sensors

Article
Muscular Damping Distribution Strategy for Bio-Inspired, Soft
Motion Control at Variable Precision †
Patrick Vonwirth * , Karsten Berns

Department of Computer Science, RPTU Kaiserslautern-Landau, 67663 Kaiserslautern, Germany

* Correspondence: [email protected]
† This paper is an extended version of our paper published in 25th International Conference Series on Climbing
and Walking Robots and the Support Technologies for Mobile Machines (CLAWAR 2022), Ponta Delgada,
Portugal, 12–14 September 2022.

Abstract: Bio-inspired and compliant control approaches have been studied by roboticists for decades
to achieve more natural robot motion. Independent of this, medical and biological researchers have
discovered a wide variety of muscular properties and higher-level motion characteristics. Although
both disciplines strive to better understand natural motion and muscle coordination, they have yet
to meet. This work introduces a novel robotic control strategy that bridges the gap between these
distinct areas. By applying biological characteristics to electrical series elastic actuators, we developed
a simple yet efficient distributed damping control strategy. The presented control covers the entire
robotic drive train, from abstract whole-body commands to the applied current. The functionality
of this control is biologically motivated, theoretically discussed, and finally evaluated through
experiments on the bipedal robot C ARL. Together, these results demonstrate that the proposed
strategy fulfills all requirements that are necessary to continue developing more complex robotic
tasks based on this novel muscular control philosophy.

Keywords: bio-inspired; impedance; control; artificial muscles; series elastic actuator; humanoid

1. Introduction
Citation: Vonwirth, P.; Berns, K. Artificial cognition capabilities grow with the development of better sensors, and
Muscular Damping Distribution
computer-guided decision capabilities grow with the availability of computational power.
Strategy for Bio-Inspired, Soft Motion
Equivalently, robotic locomotion is changing with the continuous development and im-
Control at Variable Precision. Sensors
provement of actuation systems. The available drive technology is highly specialized and
2023, 23, 2428.
tailored to the exact needs of the surrounding machines. Different types of wheels have
https://doi.org/10.3390/s23052428
been developed to drive over different terrains. Today, stiff motors with large gear ratios
Academic Editor: Simon X. Yang are used to precisely control industrial robot arms on specific trajectories. Legged locomo-
tion, however, requires different properties of an actuation system than most industrial
Received: 31 January 2023
applications. Compared to other tasks, the biggest problem of legged locomotion is likely
Revised: 16 February 2023
Accepted: 17 February 2023
the physical impact of footsteps, which can damage stiff mechanical systems. Therefore,
Published: 22 February 2023
control strategies for legged locomotion should differ from the well-studied approaches
implemented in today’s stiff and inflexible industrial automation robots.
In order to develop such a different motion strategy, this paper distances itself from
classical control theory. Instead, a natural, musculoskeletal locomotor system was used as a
Copyright: © 2023 by the authors. blueprint to develop the new motion control philosophy. The introduced concept covers all
Licensee MDPI, Basel, Switzerland. involved elements of a robotic drive system, including the low-level motor control as well
This article is an open access article as the abstract, high-level input. To the best of our knowledge, no other approach exists
distributed under the terms and that tackles the problem of natural, compliant motion control in an equivalent manner. The
conditions of the Creative Commons article at hand, first published in the proceedings of the 25th International Conference on
Attribution (CC BY) license (https:// Climbing and Walking Robots (CLAWAR 2022), is an extended description of the concept
creativecommons.org/licenses/by/
introduced in [1]. Special focus is given to the distributed muscular control details.
4.0/).

Sensors 2023, 23, 2428. https://doi.org/10.3390/s23052428 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 2428 2 of 25

Within the following, an introduction to the variety of existing robotic actuation

systems, including control approaches, is given. We focus on the bipedal robot C ARL in
Section 1.2. The development description of the new, compliant control begins in Section 2
at the muscular level. With the coverage of higher abstraction levels, Section 3 finalizes this
description. Section 4 contains a theoretical, in-depth analysis of the capabilities and limits
of the proposed strategy. An experimental evaluation and discussion are given in Section 5.
Finally, Section 6 concludes the findings and offers an outlook on further developments
that can be based on the proposed, novel control strategy.

1.1. Drive Technology for Artificial Legged Locomotion

In the past decades, many different approaches have been developed to make robots walk.
Aside from classic concepts, such as the so-called divergent component of motion [2–5] or hybrid
zero dynamics [6–9], natural, bio-inspired concepts, such as muscle-reflex strategies [10–12],
machine learning approaches [13,14], or the B4LC (Bio-inspired Behavior Based Bipedal
Locomotion Control) [15–17] system were developed and evaluated in simulations and on
real robotic systems. Despite the fact that these strategies are opening up a wide variety of
approach basics, all of them share one property—all approaches are developed upon (and,
hence, rely on) a concrete actuation system, e.g., position or torque controllability, or stiff
or compliant mechanical structures. Locomotion strategies advance with the progress of
motor technology and cannot be separated from the controlled hardware. Therefore, the
new control philosophy introduced in this article also includes a novel physical actuation
concept derived from natural muscle properties.
Unfortunately, due to the past weaknesses of actuation systems (w.r.t. the needs of
legged locomotion), the actual hardware capabilities have strongly limited the research
possibilities. As a result, legged locomotion only played a minor role in past mobile robotic
research. With the invention of soft actuation systems, however, this area of research
is about to receive more interest. Research on legged locomotion has progressed in the
context of quadruped robots, including research on MIT Cheetah [18,19], HyQ [20,21],
ANYmal [22,23], and Spot [24].
Ficht and Behnke [25], within their short overview on the technological and devel-
opment process of humanoid robots, clearly show the recent trend toward so-called soft
robotics. Obviously, this directly addresses the above-mentioned problem with external
impacts. In most cases, the softness of a robot is achieved by using so-called artificial
muscles. The most popular types of artificial muscles used for locomotion are pneumatic,
hydraulic, and electric drives, or a combination of these. They all come with physically
integrated, serially attached, or high-speed controlled elasticity and/or transparency. A
quite unorthodox example of a pneumatic drive system is a fully inflatable upper-body
humanoid introduced by Best et al. [26]. The hydraulic, so-called integrated smart actuator
introduced by Barasuol et al. [27] is one of the latest developments in hydraulic drive trains
that is meant to be integrated into the next generation of the quadruped robot HyQ. Electri-
cally powered artificial muscles are, in most cases, quasi-direct drives [25]. This architecture
considers the required force, velocity, and actuator transparency in accordance with the
needs and physical limits. A rotational example of such an actuator, introduced by Seok
et al. [28], has been built into the MIT Cheetah robot. In contrast, the compliant robotic leg
C ARL [29,30] (which is used as an evaluation platform throughout this article) is equipped
solely with linear series electric actuators (S EAs), the so-called RRL AB S EAs [31]. A hybrid
electric–pneumatic actuation concept was proposed by Mohseni et al. in [32]. This actuation
concept is also used to better reproduce natural leg construction and control characteristics.
Moreover, artificial muscles include a large variety of large and small-scale technolo-
gies suitable for all types of motion, e.g., grasping. A recent overview of such technologies
is given by Mirvakili and Hunter in [33]. However, in comparison to humans and animals,
it is clear that there is still a lot of room to improve current technical systems and to learn
from nature.
Sensors 2023, 23, 2428 3 of 25

1.2. Bio-Inspired, Muscular Actuation of C ARL

The bio-inspired locomotion strategy B4LC, developed by Luksch [16] in 2010, is
proven to be principally functional in a pure simulative environment. The basic idea of this
concept is to combine simple, torque-based feed-forward control patterns with reactive,
impedance-based reflexes. By extending this approach with more advanced skills, such as
upslope walking [17], walking on rough terrain [34], and stepping over obstacles [35], the
basic idea of the B4LC approach seemed to be suitable for real-world evaluation. Due to the
high demand on the actuation system, the first successful attempt to construct and transfer
this pure virtual control concept to a planar robotic leg was achieved by Schütz in 2020 [36].
Equal to the natural motorization principles, the planar, bipedal robot C ARL (Figure 1) [29]
is equipped with both, mono- and biarticular electrical S EAs (Figure 1b,c) [31].

Biarticular Knee–Ankle S EA

Monoarticular Ankle S EA

(b) Construction view of C ARL’s shank

(a) Bipedal Version of C ARL (c) S EAs mounted along C ARL’s shank

Figure 1. The bipedal version of the robotic leg C ARL consists of a short trunk and two planar,
muscular-driven legs with medical standard prosthetic feet.

Stand-alone experiments on the bipedal version of C ARL, although successfully demon-

strated as a single leg being mounted over a treadmill, have failed for two main reasons.
(1) The applied feed-forward torque patterns should offer basic motion characteristics. Due
to the changing environmental conditions (e.g., temperature and dust-dependent friction),
noisy measurements, and imprecise motor actions, these patterns had to be adapted perma-
nently while creating a large variety of resulting motions. (2) There are too many nested
and mechanically coupled control loops. The unmodeled and unconsidered interactions of
all individually controlled S EAs artificially lowered the stability of the overall system far
below its theoretical capabilities and far below the requirements of free walking.
Hence, with a special focus on these two issues, a new low-level, muscular control
strategy for compliant and imprecise locomotion is described as follows: It is a strongly
simplified, bio-inspired concept of distributed damping control based on series elastic
actuators that imitate muscles. In contrast to other state-of-the-art approaches, classical
control goals, such as high tracking precision, are explicitly not desired. For most natural
motions, such as walking, it is absolutely sufficient to roughly track a desired trajectory,
while the ability to react and adapt to the environment is of higher interest.
Sensors 2023, 23, 2428 4 of 25

2. Actuation Concept for Compliant Control

From the need to be impact-resistant, the requirement directly evolves to include
mechanical, compliant elements in the construction of a robotic leg. Unfortunately, every
element adds a passive, uncontrolled degree of freedom to the robot. Hence, during the
mechanical design phase of a robotic limb, impact tolerance and controllability have to be
balanced against each other.

2.1. Analysis of Impedance Control Strategies

In 2015, Semini et al. [21] proposed following the principle of stiff inside, soft outside,
a guideline for designing impact-tolerant robotic limbs. This principle suggests building
the actuated limb as stiff as possible in order to allow for classic, high-precision control.
Environmental impacts are compensated for by compliant contact points, e.g., spring-
loaded or rubber feet. The most significant advantage of this approach is its proximity
to the traditional, stiff construction design. By definition, it offers the possibility to apply
well-known, standard trajectory control algorithms on such a robot design. Hence, it
is quite common in legged robotics and has been successfully demonstrated on several
robots, e.g., on the bipedal robot DURUS [37,38] and the quadruped robots mentioned in
Section 1.1. However, the advantage of having only a single compliant element close to the
end of a limb also forms the biggest drawback of this approach. The robot is limited to
interacting with the environment only through the specially designed contact points and in
no other way than what these elements are designed for.
To overcome this limitation of interaction, a whole-limb impedance control has to be
provided. It can be realized by mechanical compliance, virtual compliance, or a combination
of both. Mechanical compliance offers impact tolerance at the cost of controllability whereas
virtual compliance does the opposite. In [21], Semini et al. focused on controllability and
suggested using pure active impedance control since “no springs are needed to protect the
actuation system”. However, their principle of stiff inside, soft outside can be interpreted as
already violating their above-mentioned statement.
Approaching the problem of impedance control from the mechanical side can be
done by using soft S EAs, such as the so-called RRL AB-S EA, introduced and successfully
evaluated by Schütz et al. [31,39,40]. However, although successfully tested as a single,
isolated actuator, the combined control of five of them mounted in the robotic leg C ARL
uncovered serious stability issues. The nested and mechanically coupled control loops form
a critically large number of unhandled oscillation circuits inside the robot. From low-level
to high-level, the mechanical and digital oscillating systems are as follows:
• The electric motor’s mass against the serial spring;
• The unconventionally by-passed PID spring-deflection controller;
• The outer-loop impedance controller of the whole actuator surrounding the inner-loop
force controller;
• The mechanical and virtual springs of several S EAs that are connected to one shared
joint;
• The high-level force distribution strategy of all available actuators;
• The physical coupling of both legs on the bipedal version of C ARL.
Details about the realization of the mentioned RRL AB-S EA controller are given in [36].
Information about the tested force distribution strategies can be found in [30,41]. To keep
the overall system stable, the capabilities of all S EAs had to be limited far below the
theoretical requirements needed to execute stable locomotion with C ARL.
From these points of view, it seems better to compromise between physical and
virtual impedance.

2.2. Natural Muscle Emulation on a Series Elastic Actuator

A natural muscle can be described as a composition of two main parts, i.e., a motor
part, which is the contractile element, and an elastic element, which is the serially attached
tendon. In between both, the so-called Golgi tendon organ is located, sensing changes in the
Sensors 2023, 23, 2428 5 of 25

muscle tendon tension. This organ is used to control and protect muscular tension, e.g., the
commonly known muscle stretch reflex is triggered by that organ [42].
From the mechanical point of view, a S EA is a nearly perfect technical platform to
emulate muscular behavior. Using an electrical S EA, the only disadvantage compared to a
biological muscle is the reflected inertia of the rotating electric motor. However, the above
Section 2.1 clearly points out the big problems that arise from the limited possibility of
stably controlling such a S EA to mimic an ideal impedance. As the only consequent result,
we propose a trade-off of the introduced pure virtual impedance control (i.e., following the
principle of stiff inside, soft outside) against the usage of series elastic actuation. That is to
shift the principle of stiff inside, soft outside from the abstract limb level into the limb itself
to the actuator level, i.e., to use the mechanical structure of the S EA but without trying to
control its elasticity. By doing so, most of the previously described instability causes can
be avoided by design. In fact, they are simply removed from the design of the controller.
Obviously, this comes at the cost of control precision. The need for high-precision force
sensing vanishes, and the elastic element, i.e., the soft-outside part degenerates to only
protect the stiff, internal drive train from physical impacts. As a direct result, much higher
stiffnesses can be used, extending the range of controllable frequencies. However, due to
the lack of feedback and control of the elastic element, impacts have to be handled by the
actuator’s transparency, i.e., its backdrivability, which becomes a mandatory property.
In the mid-1980s, Mussa-Ivaldi et al. [43] found that the short-range end-point stiffness
of a human arm features three important properties:
• Measured stiffness matrices at the end-point of a human arm can be treated as sym-
metric, i.e., elliptical;
• The general shape/orientation of the end-point stiffness ellipse mainly depends on
the arm configuration, not the muscle activities;
• Muscular co-contraction strongly increases the overall size of the measured stiffness
ellipse, i.e., its spring rate.
These insights support the assumption that the (short-range) end-point stiffness is
generated on a muscular level, exploiting the physically inherent dynamics of the con-
tractile elements, tendons, and mounting positions of the muscles. Recent studies on
the effects of having muscle-like dynamics built inside actuators support this hypothesis.
The pure existence of muscle-like force characteristics significantly reduces the amount of
control information required from some central intelligence [44]. The general types and
shapes of these characteristics are well known and create a muscle-level, zero-delay feedback
system, the so-called preflexes [45,46]. There are three main dependencies: time, length, and
motion. While the temporal force dependency (i.e., muscular fatigue) can be ignored as an
unnecessary limitation, the two others construct the two impedance properties, i.e., stiffness
and damping, respectively. In general, the emitted force of a muscle with respect to these
two separated impedance properties can be estimated as follows:

F (l ) ∼ FMax − K (l ) l F (v) ∼ FMax − D (v) v (1)

with FMax being the maximum muscle force at rest and K (·) and D (·) representing stiffness
and damping, respectively. Both impedance functions are modeled in such a way that
the individually resulting overall emitting forces meet their known shapes, given in [46].
Figure 2 sketches these two properties together, interpreted as intrinsic impedance.
Sensors 2023, 23, 2428 6 of 25

Force/Stiffness Force/Damping

l0 0
Longer Shorter Lengthening Shortening

(a) Muscular Force-Length Relation (b) Muscular Force-Velocity Relation

Figure 2. Sketches of the two principal impedance properties of a natural muscle. On the left, the
(a) stiffness characteristic is shown, on the right, the (b) damping. The overall emitted forces are
represented by the blue curves. The dashed line indicates the maximum force at rest. In red, the
derived shapes of the muscle-intrinsic stiffness and damping values are sketched.

It can be seen that muscular stiffness contains both positive and negative sections. The
muscular stiffness is partially unstable and not really suitable to be taken as additional
muscular control property. Furthermore, stiffness requires an additional equilibrium
position as a secondary control. Damping, on the other hand, stays positive, i.e., stable,
over the whole range of motion. From this technical point of view, the muscular preflexes
should be reduced to only cover muscle-intrinsic damping. That reduction is also supported
by Brown and Loeb in their publication about preflexes [45]. They explicitly observed the
force-length relation of a muscle to be, more or less, irrelevant compared to the effect of the
force–velocity relation.
Muscular damping can be viewed as a decentralized stabilization mechanism of the
natural muscular actuation system. Additionally, unlike force, damping is a non-directional
property. Forces can be offset by antagonistic muscles, whereas antagonistic damping
always adds up. Therefore, muscular co-contraction can be employed to increase damping,
irrespective of the force that serves as a lower bound for damping.
With the increment of damping of the contractile element of a muscle, the stiffness of
its attached tendon also increases to dominate the overall muscular behavior. While zero
damping results in no spring-like behavior at all, infinite damping, i.e., no motion of the
contractile element, leads to the muscle becoming its tendon. The observable end-point
stiffness of a limb can be stiffened using muscular co-contraction while its general shape
mostly stays related to the limb posture as already mentioned above [43]. Muscular co-
contraction, i.e., damping, can be used to stiffen up a limb, to increase tracking precision and
resistance to external disturbances. Unfortunately, increasing damping comes at the cost of
higher energy consumption. Various studies on disturbed natural motions support this
theory by observing a rise of muscular co-contraction to counteract unexpected disturbances
or to increase accuracy [47–49].
With the introduction of damping as the only secondary control property next to force,
an overall consistent concept of artificial, muscular control of a series elastic actuator is
defined. One S EA represents a pair of two antagonistic muscles. The two independent
muscular activations are represented by the two individual controls offered by a single
S EA: force and damping.
D Max D
F ( a1 , a2 ) ∼ a1 FMax − a2 FMax D ( a1 , a2 ) ∼ a1 + a2 Max (2)
2 2
Single muscle activation is mapped to positive or negative force generation F ( a1 , a2 ).
Simultaneous activation of both antagonists, i.e., co-contraction, is mapped to a generated
damping characteristic D ( a1 , a2 ). With this simple control principle, the required control
logic (including feedback loops) of an electrical S EA drastically reduces compared to a
complete impedance controller, e.g., [31,50]. Figure 3 sketches the muscular control scheme
from digital input commands to the physical current output of the example from 2nd
generation RRL AB-S EA hardware.
Sensors 2023, 23, 2428 7 of 25

Proximal F Series F Brushless τ Ball F Distal

Load Elasticity DC Motor Gear Load
I A,B,C
Rotation
Motor Prot Position
Driver Encoder
Physical Hardware Components

Custom Compute Unit

I Prot
+ F
F ∑ Gear & Limit Filter

Control −
Commands P̃rot
ṽ ṽrot
D ∏ Reverse Gear d/dt

Figure 3. Sketch of the muscular control of an electrical series elastic actuator. On the top, the
mechanical construction of the 2nd generation RRL AB-S EA is pictured. Attached below, a block
diagram of the introduced muscular control concept is given. The two control inputs, force and
damping, are highlighted in the lower left corner.

3. Distributed Damping Control

In the previous section, muscular control of a series elastic actuator was introduced. In
this section, an appropriate control and coordination strategy is developed. The proposed
S EA control offers force and damping generation (independent of each other). Hence, an
overall actuator coordination strategy can also be split into individual force and damping
distribution strategies. In the context of redundant actuation, force distribution strategies
were discussed in detail by Nejadfard et al. [41,51,52]. Damping distribution, on the other
hand, is mentioned in a quite primitive manner within the original conference publication
underlying the article at hand [1]. Hence, in the following, this section provides a de-
tailed analysis of the individual contribution of local muscular damping in the (Cartesian)
workspace. Afterward, a new recruitment strategy is introduced to distribute the desired
workspace damping among available muscles.

3.1. Workspace, Configuration Space, and Actuation Space

In order to control a robot in an easy and intuitive manner, commands might be given
by a desired TCP (Tool Center Point, i.e., an arbitrary, maybe virtual, point at the end of the
robot’s kinematic chain. In case of controlling a lower limb, this could be, for example, the
ankle joint, the heel position, or the tip of the toes) position and motion in the Cartesian
workspace representation of the real world. The exact robot motion, however, is executed
in the so-called configuration space, representing all individual DoFs that are the joints of
the robot. As it is not always the case that every joint is mapped to one single actuator,
the desired joint motion has to be mapped further to the available actuators, the so-called
actuation space. Unfortunately, the dimensionalities of all three spaces might increase in
the described order, i.e., several motors might operate on the same joint(s) and several
joints might operate in the same workspace dimension(s). For a better overview, Figure 4
sketches the kinematic coupling of twist and wrench vectors between these three spaces,
together with the respective embedded damping.
Sensors 2023, 23, 2428 8 of 25

(Cartesian) Configuration Actuation

≤ ≤
Workspace (Joint) Space (Muscle) Space

Twist    
ṖW JW θ̇ J AM L̇ M
[Vector]

Desired Generated
Damping    T       T     
DW JW DW JW = D J = A M DM AM DM
[Matrix]

Wrench T T
FW
 
JW τJ AM FM
[Vector]

Figure 4. Kinematic coupling of the (Cartesian) workspace, the configuration (joint) space, and the
actuation (muscle) space of an arbitrary, robotic limb. The two matrices, [ JW ] and [ A M ], represent the
 
position Jacobian and the actuator gear matrices, respectively. All three matrices in the middle DW/J/M
represent the appropriately dimensioned, symmetric, positive semidefinite damping matrices. In
blue, the desired and generated damping representations are highlighted.

From left to right, the abstract but intuitive (Cartesian) workspace, the configuration
space, and the actuation space are shown. The workspace on the left is the most simple
space that offers an intuitive API for abstract, higher-level limb control. In the middle, the
configuration space represents the physical limb configuration by all its joints. On the right,
the actuation space contains all actuators available at a limb. Note that the dimensionalities
of these three spaces might increase from left to right. Less-or-equal symbols between
their headlines indicate dimensional inequalities. Six description vectors, two per space,
are visualized as labeled circles. The three in the top row represent combined linear and
rotational velocity descriptions (twist vectors), while the ones on the bottom row represent
combined force and torque descriptions (wrench vectors). In between these six state vectors,
transformation matrices describe their relations.
  The two matrices thatdescribe
 the coupling between the spaces are the Jacobian matrix
JW and the gear matrix A M . These matrices do not need to be invertible due to the
different dimensionalities of the spaces. In general, they do not even need to be quadratic.
When using linear actuators as motor units, the gear matrix contains the lever arms between
the joints and the actuators. The lever arms of the two actuators mounted on the shank of
C ARL are sketched in Figure 1b. Further details about the lever arms of all actuators on
C ARL can be found in [52].
The relationship between twist and wrench vectors in all three spaces can be described
with a symmetric, positive semidefinite matrix representing non-negative   damping.
  In

Figure 4, these damping matrices areshown  in the middle row by DW , D J and DM .
The actuation space damping matrix D M is a diagonal matrix containing the individual,
non-negative actuator damping values d~M on its diagonal: D M = diag(d~M ). These
 

damping values are passed to the muscular S EA control, introduced in Section 2.2 and
visualized in Figure 3. Highlighted in blue, the relations between all three damping matrices
are shown in the following equation:

JW T DW JW = DJ = AM T DM AM (3)

Since none of the matrices within this equation and due to the above-mentioned
properties of the damping matrices, this equation, in general, is not solvable for DM ,
given an arbitrary, desired DW or DJ that meets the above requirements. Furthermore,
Sensors 2023, 23, 2428 9 of 25

the ’generatable’ damping values per actuator are constrained by physical and electrical
limits that must not be exceeded. Hence, an approximation strategy has to be applied to
distribute the desired damping to the available actuators.
A simple approach of optimizing within the configuration space has already been
introduced in the preliminary work [1] and will be presented in a more detailed and slightly
advanced version later in Section 3.3. However, the second potential approach to optimize
the actuator damping distribution within the workspace instead of the configuration space
has not been explored in previous studies. A brief discussion of this approach and an
explanation of why it is generally not feasible are presented below.

3.2. Discussion on Individual Muscular Damping in the Cartesian Workspace

To better highlight the upcoming problems when dealing with individual actuators in
the workspace, the above equation of damping matrices (3) can strongly be simplified.
Due to the fact that the actuator space damping matrix DM is diagonal, the right-hand
side of Equation (3) can be reformulated as a sum of sub-multiples of each actuator’s
individual maximum damping:
 
AM T diag(d~M ) AM = ∑ rowm (AM ) T · rowm (AM ) Max dm sm (4)
m

with the function rowm (·) returning the m-th row of its matrix argument · as a row vector
and Max dm sm representing the m-th element of the actuators damping vector d~M . Max dm is
the constant, maximum possible damping of the m-th actuator and sm ∈ [0, 1] represents
some fractional part of it.
Similarly, the left side of Equation (3) might be split into an equivalent sum of elements
representing each actuator’s contribution in the workspace.
 
JW T DW JW = ∑ JW T Max Dm JW sm (5)
m

This way, it is possible to observe the matrix equation for every available actuator
individually and unscaled:

!
∀m : JW T Max
Dm JW sm = rowm (AM ) T · rowm (AM ) Max
dm sm
!
⇒ ∀m : JW T Max
Dm JW = rowm (AM ) T · rowm (AM ) Max
dm (6)

Moreover, every actuator on its own is a one-dimensional object. Independent of

how it is integrated and the chosen level of abstraction, the resulting action has to be
one-dimensional. Hence, actuator-specific damping matrices in all three spaces only
need to have a single eigenvector with a non-zero eigenvalue. On the right-hand side of
Equation (6), this property is given by construction. On the left-hand side, this property
requires a simplified but more restrictive representation of the maximum damping matrix
in the workspace Max Dm := ~em ·~em
T . Using these representations, the matrices on both sides

of Equation (6) are characterized in total by their single, non-zero eigenvectors. Therefore,
the matrix equation can be simplified further to the following vector equation:

!
JW T ~em = rowm (AM ) T
p
Max d (7)
m

Now, it is easy to see that, as stated initially, the existence and uniqueness of a solution
on ~em in total depends on the Jacobian matrix JW .
The first case of no solution is the well-known problem of singular positions. Whenever
the Jacobian matrix JW T becomes singular, solving Equation (7) results in no and/or infinite
solutions. Figure 5a visualizes the problem of singularity at the example of the hip joint
actuator of a planar, two-link robot without TCP orientation. As worked out in detail by
Nejadfard et al. in 2018 [52], the direction of work of a hip-propelling actuator is parallel
Sensors 2023, 23, 2428 10 of 25

to the shank segment of the robot. In Figure 5a, these directions of work are pictured by
the blue and the dashed arrows, and the red bidirectional force arrows represent the two
sketched robot configurations. This alignment results from the requirement that these
forces have to exactly pass through all joints that should not be affected by the hip, i.e.,
all others. The distance between these vectors and the hip joint embodies the effective
lever arm coupling the joint torque and the force in the workspace. As highlighted in the
sketch, this lever arm vanishes in the singular position, leading to a theoretically infinite
resulting force vector in the workspace. Hence, in a singular position, the effective direction
of actuation in the workspace is undefined and cannot be used for further processing.

(Cartesian) Workspace (Cartesian) Workspace

Hip Hip Lever

Lever Lever Lever
Vanished

Parasitic
Knee Knee
Lever
Knee

kτ k
k Fk = k Lever k

Ankle
Parasitic
Lever

kτ k
k Fk = 0 =∞

(a) Singular Jacobian (b) Non-quadratic Jacobian

Figure 5. Visualization of upcoming problems of actuator-individual damping in the workspace.

Figure (a) highlights the problem of singularity in the example of a planar two-link robot. The
problem of having a higher-dimensional configuration space than the Cartesian workspace (i.e., a
non-square Jacobian) is sketched in figure (b) with the example of a planar, three-link robot.

The second reason that an actuator-individual workspace solution does not exist
is when the Jacobian matrix is non-square. This occurs when the configuration space
of a robot has a higher dimensionality than the workspace. To visualize this problem,
Figure 5b extends the planar, two-link robot of the previous Figure 5b by an ankle joint
with a foot segment. The orientation of the new TCP is still not considered. Now, there are
two main possibilities to orient the direction of work of the hip in the workspace. Either it
can pass through the knee joint as in the two-link robot, or it can pass through the appended
ankle. Figure 5b sketches both these variants in blue and red, respectively. It is easy to see
that both variants, when applied, cause parasitic torques on the joints that they do not pass
through. Hence, to really achieve such a vector of operation in the workspace, the parasitic
joint has to be activated, too. Unfortunately, this additional joint itself comes with the exact
same problem for its own direction of work which, in general, does not even match the
desired one of the original joint.
Hence, the above discussion can be concluded with the following statement. In general,
individual actuator (muscle) recruitment does not have proper or unique representation in
the workspace for arbitrary robot configurations. Obviously, this statement does not hold
for some special cases, e.g., in the case of a half-singular Jacobian. Consequently, actuator
damping distribution strategies can only be applied in the workspace when the respective
Jacobian matrix is guaranteed to be invertible. If this guarantee cannot be given by the
Sensors 2023, 23, 2428 11 of 25

robot construction, an alternative strategy has to be applied in the configuration space.

Such a suitable approximation strategy is introduced in the following Section 3.3.

3.3. Approximation of Muscular Damping in the Configuration Space

Since damping is represented as a quadratic, symmetric, positive semidefinite matrix,
the theoretical degree of freedom of a damping matrix is the number of its upper or
lower triangular elements. The requirement to only allow non-negative eigenvalues only
limits the space of valid solutions within the entire space of available DoF. Hence, to
exactly represent a configuration space damping in the actuation space 1/2 d(d + 1), many
independent actuators are required, with d being the dimensionality of the configuration
space. For the very simple, planar, three-joint robot within Figure 5b, this would already
require 6 individual actuators. It can be stated that, in general, every real robot will be
underactuated with respect to actuation space damping. An optimization approach to
find the best possible actuator damping commands is required. One solution is to find a
suitable damping matrix DS within the configuration space that minimizes the weighted
error against the desired damping matrix DJ in all possible directions.
!
∀θ̇ J ∈ {kθ̇ J k = 1} : DS θ̇ J = DJ θ̇ J (8)

This can be formulated as a quadratic minimization of the integrated, weighted,

squared error. Define:

tr (·) being the trace of matrix ·

rowi (·) being the i-th row of matrix ·


row(·) := row0 (·), . . . , rowi (·) being the concatenation of all rows

The minimization can now be derived as follows:

Z
2
min W DS − DJ θ̇ J (9)
DS kθ̇ J k=1


T


⇒ min tr W DS − DJ W DS − DJ (10)
DS
  

⇒ min tr (WDS ) T (WDS ) − 2 tr (WDS ) T WD J (11)
DS
1
T
⇒ min row(WDS ) · row(WDS ) T − row(WDS ) · row WDJ (12)
DS 2

With the help of the Equations (3) and (4), the weighted, suitable matrix DS can be
expressed as a linear combination of all individual damper actuators.
  
row(WDS ) = row W AM T DM AM
!
 
= row W ∑ T
rowm (AM ) · rowm (AM ) dm
m
!
 
= row ∑ W · rowm (AM ) T · rowm (AM ) dm
m
  T 
= ∑ row rowm AM W T · rowm (AM ) dm
m
 
T 
row row1 AM W T · row1 (AM )
..
 
= d~M (13)
 
  . 

T

T
row rowm AM W · rowm (AM )
=:L
Sensors 2023, 23, 2428 12 of 25

Now, the minimization problem can be reformulated to operate directly on the indi-
vidual actuator damping.

1 ~T
T
min d M L L T d~M − d~TM L row W DJ (14)
d~M 2
w.r.t. ~0 ≤ d~M ≤ Max ~
dM

The Hessian matrix L L T is only assured as positive semidefinite. With the side

condition on the actuator damping vector d~M , it is assured to only allow non-negative
damping commands for each individual actuator. In this way, the generated damping
matrix in the configuration space DS is, by definition, positive semidefinite.
Since the purpose of this control optimization is to deliver a continuous stream of
suitable motor control data, it is a better choice to optimize the control change than the
actual control values. This way, the optimization result will mostly stay around zero,
offering better numerical stability. Furthermore, by adding a diagonal e-matrix to the
hessian, the minimization can be guaranteed to be strictly positive definite. In the case of
optimizing the control change, this added diagonal matrix acts as a very small exponential
(low-pass) filter on the final control commands. The reformulation of Equation (14) to
minimize
 control change ~∆ M is straightforward. The control vector d~M is replaced by
the 
~M + ~∆ M , with t−1 d~M being the previous control vector.
t −1 d

1~ T   
t −1 ~

T 
min ∆ M L L T + diag(e) ~∆ M + ~∆ TM L L T d M − L row W DJ (15)
~∆ M 2
w.r.t. − t−1 d~M ≤ ~∆ M ≤ d M − t−1 d~M
Max ~

t~
d M = t−1 d~M + ~∆ M (16)

With respect to control stability, it is always safer to over-damp a system than to

under-damp it. In order to favor higher damping values over small ones, an interesting
choice for the weighting of the optimization is the desired configuration space damping
itself W = DJ . The effect of choosing such a weight is discussed in the following section
and illustrated in Figures 6–8.

4. Discussion on Capabilities and Limits of Actuator-Damped Systems

Three sets of theoretical experiments were executed to evaluate the capabilities and
limits of the herein-introduced new control philosophy. The ankle of one leg of the planar,
bipedal robot C ARL, reduced to its two leg segments, thigh and shank, was moved in
an ellipsoidal trajectory. Both segments have identical, natural lengths of 42 cm. The hip
and knee joints are limited in their range of motion from each 0 ◦ to −120 ◦ and −90 ◦ ,
respectively. Detailed descriptions of the construction of C ARL can be found in [29,36].
During the ellipsoidal motions, the introduced distribution strategy is applied to approxi-
mate four individual, one-dimensional unit-damping vectors in the workspace. The four
directions are 0 ◦ , −65 ◦ , −90 ◦ , and −115 ◦ . They are visualized as arrows attached to the
robot’s TCP in Figure 6a. The damping matrix in the workspace is the outer product of
the damping vectors d~W with themselves. Equation (3) is used to translate the workspace
damping matrix DW into the configuration space damping matrix DJ that is required by
the damping distribution algorithm.
Sensors 2023, 23, 2428 13 of 25

DJ = JW T DW JW (17)
= JW d~W · d~W
T T
JW
= d~J · d~T
J (18)
with: d~J : = JW T d~W

By reducing C ARL to a two-link configuration, we can design the simplest possible

robot that can feature both mono- and multiarticular actuators. This reduction is necessary
to limit the resulting data amount and complexity of the algorithm. Accordingly, the four
directions of workspace damping are specifically chosen to highlight certain properties
of the applied damping distribution algorithm. The horizontal and vertical damping
directions represent the primary axes required for upright (bipedal) balance and gravity
resistance. Moreover, the horizontal damping direction emphasizes the advantages of
properly chosen biarticular actuation. The vertical damping direction also underscores the
algorithm’s resistance against singularity, which often causes numerical instabilities. The
damping generation at the angle of −115 ◦ is chosen to show the effect of being aligned with
the shank. In the case of this two-segment robot, this alignment represents the principle
direction of work of a monoarticular actuator acting on the hip. The problem of singularity
and the principle direction of work are discussed in detail in Section 3.2 and sketched in
Figure 5a. The damping direction of −65 ◦ highlights the big differences between equally
forward and backward (−115 ◦ )-pointing desired workspace damping arising from the
one-directional knee joint. All results are explained in detail in the following.
The motion starts at the topmost position, moving forward first. In the motion path
sketch within Figure 6b, this is a clockwise execution. After half of the motion, the robot
passes its singular position with a backward motion direction. In order to keep the influence
of the changing lever arm low, the motion is kept small with a horizontal diagonal of 20 cm
and a vertical height of 10 cm. The experiment is executed using three different configura-
tions, which highlight important properties of the distributed damping coordination, given
in Equation (15). First, the optimization is executed without using an optimization weight
matrix: W = 1. Second, the desired damping vector in the configuration space d~J itself
is taken as additional weight, i.e., W = (d~J ·d~TJ )/(d~TJ ·d~J ). Moreover, the optimization is again
executed without weighting, but the lever arm ratio of the biarticular hip–knee actuator is
forced to always be 2-to-1 (hip-to-knee) as proposed by Nejadfard et al. in [52]. Figure 6
shows the experimental setup, the definitions, the execution, and the optimization results in
the configuration space. Three different optimization results at three concrete Cartesian po-
sitions were plotted for the desired Cartesian damping direction of −115◦ . Figures 7–9 plot
all optimized actuator damping values for all four desired Cartesian directions, respective
to the three above-mentioned optimization configurations.
As discussed in Section 3.2, and sketched in Figure 5a, the direction of work of the
hip joint always aligns with the shank of the two-segment robot. This phenomenon is
clearly visible in Figure 6b in the back position. The desired joint damping is a horizontal
line. Hence, in this situation, the hip is the only required actuator to perfectly generate
the desired Cartesian damping. In the two other positions, there is a clear vertical (knee)
component present. Note that the singular position in the middle of the motion does not
cause any numerical artifacts or instabilities. This can also be observed in all of the detailed
result plots.
Except for the specially constructed back position, the proposed damping distribution
strategy in general is not able to exactly match the desired damping. (Note that the blue
ellipse of the back position is covered by the green one and, hence, is not visible in the
figure.) It can be seen that both the blue and the green ellipses are optimized to be close
to the desired one. The errors are ellipses with slightly different radii and orientations.
This inability is due to the fact that the generated damping cannot be different from
a non-negative, linear combination of the actuator-individual dampings as explained
in Equation (4). By definition, the direction of work of monoarticular actuators in the
Sensors 2023, 23, 2428 14 of 25

configuration space is fixed. Therefore, the lever arm ratio of biarticular (or multiarticular)
actuators is the only possibility to tune the capability of a robot to render the concrete
damping at the actuator level. Note that this is a mechanical design decision. It is clearly
visible that the green ellipses, generated using the improved biarticular lever arm ratio of
2-to-1 for the biarticular hip–knee actuator, are much closer to the desired damping than
the blue ones that are using the original, existing levers of C ARL. Details about the exact
lever arms of C ARL can be found in [29,52].

Y (Cartesian) Configuration

Knee
Workspace (Joint) Space
Hip Hip
−θ Thigh
−θCombined

Knee

−θShank

Ankle

−115◦

Desired Workspace
Damping Directions

θ X Hip

(a) Kinematic definitions of the reduced, (b) Examples of the desired and generated
two-segment C ARL in Cartesian coordinates damping ellipses in the configuration space

Figure 6. Theoretical experiment setup and result examples of the reduced, two-segment C ARL robot.
Figure (a) pictures the kinematic definitions of the robot in Cartesian coordinates. At the ankle, the
four desired damping directions that are approximated individually throughout the experimental
ellipsoidal motion are shown. The motion itself is visualized in the background in Figure (b). In the
foreground, Figure (b) shows the desired (solid red double-sided arrows) and generated (dashed,
colored) damping ellipses in the configuration space (hip/knee) at three different robot configurations.
The desired damping is calculated from the Cartesian damping direction of −115◦ that is highlighted
in red within the Cartesian definitions. The blue and orange ellipses are related to the optimizations
without weighting and with weighting by the desired damping itself, respectively. In green, the result
ellipse is shown, which emerges from the optimized, biarticular lever arm ratio without any further
weighting. Muscular damping results over the whole elliptical motion for all three optimization
configurations are plotted in Figures 7–9.

Comparing the weighted orange damping ellipses against the others, it is not difficult
to see that weighting highly influences the optimization results. In the case of the un-
weighted experimental optimization runs, the algorithm balances both the length and the
width of the resulting ellipse against the desired one. Consequently, the emerging damping
is a little shorter but thicker than the desired one. Against these observations, weighting
the optimization with the damping itself results in an optimization that only is interested
in matching this single length of the desired damping vector. Everything else is ignored.
This effect is clearly visible in Figure 6b. The orange ellipses in all three positions perfectly
match the tips of the desired, double-sided vectors shown in red. The desired thickness
(zero) is not taken into account at all. Hence, the width of these orange ellipses emerges
from the mostly unchanged previous one. However, having one dimension weighted to
zero is an extreme scenario that most likely will not occur in real motion scenarios.
Sensors 2023, 23, 2428 15 of 25

−60 Segment Angles

Angle [degree]
−80 Thigh
Shank
−100 Combined
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Horizontal Damping
Damping [N/(m/s)]

400
300 Hip
200 Knee
Hip-Knee
100
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−65◦ Damping
Damping [N/(m/s)]

100

Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

40 Vertical Damping
Damping [N/(m/s)]

30
Hip
20 Knee
Hip-Knee
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−115◦ Damping
Damping [N/(m/s)]

100
Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Relative progress throughout one complete elliptical motion cycle

Figure 7. Muscular damping results for all four Cartesian damping directions optimized without
special weighting: W = 1.On top, the angles of the leg segments as defined in Figure 6a are
plotted together with a dashed, horizontal line highlighting the −115◦ of the last desired damping
direction. The three positions that are visualized in Figure 6b are marked with dashed, vertical
lines throughout all five plots. From left to right, these are the front, singular, and back positions.
The lower four plots show the damping recruitment of all three actuators independently for the
four independent directions of the desired workspace dampings, sketched in Figure 6a. The two-
dimensional configuration space representation of the three marked positions for the −115◦ Cartesian
damping directions are visualized in Figure 6b by the dashed blue ellipses.
Sensors 2023, 23, 2428 16 of 25

−60 Segment Angles

Angle [degree]
−80 Thigh
Shank
−100 Combined
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Horizontal Damping
Damping [N/(m/s)]

400
300 Hip
200 Knee
Hip-Knee
100
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−65◦ Damping
Damping [N/(m/s)]

100

Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

40 Vertical Damping
Damping [N/(m/s)]

30
Hip
20 Knee
Hip-Knee
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−115◦ Damping
Damping [N/(m/s)]

100
Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Relative progress throughout one complete elliptical motion cycle

Figure 8. Muscular damping results for all four Cartesian damping directions taking the desired
damping vector d~J itself as optimization weight, W = (d~J ·d~TJ )/(d~TJ ·d~J ). On top, the angles of the leg
segments as defined in Figure 6a and are plotted together with a dashed, horizontal line highlighting
the −115◦ of the last desired damping direction. The three positions that are visualized in Figure 6b
are marked with dashed, vertical lines throughout all five plots. From left to right, these are the front,
singular, and back positions. The lower four plots show the result damping recruitment of all three
actuators independently for the four independent directions of the desired workspace dampings,
sketched in Figure 6a. The two-dimensional configuration space representation of the three marked
positions for the −115◦ Cartesian damping directions are visualized in Figure 6b by the dashed
orange ellipses.
Sensors 2023, 23, 2428 17 of 25

−60 Segment Angles

Angle [degree]
−80 Thigh
Shank
−100 Combined
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Horizontal Damping
Damping [N/(m/s)]

400
300 Hip
200 Knee
Hip-Knee
100
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−65◦ Damping
Damping [N/(m/s)]

100

Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

40 Vertical Damping
Damping [N/(m/s)]

30
Hip
20 Knee
Hip-Knee
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

−115◦ Damping
Damping [N/(m/s)]

100
Hip
50 Knee
Hip-Knee

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Relative progress throughout one complete elliptical motion cycle

Figure 9. Muscular damping results for all four Cartesian damping directions without special
weighting, but with the lever arm ratio of the biarticular hip–knee actuator forced to be 2-to-1 (hip-
to-knee) as proposed by Nejadfard et al. [52]. On top, the angles of the leg segments as defined in
Figure 6a are plotted together with a dashed, horizontal line highlighting the −115◦ of the last desired
damping direction. The three positions that are visualized in Figure 6b are marked with dashed,
vertical lines throughout all five plots. From left to right, these are the front, singular, and back
positions. The lower four plots show the damping recruitment of all three actuators independently
for the four independent directions of the desired workspace dampings, sketched in Figure 6a. The
two-dimensional configuration space representation of the three marked positions for the −115◦
Cartesian damping directions are visualized in Figure 6b by the dashed green ellipses.
Sensors 2023, 23, 2428 18 of 25

From the technical description of C ARL in [29], the moment arm ratio of the biarticular
hip–knee actuator can be estimated notably below 1 throughout the whole elliptical motion.
The detailed analysis of this ratio by Nejadfard et al. [52] gives insight that the direction of
work of a biarticular actuator at the hip–knee position with such a small lever arm ratio
is quite close to the knee’s one. Consequently, when looking through the result plots in
Figures 7 and 8 using the original lever arms of C ARL, the knee actuator is mostly unused.
Except for the special case of the vertical desired damping, the biarticular hip–knee actuator
takes over all responsibilities of the knee. A second exception is the desired damping at
−115◦ with damping-weighted optimization (see Figure 8). In this case, it is just the other
way around. The monoarticular knee is recruited, while the biarticular hip–knee actuator is
more or less unused. However, this can be treated as a side effect of the massively deformed
generated damping ellipse due to the zero-weighting, as described above.
Comparing both unweighted experiments (Figures 7 and 9), the effect of the changed
lever arm ratio of the hip–knee actuator can be observed. The most obvious effect takes
place at horizontal damping. Since the lever arm ratio of 2-to-1 works orthogonally to
the hip–ankle axis (the herein-named combined angle) [52], the horizontal damping is
most solely taken over by the hip–knee actuator. Inspecting the muscle recruitment for
the slightly forward pointing damping (at −65◦ ), the updated lever arm ratio leads to a
recruitment of the knee actuator instead of the hip one, used with the original lever arms.
Keeping in mind that the knee actuator acts at the so-called combined angle while the hip
actuator works at the angle of the shank, it is obvious that recruiting the knee is a much
better choice than using the hip actuator for approximating −65◦ damping. Unfortunately,
in the first experiment (Figure 7), the knee is quite useless due to the working direction
of the hip–knee actuator. The only leftover option for the optimization is to use the hip
in order to push the resulting damping ellipse toward the desired direction, at the cost of
a much higher deviation from the desired. In the case of pure vertical damping, there is
no relevant change at all. The monoarticular knee actuator remains the optimal choice for
that direction. The same holds for the slightly backward-pointing damping. Whenever
the shank is aligned with the desired damping at −115◦ , the hip actuator takes over that
generation solely. The additional knee recruitment in Figure 8 is due to the zero weighting
that does not care about any side effects.
At this point, it is not difficult to formulate two main results from these observations:
• The capabilities and limits of a distributed damping strategy at the actuator level are
only determined by the mechanical design of the actuator mounts within a robot.
• The higher the variability in the set of available actuators, the higher the possible
precision of an approximated distributed damping.
These findings correlate with Schumacher et al.’s conclusion in their review article,
i.e., biarticular muscles in light of template models, experiments and robotics: a review, from 2020:
“Thus, the muscle’s mechanical function is strongly influenced by the leg architecture.
This coupling also affects the neural coordination of the muscles.” [53]
Note that the latter observation result also covers the number of available actuators. Hence,
the herein-introduced concept of the distributed damping control at the actuator level offers
a novel sense of the existence of muscular variety and the massive over-actuation, present
in nature. In addition to the known energy reduction opportunity on torque generation,
multiarticular muscles offer the possibility to better generate concrete, desired damping. In
this way, unnecessary overdamping in irrelevant directions can be avoided, contributing to
less energy-consuming control.

5. Actuator-Damping Applied on the Planar Robotic Leg C ARL

With controlling damping, the precision of a desired motion (velocity or twist) can be
controlled. This way, the damping acts as the proportional gain of a velocity P-controller.
However, due to the over-actuation of C ARL, it is better to reformulate the controller to use
a desired force instead of a desired velocity along with the damping.
Sensors 2023, 23, 2428 19 of 25

 
~F = D0 V~0 − V
~ = D0 V
~ 0 − D0 V
~ =: ~F0 − D0 V
~ (19)

The main reasons for going with a desired force are as follows. As shown in Figure 4,
the gear matrix AM couples the configuration space and the actuation space. On the
velocity side, the calculation of all actuator velocities, given the desired joint velocities, is
straightforward. The coupling between joint and actuator motion is mechanically defined,
so increasing the dimensionality of the configuration space with respect to the actuation
space does not generate additional information. However, when it comes to forces, the
coupling goes in the opposite direction, so AM describes a dimensional reduction that may
result in information loss. This loss presents the opportunity to choose from various sets
of actuator forces without affecting the torques in the configuration space. Different force
distribution strategies can be applied to optimize the control effort, energy loss, or other
factors. Additionally, passive actuators (e.g., springs) can easily be integrated as additional
sources of force or torque. Having passive springs in parallel to active actuators can be used
to shift an actuator’s working point out of the zero force. In nature, this effect is achieved
by having antagonistic muscles of different sizes and strengths. Two examples of possible
distribution strategies were introduced and discussed by Nejadfard et al. in [51].
Since the generation of damping is only an approximation, proper actuator forces ~F0
can only be calculated once the actually applied damper actuators are known. Therefore,
within the configuration space, the desired control, again, has to be the velocity instead of
the force. Given the optimized, diagonal actuator damping matrix DM and some arbitrary,
˙
desired joint velocities vector ~θ J , appropriate actuator forces are calculated by solving the
following system of equations for ~FM :
˙
AM T DM AM ~θ J = ~τJ = AM T ~FM (20)

Using damping and force, a very simple, so-called compliant trajectory control can be
set up, i.e., the equilibrium position of a virtual spring–damper element is moved along a
desired trajectory. The actual object to be moved is attached to that virtual spring–damper
element. A sketch showing the principle of such a compliant trajectory, including the virtual
spring–damper element, is given in Figure 10a. Due to the usage of velocity instead of force,
as explained above, the stiffness of the virtual spring has to be split into a velocity-stiffness
followed by the damping
K := DW Kv (21)
with DW being the desired Cartesian workspace damping. The control velocity ~vW is then
calculated from the position delta ~∆W in the Cartesian workspace using the velocity-stiffness
matrix Kv
~vW = Kv ~∆W (22)
To keep the control stable, Kv should be defined to at least achieve critical damping.
Under the simplification of considering the whole system as a simple, damped spring-mass
system with mass m, Kv can be derived as follows:
√
2 K m = DW (23)
p 1
DW Kv m = DW
2
1
DW Kv m = DW 2
4
1
Kv = DW (24)
4m
Note that this definition requires the exact desired Cartesian damping DW . Unfor-
tunately, the damping is known not to be generated precisely by the actuators. Hence,
the estimated mass m should be a little overestimated in order to keep the stiffness small
enough for proper damping values.
Sensors 2023, 23, 2428 20 of 25

To evaluate the functionality of the herein-introduced distributed damping strategy,

two sets of differently damped, circular, compliant trajectories have been executed at the
ankle of C ARL [1]. Kinematic details of C ARL are introduced at the beginning of Section 4. A
sketch of the principle setup and the experimental outcomes are visualized in Figure 10. In
all experiments, a virtual, circular compliant trajectory with a radius of 5 cm was executed
about 63 cm straight below the hip at two different speeds, with five different desired
are: Blue: 50
 0
damping  50matrices DW . The individual, color-coded damping matrices 0 50 ,
0 , Orange: 250 0 , Green: 150 100 , and Black: 500 0 .
     
Purple: 0 250 0 50 100 150 0 500 h i
All damping matrices are given in the standard damping SI-unit, force per velocity mN/s .
The respective configuration space damping matrices DJ are calculated as described
by Equation (17) in Section 4.
The captured, real executed ankle trajectories at all five desired damping matrices
are shown for the two different trajectory execution speeds of 30 RPM and 120 RPM in
Figure 10b and Figure 10c, respectively. The ankle positions for all experiments were
recorded at a frequency of 1 kHz and were plotted without any post-processing. The
estimated mass parameter that was used for calculating the virtual stiffness, as described in
Equation (24), is commonly set to 15 kg. That is about half the weight of the whole bipedal
robot C ARL.

Y Y [m]
(Cartesian) 0
Workspace
Hip
−0.1

X [m]
−0.2 −0.1 0 0.1
(b) At 30 RPM
Virtual
0 Y [m]
Trajectory
0
Virtual
Impedance
−0.1
Real
Trajectory X [m]
θ 0 X −0.2 −0.1 0 0.1
(a) Compliant Trajectory (c) At 120 RPM

Figure 10. Sketch and experimental results of the so-called compliant trajectory control applied
to the ankle of the planar, compliant robotic leg C ARL as first presented in [1]. On the left, figure
(a) visualizes the working principle of the compliant trajectory control. In red, the commanded circular
reference trajectory is shown. The virtual spring–damper element and a potential real resulting ankle
trajectory are sketched in blue underneath. On the right-hand side, Figures (b,c) plot the captured,
real ankle trajectories of C ARL in workspace coordinates at different virtual impedances at speeds of
30 RPM and 120 RPM respectively. Centered at the origin, the shared circular trajectory is given in
red. The two-dimensional damping matrices of each of the five individual experimental runs per
speed are plotted as dashed ellipses of the same color as the appropriate resulting ankle trajectories.
The five damping matrices are: Blue: 50
 0  50 0   250 0   150 100 
 500 0  0 50 , Purple: 0 250 , Orange: 0 50 , Green: 100 150 , and
Black: 0 500 . All damping matrices are given in standard damping SI-units, i.e., force per velocity
h i
N
m/s . Due to gravity, the real ankle trajectory is expected to be below the virtual reference one.

Both figures, Figures 10b,c show the effects of the different impedances on the real
ankle trajectories. As expected, due to gravity, the real trajectories are, in general, below the
virtual reference one that is pictured as a red circle. Moreover, it can be seen that increasing
Sensors 2023, 23, 2428 21 of 25

damping reduces the error in that direction. To do that, the blue trajectory (i.e., the lowest
one that refers to the smallest damping ellipse) is taken as a reference trajectory. Indepen-
dent of the execution speed, the purple, orange, and green experiments visualize the effects
of increased impedance in vertical, horizontal, and diagonal directions, respectively.
However, the purple trajectory is of special interest. It is clearly recorded that not only
the vertical direction is increased, as desired, but the horizontal error increases at the same
time. An explanation of this effect delivers a closer look at the physical leg configuration in
that situation. In order to move the ankle of the two-link system upwards, the knee joint
has to be bent forward. That, however, leads to the side effect of moving about half of the
mass of the leg forward, too. As a direct consequence, the gravity force that acts on the
leg’s shifted center of mass induces a backward torque of the whole leg at the hip. Due to
the absence of any gravity-compensation mechanism or similar, this change of the physical
properties of the leg hence directly influences the real executed trajectory.
The opposite behavior can be observed in the orange experiment. Unlike the purple
trajectory, the orange experiment shows an error reduction in the direction where the
desired impedance is not increased. This time, a proper explanation for that effect can be
found with a closer look at the damping distribution strategy, as discussed in detail in the
previous Section 4. By construction, the generation of horizontal damping always comes at
the cost of an inherent vertical one. Consequently, the desired reduction of the horizontal
error is achieved by the optimization strategy at the cost of additionally reducing the
vertical error, although not desired. Note that this effect is only visible due to the absence
of gravity compensation.
The last experiment, the black one, encodes a substantial increment of the desired
impedance uniformly in both directions. It was chosen high enough to partially push
the actuators and the links of C ARL to their physical limits. The flat upper part of the
black trajectory at 30 RPM occurs due to the knee joint limit of C ARL that occurs at −90◦ .
Aside from that, the desired effect of a much more precise motion is achieved clearly. The
recorded trajectory is very close to the commanded virtual one.
Aside from gravity, dynamic effects influence the executed real trajectory. Such dy-
namic effects can be observed by comparing the two plots of different velocities against
each other. First of all, it can be observed that the velocity does not have that much of an
effect on the general shape and location of the executed trajectory. The main difference is
the variety of motion. This effect is best visible in the blue, the most compliant trajectory. It
can be seen that the blue plot at 120 RPM comes at a much higher divergence within the
repetitive circular motion compared to the one at 30 RPM.
Moreover, the massive deformation in the most precise trajectory execution, the black
experiment, is noticeable at a high speed. An explanation for this observation can be given
by the robot’s segment configuration. As mentioned earlier, during the execution of the
black trajectory, the robot operates at the boundary of having a maximum possible bent
knee. Therefore, a lot of mass is widely extended to the front, which is out of the primary
leg axis from hip to ankle. Unfortunately, this creates a long lever for undesired dynamics to
act on the robot. On having a closer look at the orange and green trajectories in Figure 10c,
this effect, although much smaller, can already be observed on both of these trajectories.
Overall, within this paper, the distributed damping control was evaluated to behave
as expected. It is shown that it is possible to generate and control compliant motions in a
predictable manner.

6. Conclusions and Outlook

Previous work on the bio-inspired actuation system of C ARL already provides great
insight into the individual contribution of force-emitting muscles [51,52]. Every available
actuator, including monoarticular and multiarticular actuators, offers a specific direction
of work that can be recruited by any arbitrary motor coordination system. Obviously, the
more different types of actuators are present, the more opportunities for recruitment exist.
Hence, at the cost of control effort, over-actuation can be used to improve a robot’s overall
Sensors 2023, 23, 2428 22 of 25

energy efficiency. However, although energy reduction is an important feature, it does not
really contribute to a robot’s movement capabilities. Furthermore, the possibility of energy
reduction by over-actuation strongly depends on the actuator’s mounting positions, the
internal state of the limb configuration, and the requested control. Hence, energy efficiency
can be treated as not being the only reason for the existence of the massive over-actuation of
natural limbs. The capability of precisely generating distributed impedance at the muscular
level requires a variety of available actuators. The larger and more diverse the set of
available actuators, the better the desired damping can be approximated. However, when
compensating for missing damping capabilities by over-damping, it again breaks down to
energy reduction.
Analogous to the analysis of muscle force contributions, the contribution of muscular
damping can be analyzed. This way, it has been shown that the possible shape of a muscle-
generated, distributed workspace impedance mostly depends on the pose of a limb (and
the internal actuator mounting). As described in Section 2, identical observations on human
arms were made by Mussa-Ivaldi et al. in 1985 [43]. Following the same argumentation
of using over-actuation for force generation, every available actuator adds its own unique
capability to the overall control of a limb. Hence, the more muscular variety that is available,
the better and more accurate a desired damping can be generated at a muscular level.
The novel muscular control principle introduced within this article is widely backed
by reasoning against observed, natural characteristics. As a result, the overall observable
behavior of the proposed muscular control strategy comes with similar characteristics as can
be observed in nature. With theoretical experiments on the kinematic structure of C ARL, the
capabilities and the limits of the novel, distributed damping strategy have been highlighted.
Experimental results on the real bipedal robot C ARL have been executed to prove the novel
control strategy to be functional and stable [1]. Observed, undesired motion perturbations
were clearly traced back to the results as well as observations of the theoretical experiments.
To the best of our knowledge, no other legged robot has been equipped and controlled
with such a distributed, muscular damping control approach. Most similar to the concept
introduced here is the control approach proposed by Sharbafi et al. in [54]. Within that
work, the so-called BioBiped 3 robot was controlled by altering the mounting positions of
springs spanning across the robot’s joints.
One next step to evaluate the general capability of distributed damping is to bring this
control approach into more complex motions. Moreover, further improvements on the basic
functionality can be made by introducing gravity or dynamics compensation mechanisms.

Author Contributions: Conceptualization , P.V.; Software, P.V.; Validation, P.V.; Supervision, K.B. All
authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2023, 23, 2428 23 of 25

References
1. Vonwirth, P.; Berns, K. Bio-Inspired Imprecise Impedance Control of Muscle-Driven Robotic Limbs. In Proceedings of the Robotics
in Natural Settings, CLAWAR 2022; Cascalho, J.M., Tokhi, M.O., Silva, M.F., Mendes, A., Goher, K., Funk, M., Eds.; Lecture Notes
in Networks and Systems; Springer International Publishing: Cham, Switzerland, 2023; Volume 530, pp. 42–53. [CrossRef]
2. Pratt, J.; Carff, J.; Drakunov, S.; Goswami, A. Capture Point: A Step toward Humanoid Push Recovery. In Proceedings of the 2006
6th IEEE-RAS International Conference on Humanoid Robots, Genova, Italy, 4–6 December 2006; pp. 200–207. [CrossRef]
3. Englsberger, J.; Ott, C.; Albu-Schäffer, A. Three-Dimensional Bipedal Walking Control Using Divergent Component of Motion.
In Proceedings of the 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Tokyo, Japan, 3–7
November 2013; pp. 2600–2607.
4. Englsberger, J.; Ott, C.; Albu-Schäffer, A. Three-Dimensional Bipedal Walking Control Based on Divergent Component of Motion.
IEEE Trans. Robot. 2015, 31, 355–368. [CrossRef]
5. Hosokawa, M.; Nenchev, D.N.; Hamano, T. The DCM Generalized Inverse: Efficient Body-Wrench Distribution in Multi-Contact
Balance Control. Adv. Robot. 2018, 32, 778–792. [CrossRef]
6. Sreenath, K.; Park, H.W.; Poulakakis, I.; Grizzle, J.W. A Compliant Hybrid Zero Dynamics Controller for Stable, Efficient and Fast
Bipedal Walking on MABEL. Int. J. Robot. Res. 2011, 30, 1170–1193. [CrossRef]
7. Ames, A.D. Human-Inspired Control of Bipedal Walking Robots. IEEE Trans. Autom. Control 2014, 59, 1115–1130. [CrossRef]
8. Grizzle, J.W.; Chevallereau, C.; Sinnet, R.W.; Ames, A.D. Models, Feedback Control, and Open Problems of 3D Bipedal Robotic
Walking. Automatica 2014, 50, 1955–1988. [CrossRef]
9. Nguyen, Q.; Hereid, A.; Grizzle, J.W.; Ames, A.D.; Sreenath, K. 3D Dynamic Walking on Stepping Stones with Control Barrier
Functions. In Proceedings of the IEEE International Conference on Decision and Control (CDC), Las Vegas, Nevada, USA, 12–14
December 2016, Volume 1001.
10. Geyer, H.; Herr, H. A Muscle-Reflex Model That Encodes Principles of Legged Mechanics Produces Human Walking Dynamics
and Muscle Activities. IEEE Trans. Neural Syst. Rehabil. Eng. 2010, 18, 263–273. [CrossRef] [PubMed]
11. Song, S.; Geyer, H. A Neural Circuitry That Emphasizes Spinal Feedback Generates Diverse Behaviours of Human Locomotion.
J. Physiol. 2015, 593, 3493–3511. [CrossRef] [PubMed]
12. Batts, Z.; Song, S.; Geyer, H. Toward a Virtual Neuromuscular Control for Robust Walking in Bipedal Robots. In Proceedings of
the 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Hamburg, Germany, 28 September–2
October 2015; pp. 6318–6323. [CrossRef]
13. Geijtenbeek, T.; van de Panne, M.; van der Stappen, A.F. Flexible Muscle-based Locomotion for Bipedal Creatures. ACM Trans.
Graph. 2013, 32, 206:1–206:11. [CrossRef]
14. Li, Z.; Cheng, X.; Peng, X.B.; Abbeel, P.; Levine, S.; Berseth, G.; Sreenath, K. Reinforcement Learning for Robust Parameterized
Locomotion Control of Bipedal Robots. arXiv 2021, arXiv:2103.14295.
15. Luksch, T.; Berns, K. Control of Bipedal Walking Exploiting Postural Reflexes and Passive Dynamics. In Proceedings of the IEEE
International Conference on Applied Bionics and Biomechanics (ICABB), Venice, Italy, 14–16 October 2010.
16. Luksch, T. Human-like Control of Dynamically Walking Bipedal Robots; RRLab Dissertations; Verlag Dr. Hut: Munchen, Germany
2010.
17. Zhao, J.; Liu, Q.; Schütz, S.; Berns, K. A Bio-inspired Behavior Based Bipedal Locomotion Control—B4LC Method for Bipedal
Upslope Walking. In Proceedings of the 18th International Conference on Climbing and Walking Robots (CLAWAR), Hangzhou,
China, 6–9 September 2015.
18. Seok, S.; Wang, A.; Chuah, M.Y.; Otten, D.; Lang, J.; Kim, S. Design Principles for Highly Efficient Quadrupeds and Imple-
mentation on the MIT Cheetah Robot. In Proceedings of the 2013 IEEE International Conference on Robotics and Automation,
Karlsruhe, Germany, 6–10 May 2013; pp. 3307–3312. [CrossRef]
19. Seok, S.; Wang, A.; Chuah, M.Y.; Hyun, D.J.; Lee, J.; Otten, D.M.; Lang, J.H.; Kim, S. Design Principles for Energy-Efficient Legged
Locomotion and Implementation on the MIT Cheetah Robot. IEEE/ASME Trans. Mechatronics 2015, 20, 1117–1129. [CrossRef]
20. Semini, C.; Tsagarakis, N.G.; Guglielmino, E.; Focchi, M.; Cannella, F.; Caldwell, D.G. Design of HyQ—A Hydraulically and
Electrically Actuated Quadruped Robot. Proc. Inst. Mech. Eng. Part I J. Syst. Control Eng. 2011, 225, 831–849. [CrossRef]
21. Semini, C.; Barasuol, V.; Boaventura, T.; Frigerio, M.; Focchi, M.; Caldwell, D.G.; Buchli, J. Towards Versatile Legged Robots
through Active Impedance Control. Int. J. Robot. Res. 2015, 34, 1003–1020. [CrossRef]
22. Hutter, M.; Gehring, C.; Jud, D.; Lauber, A.; Bellicoso, C.D.; Tsounis, V.; Hwangbo, J.; Bodie, K.; Fankhauser, P.; Bloesch, M.; et al.
ANYmal—A Highly Mobile and Dynamic Quadrupedal Robot. In Proceedings of the 2016 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), Daejeon, Republic of Korea, 9–14 October 2016; pp. 38–44. [CrossRef]
23. Fankhauser, P.; Hutter, M. ANYmal: A Unique Quadruped Robot Conquering Harsh Environments. Res. Featur. 2018, 126, 54–57.
[CrossRef]
24. Bouman, A.; Ginting, M.F.; Alatur, N.; Palieri, M.; Fan, D.D.; Touma, T.; Pailevanian, T.; Kim, S.K.; Otsu, K.; Burdick, J.; et al.
Autonomous Spot: Long-Range Autonomous Exploration of Extreme Environments with Legged Locomotion. In Proceedings of
the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Las Vegas, NV, USA, 24 October 2020–24
January 2021; pp. 2518–2525. [CrossRef]
25. Ficht, G.; Behnke, S. Bipedal Humanoid Hardware Design: A Technology Review. Curr. Robot. Rep. 2021, 2, 201–210. [CrossRef]
Sensors 2023, 23, 2428 24 of 25

26. Best, C.M.; Wilson, J.P.; Killpack, M.D. Control of a Pneumatically Actuated, Fully Inflatable, Fabric-Based, Humanoid Robot. In
Proceedings of the 2015 IEEE-RAS 15th International Conference on Humanoid Robots (Humanoids), Seoul, Republic of Korea,
3–5 November 2015; pp. 1133–1140. [CrossRef]
27. Barasuol, V.; Villarreal-Maga na, O.A.; Sangiah, D.; Frigerio, M.; Baker, M.; Morgan, R.; Medrano-Cerda, G.A.; Caldwell, D.G.;
Semini, C. Highly-Integrated Hydraulic Smart Actuators and Smart Manifolds for High-Bandwidth Force Control. Front. Robot.
AI 2018, 5, 51. [CrossRef] [PubMed]
28. Seok, S.; Wang, A.; Otten, D.; Kim, S. Actuator Design for High Force Proprioceptive Control in Fast Legged Locomotion. In
Proceedings of the 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, Vilamoura-Algarve, Portugal,
7–12 October 2012; pp. 1970–1975. [CrossRef]
29. Schütz, S.; Nejadfard, A.; Mianowski, K.; Vonwirth, P.; Berns, K. CARL—A Compliant Robotic Leg Featuring Mono- and
Biarticular Actuation. In Proceedings of the IEEE-RAS International Conference on Humanoid Robots, Birmingham, UK, 15–17
November 2017.
30. Nejadfard, A.; Schütz, S.; Mianowski, K.; Vonwirth, P.; Berns, K. Design of the Musculoskeletal Leg CARL Based on the Physiology
of Mono-Articular and Biarticular Muscles in the Human Leg. Bioinspir. Biomimetics 2019, 14, 066002. [CrossRef]
31. Schütz, S.; Mianowski, K.; Kotting, C.; Nejadfard, A.; Reichardt, M.; Berns, K. RRLAB SEA—A Highly Integrated Compliant
Actuator with Minimised Reflected Inertia. In Proceedings of the 2016 IEEE International Conference on Advanced Intelligent
Mechatronics (AIM), Banff, AB, Canada, 12–15 July 2016; pp. 252–257.
32. Mohseni, O.; Rashty, A.M.N.; Seyfarth, A.; Hosoda, K.; Sharbafi, M.A. Bioinspired Legged Robot Design via Blended Physical
and Virtual Impedance Control. J. Intell. Robot. Syst. 2022, 105, 22. [CrossRef]
33. Mirvakili, S.M.; Hunter, I.W. Artificial Muscles: Mechanisms, Applications, and Challenges. Adv. Mater. 2018, 30, 1704407.
[CrossRef]
34. Liu, Q.; Zhao, J.; Schütz, S.; Berns, K. Adaptive Motor Patterns and Reflexes for Bipedal Locomotion on Rough Terrain. In
Proceedings of the 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Hamburg, Germany, 28
September–2 October 2015.
35. Liu, Q.; Zhao, J.; Berns, K. Dynamically Stepping Over Large Obstacle Utilizing PSO Optimization in the B4LC System. In
Proceedings of the Advances in Cooperative Robotics: Proceedings of the 19th International Conference on Clawar 2016, London,
UK, 12–14 September 2016; p. 428.
36. Schütz, S. CARL—A Compliant Robotic Leg Designed for Human-like Bipedal Locomotion. Ph.D. Thesis, University of
Kaiserslautern, Kaiserslautern, Germany, 2020.
37. Reher, J.; Cousineau, E.A.; Hereid, A.; Hubicki, C.M.; Ames, A.D. Realizing Dynamic and Efficient Bipedal Locomotion on
the Humanoid Robot DURUS. In Proceedings of the 2016 IEEE International Conference on Robotics and Automation (ICRA),
Stockholm, Sweden, 16–21 May 2016; pp. 1794–1801. [CrossRef]
38. Hereid, A.; Cousineau, E.A.; Hubicki, C.M.; Ames, A.D. 3D Dynamic Walking with Underactuated Humanoid Robots: A Direct
Collocation Framework for Optimizing Hybrid Zero Dynamics. In Proceedings of the IEEE International Conference on Robotics
and Automation (ICRA), Stockholm, Sweden, 16–21 May 2016; pp.1447–1454. [CrossRef]
39. Schütz, S.; Nejadfard, A.; Kötting, C.; Berns, K. An Intuitive and Comprehensive Two-Load Model for Series Elastic Actuators. In
Proceedings of the 2016 IEEE 14th International Workshop on Advanced Motion Control (AMC), Auckland, New Zealand, 22–24
April 2016; pp. 573–580.
40. Schütz, S.; Nejadfard, A.; Berns, K. Influence of Loads and Design Parameters on the Closed-Loop Performance of Series Elastic
Actuators. In Proceedings of the 2016 IEEE International Conference on Robotics and Biomimetics (ROBIO), Qingdao, China, 3–7
December 2016; pp. 2075–2080.
41. Nejadfard, A.; Berns, K.; Vonwirth, P. Technical Advantages and Disadvantages of Biarticular Actuators in Bipedal Robots. In
Proceedings of the Robots in Human Life–Proceedings of the 23rd International Conference on Climbing and Walking Robots and the Support
Technologies for Mobile Machines; CLAWAR Association Ltd.: Moscow, Russia, 2020; pp. 166–174. [CrossRef]
42. Mihailoff, G.; Haines, D. Chapter 24—Motor System I: Peripheral Sensory, Brainstem, and Spinal Influence on Anterior Horn
Neurons. In Fundamental Neuroscience for Basic and Clinical Applications, 5th ed.; Elsevier: Amsterdam, The Netherlands, 2018;
pp. 346–359.e1. [CrossRef]
43. Mussa-Ivaldi, F.; Hogan, N.; Bizzi, E. Neural, Mechanical, and Geometric Factors Subserving Arm Posture in Humans. J. Neurosci.
1985, 5, 2732–2743. [CrossRef]
44. Haeufle, D.F.B.; Wochner, I.; Holzmüller, D.; Driess, D.; Günther, M.; Schmitt, S. Muscles Reduce Neuronal Information Load:
Quantification of Control Effort in Biological vs. Robotic Pointing and Walking. Front. Robot. AI 2020, 7, 77. [CrossRef]
45. Brown, I.E.; Loeb, G.E. A Reductionist Approach to Creating and Using Neuromusculoskeletal Models. In Biomechanics and Neural
Control of Posture and Movement; Winters, J.M., Crago, P.E., Eds.; Springer New York: New York, NY, USA, 2000; pp. 148–163.
[CrossRef]
46. Winter, D.A. Biomechanics and Motor Control of Human Movement, 4th ed.; John Wiley & Sons: Hoboken, NJ, USA; Toronto, ON,
Canada, 2009.
47. Milner, T.E. Adaptation to Destabilizing Dynamics by Means of Muscle Cocontraction. Exp. Brain Res. 2002, 143, 406–416.
[CrossRef]
Sensors 2023, 23, 2428 25 of 25

48. Gribble, P.L.; Mullin, L.I.; Cothros, N.; Mattar, A. Role of Cocontraction in Arm Movement Accuracy. J. Neurophysiol. 2003, 89,
2396–2405. [CrossRef]
49. Cashaback, J.G.; Cluff, T. Increase in Joint Stability at the Expense of Energy Efficiency Correlates with Force Variability during a
Fatiguing Task. J. Biomech. 2015, 48, 621–626. [CrossRef] [PubMed]
50. Paine, N.; Mehling, J.S.; Holley, J.; Radford, N.A.; Johnson, G.; Fok, C.L.; Sentis, L. Actuator Control for the NASA-JSC Valkyrie
Humanoid Robot: A Decoupled Dynamics Approach for Torque Control of Series Elastic Robots. J. Field Robot. 2015, 32, 378–396.
[CrossRef]
51. Nejadfard, A.; Schütz, S.; Vonwirth, P.; Mianowski, K.; Berns, K. Coordination of the Biarticular Actuators Based on Instant Power
in an Explosive Jump Experiment. In Proceedings of the IEEE International Conference on Advanced Intelligent Mechatronics
(AIM), Auckland, New Zealand, 9–12 July 2018.
52. Nejadfard, A.; Schütz, S.; Vonwirth, P.; Mianowski, K.; Karsten, B. Moment Arm Analysis of the Biarticular Actuators in
Compliant Robotic Leg CARL. In Proceedings of the Conference on Biomimetic and Biohybrid Systems, Paris, France, 17–20 July
2018; pp. 348–360.
53. Schumacher, C.; Sharbafi, M.; Seyfarth, A.; Rode, C. Biarticular Muscles in Light of Template Models, Experiments and Robotics:
A Review. J. R. Soc. Interface 2020, 17, 20180413. [CrossRef] [PubMed]
54. Sharbafi, M.A.; Rode, C.; Kurowski, S.; Scholz, D.; Möckel, R.; Radkhah, K.; Zhao, G.; Rashty, A.M.; von Stryk, O.; Seyfarth, A.
A New Biarticular Actuator Design Facilitates Control of Leg Function in BioBiped3. Bioinspir. Biomimetics 2016, 11, 046003.
[CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.