robot-environment interface, such as increased safety and high-fidelity force control [74][75][76][77]. SEAs have been used to develop exoskeletons, humanoids, and other complex robotic systems, as well as compliant industrial robots. ...
Article
Full-text available
The goal of this study is to offer an operational model for a group of elasticity actuation which could satisfy the morphological and payload criteria of industrialised cooperative robotic tyrants Flexible actuation defy the maxim that "stiffer is better" in machine layout, whereas being widely advertised because its benefits like excellent pressure loyalty, exceptionally low resistance, decreased abrasion, and a broad power regulation bandwidth. For integrated monitoring of robotics in broad spaces, Sequential Stretch Actuators (SEAs) are especially appropriate. We describe the output-feedback torque management approach for SEAs, that employs a form of filtering to predict mobility indicators and systemic aggregated disruptions, to address parameter uncertainties and external disturbances. By doing this, the controller's resistance to system uncertainty is strengthened. Projects that employ series elasticity actuators, especially position control, can frequently have performance constraints as a result of poor controller design. However, because of its low impedance, great force fidelity, and built-in shock resistance, SEAs are perfect for a wide range of applications. Legged robots, exoskeletons, robotic arms, haptic interfaces, and adjustable suspension are some of these uses. We also demonstrate that the control process and controller types for SEAs differ depending on the application of the SEA. In this study, we aim to compare and analyse six existing controller theories or control type designs for SEAs that have been previously published. We will cover topics such as the design of mechanical systems, choice of actuators, motors, and drive systems, and power transfer from actuators to moving parts.
Article
Full-text available
Compliant manipulators with variable stiffness actuation systems are crucial for safety in physical human-robot interactions, improving performance during unexpected collisions. However, their inherent compliance poses motion control challenges, especially with rapid stiffness changes and uncertain end-effector loads, adversely affecting controller stability and accuracy. To address uncertainties from stiffness variations and changing payloads, this paper proposes the Bidirectional Fuzzy Brain Emotional Learning controller—a fuzzy neural network with a unique bidirectional brain emotional learning algorithm for weight adaptation—that effectively handles varying stiffness and payload uncertainties. Additionally, a novel algorithm for systematic establishment of fuzzy layers is presented which significantly reduces the effort and time for implementation. Simulation and experimental results demonstrate the proposed controller’s superior adaptability and tracking performance under varying stiffness and payload uncertainties compared to the conventional PID controller. This new algorithm for fuzzy layer setup can serve as a default for all fuzzy neural network-based controllers, enhancing their ease of use across various applications. The controller can also be directly extended to grasping uncertain objects using variable stiffness actuated systems, improving safety and reliability in physical human-robot interactions.
Thesis
Full-text available
In recent years, the field of robotics has witnessed substantial advancements in the design, development, and control of anthropomorphic robotic manipulators, catering to diverse research, assistive, commercial, service, and industrial applications. Despite the increased interest, most platforms are still highly actuated, bulky, complex, expensive, rigid, and dangerous around humans. This complicates human-robot interaction (HRI) and in particular physical HRI and prevents robots from delivering impact in everyday life. This work focuses on the design, development, and experimental investigation of modular, bioinspired, anthropomorphic, adaptive robot arm-hand systems targeting research and assistive tasks that necessitate human-robot interaction and physical HRI. The thesis details crucial aspects of the manipulation systems, including the hardware, the electrical systems, the electronics, the low-level control software, and the human-robot interaction oriented skill transfer frameworks. In particular, this work introduces modular, lightweight, anthropomorphic arm designs and various adaptive robot grippers and hands that take advantage of alternative uses of structural compliance and underactuated transmission schemes to offer affordable dexterity tailored to diverse tasks. These include the manipulators for the New Dexterity Autonomous Robotic Assistant (ARoA) platform and the arm hand systems and grippers used for the Open Bimanual Manipulation Platform (OpenBMP), which is an open-source hardware testbed that allows replication by other research groups and roboticists, facilitating affordable experimentation in the research domain of bimanual robotic manipulation. The proposed platforms are bio-inspired, introducing low inertial and modular designs that offer flexibility, customization, and adaptability to various applications, including dexterous bimanual manipulation and physical HRI such as tactile American Sign Language (t-ASL). The thesis also introduces several adaptive robot grippers and hands that are characterised by: i) their inherent structural compliance, ii) their ability to conform to the object shapes, maximizing grasp stability, iii) their ability to enable a robust execution of dexterous manipulation tasks, iv) their lightweight structures, and finally, in some cases, v) their highly anthropomorphic kinematics and form factors. Integrating all the components, the thesis examines various optimised torso configurations for the ARoA and the OpenBMP robotic platforms through extensive simulations that enable a more efficient execution of bimanual manipulation tasks, increasing the available dexterous workspaces. In the area of human-to-robot skill transfer, the thesis proposes: i) wearable human-machine interfaces that can facilitate a human-in-the-loop, fast, scalable, and intuitive collection of big data of robotic manipulation so as to offer an efficient task replication by the robot exploiting human intelligence, ii) a human to robot skill transfer framework for the execution of complex telemanipulation tasks with a bimanual platform, utilizing wearable interfaces and sensorized gloves to replicate human motions, and iii) a physical HRI oriented human to robot skill transfer framework that enables tactile ASL letter signing and Deaf-Blind communication.
Article
Full-text available
This paper discusses design and control of a prismatic series elastic actuator with high mechanical power output in a small and lightweight form factor. A design is introduced that pushes the performance boundary of electric series elastic actuators by using high motor voltage coupled with an efficient drivetrain to enable large continuous actuator force while retaining speed. Compact size is achieved through the use of a novel piston-style ball screw support mechanism and a concentric compliant element. Generic models for two common series elastic actuator configurations are introduced and compared. These models are then used to develop controllers for force and position tracking based on combinations of PID, model-based, and disturbance observer control structures. Finally, our actuator’s performance is demonstrated through a series of experiments designed to operate the actuator at the limits of its mechanical and control capability.
Conference Paper
Full-text available
This paper presents a generic stabilization framework which is applicable for both compliant and stiff humanoids. The proposed control framework is applied to the passive compliant humanoid robot COMAN which is equipped with series elastic actuators. The stabilization control framework combines the compliance control and the intrinsic angular momentum modulation to achieve an agile and compliant interaction against external perturbations. The admittance based compliance control uses the force/torque sensing in both feet to regulate the active compliance for the position controlled system. The physical elasticity in the new full body COMAN is exploited for the reduction and absorption of the instantaneous impacts while the admittance control further dissipates the excessive elastic energy. The angular momentum controller reduces the overall inertia effect for providing more rapid reactions. Both the theoretical work and experimental validation were presented. The effectiveness of the control scheme is demonstrated by COMAN's capabilities of withstanding various types of perturbations applied over the body, balancing on a moving platform and stabilizing while walking. Experimental data of the ground reaction force/torque, center of mass references and estimations, and the stored elastic energy are presented and analyzed.
Article
Full-text available
In this work, we implement the floating base prioritized whole-body compliant control framework described in Sentis et al. (IEEE Transactions on Robotics 26(3):483–501, 2010) on a wheeled humanoid robot maneuvering in sloped terrains. We then test it for a variety of compliant whole-body behaviors including balance and kinesthetic mobility on irregular terrain, and Cartesian hand position tracking using the co-actuated (i.e. two joints are simultaneously actuated with one motor) robot’s upper body. The implementation serves as a hardware proof for a variety of whole-body control concepts that had previously been developed and tested in simulation. First, behaviors of two and three priority tasks are implemented and successfully executed on the humanoid hardware. In particular, first and second priority tasks are linearized in the task space through model feedback and then controlled through task accelerations. Postures, on the other hand, are shown to be asymptotically stable when using prioritized whole-body control structures and then successfully tested in the real hardware. To cope with irregular terrains, the base is modeled as a six degree of freedom floating system and the wheels are characterized through contact and rolling constraints. Finally, center of mass balance capabilities using whole-body compliant control and kinesthetic mobility are implemented and tested in the humanoid hardware to climb terrains with various slopes.
Article
In this work a novel type of impedance controllers for flexible joint robots is proposed. As a target impedance a desired stiffness and damping are considered without inertia shaping. For this problem two controllers of different complexity are proposed. Both have a cascaded structure with an inner torque feedback loop and an outer impedance controller. For the torque feedback, a physical interpretation as a scaling of the motor inertia is given, which allows to incorporate the torque feedback into a passivity based analysis. The outer impedance control law is then designed differently for the two controllers. In the first approach the stiffness and damping terms and the gravity compensation term are designed separately. This outer control loop uses only the motor position and velocity, but no noncollocated feedback of the joint torques or link side positions. In combination with the physical interpretation of torque feedback, this allows us to give a proof of the asymptotic stability of the closed-loop system based on the passivity properties of the system. The second control law is a refinement of this approach, in which the gravity compensation and the stiffness implementation are designed in a combined way. Thereby, a desired static stiffness relationship is obtained exactly. Additionally, some extensions of the controller to visco-elastic joints and to Cartesian impedance control are given. Finally, some experiments with the DLR lightweight robots verify the developed controllers and show the efficiency of the proposed control approach.
Patent
An SEA architecture for controlling the torque applied by an SEA that has particular application for controlling the position of a robot link. The SEA architecture includes a motor coupled to one end of an elastic spring and a load coupled to an opposite end of the elastic spring, where the motor drives the load through the spring. The orientation of the shaft of the motor and the load are measured by position sensors. Position signals from the position sensors are sent to an embedded processor that determines the orientation of the load relative to the motor shaft to determine the torque on the spring. The embedded processor receives reference torque signals from a remote controller, and the embedded processor operates a high-speed servo loop about the desired joint torque. The remote controller determines the desired joint torque based on higher order objectives by their impedance or positioning objectives.
Article
On March 11, 2011, a massive earthquake (magnitude 9.0) and accompanying tsunami hit the Tohoku region of eastern Japan. Since then, the Fukushima Daiichi Nuclear Power Plants have been facing a crisis due to the loss of all power that resulted from the meltdown accidents. Three buildings housing nuclear reactors were seriously damaged from hydrogen explosions, and, in one building, the nuclear reactions became out of control. It was too dangerous for humans to enter the buildings to inspect the damage because radioactive materials were also being released. In response to this crisis, it was decided that mobile rescue robots would be used to carry out surveillance missions. The mobile rescue robots needed could not be delivered to the Tokyo Electric Power Company (TEPCO) until various technical issues were resolved. Those issues involved hardware reliability, communication functions, and the ability of the robots' electronic components to withstand radiation. Additional sensors and functionality that would enable the robots to respond effectively to the crisis were also needed. Available robots were therefore retrofitted for the disaster reponse missions. First, the radiation tolerance of the electronic componenets was checked by means of gamma ray irradiation tests, which were conducted using the facilities of the Japan Atomic Energy Agency (JAEA). The commercial electronic devices used in the original robot systems operated long enough (more than 100 h at a 10% safety margin) in the assumed environment (100 mGy/h). Next, the usability of wireless communication in the target environment was assessed. Such tests were not possible in the target environment itself, so they were performed at the Hamaoka Daiichi Nuclear Power Plants, which are similar to the target environment. As previously predicted, the test results indicated that robust wireless communication would not be possible in the reactor buildings. It was therefore determined that a wired communication device would need to be installed. After TEPCO's official urgent mission proposal was received, the team mounted additional devices to facilitate the installation of a water gauge in the basement of the reactor buildings to determine flooding levels. While these preparations were taking place, prospective robot operators from TEPCO trained in a laboratory environment. Finally, one of the robots was delivered to the Fukushima Daiichi Nuclear Power Plants on June 20, 2011, where it performed a number of important missions inside the buildings. In this paper, the requirements for the exploration mission in the Fukushima Daiichi Nuclear Power Plants are presented, the implementation is discussed, and the results of the mission are reported. © 2013 Wiley Periodicals, Inc. (Webpage: http://www.astro.mech.tohoku.ac.jp/)
Conference Paper
We propose a balancing control framework for a torque-controlled biped robot, Roboray. Roboray has two 6 DOF legs and torque sensors are integrated at all the leg joints. It has a new cable-driven joint module as a pitch joint drive, which is highly back-drivable and elastic. Using these hardware characteristics, we propose a new balancing control algorithm. This algorithm is the combination of gravity compensation, virtual gravity control and damping control. A friction compensation technique is also introduced in order to eliminate the nonlinearity of damping and to improve the performance of torque tracking. Our proposed method is applied to a simple inverted pendulum system and Roboray. Experimental results show that these two system keep their balance when they are pushed slightly.
Conference Paper
Current control approaches to robotic legged loco-motion rely on centralized planning and tracking or motion pat-tern matching. Central control is not available to robotic assistive devices that integrate with humans, and matching predefined patterns severely limits user dexterity. By contrast, biological systems show substantial legged dexterity even when their central nervous system is severed from their spinal cord, indicating that neuromuscular feedback controls can be harnessed to encode stability, adaptability, and maneuverability into legged systems. Here we present the initial steps to develop a robotic gait testbed that can implement and verify neuromuscular controls for robotic assistive devices. The initial stage consists of an antagonistically actuated two segment leg with a floating compliant joint. We detail its electromechanical design and low level, velocity-based torque control. Additionally, we present experiments that test the leg's performance during human-like high fidelity motions. The results show that the robot can track fast motions corresponding to 87% of the maximum performance limit of human muscle. The experiments also reveal limitations of our current implementation and we discuss solutions to overcoming them.