arXiv:1710.08557v1  [cs.RO]  24 Oct 2017
On Neuromechanical Approaches for the Study of Biological and Robotic Grasp
and Manipulation
Francisco J Valero-Cuevas1,2,∗and Marco Santello3
1 Biomedical Engineering Department and
2 Division of Biokinesiology & Physical Therapy
University of Southern California, Los Angeles, CA, USA
3 School of Biological and Health Systems Engineering
Arizona State University, Tempe, AZ, USA
∗corresponding author
Abstract
Biological and robotic grasp and manipulation are undeniably similar at the level of mechanical task performance.
However, their underlying fundamental biological vs. engineering mechanisms are, by deﬁnition, dramatically diﬀer-
ent and can even be antithetical. Even our approach to each is diametrically opposite: inductive science for the study
of biological systems vs. engineering synthesis for the design and construction of robotic systems. The past 20 years
have seen several conceptual advances in both ﬁelds and the quest to unify them. Chief among them is the reluctant
recognition that their underlying fundamental mechanisms may actually share limited common ground, while exhibit-
ing many fundamental diﬀerences. This recognition is particularly liberating because it allows us to resolve and move
beyond multiple paradoxes and contradictions that arose from the initial reasonable assumption of a large common
ground. Here, we begin by introducing the perspective of neuromechanics, which emphasizes that real-world behavior
emerges from the intimate interactions among the physical structure of the system, the mechanical requirements of a
task, the feasible neural control actions to produce it, and the ability of the neuromuscular system to adapt through in-
teractions with the environment. This allows us to articulate a succinct overview of a few salient conceptual paradoxes
and contradictions regarding under-determined vs. over-determined mechanics, under- vs. over-actuated control, pre-
scribed vs. emergent function, learning vs. implementation vs. adaptation, prescriptive vs. descriptive synergies,
and optimal vs. habitual performance. We conclude by presenting open questions and suggesting directions for fu-
ture research. We hope this frank and open-minded assessment of the state-of-the-art will encourage and guide these
communities to continue to interact and make progress in these important areas at the interface of neuromechanics,
neuroscience, rehabilitation and robotics.
Keywords: Neuromuscular control, Hand, Prosthetics
Introduction
Grasp and manipulation have captivated the imagina-
tion and interest of thinkers of all stripes over the millen-
nia; and with enough reverence to even attribute the in-
tellectual evolution of humans to the capabilities of this
appendage [1, 2, 3]. Simply put, manipulation function
is one of the key elements of our identity as a species
(for an overview see [4]). This is a natural response to
the fact that much of our physical and cognitive abil-
ity and well-being is intimately tied to the use of our
hands. Importantly, we have shaped our tools and en-
vironment to match its capabilities (straightforward ex-
amples include lever handles, frets in string instruments,
and touch-screens). This co-evolution between hand-
and-world reinforces the notion that our hands are truly
amazing and robust manipulators, as well as providing
rich sensory, perceptual and even social information.
It then comes as no surprise that engineers and
physicians have long sought to replicate and restore
this functionality in machines—both as appendages
to robots and prostheses attached to humans with
missing upper limbs [5]. Robotic hands and prostheses
have a long and illustrious history, with records of
sophisticated articulated hands as early as Gottfried
‘G¨otz’ von Berlichingen’s iron hand in 1504 [6]. Other
Preprint submitted to Journal of Neuroengineering and Rehabilitation
November 9, 2018
eﬀorts [7, 8, 9, 10, 11] were often fueled by the injuries
of war [12, 13, 14, 15] and the Industrial Revolution
[16].
The higher survival rate in soldiers who lose
upper limbs [17, 18] and the continual emergence
of artiﬁcial intelligence [19, 20] are but the latest
impetus. Thus, the past 20 years have seen an explo-
sion in designs, fueled by large scale governmental
funding (e.g., DARPA’s Revolutionizing Prosthetics
and HAPTIX projects, EU’s INPUT and SOFTPRO
projects) and private eﬀorts such as DeepMind.
A
new player in this space are the potentially more
revolutionary social networks of high-quality amateur
scientists as exempliﬁed by the FABLAB movement
[21].
They are enabled by ubiquitously accessible
and inexpensive 3D printing and additive manu-
facturing tools [22], collaborative design databases
(www.eng.yale.edu/grablab/openhand/
and
others),
and
communities
with
formal
journals
(www.liebertpub.com/overview/3d-printing-and-additive-manufacturing/621/
and www.journals.elsevier.com/additive-manufacturing/).
Grassroots
communities
have
also
emerged
that
can,
for
example,
compare
and
contrast
the
functionality
of
prosthetic
hands
whose
price
diﬀers
by
three
orders
of
magnitude
(3dprint.com/2438/50-prosthetic-3d-printed-hand/).
For all the progress that we have seen, however, (i)
robotic platforms remain best at pre-sorted, pick-and-
place assembly tasks [23]; and (ii) many prosthetic users
still prefer simple designs like the revered whole- or
split-hook designs originally developed centuries ago
[24, 25].
Why have robotic and prosthetic hands not come of
age? This short review provides a current attempt to
tackle this long-standing question in response to the cur-
rent technological boom in robotic and prosthetic limbs.
Similar booms occurred in response to upper limb in-
juries [25] after the Napoleonic [26], First [12], and
Second World Wars [8], and—with the advent of pow-
erful inexpensive computers—in response to industrial
and space exploration needs in the 1960’s, 1970’s and
1980’s [27, 28, 29, 30, 31, 32]. We argue that a truly bio-
inspired approach suﬀers, by deﬁnition, from both gaps
in our understanding of the biology, and technical chal-
lenges in mimicking (what we understand of) biological
sensors, motors and controllers. Although biomimicry
is often not the ultimate goal in robotics in general, it
is relevant for (1) humanoids and (2) prostheses. Thus,
our approach is to clarify when and why a better un-
derstanding of the biology of grasp and manipulation
would beneﬁt robotic grasping and manipulation.
Similarly, why is our understanding of the na-
ture, function and rehabilitation of biological arms and
hands incomplete? Jacob Benignus Winsløw (Jacques-
B´enigne Winslow, 1669—1760) noted in his Exposi-
tion anatomique de la structure du corps humain (1732)
that ‘The coordination of the muscles of the live hand
will never be understood’ [33]. Interestingly, he is still
mostly correct. As commented in detail before [4], there
has been much work devoted to inferring the anatomi-
cal, physiological, neural and cognitive processes that
produce the upper limb function we so dearly appreci-
ate and passionately work to restore following trauma
or pathology. We argue, as Galileo Galilei did, that
mathematics and engineering have much to contribute
to the understanding of biological system. Without such
a ‘mathematical language,’ we run the risk, as Galileo
put it, of ‘wandering in vain through a dark labyrinth’
[34]. Thus, this short review also attempts to point out
important mathematical and engineering developments
and advances that have helped our understanding of our
hands.
This review ﬁrst contrasts the fundamental dif-
ferences between engineering and neuroscience ap-
proaches to biological vs. robotic systems. Whereas the
former applies engineering principles, the latter relies
on scientiﬁc inference. We then discuss how the physics
of the world provides a common ground between them
because both types of systems have similar functional
goals, and must abide by the same physical laws. We go
on to evaluate how biological and robotic systems im-
plement the necessary sensory and motor functions us-
ing the dramatically diﬀerent anatomy, morphology and
mechanisms available to each. This inevitably raises
questions about diﬀerences in their sensorimotor con-
trol strategies. Whereas engineering system can be de-
signed and manufactured to optimize well-deﬁned func-
tional features, biological systems evolve without such
strict tautology. Biological systems likely evolve by im-
plementing ecologically and temporally good-enough,
sub-optimal or habitual control strategies in response
to the current multi-dimensional functional constraints
and goals in the presence of competition, variability, un-
certainty, and noise. We conclude by exploring the no-
tion that the functional versatility of biological systems
that roboticists admire is, in fact, enabled by the very
nonlinearities and complexities in anatomy, sensorimo-
tor physiology, and neural function that engineering ap-
proaches often seek to avoid.
2
Deductive and inductive science for the study of bio-
logical systems vs. engineering synthesis for the con-
struction of robotic systems
As we have discussed before [35], any understanding
of a biological system is done by a diﬃcult scientiﬁc
process of logical inference [36] 1. Scientiﬁc inquiry,
in particular, is a combination of a deductive (top-down
logic) approach that invokes laws of physics—with in-
ductive (bottom-up logic) reasoning that uses speciﬁc
instances of observed behavior in the complete sys-
tem (e.g., gait, ﬂight, manipulation) to build concep-
tual, analytical or computational models (i.e., hypothe-
ses) about how constitutive parts interact to produce the
overall behavior. In contrast, the engineering perspec-
tive is to use proven physical laws (i.e., mechanisms)
to synthesize and build complex systems, like robots,
in a bottom-up way. The multiple technological suc-
cesses of this engineering approach naturally encourage
us to identify such fundamental mechanisms in biologi-
cal systems, and assemble models and hypotheses about
how they interact in biological systems to explain verte-
brate function and dysfunction—and apply them to rev-
olutionize rehabilitation medicine and build better ma-
chines that are either bio-inspired or able to interact with
humans (e.g., exoskeletons).
Thus, this engineering bottom-up approach has been
used by biomechanists, kinesiologists, clinicians, and
computational- and neuro-scientists to build models to
address questions such as ‘What is the control strategy
the nervous system uses to (i) move the arm to place the
hand in a sequence of locations in space? Or (ii) reach to
an object and grasp an object? Or (iii) use the ﬁngers to
use a hand held tool?’ Similarly, ’How do speciﬁc parts
of the brain contribute to produce the speciﬁc features
of reach, grasp and manipulation?’ (for recent reviews
see, e.g., [40, 4, 41, 42, 43, 44]).
For example, neuroscientists have characterized the
cortical networks responsible for selection of hand pos-
tures, forces or dexterous manipulation [45, 46]. Specif-
ically, invasive and noninvasive recordings from cor-
tical regions in non-human and human primates, and
1Quoted from [37]: ‘Deductive reasoning (top-down logic) con-
trasts with inductive reasoning (bottom-up logic) in the following
way: In deductive reasoning, a conclusion is reached reductively by
applying general rules that hold over the entirety of a closed domain of
discourse, narrowing the range under consideration until only the con-
clusion(s) is left.’. Conversely [38], ’Inductive reasoning is reasoning
in which the premises are viewed as supplying strong evidence for the
truth of the conclusion. While the conclusion of a deductive argument
is certain, the truth of the conclusion of an inductive argument may be
probable, based upon the evidence given [39].’
non-invasive brain stimulation studies in humans have
revealed the functional role of sensory, premotor, mo-
tor, and associated areas in motor planning, sensorimo-
tor adaptation, grip type selection, storing and retriev-
ing sensorimotor memories of hand-object interactions,
and controlling grasping and manipulation (for review
see[43]).
However, inferring valid models of biological sys-
tems is not trivial. It remains reasonable to ask whether
it is even possible for us to produce robust insights
about the mechanisms underlying complex neuromus-
cular systems [47]. Scientiﬁc inquiry requires that we
trust current (imperfect) theories of the mechanisms be-
hind the material properties, sensors, muscles, and neu-
ral processing in biological systems which, when in-
teracting with physical laws in a particular functional
regime (e.g., turbulent and laminar ﬂow, continuum and
rigid body mechanics, stable and unstable dynamical
domains, information theory, etc.), give rise to the ob-
served biological behavior. Unfortunately, the diﬀer-
ences that invariably emerge between model predic-
tions and experimental data can be attributed to a va-
riety of sources, ranging from the validity of the sci-
entiﬁc hypothesis being tested to the choice of each
constitutive element, or even their numerical imple-
mentation [48, 35].
Even when conceptual, analyti-
cal or computational models are carefully assembled,
the modeler must make arbitrary choices that often af-
fect the predictions of the model in counterintuitive
ways. Some examples of unavoidable choices are the
types of models for joints (e.g., a hinge vs.
articu-
lating surfaces), muscles (e.g., simple low pass ﬁlters,
Hill-type, populations of motor units), controllers (e.g.,
proportional-derivative, Bayesian, internal models, op-
timal control), and solution methods (e.g., forward, in-
verse) [35].
These choices are driven, at best, by a
comprehensive distillation of the vast literature, a fo-
cused research scope on simpliﬁed or speciﬁc scales
or domains of function, or intentional sidestepping of
unknown aspects of model elements and their inter-
actions that cannot be easily or accurately measured
experimentally—and thus, cannot be conﬁdently in-
cluded in the model. At worst, (over)simpliﬁcations are
driven by biological/mathematical/computational con-
venience or expediency.
A salient example of applying scientiﬁc ﬁndings to
robotic systems is the application of kinematic (postu-
ral) hand synergies to the design of the controller for a
robotic hand, such as the Pisa/IIT SoftHand [49]. This
design is based on using the ﬁrst principal component
computed from a set of static hand postures recorded in
human subjects while grasping imagined objects [50].
3
Here, the observed covariation in joint angles of the dig-
its was captured by the ﬁrst principal components ac-
counting for over 50% of variance of kinematic data.
This inspired the robotic design of the Pisa/IIT Soft-
Hand where a single actuator drives the motion across
multiple digits along that main joint-angle coordination
pattern.
It is important to note that it is perhaps the compli-
ance built into this robotic hand by design that allows
a passive (i.e., uncontrolled) adaptation to the speciﬁcs
of each grasp [51].
Such counter-intuitive and often
overlooked contributions of ‘passive’ structures to static
and dynamic functional versatility falls within the realm
of morphological computation—a longstanding concept
that is being revisited as an important contributor to the
versatility of biological systems, robots and prosthetics
(e.g., [52, 53, 54, 55, 56]).
The status quo in traditional robotics is a prime exam-
ple of the converse, i.e., engineering synthesis. That is,
building robotic systems that are the embodiment and
application of the closed-form mathematical, design,
and control principles we have and understand well. To
give an example discussed in detail below, many robots
are designed to have rotational motors. This comes from
the fact that the torque-control formulation for robotics
is well developed [57].
However, the control of robotic systems can also de-
pend heavily on a deductive approach. Namely, most
control architectures include an idealized mathematical
model of the system they are controlling (i.e., the so-
called plant2) [58, 59]. It is common practice to, for
example, use system identiﬁcation techniques [60, 61]
to use experimental data to infer a data-driven model,
i.e. input/output transfer function of the plant ‘as built
[62, 63, 64]. These models range from linear models
through to nonlinear and probabilistic approximations
to the robotic systems dynamics [65, 66, 67, 68, 69].
This is especially necessary when using optimal con-
trol formulations which depend critically on an accurate
model of the plant [70]. These limitations are currently
being addressed by real-time model predictive control
strategies that continually update families of possible
models of the plant (and its state), and operate only over
a limited time horizon, e.g., [71].
2A note on nomenclature. This does not mean to tip over a pot-
ted houseplant. This term comes from control engineering where the
process to be controlled was usually an industrial or manufacturing
plant.
Mechanics and neuromechanics as the common
ground between biological and robotic systems for
grasp and manipulation
The ﬁelds of biological and robotic behavior are, nev-
ertheless, fortunate in that principles of mechanics are at
the root of both evolutionary biology and robotics. Dar-
winian evolution and Newtonian mechanics are unfor-
giving arbiters that continually shape what is possible
and successful in the physical world. Thus, even though
animals have had to evolve whereas robots have had to
be designed and built, both had to successfully with-
stand and exploit the laws of mechanics [57]. There-
fore, studying biological systems in the context of the
physical function of grasp and manipulation does have
the possibility of providing insights into how the struc-
ture of the body, information processing in the nervous
system, and their interactions give rise to complex be-
havior.
An appropriate name for this approach is Neurome-
chanics. To our knowledge, this term was ﬁrst coined
by Enoka in his 1988 book Neuromechanical Basis of
Kinesiology [72]. We use the term neuromechanics as
describing the functional co-adaptations of the nervous,
motor, sensory and musculoskeletal systems to produce
eﬀective and versatile mechanical behavior with a phys-
ical body in the physical world. It aptly emphasizes that,
in a ﬁeld that mostly emphasizes the cognitive capa-
bilities of the mammalian brain, it is easy to overlook
that the nervous system and body co-evolved well be-
fore mammals appeared [73]. Thus, there is much to
be learned when studying neural function using physi-
cal behavior by the periphery (i.e., limbs) as a means to
understand central (i.e., neural) function [57]. But how
can we move away from the diﬃculties in deductive in-
ference mentioned in the prior section?
One promising approach is to use synthetic anal-
ysis to build neuromorphic neuromechanical systems
that exploit physical reality as the common ground
between biological and robotic systems.
The neuro-
morphic approach reﬂects the sentiment expressed by
Richard Feynman, ‘What I cannot build I do not un-
derstand. Know how to solve every problem that has
been solved.’ 3 In our context, it can be taken to mean
that, if we have over one hundred years of sensorimo-
tor neuroscience since Sir Charles Sherrington [74], and
3This
was
found
written
on
his
blackboard
at
the
time
of
his
death
in
February
1988,
https://archives.caltech.edu/pictures/1.10-29.jpg. It
is thought he meant to suggests that one should only use mathematical
concepts one has derived, and therefore proven, to oneself.
4
if the principles we have deduced are sound, then we
should be able to build components that embody those
mechanisms in such a way that when assembled they
behave like biological systems [75, 76]. One example
of such a neuromorphic approach uses ultra-fast com-
puter processors to simultaneously implement popula-
tions of autonomous, interconnected spiking neurons in
real time that follow Hodgkin-Huxley rules of how ac-
tion potentials in neurons are initiated and propagated
[77]. As mentioned in [75, 78], this general approach
has also been successfully applied to understand mech-
anisms of memory, visual representation, and recently,
cognitive function. Note that neuromorphic is distinct
from neuromimetic or neuroinspired. Biomimetic (neu-
romimetic) and bioinspired (neuroinspired) work seeks
to copy or replicate the biological (neural) behavior by
any engineering means—like prosthetic hands that have
no muscles or tendons, or airplanes that ﬂy without ﬂap-
ping wings. In contrast, neuromorphic approaches use
engineering means to implement the biological mecha-
nisms themselves.
As explained in [75], we have taken this approach
one step further by combining neuromorphic and neu-
romechanical approaches, we seek to implement the the
neural control of the body—eﬀectively merging biol-
ogy and robotics in the arena of physical function. We
have coupled real-time neuromorphic implementations
of stretch reﬂex circuitry in populations of spinal neu-
rons, to electric motors controlled by real-time models
of muscle function to apply forces to the tendons of ac-
tual human cadaveric ﬁngers. This is the ﬁrst neuromor-
phic neuromechanical system, to our knowledge, that
has put our understanding of fundamental sensorimo-
tor mechanisms in the spinal cord to the ultimate test
of physical implementation. Importantly, the behavior
emerges from the system as it is not prescribed beyond
the nature and connectivity of its elements. An added
advantage is that one can also ‘record’ from single or
multiple neurons, motor units, aﬀerent nerves, etc. to
explore emergent behavior at truly multiple scales. So
far, this approach has allowed us to begin to understand
cardinal features of aﬀerent muscles of human ﬁngers to
replicate fundamental features of healthy muscle tone,
hypo and hypertonia [76].
While still an imperfect approximation, this neo-
Sherringtonian approach helps us test arguments about
which speciﬁc features of spiking neurons and their con-
nectivity, spindle function, fusimotor drive, descending
commands, ﬁnger anatomy and tendon/skin/joint tissue
properties suﬃce to produce realistic healthy and patho-
logic behavior in aﬀerent muscles acting on anatomical
ﬁngers. Moreover, this has the advantage of using phys-
ical behavior as the ground truth for the evaluation of
functional performance.
A combined neuromorphic and neuromechanical ap-
proach, although grounded and developed for neurosci-
entiﬁc applications, could potentially inspire robotics
research and design by revealing insights into how com-
plex behaviors emerge from adaptation of neural con-
trollers to mechanical properties of physical systems.
Fundamental diﬀerences between biological and
robotic systems
Sensory diﬀerences
From the perspective of biological vs. robotic closed-
loop behavior, a striking diﬀerence is the superior abil-
ity of the nervous system to utilize and eﬀectively inte-
grate information acquired through an incredibly wide
array of nonlinear, delayed, noisy, non-collocated and
distributed sensors (for review, see [79]). Important ad-
vances have been made in our understanding of how
multimodal sensory information is integrated to make
possible the ability of humans to extrapolate sensory
information based on the statistical properties of stim-
uli [80], as well as its vulnerability to sensory illusions
[81].
Remarkably, the hand’s sensory system endows hu-
mans not only the ability to perform online sensing
of the state of the system (e.g., contact onset and oﬀ-
set) [82, 83], but also to use sensory information ac-
quired through past hand-object interactions to predict
sensory consequences stemming from planned interac-
tions [84, 85]. Even more impressive is the exquisite
ability of vertebrates to perform ‘active sensing’ (e.g.,
whisking in rodents [86, 87]) and tactile exploration in
humans of the shape, texture, features and mechanical
properties of objects [88, 89] 4. In active sensing, motor
actions are explicitly driven to extract relevant sensory
information [91]. This has led some to express that the
human hand is as much a sensory organ as a motor ef-
fector [92].
In robotics, sensory data are central to feedback con-
trol [93]. The idea of extending the use of sensory data
beyond feedback to also extract the properties of a sys-
tem is called system identiﬁcation [61] or ‘plant inver-
sion’ [82, 94]. That is, characterizing the eﬀects of com-
mand signals (i.e., the input-output characteristics of the
4Other important forms of active learning are active vision and
saccades in mammalian vision [90]
5
plant) can be done when the process is inverted mathe-
matically (i.e., ﬁnd the output-input characteristics), an-
alytically or experimentally. Hence the term ‘plant in-
version.’ System identiﬁcation has been an extensive
ﬁeld since the middle of last century [95, 60], and con-
tinues to make progress as sensors, algorithms and pro-
cessing power increase [96, 97, 98, 99].
An important challenge in tactile sensors is their
problematic placement at the ﬁngertips, where they are
exposed to damage [100, 101]. A notable advance has
been the biomimetic idea by Loeb and Johansson to
develop the Biotac [102, 103], where the ﬁngertip it-
self is a rugged and ribbed rubber balloon (much like
a ﬁnger pad) inﬂated with conductive ﬂuid over the
distal bone. Static and dynamic contacts with objects
produce changes in hydrostatic pressure and electrical
impedance in the ﬂuid that produce a rich and time-
varying multi-dimensional sensory signal.
The com-
mercial version of this system is now being incorporated
into multiple robotic hands [104]. It is even being de-
ployed to create industrial standards for the perceptual
aspects of touch such as smoothness, roughness, silk-
iness, richness, quality, etc. Solving the diﬃcult sig-
nal processing challenges that this type of tactile sensor
presents has the potential to imbue robotic hands with
truly robust and useful synthetic touch.
Solving the
computational challenges of sensory fusion (let alone
the active sensing) in robotic and prosthetic systems is a
critical frontier. There are even eﬀorts at the interface of
prosthetics and robotics to translate touch information
from prosthetic hands into neurostimulation to restore
the sense of touch (e.g., [105, 106, 107]).
Artiﬁcial feedback approaches, of course, extend be-
yond these examples of sensors located on the ﬁnger-
tips, and include artiﬁcial skins with embedded strain
gauges (e.g., [108]) and vibrotactile feedback used
in prosthetics delivering information about mechanical
events (i.e., contact) to the residual limb (e.g., [109]).
Nevertheless, the necessity of sensory information for
manipulation has been challenged by practical examples
of sensor-less, fully open-loop grasp [110, 51, 111, 49,
56], and pre-planned manipulation [112]. It is therefore
most likely that in biological systems—and by exten-
sion in robots—sensory information is most useful dur-
ing learning [83] (see section on learning below).
Motor diﬀerences
Chief among these is our inability, so far, to match
the power:weight ratio, mechanical eﬃciency, versatil-
ity, adaptability and self-repair properties of muscle.
A pneumatic analog to muscle was ﬁrst developed by
McKibben in the 1950’s [113, 114, 115] and contin-
ues to be used and studied [116][117], but the need
for compressors/pressurized tanks, valves, cables, muf-
ﬂers, etc.
remains a challenge to its portability and
versatility.
Electric, hydraulic and pneumatic actua-
tors are, of course, the mainstay of robotics. The last
two decades have seen great progress in the technol-
ogy to control [118] and power useful and portable ex-
oskeletons and prostheses [119, 120, 121]. Moreover,
battery life has improved by several orders of magni-
tude [122, 123, 124].
However, muscles remain un-
matched in the continually surprising variety of me-
chanical functions they accomplish in locomotion and
manipulation; serving as motors, brakes, springs, struts,
etc. [125, 126].
From the architectural perspective, contractile pro-
teins in muscle can only make muscles actively pull on
their tendons, which attach to bones after crossing joints
[127, 128]. Consequently, limb function in vertebrates
is tendon-driven, not torque-driven as in the mathemat-
ics and dominant practice of robotics.
A recent conceptual advancement in the study of sen-
sorimotor control in vertebrates comes from embracing
and emphasizing the fact that muscles act on the body
via tendons [57]. While this has been obvious since the
very ﬁrst anatomical studies of antiquity, most modern
engineering, neuroscience, biomechanical, and mathe-
matical analyses have tended to prefer the torque-driven
abstraction. To be fair, the torque-driven phrasing of the
problem is attractive, mathematically correct, and well
developed from the conceptual, analytical and computa-
tional perspectives. In it, rotational motors at each hinge
joint produce torque, angle, or angular velocity directly,
which give rise to the kinematics and kinetics of the
limbs and ﬁngers they control. The actions of muscles
in a simulated biological system are, thus, collapsed into
net joint torques at each joint. Then, their analysis pro-
ceeds as with any other robotic system. But there are
several arguments concluding that this abstraction can
be misleading; as it does not represent the actual me-
chanical problem of controlling tendons, which is the
actual problem the nervous system confronts. Due to re-
cent advances in the tendon-driven formulation of limb
and ﬁnger function (e.g., [129, 57, 130, 131, 132, 133]),
we are now better able to focus on the actual tendon-
driven mechanical problem that confronts the nervous
system. In fact, as argued in [57], many of the concep-
tual/mathematical problems associated with the analysis
of the neuromuscular control of biological limbs can be
clariﬁed by using this perspective.
6
Sensorimotor control
Due to the above-described limitations of scientiﬁc
inference, it is not surprising that competing views ex-
ist on the model (i.e., hypothesis, strategy) that best ex-
plains the nervous system’s exquisite ability to control
movement, locomotion and manipulation; as well as its
uncanny ability to generalize and adapt learned motor
behaviors. For example, the concepts of ‘internal mod-
els’ and ‘optimal control’ can capture signiﬁcant fea-
tures of motor behavior, but it has been challenging
to associate these theoretical frameworks with speciﬁc
anatomical structures or physiological mechanisms. It
is only natural that some of these models have been in-
spired by the formalism and successes of robotics and
control theory, which in turn have been inﬂuential in
driving experimental approaches and theoretical frame-
works (for review, see [134]).
However, even a cursory comparison between how
robotic and human hands reach, grasp and manipu-
late objects reveals major diﬀerences between them.
Some diﬀerences stem from ability of the neuromus-
cular system to implement versatile transitions, adjust-
ments and adaptations of control strategies. One ex-
ample is the ability of the hand to swiftly change con-
trol strategies when transitioning from ﬁnger motion to
force application [135]. Another example is humans’
ability to modulate ﬁnger force distribution shortly af-
ter contact and prior to onset of manipulation to ac-
count for trial-to-trial variability in ﬁnger placement
[136, 137, 138, 139, 140]. Such problems are extremely
challenging to replicate in a robotic system.
Other
diﬀerences emerge from the hand’s unique anatomical
structure, which allows it to adapt to task demands,
including passively shaping itself to object geometry.
These are diﬀerences that robotic designs can partially
address (see section ‘Under- vs.
Over-actuated con-
trol’), but cannot fully match given the limitations of
robotic systems in actively integrating sensory feedback
with motor commands, or passively adapting to objects
and the environment.
Advantages and limitations of control theoretic ap-
proaches to biological sensorimotor control
These fundamental diﬀerences motivate and justify
a candid evaluation of the extent to which our concep-
tual approaches to robotics are appropriate for the study
of sensorimotor control in biological systems. As men-
tioned above, the approach to biological sensorimotor
control has, for historical and practical reasons, lever-
aged the formalism of robotics and control theory (e.g.,
[29, 57, 141, 142]). At the risk of oversimplifying a
large ﬁeld for the sake of succinctness, we can describe
real-time feedback control as follows: sensors transduce
physical signals to estimate the performance of the sys-
tem, which the controller considers as it applies control
laws to take the next actions to correct mistakes, reject
perturbations, and meet the constraints of the task to
achieve a goal. It stands to reason that the neuromus-
cular control of biological systems can perform all of
these functions. Therefore, control-theoretic constructs
likely apply (e.g., [143, 144, 145, 146, 147]). Such an
approach is justiﬁed by the successes in identifying neu-
ral circuits that perform closed-loop feedback control
such as homeostasis in physiological control systems
[148], muscle stretch reﬂexes [149], and vestibuloocular
reﬂexes for eye tracking in the presence of head rotation
[150, 151].
Why is it that such a well-founded approach has
failed to produce conclusive theories for sensorimo-
tor control in humans for grasp and manipulation?5
One possibility is that sensorimotor control for grasp
and manipulation is unlike other forms of motor con-
trol because its motor actions involve perception, as
well as more complex sensorimotor transformation pro-
cesses than, say, cyclical movements for locomotion.
The importance of the interaction between perception
and action has been convincingly argued by, for ex-
ample, Prinz, Iberall and Arbib [155, 156, 157, 158].
This is further exempliﬁed by other work bridging this
perception-action gap [159, 160, 161, 162, 163, 164].
This conceptual divergence across biological and
robotics problems can perhaps be explained by assess-
ing some fundamental features of the robotics perspec-
tive, and their appropriateness for the study of sensori-
motor control in biological systems. Earlier, we spoke
of the control-theoretic framework that deﬁnes interac-
tions among sensors, control laws and the goal of the
task.
A more formal presentation is that the control
laws operate to change the ‘state’ of the system. The
formal deﬁnition of the state of a system is the minimal
set of variables to characterize its governing equations,
or that can be used to describe its dynamical evolution
[58, 165, 29]. Thus, control theory is based on actions
that will eﬀectively and appropriately change the state
to reach a goal, track an external process or reject a per-
turbation on the basis of a user-deﬁned optimization cri-
terion (often called objective or cost function). In fact,
the deﬁnition of ‘controllability’ is to be able to arbi-
trarily move a system from any state to any other state
in ﬁnite time. Its counterpart is ‘observability,’ which
5One can argue that work in invertebrates or locomotor patterns in
vertebrates has been more successful, e.g., [152, 153, 154].
7
is the formal quantiﬁcation of the ability to arbitrarily
extract the state of the system from measurements ob-
tained by sensors [166, 167]. When designing a robotic
system, its state variables are explicitly chosen by the
user from the many potential options, such as joint an-
gles, angular velocities, endpoint locations, velocities,
etc. Implementing observability via methods of state
estimation is a form of system identiﬁcation that spans
a wide set of techniques, and which is central to any
control-theoretic approach applied to robotics and mod-
els of biological function [168, 165, 35].
The ﬁeld of neuromuscular control has come to the
realization that it is still an open question whether and
how the nervous system adopts the concept of state,
and even whether it optimizes an explicit or implicit
cost function [169, 170, 57]. This realization has been
slow and reluctant because abandoning the sound engi-
neering formalism is both unappealing and unnecessar-
ily extreme. Thankfully, the past several decades have
seen, as described below, the development of equally
well-founded alternatives or complements to the classi-
cal control theoretic perspective that mitigate the draw-
backs of state estimation, cost functions and even con-
trol laws. Furthermore, there are many other techniques
and approaches that have yet to be applied to biological
problems [35].
Feasible rather than optimal function
Optimization is a computationally eﬃcient means to
use convex, preferably quadratic, cost functions to se-
lect a speciﬁc control action from among all feasible
actions within a high-dimensional solution space. For
example, if you have N muscles crossing a joint, opti-
mization can be used to search an N-dimensional space
to ﬁnd a point in it (i.e., a combination of muscle acti-
vations) to produce a given net joint torque while mini-
mizing, say, the sum of squares of activations.
Roboticists have always used optimization to con-
trol robots. Be it to tune gains in the P, PD or PID
controllers to implement force, impedance and position
control [58], plan paths (e.g., [171]), etc. In fact, the
ubiquitous use of state estimators (such as Kalman Fil-
ters) are, in fact, optimal solutions in the least-square
sense [172]. In addition, this emphasis on optimality
is central to Optimal Control [165] in its various mod-
ern forms as LQR, LQG, iLQR, iLQG (e.g., [173].).
From its inception, however, the emphasis on optimal-
ity has proven problematic to stability margins, which
has led to other forms of control such as robust con-
trol, path integral control, model predictive control, etc.
[70, 174, 175, 176].
Biomechanists and neuroscientists have, neverthe-
less, adopted the well-founded mathematical concept of
optima to cast the problem of neuromuscular control as
one of numerical optimization [143, 177, 178]. How-
ever, it is unlikely that the nervous system acts strictly
like a computer running optimal control, gradient de-
scent or policy gradient algorithms. Rather, optimiza-
tion has and should be used as a metaphor, but one that
should not be taken too literally when working with bi-
ological systems [169, 170].
We must recognize that, as with any metaphor, there
are limits to its validity.
It is important to explic-
itly acknowledge a bifurcation in the approaches we
use to build robots vs. understand biological systems.
For example, recent advances in control have begun to
yield very impressive real-time performance in physical
robots [71, 179, 180, 181]—but there is no need to in-
sist that biological systems use those methodologies or
algorithms.
Nevertheless, scientists studying biological systems
must ask themselves how faithful they want to adhere to
physiological realism and, on the basis of that decision,
select appropriate problems and solution methods—
while also avoiding the temptation to necessarily im-
bue biological systems with mathematical algorithms,
or robotic solutions with physiological meaning.
An approach we can call Feasibility Theory is an
alternative to optimization.
It tackles and solves the
computationally expensive problem of explicitly deﬁn-
ing and ﬁnding families of valid solutions (i.e., a ‘fea-
sible muscle activation set’)[57]. In doing so, we can
characterize the set of options open to the nervous sys-
tem without having to advocate a particular (and debat-
able) cost function [57, 182, 183]. These families of
feasible muscle activations have a well-deﬁned, low-
dimensional structure because they emerge, unavoid-
ably and naturally, from the interactions among the
known mechanical and physiological properties of the
limb, and the functional constraints of the task (which
can be mechanical, metabolic, physiological, etc.). That
is, if you have a limb with 9 muscles and you are pro-
ducing a task with ﬁve constraints (e.g., the x, y, z
magnitude of the force vector produced by the end-
point of the limb, and the stiﬀness of the endpoint in
the x and y directions), then the solution space is a 4-
dimensional (i.e., 4=9-5) subset of the 9-dimensional
activation space [184, 185].
More generally, thinking of the problem of control-
ling muscles as one of exploring and exploiting solution
spaces is perhaps biologically tenable. Biological sys-
tems could use sparse trial-and-errorlearning to ﬁnd and
explore feasible activation sets. That is, implementing
8
any point within the low-dimensional feasible solution
space will be adequate to perform the task. Memory and
pattern matching could be used to exploit the ‘informa-
tion’ collected from prior experience in the context of
the current environment and task goals [57].
Probabilistic sensorimotor control
What information does the nervous system use to
produce physical behavior? And how does it assem-
ble, encode, store, access and use that information?
A promising approach is a probabilistic one, where
trial-and-error can be combined with memory to form
probabilistic representations of actions in the physical
world. Bayesian inference [186, 187] or stochastic con-
trol [143] are formal approaches to describe the emer-
gence of probability density functions of the mapping
from perception/intention to action in the presence of
unavoidable uncertainty, noise, risk and variability in
the real world. Moreover, there are probabilistic ap-
proaches that can be used with populations of spiking
neurons to produce physical behavior in a cost-agnostic,
emergent, model-free way [188, 189, 75, 76].
Probabilistic approaches to control have their origins
in robotics, where noise, uncertainty about the state,
properties of the plant, dynamics of the coupled robot-
world system, etc. have always posed challenges. One
can argue that such probabilistic models can be consid-
ered model-free or semi-model free because there is no
explicit representation of the body or world, other than
from the statistical representation of the results of trial-
and-error experiments that inform the learning of a suc-
cessful policy (i.e., control law or motor habit) that is
valid in the neighborhood of some initial conditions to
meet a speciﬁc goal [176, 190].
Along these same lines, it is also possible that the
controller itself is ‘embedded’ in the structure of the
hand. Thus, there would be no need to regulate the ma-
nipulation task, but rather that the mechanical architec-
ture of the hand naturally leads to adequate and versa-
tile grasping function. This is sometimes called embed-
ded logic or underactuated control, as described below
[110, 191, 51, 111, 49, 56].
Paradoxes and insights
Under-determined vs. Over-determined mechanics
One of the central tenets of of motor control has been
the concept that the control of biological systems is
under-determined, meaning that they have ‘too many’
kinematic or muscular degrees of freedom. Therefore,
the nervous system faces a problem of selecting and
implementing a solution from among an inﬁnite set of
choices.
Such kinematic redundancy can be demonstrated by
simple examples such as the possibility of using any one
of multiple arm trajectories to hammer the same loca-
tion in space [192], or one of multiple types of grasps
to hold the object just as well [193]. From the muscle
control perspective, vertebrates have multiple tendons
crossing each joint, thus, there are multiple individual
muscle forces that can produce a given net joint torque
[177]. In contrast, roboticists have emphasized design
architectures to reduce kinematic and actuation redun-
dancy and typically build robots with as few kinematic
degrees of freedom or tendons to be controllable.
This begs the question why the evolutionary pro-
cess has tended to converged on such so-called under-
determined mechanical systems for vertebrates. As re-
viewed in [57], thinking of biological systems as under-
determined is paradoxical with respect to the evolu-
tionary process and clinical reality. For example, why
would organisms evolve, encode, grow, maintain, re-
pair, and control unnecessarily many muscles when a
simpler musculoskeletal system would suﬃce, and thus,
have phenotypical and metabolic advantages? Why do
people seek clinical treatment for measurable dysfunc-
tion even after injury to a few muscles, or mild neu-
ropathology? Which muscle would you donate to im-
prove your neural control?
Somehow, however, many muscles are a good thing.
Given the evolutionary process, we probably have close
to the right number of muscles to allow us to produce
useful behavior in the real world 6. One approach to
explain the apparent paradox that we have ‘too many’
muscles in vertebrates is that every muscle expands
our abilities and provides an additional degree of free-
dom for control. Behavior in the real world 7 consists
of satisfying multiple—at times competing—demands.
Therefore, a mathematical argument can be made [57]
that behavior in the real world, by virtue of needing to
satisfy multiple demands or constraints, requires multi-
ple muscles [185]. And, by extension, that dysfunction
of even a few muscles will make the limb less versatile
[196].
Similarly, kinematic redundancy loses its relevance
when we consider that limbs are actuated by muscles
that pull on tendons. It is clear that, if multiple ten-
dons cross each joint, then there is redundancy in the
6Note that this is not an argument for optimality of anatomical
architecture, but only for suﬃciency.
7Also called neuroethology to distinguish it from reductionist lab-
oratory work [194, 195].
9
sense that multiple individual muscle forces at those
tendons that can produce a given joint torque. How-
ever, the same is not true when we consider movement.
The rotation of that single joint deﬁnes the lengths of all
muscles that cross it [183, 57, 197]. While in principle
muscles can go slack, muscles with tone will shorten
appropriately.
However, muscles that lengthen must
lengthen by a prescribed amount. Thus, the relation-
ship where a few joint angles and angular velocities
for a given limb movement determine the lengths and
velocities of all muscles is over-determined—the very
opposite of redundant. That is, if any one muscle that
needs to lengthen to accommodate the movement fails
to do so (because, for example, it received the incor-
rect neural command or its stretch reﬂex fails to be
appropriately modulated in time), the movement will
be disrupted [183, 57, 197]. Therefore, while multi-
ple limb kinematics may be equivalent from a task per-
spective (i.e., reaching a cup or throwing a ball), they
are far from equivalent from the perspective of neuro-
physiological control, robustness to sensorimotor noise,
or time-critical modulation of activation and reﬂexes
[197].
Grasp vs. Manipulation
Another recent evolution in the biological side has
been the explicit distinction between grasp and manipu-
lation. Although these terms are often used interchange-
ably in the biological literature, there is a long tradi-
tion of creating clear taxonomies and descriptions of
hand actions that clearly distinguish between the two
[27, 156, 198, 29]. Speciﬁcally, grasp in general relates
to the act of seizing an object by wrapping the ﬁngers
around it. Manipulation has a more general connotation
of imparting change to an object or process. However,
precision or dexterous manipulation is a more speciﬁc
term reserved for cases where only the ﬁngertips make
contact with the object, not simply to grip the object,
but to be able to act independently to produce in-hand
manipulation.
Interestingly, most biological research has focused on
grasp [51, 29, 27]. Largely because studying precision
or dexterous manipulation has important practical dif-
ﬁculties with motion capture of individual ﬁnger mo-
tions, and measurement of individual ﬁngertip forces.
Similarly, in spite of the mathematics of dexterous ma-
nipulation being well developed [29], robotic hands and
prostheses tend to focus on grasp because of the dif-
ﬁculties in designing, building, and controlling ﬁngers
and ﬁnger contacts independently.
There have been some important advances, however.
For example, it is possible to begin to simulate contact
forces that go beyond physics engines for gaming or an-
imation [181]. Similarly, some experimental methods
have been developed to quantify dynamic dexterous ma-
nipulation, which has revealed novel aspects about the
neuromechanical control of dexterous manipulation in
development, adulthood, healthy aging and neurologi-
cal conditions (for overviews see [199, 200, 201, 202,
203, 45]), as well as how the interaction between cogni-
tive and biomechanical factors aﬀects dexterous manip-
ulation performance (e.g., [136, 137, 204]).
Under- vs. Over-actuated control
Similarly, it is reasonable to ask to what extent the
nervous system is necessary for grasp (and perhaps even
manipulation).
It is an analogous question to what
has also been recognized for passive dynamic walkers
[205, 206]. While eﬀective mechanical function can be
found in many vertebrates, primates (and humans) are
the beneﬁciaries of highly specialized neuroanatomical
coevolution of brain and hand (e.g., [44, 207, 208, 2]).
Understanding the contributions of a neural controller,
or speciﬁc neuroanatomical areas of the brain, to grasp
and manipulate remains an active area of study. In fact,
it is critical to consider moving away from a strictly so-
matotopic [209] and cortico-centric view of manipula-
tion, especially in the cases of dynamic dexterous ma-
nipulation where time delays preclude active involve-
ment [200].
After all, the current concept of cortical control is
not the exclusive micromanagement of individual mus-
cle activations, but rather includes the ‘binding’ of mo-
tor neurons into ﬂexible, context-dependent functional
groups [210, 211, 212], the utilization of primitive ‘syn-
ergies’ prepared by networks of spinal interneurons
[213], adjustment of sensory feedback gains [214], and
the formation/recall of motor memories [215], to name
just a few. Synergies are discussed in more detail in
subsequent sections.
Nevertheless, as in the case of passive dynamic walk-
ers, robotics provides counterexamples to such micro-
management of muscle actions by the brain, or even
the nervous system in general. A class of robotic hand
designs is called under-actuated because few motors
drive multiple degrees of freedom (this is in contrast
with over-actuated hands that have enough muscles to
control every degree of freedom independently). Such
hands can display multiple versatile grasp functions,
without requiring a controller [56, 111] or even ﬁngers
[216]. Such developments are alternatives that promise
to develop multiple designs along the spectrum between
under- and over-actuated robotic hands. This is espe-
cially useful in cases of brain-machine interfaces for
10
hand prostheses, where only a few degrees of freedom
of control can be extracted from the human pilot’s ner-
vous system (e.g., [217]).
Learning vs. Implementation vs. Adaptation
Human sensorimotor learning has been extensively
studied [134, 83]. One view posits that that humans’
ability to perform skilled motor behaviors relies on
learning both control and prediction through inverse and
forward internal models (implicit, explicit, probabilis-
tic or otherwise).
Speciﬁcally, a given control strat-
egy generates motor commands needed to create de-
sired consequences (e.g., a given reach trajectory or
grasping an object at speciﬁc locations), whereas pre-
diction maps motor commands (i.e., eﬀerence copy)
into expected sensory consequences (e.g., object con-
tact or onset of acceleration at object lift oﬀ[218, 219]).
The mechanisms proposed to account for updating of
these internal models may or may not include errors that
would occur when a mismatch between sensed and pre-
dicted sensory outcome occur, i.e., error-based learning
(e.g., [220, 221, 222]) and use-dependent learning (e.g.,
[223, 224, 225, 226]). However, most of what we know
about human sensorimotor learning for reach has been
derived from studies of reaching movements over dis-
tances of ± 10—14 cm, and their adaptations to force
ﬁelds or visuomotor rotations. Relatively little is known
about mechanisms underlying sensorimotor learning of
grasping and manipulation.
We have known for decades that ﬁnger force control
used in previous manipulation can inﬂuence how forces
are coordinated on the current manipulation through the
memory of an object’s physical properties [84, 85, 227].
More recently, it has been shown that humans may ac-
quire and retain multiple internal representations of ma-
nipulation [228, 229]. Later studies provided further
evidence supporting the concept of multiple sensori-
motor mechanisms and how their diﬀerent time scales
may interfere with generalization or retrieval of previ-
ously learned manipulation [230]—even when the ob-
ject being manipulated is the same.
A recent study
has provided evidence for the co-existence of context-
dependent and independent learning processes [204],
which would operate similarly to those described for
adaptation of reaching movements [231]. The advan-
tage of context-dependent representations of manipu-
lation is that they can be recalled when the object has
strong contextual cues (i.e., object geometry and per-
haps other perceptual attributes). In contrast, context-
independent representations are more sensitive to the
practice schedule used to learn a given manipulation,
but might be particularly advantageous when the up-
coming context has no context cues.
That is, in the
absence of information to the contrary, it is preferable
to repeat the most recent manipulation strategy even
though it is not guaranteed to be the correct one.
When considering parallels between the above-
described framework for learning of dexterous manipu-
lation in humans with learning manipulation by robotics
systems, it has been suggested that artiﬁcial controllers
could take advantage of select features of the biological
framework. Speciﬁcally, and as reviewed in [232], mul-
tiple parallel learning mechanisms could beneﬁt robotic
learning of manipulation tasks to aﬀord to deal with
structured and unstructured environments. At the same
time, the detrimental eﬀects or interference of neural
representation built through learning in one manipula-
tion context, and then transferring it to another con-
text can be theoretically minimized or bypassed when
designing an artiﬁcial controller.
Some examples of
successful robotic learning for grasp and manipula-
tion show that this is possible [180, 179, 233, 234].
Of course, these theoretical considerations assume that
building multiple representations of learned manipula-
tions allows them to operate independently, something
that—as described above—clearly also challenges bio-
logical controllers.
Another
biologically-inspired
phenomenon
that
could be of value to robotic manipulators is ﬁnger
force-to-position modulation.
Brieﬂy, it has been
shown that humans are able to modulate manipulative
forces in an anticipatory fashion, i.e., between contact
and onset of manipulation, according to where the ob-
ject is grasped [136, 137]. This phenomenon, which has
been conﬁrmed by several studies [139, 138, 235, 140],
ensures attainment of the manipulation goal despite
trial-to-trial variability in ﬁnger placement that may
naturally occur while using the same or diﬀerent
number of ﬁngers ([136] and [137], respectively).
Finger force-to-position modulation is a phenomenon
that is very useful for inferring its underlying neural
control mechanisms.
Speciﬁcally, for humans to be
able to adjust ﬁnger forces as a function of variable
position, a ‘high-level’ representation of the task (e.g.,
a given compensatory torque) is required, rather than
learning a ﬁxed ﬁnger force distribution.
Addition-
ally, such high-level representation has to drive how
sensing of the relative position of the ﬁngers is used
to implement the appropriate ﬁnger force distribution
by the time the learned manipulation is initiated. As
ﬁnger force-to-position modulation aﬀords biological
systems to be very adaptive—a given manipulation can
be performed without having to grasp an object exactly
11
in the same way each time it is being manipulated—one
can envision important robotics applications.
These
include controllers that are designed to build, through
extensive training, the high-level representation of a
task performed in many diﬀerent ways. If such high-
level representation could be built, stored, retrieved,
and designed to interact with artiﬁcial sensing of
ﬁnger positions, such a controller should, theoretically,
be able to be adaptable to manipulators that widely
diﬀer in terms of number of joints or ﬁngers. Such a
controller could be shared by multiple representations
learned through training of manipulation in structured
and unstructured contexts.
Another important distinction to be made is that, as
roboticists, we marvel at the learned capabilities of bi-
ological systems.
However, we tend to forget how
diﬃcult it is for organisms to learn and maintain that
level of performance. Recent work has begun to elu-
cidate why learning to produce accurate, smooth and
repeatable movements takes immense amounts of prac-
tice even in typically developing children [236], why so
few of us can become elite musicians or athletes [237],
and why rehabilitation requires very intensive practice
[238]. That is, controlling our bodies is not as easy as
it appears. We are seeing the result of millennia of co-
evolution and years of development, training and learn-
ing. Moreover, in the case of manipulation, we have
co-evolved environments, objects and tools to match the
capabilities of our hands. The design of airplane cock-
pits, left- and right-handed scissors, frets in string in-
struments, the key system in clarinets, and touch screens
are but a few examples.
Thus, biological hands in particular have an unfair ad-
vantage over robotic hands and prosthetics. Engineers
should explicitly begin to decide what functionality and
control to embed in the mechanics of the system, what
control algorithms to use for learning vs. standard per-
formance vs. elite performance vs. adaptation. It is
not unreasonable to propose that robotic hands, once
built, should undergo a developmental learning process
(a ‘robot kindergarten?’) to learn the speciﬁc control al-
gorithms, motor habits, and statistically useful anticipa-
tory strategies deﬁned by their intended use or—in the
case of prosthetics—the environment, job and prefer-
ences of their human pilot. Insisting on a one-size-ﬁts-
all, real-time control approach to robotic hands has been
shown to be overly ambitious, and even unnecessary as
demonstrated by the capabilities of the under-actuated
hands mentioned above (e.g., [49, 111]), as well as
grippers with no ﬁngers at all [216]. Some salient ex-
amples of such learning (and re-learning) come from
[239, 240, 241].
Prescriptive vs. Descriptive synergies
What are the debates in the study of synergies in bi-
ological systems? A root cause of the debates is the
nature of scientiﬁc inference based on experimental ob-
servations. The fact that experimental recordings detect
dimensionality reduction is not surprising because sen-
sorimotor control must, by deﬁnition, select motor ac-
tions from within the low-dimensional subspace of fea-
sible actions [184]. Therefore, disambiguating prescrip-
tive synergies of neural origin (those that are prescribed
by the nervous system as a control strategy) from de-
scriptive synergies (those that describe the expected di-
mensionality reduction) is diﬃcult [184, 51]. Thus, the
main question is not whether the nervous system inhab-
its a low-dimensional solution space to perform tasks,
but rather how it does so [242, 243, 184, 244]. More-
over, although several tasks can share the same general
features captured by such dimensionality reduction, it is
perhaps the ﬁne details particular to each task that may
be critical to their performance [51].
Biological ‘controllers’ co-evolved with mechanical
systems whose operations are characterized by a very
large number of elements—e.g., motor units, muscles,
and joints—while relying on their spatiotemporal co-
ordination and adaptability to task demands. The re-
ﬁnements aﬀorded by such evolution can be appreci-
ated when examining the eﬃcacy of the neural con-
trol of several complex motor behaviors, including, but
not limited to, speech production, locomotion, and ma-
nipulation. When examined in detail, researchers were
surprised and intrigued to see that those functionally
complex behaviors that involve the control of many
variables (like the 20+ angles for the joints of all ﬁn-
gers) in reality evolve in a lower-dimensional space
(i.e., can be well approximated by roughly 5 variables)
[50]. Such motor synergies have also been observed
in the phase-locked coordination (or correlated action)
of multiple muscles that produce complex behaviors
[245, 246, 194]
The theoretical framework of synergies has been ex-
tensively used and tested to account for the nervous sys-
tem’s ability to control multiple muscles and multi-joint
movements (for reviews see [247, 248, 249]). Synergies
would operate by constraining the spatial and temporal
activation of multiple muscles. Therefore, the existence
of consistent covariation patterns in electromyographic
(EMG) activity or joint excursions, whose structure can
be spatially and/or temporally modulated according to
task requirements, would be compatible with the syn-
ergy framework. Synergies have also been used as a
framework to understand pathological coordination of
12
movement (for reviews see [250, 251]). However, long-
standing issues remain regarding the extent to which
synergies can be considered ‘ﬁxed’ building blocks of
movements, the extent to which they are modiﬁable as
a function of task demands, adaptation and perceptual
context, as well as their very role in facilitating senso-
rimotor learning in tasks that may beneﬁt from, or be
penalized by a synergy-like control structure [157] (for
a review see [252]).
When considering the biomechanics of hand mus-
cles, the existence of anatomical constraints would sup-
port synergistic actions of the ﬁngers.
These con-
straints can come from, for example, ﬁnger muscle-
tendon complexes spanning several joints and passive
linkages among tendons [208]. Such synergistic actions
have been described as subject-independent ﬁnger kine-
matic patterns for grasping [50, 253] (for review see
[252]), as well as coupling of ﬁnger movement or forces
among non-instructed ﬁngers when humans are asked to
move or exert force with one ﬁnger [254, 255] (for re-
view see [207]). Early attempts to deﬁne the control of
individuated ﬁnger forces in cortical neuron activity re-
vealed a much more complex picture characterized by
broadly distributed activity [256]. More recent work in
non-human primates, however, supports an organization
of cortical activity that is compatible with the synergy
framework [257]. When searching for neural correlates
of synergies in humans, a recent study revealed that the
cortical representation of hand postures can be better ac-
counted for by using a synergy-based network than so-
matotopic or muscle-based models [258], which is com-
patible with the view of cortical organization of ﬁnger
movement being shaped by habitual use [259], and even
goal equivalence in ﬁnger actions being implemented at
a cortical level [260].
Many studies have attempted to identify synergies at
diﬀerent levels of biological systems and species, in-
cluding primary motor cortex [258], spinal cord [261],
motor units [262, 263, 264], motion [265, 50], and
forces [266, 267, 268]. However, the functional role of
synergies has been debated for decades, partly due to
the fact that the operational deﬁnition of synergies can
vary signiﬁcantly depending on several factors, includ-
ing the level of the system at which they are analyzed,
the methods used to quantify them, and the tasks used
to prove or negate their existence (as discussed in [252];
see also [269]). Among the conceptual frameworks that
have been proposed, synergies would be instrumental in
reducing the number of independent degrees of freedom
that the nervous system has to control as originally pro-
posed by Bernstein [192], or ensure attainment of task
goals by minimizing the variance that would be detri-
mental to performance [246, 270, 271]. An alternative
interpretation of the role of synergies, however, points
out the diﬃculty of interpreting synergies as the root
cause of multi-muscle coordination or a byproduct of
mechanical interactions between the biological system
and the environment [244].
Robotics, in contrast, synthetically designs, assem-
bles and operates engineered systems where synergies
can be prescribed. Over the past two decades, roboti-
cists have exploited the concept of (prescriptive) syner-
gies to design robotic hands (for review see [269]). Ex-
amples of these designs include the Pisa/IIT SoftHand
[49], whose design was based on the kinematic syner-
gies extracted from grasping a set of imagined objects
[50], as well as devices to constrain motion of human
ﬁngers for rehabilitation of sensorimotor function [272].
Here, the underlying design motivation is to capture
human-like kinematic features, i.e, simultaneous mo-
tion of all ﬁngers, by using a signiﬁcantly smaller num-
ber of actuators than joints. Preliminary clinical appli-
cations of this approach for prosthetic applications have
shown that individuals with upper limb loss can quickly
adopt such synergy-based design with minimal training
[273]. A major goal and challenge for robotic grasping
and manipulation is the implementation of force con-
trol using kinematic synergies. The results of compu-
tational modeling suggest that the ﬁrst few hand postu-
ral synergies may play an important role for attaining
force closure [274]. Nevertheless, it remains to be in-
vestigated the extent to which robotic motion-to-force
transition can fully leverage a synergy-based motion-to-
force coordination. Experimental evidence and theoret-
ical frameworks developed by studies of human multi-
ﬁnger synergies might potentially be used to inspire a
hierarchical control of high- and low-level grasp vari-
ables (i.e., task goal and distribution of individual ﬁn-
gertip forces, respectively) [137], as well as ‘default’
vs. task-dependent modulation of ﬁngertip force dis-
tributions [267]. Nevertheless, a major challenge, both
from neuroscientiﬁc and robotics perspective, is evalu-
ating the role of sensory feedback elicited by object con-
tact and force production on the coordination of multi-
ple (human and robotic) ﬁngers.
Conclusions
The literature on the biological and robotic ap-
proaches to grasp and manipulation is large, and has ex-
perienced exponential growth in the past two decades.
The reinvigorated interest in this topic has come, as in
past conﬂicts, from governments attending to wounded
soldiers and civilians who survive traumatic loss of
13
limbs (e.g., the DARPA program to Revolutionize Pros-
thetics), from the need to improve the quality of life
of individuals with stroke, cerebral palsy and other
neurological conditions that now have greater survival
rates, and from recent advances in autonomous and hu-
manoid robots for whom manipulation remains a litmus
test for performance. A positive feature of the latest
developments has been the greater and fruitful cross-
fertilization between biology and robotics approaches.
It is no longer an aberration to have engineers who are
well versed in neuroscience working in robotics or neu-
roscience, or neuroscientists who are well versed in en-
gineering working on scientiﬁc problems or robotics.
This greater interaction across ﬁelds, however, has the
added burden of needing to understand and keep cur-
rent on two vast and rapidly growing ﬁelds—which has
led to confusion of terms and principles, duplication of
eﬀorts, loss of nuance in translation, and lack of famil-
iarity with fundamental concepts.
Thus, our goal here is to provide researchers work-
ing in these ﬁelds the briefest of overviews of advances
(conceptual and material), pitfalls, and open questions.
In particular, we aimed to provide speciﬁc examples to
clarify when and why a better understanding of the bi-
ology of grasp and manipulation would beneﬁt robotic
grasping and manipulation. Namely, the evolutionary
process has yielded organisms that can produce ver-
satile function even when cherished engineering prin-
ciples are not present and biological systems operate
well in spite of having noise, delays, nonlinearities, etc.
There are many lessons learned, described above, such
as hierarchical and open-loop control, morphology that
simpliﬁes the control problem, the utility of having mul-
tiple muscles, etc. Conversely, engineering thinking has
provided well-founded principles of mechanics, mathe-
matics and control engineering to aid the scientiﬁc work
aimed at understanding the abilities of human hand.
Some salient points are:
• Deductive vs. synthetic science
Engineering approaches to understand biology—
and biology as an inspiration to engineering—have
generated a rich repertoire of experimental and the-
oretical advances in both areas. However, it is im-
portant to be aware of fundamental diﬀerences and
inherent limitations of scientiﬁc vs. synthetic ap-
proaches. It is the recognition of these diﬀerences
that needs to guide work in each ﬁeld and in their
interactions. This will allow those interactions to
be as fruitful as possible.
• Redundancy
The mathematics of robotics has contributed
greatly to our ability to study biological limbs
and hands.
Unfortunately,
the joint-torque
formulation—while valid and correct—can over-
simplify the nature of the problem that confronts
the nervous system. At its core, the control prob-
lem the nervous system faces is one of linear actu-
ators (muscles) that can only actively pull on ten-
dons. We have revisited the redundancy problem
while also emphasizing that the control of forces
is distinct from the control of movements. This
has allowed us to clarify many aspects of biolog-
ical function and dysfunction. For example, it is
now clear that muscle redundancy to produce a set
of net joint torques is not as great as once thought.
This is because the set of feasible muscle activa-
tions has a very strong structure. Similarly, kine-
matic redundancy is greatly constrained because
diﬀerent movements are not entirely equivalent.
From the perspective of muscle excursions and the
regulation of stretch reﬂexes, each natural move-
ment in fact represents an over-determined prob-
lem (the opposite of redundant), where muscle ac-
tivations and reﬂex modulation must follow very
speciﬁc spatiotemporal patterns. This all begins
to explain why it is so diﬃcult to learn to use our
limbs and hands (i.e., it takes years of practice),
why dysfunction arises even from minor neuro-
muscular damage, why rehabilitation is so diﬃcult,
and why limbs have evolved to have ‘so many’
muscles. Roboticists can use these lessons to re-
visit their design speciﬁcations, which often prefer
torque motors or a sparse architecture where each
joint is controlled by two dedicated tendons.
• From motor towards sensorimotor
Sensory feedback plays critical roles in human
grasp and manipulation for building internal rep-
resentations that are considered critical for pre-
dicting sensory consequences of hand-object in-
teractions; and for online sensing of those inter-
actions.
This dual nature of sensory feedback
(e.g., exploration-exploitation; learning-execution)
is particularly challenging to capture in robotics.
From the technical perspective, replicating human
tactile function has proven diﬃcult, although there
have been recent advances not only in sensor hard-
ware development; but also in the expansion of
prostheses to explicitly add hand proprioception
and touch interfaces (i.e., DARPA’s HAPTIX pro-
gram).
But equally important, the ﬁeld has be-
gun to recognize that the use and interpreting sen-
sory function may change as the learning pro-
14
cess evolves. Thus, as in biological systems, on-
line monitoring and processing of sensory infor-
mation is necessary at ﬁrst, but can actually be
counterproductive and slow down the performance
once learning has stabilized. Thus, hierarchical,
distributed and state-dependent gating, processing
and use of sensory information is yet to be well un-
derstood in human hands, and should be a goal of
robotic systems.
• Grasp is not manipulation
Grasp and dexterous manipulation, although often
interchangeably referred to in neuroscience litera-
ture, are fundamentally diﬀerent and challenge bi-
ological and robotic controllers in unique ways.
In robotics, greater advances have been made in
grasping than manipulation. Some important ad-
vances are being made to understand and imple-
ment the ability of ﬁngertips to produce motion-
to-force transitions, force-to-position modulation,
and in-hand object re-orientation/re-conﬁguration,
mostly in simulation [275, 276]. The challenge is
now how to design and build robotic hands with
the mechanical, motor and sensory abilities to im-
plement those behaviors. This is a very promising
and necessary direction for research. On the bio-
logical front, the new approaches to understand dy-
namic dexterous manipulation with the ﬁngertips
allows the quantiﬁcation of performance. But more
importantly, those ﬁndings challenge the dominant
cortico-centric view of manipulation, and highlight
the importance of considering subcortical, spinal
and neuromechanical contributors.
This further
motivates the exploration of hierarchical and dis-
tributed mechanisms that allow such versatile ca-
pabilities in the presence of delays, noise and un-
certainty. Such evolution in our thinking promises
to revolutionize our understanding of the mecha-
nisms that enable healthy function, explain disabil-
ity, and inform rehabilitation. Combining future
developments along these parallel lines of work
also promises important developments in the evo-
lution of the design and control of robotic and pros-
thetic hands.
• Neuromorphic vs. neuromimetic
Unlike neuromimetic approaches aimed at cap-
turing aspects of neural function without adher-
ing to the replication of biological structures or
processes, neuromorphic approaches seek to repli-
cate (to the extent possible possible) the biological
components and mechanisms at a particular scale.
This allows us to ask whether and how neurome-
chanical function can emerge without being de-
ﬁned a priori. This allows us to understand the ex-
tent to which the implementation of the structures
and processes (i.e., spiking neurons, delayed trans-
missions, muscle nonlinearities) deﬁne and con-
tribute to function, and the diﬀerent presentations
of dysfunction. Thus, neomorphism is proposed as
a means to reconcile engineering and biology, as
well as accelerate their cross-fertilization.
• Prescriptive vs. descriptive synergies
Synergies have been deﬁned and studied by neu-
roscientists as building blocks of complex move-
ments based on observations made at diﬀerent lev-
els of the neuromuscular system. Although their
fundamental characteristics and functional role re-
main to be established, the concept of synergies
has inspired robotic and prosthetic hand design
not only to simplify control, but also to capture
human-like motion features—the latter objective
being particularly important for the assistive tech-
nologies where there are few degrees of freedom
for control and the prosthetics are underactuated.
From the conceptual perspective, however, it is im-
portant that we recognize that the execution of a
task will always exhibit dimensionality reduction
in the control, kinetic and kinematic variables be-
cause they are, by deﬁnition, inhabiting the sub-
space of feasible actions that satisfy the constraints
that deﬁne the task. Therefore, it is to be expected
that (descriptive) synergies will be detected when
studying human function. Nevertheless, interest-
ing, valid and open questions remain in our un-
derstanding of whether and how the healthy and
damaged nervous system implements (prescrip-
tive) synergies to inhabit those subspaces (i.e., the
feasible activation manifold for a given task).
It is our hope and expectation that these recent con-
ceptual clariﬁcations, advances and newly-deﬁned open
problems will accelerate our understanding of healthy
and pathologic hand function (and biological function
in general), and will catalyze the creation of truly versa-
tile and dexterous robotic hands and prostheses.
Declarations
Ethics approval and consent to participate
Not applicable in this review article.
Consent for publication
The authors consent.
15
Availability of data and material
Not applicable in this review article.
Competing interests
FV-C holds US Patent No. 6,537,075 on some of the
technology reviewed in this article that is commercial-
ized by Neuromuscular Dynamics, LLC. MS has no ﬁ-
nancial or personal relationships with other people or
organizations that could inappropriately inﬂuence this
work.
Funding
Research reported in this publication was supported
by the National Institute of Arthritis and Musculoskele-
tal and Skin Diseases of the National Institutes of Health
(NIH) under award numbers AR050520 and AR052345
to FV-C, and Eunice Kennedy Shriver National Insti-
tute Of Child Health and Human Development of the
NIH under award number HD081938, National Science
Foundation (NSF) under award number BCS-1455866,
and the Grainger Foundation to MS. The content is
solely the responsibility of the authors and does not nec-
essarily represent the oﬃcial views of the NIH or NSF.
Authors’ contributions
FV-C and MS contributed to the conceptualization
and writing of this review article.
Acknowledgements
We thank Dr. Qiushi Fu, Dr. Christopher Laine, Dr.
Kian Jalaleddini, Suraj Chakravarthi Raja, Ali Marja-
ninejad, Melisa Osborne, and Jamie Flores for their help
with critical reading of this manuscript.
References cited
References
[1] M. Nussbaum, Aristotle’s De motu animalium, Princeton Univ
Pr, 1985.
[2] F. R. Wilson, The hand: How its use shapes the brain, lan-
guage, and human culture, Vintage, 2010.
[3] C. Bell, The hand: its mechanism and vital endowments, as
evincing design, volume 4, Bell & Daldy, 1865.
[4] F. J. Valero-Cuevas, Why the hand?, Adv. Exp. Med. Biol. 629
(2009) 553–557.
[5] F. Cordella, A. L. Ciancio, R. Sacchetti, A. Davalli, A. G. Cutti,
E. Guglielmelli, L. Zollo, Literature review on needs of upper
limb prosthesis users, Frontiers in neuroscience 10 (2016).
[6] L. L¨oﬄer, Der Ersatz f¨ur die Obere Extremit¨at: die Entwick-
lung von den ersten Zeugnissen bis heute, Ferdinand Enke Ver-
lag, 1984.
[7] A. J. Thurston, Par´e and prosthetics: the early history of artiﬁ-
cial limbs, ANZ journal of surgery 77 (2007) 1114–1119.
[8] K. Norton, A brief history of prosthetics, InMotion 17 (2007)
11–13.
[9] M. Hillman,
2 rehabilitation robotics from past to present–
a historical perspective,
in:
Advances in Rehabilitation
Robotics, Springer, 2004, pp. 25–44.
[10] I. G. Schlesinger,
Der mechanische aufbau der k¨unstlichen
glieder, in: Ersatzglieder und Arbeitshilfen, Springer, 1919,
pp. 321–661.
[11] C. A. F. Kluge,
H. E. Fritze,
Arthroplastik oder die
s¨ammtlichen, bisher bekannt gewordenen k¨unstlichen H¨ande
und F¨usse, zum Ersatz dieser verloren gegangenen Glied-
massen: mit 26 in Stein gravirten Tafeln, Meyer, 1842.
[12] R. Panchasi, Reconstructions: prosthetics and the rehabilita-
tion of the male body in world war i france, Diﬀerences: A
Journal of Feminist Cultural Studies 7 (1995) 109–141.
[13] D. Serlin, Artiﬁcial parts, practical lives: modern histories of
prosthetics, NYU Press, 2002.
[14] K. McSorley, War and the body: militarisation, practice and
experience, Routledge, 2012.
[15] G. S. Ling, P. Rhee, J. M. Ecklund, Surgical innovations arising
from the iraq and afghanistan wars, Annual review of medicine
61 (2010) 457–468.
[16] J. Riskin, Eighteenth-century wetware, Bernadette Bensaude-
Vincent and William R. Newmaneds. The Artiﬁcial and the
Natural: an Evolving Polarity, Cambridge, Mass (2007) 239–
74.
[17] J. Feinglass, W. H. Pearce, G. J. Martin, J. Gibbs, D. Cowper,
M. Sorensen, W. G. Henderson, J. Daley, S. Khuri, Postoper-
ative and late survival outcomes after major amputation: ﬁnd-
ings from the department of veterans aﬀairs national surgical
quality improvement program, Surgery 130 (2001) 21–29.
[18] M.
T.
Kristensen,
G.
Holm,
K.
Kirketerp-Møller,
M. Krasheninnikoﬀ, P. Gebuhr,
Very low survival rates
after non-traumatic lower limb amputation in a consecutive
series: what to do?,
Interactive cardiovascular and thoracic
surgery 14 (2012) 543–547.
[19] P. McCorduck, M. Minsky, O. G. Selfridge, H. A. Simon, His-
tory of artiﬁcial intelligence., in: IJCAI, pp. 951–954.
[20] A. Benko, C. S. L´anyi, History of artiﬁcial intelligence, in:
Encyclopedia of Information Science and Technology, Second
Edition, IGI Global, 2009, pp. 1759–1762.
[21] N. Gershenfeld, Fab: the coming revolution on your desktop–
from personal computers to personal fabrication, Basic Books,
2008.
16
[22] J. D. Prince, 3d printing: an industrial revolution, Journal of
electronic resources in medical libraries 11 (2014) 39–45.
[23] M. Honarpardaz, M. Tarkian, J. ¨Olvander, X. Feng,
Fin-
ger design automation for industrial robot grippers: A review,
Robotics and Autonomous Systems 87 (2017) 104 – 119.
[24] R. F. Weir, J. W. Sensinger, Design of artiﬁcial arms and hands
for prosthetic applications, 2003.
[25] D. S. Childress, Historical aspects of powered limb prostheses,
Clin Prosthet Orthot 9 (1985) 2–13.
[26] T. L. Wellerson, A Manual for Occupational Therapists on the
Rehabilitation of Upper Extremity Amputees, Kendall/Hunt
Publishing Company, 1958.
[27] M. R. Cutkosky, On grasp choice, grasp models, and the de-
sign of hands for manufacturing tasks, IEEE Transactions on
robotics and automation 5 (1989) 269–279.
[28] J. E. Colgate, N. Hogan, Robust control of dynamically in-
teracting systems, International journal of Control 48 (1988)
65–88.
[29] R. Murray, Z. Li, S. Sastry, A mathematical introduction to
robotic manipulation, CRC, 1994.
[30] M. Uchiyama, A. Konno, T. Uchiyama, S. Kanda,
Devel-
opment of a ﬂexible dual-arm manipulator testbed for space
robotics, in: Intelligent Robots and Systems’ 90.’Towards a
New Frontier of Applications’, Proceedings. IROS’90. IEEE
International Workshop on, IEEE, pp. 375–381.
[31] B. A. Aikenhead, R. G. Daniell, F. M. Davis, Canadarm and
the space shuttle, Journal of Vacuum Science & Technology
A: Vacuum, Surfaces, and Films 1 (1983) 126–132.
[32] C. C. Nguyen, F. J. Pooran, Dynamic analysis of a 6 dof ckcm
robot end-eﬀector for dual-arm telerobot systems,
Robotics
and Autonomous Systems 5 (1989) 377–394.
[33] J. Winslow, Exposition anatomique de la structure du corps hu-
main, chez Guillaume Desprez... et Jean Desessartz, 1732.
[34] G. Galilei, Il saggiatore, Italian translation by L. Sosio (Mi-
lano, Feltrinelli, 1979).
[35] F. Valero-Cuevas,
H. Hoﬀmann,
M. Kurse,
J. Kutch,
E. Theodorou, Computational models for neuromuscular func-
tion, Biomedical Engineering, IEEE Reviews in 2 (2009) 110–
135.
[36] J. W. Krakauer, A. A. Ghazanfar, A. Gomez-Marin, M. A.
MacIver, D. Poeppel, Neuroscience needs behavior: correcting
a reductionist bias, Neuron 93 (2017) 480–490.
[37] Wikipedia
contributors,
Deductivereason-
ing,
Wikipedia,
The
Free
Encyclopedia.
http://en.wikipedia.org/wiki/Deductive_reasoning,
2017. Accessed 19 August 2017.
[38] Wikipedia
contributors,
Inductivereason-
ing,
Wikipedia,
The
Free
Encyclopedia.
https://en.wikipedia.org/wiki/Inductive_reasoning,
2017. Accessed 19 August 2017.
[39] I. Copi, Essentials of logic, Taylor & Francis, 2016.
[40] H. van Duinen, S. C. Gandevia, Constraints for control of the
human hand,
The Journal of physiology 589 (2011) 5583–
5593.
[41] M. Santello, Getting a grasp of theories of sensorimotor control
of the hand: Identiﬁcation of underlying neural mechanisms,
Motor control 19 (2015) 149–153.
[42] S. Jacobs, C. Danielmeier, S. H. Frey, Human anterior intra-
parietal and ventral premotor cortices support representations
of grasping with the hand or a novel tool, Journal of Cognitive
Neuroscience 22 (2010) 2594–2608.
[43] M. Davare, A. Kraskov, J. C. Rothwell, R. N. Lemon, Interac-
tions between areas of the cortical grasping network, Current
opinion in neurobiology 21 (2011) 565–570.
[44] R. N. Lemon, Descending pathways in motor control, Annu.
Rev. Neurosci. 31 (2008) 195–218.
[45] K. Mosier, C. Lau, Y. Wang, M. Venkadesan, F. J. Valero-
Cuevas, Controlling instabilities in manipulation requires spe-
ciﬁc cortical-striatal-cerebellar networks,
Journal of neuro-
physiology 105 (2011) 1295–1305.
[46] N. Ejaz, M. Hamada, J. Diedrichsen, Hand use predicts the
structure of representations in sensorimotor cortex,
Nature
neuroscience 18 (2015) 1034–1040.
[47] E. Jonas, K. P. Kording,
Could a neuroscientist understand
a microprocessor?, PLOS Computational Biology 13 (2017)
e1005268.
[48] F. J. Valero-Cuevas, V. V. Anand, A. Saxena, H. Lipson, Be-
yond parameter estimation: extending biomechanical model-
ing by the explicit exploration of model topology, IEEE Trans
Biomed Eng 54 (2007) 1951–1964.
[49] M. G. Catalano, G. Grioli, E. Farnioli, A. Serio, C. Piazza,
A. Bicchi, Adaptive synergies for the design and control of
the pisa/iit softhand, The International Journal of Robotics Re-
search 33 (2014) 768–782.
[50] M. Santello, M. Flanders, J. F. Soechting, Postural hand syn-
ergies for tool use, Journal of Neuroscience 18 (1998) 10105–
10115.
[51] O. Brock, F. Valero-Cuevas, Transferring synergies from neu-
roscience to robotics: Comment on “hand synergies: Integra-
tion of robotics and neuroscience for understanding the control
of biological and artiﬁcial hands” by m. santello et al., Physics
of life reviews 17 (2016) 27–32.
[52] R. Pfeifer, G. G´omez, Morphological computation–connecting
brain, body, and environment, Creating brain-like intelligence
(2009) 66–83.
[53] F. J. Valero-Cuevas, J.-W. Yi, D. Brown, R. V. McNamara,
C. Paul, H. Lipson, The tendon network of the ﬁngers performs
anatomical computation at a macroscopic scale, IEEE Trans-
actions on Biomedical Engineering 54 (2007) 1161–1166.
[54] R. M. Alexander, Animal mechanics (1968).
[55] K. L. Mardula, R. Balasubramanian, C. H. Allan, Implanted
passive engineering mechanism improves hand function af-
ter tendon transfer surgery: a cadaver-based study, Hand 10
(2015) 116–122.
[56] L. U. Odhner, L. P. Jentoft, M. R. Claﬀee, N. Corson, Y. Tenzer,
R. R. Ma, M. Buehler, R. Kohout, R. D. Howe, A. M. Dollar,
A compliant, underactuated hand for robust manipulation, The
International Journal of Robotics Research 33 (2014) 736–752.
[57] F. J. Valero-Cuevas, Fundamentals of neuromechanics, vol-
ume 8 of Biosystems and Biorobotics, Springer Verlag London,
2015.
[58] K. Ogata, Modern control engineering, Prentice Hall, Upper
Saddle River, N.J., 3rd edition, 1997. Katsuhiko Ogata. In-
cludes bibliographical references (p. 983-986) and index.
[59] R. F. Stengel, Optimal control and estimation, Courier Corpo-
ration, 2012.
[60] M. Verhaegen, V. Verdult, Filtering and system identiﬁcation:
a least squares approach, Cambridge university press, 2007.
[61] L. Ljung, System identiﬁcation, Wiley Online Library, 1999.
[62] F. C. Van der Helm, A. C. Schouten, E. de Vlugt, G. G.
Brouwn, Identiﬁcation of intrinsic and reﬂexive components
of human arm dynamics during postural control, Journal of
neuroscience methods 119 (2002) 1–14.
[63] K. Jalaleddini, E. S. Tehrani, R. E. Kearney, A subspace ap-
proach to the structural decomposition and identiﬁcation of an-
kle joint dynamic stiﬀness, IEEE transactions on biomedical
engineering 64 (2017) 1357–1368.
[64] E. J. Perreault, R. F. Kirsch, A. M. Acosta,
Multiple-input,
multiple-output system identiﬁcation for characterization of
limb stiﬀness dynamics, Biological cybernetics 80 (1999) 327–
17
337.
[65] J. M. Hollerbach, D. M. Lokhorst, Closed-loop kinematic cal-
ibration of the rsi 6-dof hand controller, IEEE Transactions on
Robotics and Automation 11 (1995) 352–359.
[66] J. E. Bobrow, B. W. McDonell, Modeling, identiﬁcation, and
control of a pneumatically actuated, force controllable robot,
IEEE transactions on Robotics and Automation 14 (1998) 732–
742.
[67] R. Johansson, A. Robertsson, K. Nilsson, M. Verhaegen, State-
space system identiﬁcation of robot manipulator dynamics,
Mechatronics 10 (2000) 403–418.
[68] F. Lewis, S. Jagannathan, A. Yesildirak, Neural network con-
trol of robot manipulators and non-linear systems, CRC Press,
1998.
[69] R. Ortega, M. W. Spong,
Adaptive motion control of rigid
robots: A tutorial, Automatica 25 (1989) 877–888.
[70] J. C. Doyle, Guaranteed Margins for LQG Regulators, IEEE
Transactions on Automatic Control AC-23 (1978) 756–757.
[71] G. Williams, P. Drews, B. Goldfain, J. M. Rehg, E. A.
Theodorou, Aggressive driving with model predictive path in-
tegral control,
in: Robotics and Automation (ICRA), 2016
IEEE International Conference on, IEEE, pp. 1433–1440.
[72] R. Enoka, Neuromechanical Basis of Kinesiology, Human Ki-
netic Books, Champaign, Illinois, 1988.
[73] S. Lisberger, W. Thach, E. Kandel, J. Schwartz, T. Jessell,
S. Siegelbaum, A. Hudspeth, Principles of neural science, New
York City, NY: McGraw-Hill Companies (2013).
[74] C. S. Sherrington,
Reﬂex inhibition as a factor in the co-
ordination of movements and postures, Experimental Physi-
ology 6 (1913) 251–310.
[75] C. M. Niu, K. Jalaleddini, W. J. Sohn, J. Rocamora, T. D.
Sanger, F. J. Valero-Cuevas, Neuromorphic meets neurome-
chanics, part i: the methodology and implementation, Journal
of Neural Engineering 14 (2017) 025001.
[76] K. Jalaleddini, C. M. Niu, S. C. Raja, W. J. Sohn, G. E. Loeb,
T. D. Sanger, F. J. Valero-Cuevas, Neuromorphic meets neu-
romechanics, part ii: the role of fusimotor drive, Journal of
Neural Engineering 14 (2017) 025002.
[77] A. L. Hodgkin, A. F. Huxley, The dual eﬀect of membrane
potential on sodium conductance in the giant axon of loligo,
The Journal of physiology 116 (1952) 497.
[78] B. A. Olshausen, D. J. Field,
Emergence of simple-cell re-
ceptive ﬁeld properties by learning a sparse code for natural
images, Nature 381 (1996) 607.
[79] R. S. Johansson, J. R. Flanagan,
Coding and use of tactile
signals from the ﬁngertips in object manipulation tasks, Nature
Reviews Neuroscience 10 (2009) 345–359.
[80] M. O. Ernst, M. S. Banks, Humans integrate visual and hap-
tic information in a statistically optimal fashion, Nature 415
(2002) 429–433.
[81] D. Shibata, A. M. Kappers, M. Santello, Digit forces bias sen-
sorimotor transformations underlying control of ﬁngertip posi-
tion, Frontiers in human neuroscience 8 (2014) 564.
[82] D. M. Wolpert, Z. Ghahramani, Computational principles of
movement neuroscience, nature neuroscience 3 (2000) 1212–
1217.
[83] D. M. Wolpert, Z. Ghahramani, J. R. Flanagan, Perspectives
and problems in motor learning, Trends in cognitive sciences
5 (2001) 487–494.
[84] R. Johansson, G. Westling, Roles of glabrous skin receptors
and sensorimotor memory in automatic control of precision
grip when lifting rougher or more slippery objects,
Experi-
mental brain research 56 (1984) 550–564.
[85] R. S. Johansson, G. Westling, Coordinated isometric muscle
commands adequately and erroneously programmed for the
weight during lifting task with precision grip, Experimental
Brain Research 71 (1988) 59–71.
[86] T. Celikel, B. Sakmann, Sensory integration across space and
in time for decision making in the somatosensory system of
rodents,
Proceedings of the National Academy of Sciences
104 (2007) 1395–1400.
[87] J. H. Solomon, M. J. Hartmann,
Biomechanics:
robotic
whiskers used to sense features, Nature 443 (2006) 525–525.
[88] S. J. Lederman, R. L. Klatzky, Hand movements: A window
into haptic object recognition, Cognitive psychology 19 (1987)
342–368.
[89] S. J. Lederman, R. L. Klatzky,
Extracting object properties
through haptic exploration, Acta psychologica 84 (1993) 29–
40.
[90] B. A. Olshausen, 20 years of learning about vision: Questions
answered, questions unanswered, and questions not yet asked,
in: 20 Years of Computational Neuroscience, Springer, 2013,
pp. 243–270.
[91] E. Eggermann, Y. Kremer, S. Crochet, C. C. Petersen, Cholin-
ergic signals in mouse barrel cortex during active whisker sens-
ing, Cell reports 9 (2014) 1654–1660.
[92] Å. B. Vallbo, R. S. Johansson, et al., Properties of cutaneous
mechanoreceptors in the human hand related to touch sensa-
tion, Hum Neurobiol 3 (1984) 3–14.
[93] B. Siciliano, L. Sciavicco, L. Villani, G. Oriolo, Robotics:
modelling, planning and control, Springer Science & Business
Media, 2010.
[94] T. Uhl, The inverse identiﬁcation problem and its technical
application,
Archive of Applied Mechanics 77 (2007) 325–
337.
[95] D. T. Westwick, R. E. Kearney, Identiﬁcation of nonlinear
physiological systems, volume 7, John Wiley & Sons, 2003.
[96] Y. Yang, T. Solis-Escalante, J. Yao, F. C. van der Helm, J. P.
Dewald, A. C. Schouten, Nonlinear connectivity in the human
stretch reﬂex assessed by cross-frequency phase coupling, In-
ternational Journal of Neural Systems 26 (2016) 1650043.
[97] K. Jalaleddini, R. E. Kearney, Subspace identiﬁcation of siso
hammerstein systems: application to stretch reﬂex identiﬁca-
tion, IEEE Transactions on Biomedical Engineering 60 (2013)
2725–2734.
[98] E. de Vlugt, A. C. Schouten, F. C. van der Helm, Closed-loop
multivariable system identiﬁcation for the characterization of
the dynamic arm compliance using continuous force distur-
bances: a model study, Journal of neuroscience methods 122
(2003) 123–140.
[99] D. Ludvig, E. J. Perreault, System identiﬁcation of physiolog-
ical systems using short data segments, IEEE Transactions on
Biomedical Engineering 59 (2012) 3541–3549.
[100] V. Shamanna, S. Das, Z. C¸ elik-Butler, D. P. Butler, K. L.
Lawrence,
Micromachined integrated pressure–thermal sen-
sors on ﬂexible substrates,
Journal of Micromechanics and
Microengineering 16 (2006) 1984.
[101] M. Lowe, A. King, E. Lovett, T. Papakostas, Flexible tactile
sensor technology: bringing haptics to life, Sensor review 24
(2004) 33–36.
[102] G. E. Loeb, R. Johansson, Biomimetic tactile sensor, 2010. US
Patent 7,658,119.
[103] N. Wettels, V. J. Santos, R. S. Johansson, G. E. Loeb,
Biomimetic tactile sensor array, Advanced Robotics 22 (2008)
829–849.
[104] J. Fishel, G. Lin, G. Loeb, Syntouch llc biotac product manual,
v. 16, Tech. Rep. (2013).
[105] K. Horch, S. Meek, T. G. Taylor, D. T. Hutchinson, Object dis-
crimination with an artiﬁcial hand using electrical stimulation
of peripheral tactile and proprioceptive pathways with intrafas-
18
cicular electrodes, IEEE Transactions on Neural Systems and
Rehabilitation Engineering 19 (2011) 483–489.
[106] G. A. Tabot, J. F. Dammann, J. A. Berg, F. V. Tenore, J. L.
Boback, R. J. Vogelstein, S. J. Bensmaia, Restoring the sense
of touch with a prosthetic hand through a brain interface, Pro-
ceedings of the National Academy of Sciences 110 (2013)
18279–18284.
[107] S. Raspopovic, M. Capogrosso, F. M. Petrini, M. Bonizzato,
J. Rigosa, G. Di Pino, J. Carpaneto, M. Controzzi, T. Boretius,
E. Fernandez, G. Granata, C. M. Oddo, L. Citi, A. L. Cian-
cio, C. Cipriani, M. C. Carrozza, W. Jensen, E. Guglielmelli,
T. Stieglitz, P. M. Rossini, S. Micera, Restoring Natural Sen-
sory Feedback in Real-Time Bidirectional Hand Prostheses,
Science Translational Medicine 6 (2014) 222ra19–222ra19.
[108] J. Kim, M. Lee, H. J. Shim, R. Ghaﬀari, H. R. Cho, D. Son,
Y. H. Jung, M. Soh, C. Choi, S. Jung, et al., Stretchable silicon
nanoribbon electronics for skin prosthesis, Nature communi-
cations 5 (2014) 5747.
[109] C. Antfolk, M. D’Alonzo, M. Controzzi, G. Lundborg,
B. Ros´en, F. Sebelius, C. Cipriani,
Artiﬁcial redirection of
sensation from prosthetic ﬁngers to the phantom hand map on
transradial amputees: vibrotactile versus mechanotactile sen-
sory feedback, IEEE transactions on neural systems and reha-
bilitation engineering 21 (2013) 112–120.
[110] F. J. Valero-Cuevas, J. W. Yi, D. Brown, R. V. McNamara III,
C. Paul, H. Lipson, The tendon network of the ﬁngers performs
anatomical computation at a macroscopic scale, IEEE Trans
Biomed Eng 54 (2007) 1161–6.
[111] R. Deimel, O. Brock, A novel type of compliant and underac-
tuated robotic hand for dexterous grasping, The International
Journal of Robotics Research 35 (2016) 161–185.
[112] Y. Imai, A. Namiki, K. Hashimoto, M. Ishikawa, Dynamic ac-
tive catching using a high-speed multiﬁngered hand and a high-
speed vision system, in: Robotics and Automation, 2004. Pro-
ceedings. ICRA’04. 2004 IEEE International Conference on,
volume 2, IEEE, pp. 1849–1854.
[113] H. Schulte, The characteristics of the mckibben artiﬁcial mus-
cle, The application of external power in prosthetics and or-
thotics 874 (1961) 94–115.
[114] M. Gavrilovi´c, M. Mari´c,
Positional servo-mechanism acti-
vated by artiﬁcial muscles, Medical and Biological Engineer-
ing and Computing 7 (1969) 77–82.
[115] C.-P. Chou, B. Hannaford,
Measurement and modeling of
mckibben pneumatic artiﬁcial muscles, IEEE Transactions on
robotics and automation 12 (1996) 90–102.
[116] T. Kodama, A. Okabe, K. Kogiso, Simultaneous estimation
of contraction ratio and parameter of mckibben pneumatic ar-
tiﬁcial muscle model using log-normalized unscented kalman
ﬁlter, in: Cyber-Physical Systems, Networks, and Applications
(CPSNA), 2016 IEEE 4th International Conference on, IEEE,
pp. 44–48.
[117] K. E. Gordon, G. S. Sawicki, D. P. Ferris, Mechanical perfor-
mance of artiﬁcial pneumatic muscles to power an ankle–foot
orthosis, Journal of biomechanics 39 (2006) 1832–1841.
[118] R. W. Jackson, S. H. Collins, An experimental comparison of
the relative beneﬁts of work and torque assistance in ankle ex-
oskeletons, Journal of Applied Physiology (2015) jap–01133.
[119] R. Van Ham, T. G. Sugar, B. Vanderborght, K. W. Hollander,
D. Lefeber,
Compliant actuator designs,
IEEE Robotics &
Automation Magazine 16 (2009).
[120] J. Pons, H. Rodriguez, I. Luyckx, D. Reynaerts, R. Ceres, H. V.
Brussel, High torque ultrasonic motors for hand prosthetics:
current status and trends,
Technology and Health Care 10
(2002) 121–133.
[121] R. Bogue, Exoskeletons and robotic prosthetics: a review of
recent developments, Industrial Robot: An International Jour-
nal 36 (2009) 421–427.
[122] M. A. Rahman, X. Wang, C. Wen, A review of high energy
density lithium–air battery technology,
Journal of Applied
Electrochemistry 44 (2014) 5–22.
[123] B. Scrosati, J. Garche, Lithium batteries: Status, prospects and
future, Journal of Power Sources 195 (2010) 2419–2430.
[124] A. Yoshino, Development of the lithium-ion battery and recent
technological trends, in: G. Pistoia (Ed.), Lithium-Ion Batter-
ies, Elsevier, 2014, pp. 1–20.
[125] M. H. Dickinson, C. T. Farley, R. J. Full, M. Koehl, R. Kram,
S. Lehman, How animals move: an integrative view, Science
288 (2000) 100–106.
[126] A. A. Biewener, Locomotion as an emergent property of mus-
cle contractile dynamics, Journal of Experimental Biology 219
(2016) 285–294.
[127] R. L. Lieber, Skeletal muscle structure and function: impli-
cations for rehabilitation and sports medicine, Williams &
Wilkins, 1992.
[128] R. M. Enoka, Neuromechanics of human movement, Human
kinetics, 2008.
[129] P. Martin, E. Johnson, T. Murphey, M. Egerstedt, Constructing
and implementing motion programs for robotic marionettes,
IEEE Transactions on Automatic Control 56 (2011) 902–907.
[130] S. Shinjiro, K. Andrew, K. P. Dinesh, Musculotendon simula-
tion for hand animation, ACM Trans. Graph. 27 (2008) 1–8.
1360682.
[131] D. M. Kaufman, T. Edmunds, D. K. Pai, Fast frictional dy-
namics for rigid bodies, in: International Conference on Com-
puter Graphics and Interactive Techniques, ACM New York,
NY, USA, 2005, pp. 946–956.
[132] Y. Mao, S. K. Agrawal, Design of a cable-driven arm exoskele-
ton (carex) for neural rehabilitation,
IEEE Transactions on
Robotics 28 (2012) 922–931.
[133] S.-R. Oh, S. K. Agrawal, Cable suspended planar robots with
redundant cables: Controllers with positive tensions,
IEEE
Transactions on Robotics 21 (2005) 457–465.
[134] D. M. Wolpert, J. Diedrichsen, J. R. Flanagan,
Principles
of sensorimotor learning., Nature Reviews Neuroscience 12
(2011) 739–51.
[135] M. Venkadesan, F. J. Valero-Cuevas, Neural control of motion-
to-force transitions with the ﬁngertip, J. Neurosci. 28 (2008)
1366–1373.
[136] Q. Fu, W. Zhang, M. Santello, Anticipatory planning and con-
trol of grasp positions and forces for dexterous two-digit ma-
nipulation., Journal of Neuroscience 30 (2010) 9117–26.
[137] Q. Fu, Z. Hasan, M. Santello,
Transfer of learned manipu-
lation following changes in degrees of freedom,
Journal of
Neuroscience 31 (2011) 13527–34.
[138] Q. Fu, M. Santello,
Coordination between digit forces and
positions: interactions between anticipatory and feedback con-
trol., Journal of Neurophysiology 111 (2014) 1519–28.
[139] Q. Fu, J. Y. Choi, A. M. Gordon, M. Jesunathadas, M. San-
tello, Learned manipulation at unconstrained contacts does not
transfer across hands, PLoS ONE 9 (2014).
[140] M. Marneweck, T. Lee-miller, M. Santello, A. M. Gordon,
A. M. Gordon, Digit Position and Forces Covary during An-
ticipatory Control of Whole-Hand Manipulation, Frontiers in
Human Neuroscience 10 (2016) 1–10.
[141] G. T. Yamaguchi, F. E. Zajac,
Restoring unassisted natural
gait to paraplegics via functionalneuromuscular stimulation: a
computer simulation study, IEEE Transactions on Biomedical
Engineering 37 (1990) 886–902.
[142] R. Shadmehr, S. Mussa-Ivaldi, Biological learning and con-
trol: how the brain builds representations, predicts events, and
19
makes decisions, Mit Press, 2012.
[143] E. Todorov, M. I. Jordan, Optimal feedback control as a theory
of motor coordination, Nature neuroscience 5 (2002) 1226–
1235.
[144] S. H. Scott,
Optimal feedback control and the neural basis
of volitional motor control, Nature Reviews Neuroscience 5
(2004) 532–546.
[145] G. Loeb, I. Brown, E. Cheng, A hierarchical foundation for
models of sensorimotor control, Experimental Brain Research
126 (1999) 1–18.
[146] G. Loeb, W. Levine, J. He, Understanding sensorimotor feed-
back through optimal control, in: Cold Spring Harbor sym-
posia on quantitative biology, volume 55, Cold Spring Harbor
Laboratory Press, pp. 791–803.
[147] R. Peterka, Sensorimotor integration in human postural con-
trol, Journal of neurophysiology 88 (2002) 1097–1118.
[148] M. C. Khoo, Physiological control systems, IEEE press, 2000.
[149] J. McIntyre, E. Bizzi, Servo hypotheses for the biological con-
trol of movement, Journal of motor behavior 25 (1993) 193–
202.
[150] D. E. Angelaki, K. E. Cullen,
Vestibular system: the many
facets of a multimodal sense, Annu. Rev. Neurosci. 31 (2008)
125–150.
[151] M. Ranjbaran, H. L. Galiana,
Hybrid model of the context
dependent vestibulo-ocular reﬂex: implications for vergence-
version interactions, Frontiers in computational neuroscience
9 (2015).
[152] M. Suver, M. Dickinson, Sensory integration by descending
interneurons in the ﬂying fruit ﬂy, Integrative and Comparative
Biology 56 (2016) E216.
[153] R. M. Harris-Warrick, General principles of rhythmogenesis in
central pattern networks, Progress in brain research 187 (2010)
213.
[154] S. Grillner,
Biological pattern generation: the cellular and
computational logic of networks in motion, Neuron 52 (2006)
751–766.
[155] T. Iberall, M. A. Arbib, Schemes for the control of hand, Vision
and action: The control of grasping 2 (1990) 204.
[156] C. L. MacKenzie, T. Iberall, The grasping hand, volume 104,
Elsevier, 1994.
[157] F. Mechsner, D. Kerzel, G. Knoblich, W. Prinz,
Perceptual
basis of bimanual coordination, Nature 414 (2001) 69–73.
[158] A. Charpentier, Analyse experimentale de quelques elements
de la sensation de poids, 1891.
[159] G. Yue, K. J. Cole, Strength increases from the motor program:
comparison of training with maximal voluntary and imagined
muscle contractions,
Journal of neurophysiology 67 (1992)
1114–1123.
[160] A. M. Gordon, G. Westling, K. J. Cole, R. S. Johansson, Mem-
ory representations underlying motor commands used during
manipulation of common and novel objects, Journal of Neuro-
physiology 69 (1993) 1789–1796.
[161] D. J. Murray, R. R. Ellis, C. A. Bandomir, H. E. Ross, Charpen-
tier (1891) on the size—weight illusion, Attention, Perception,
& Psychophysics 61 (1999) 1681–1685.
[162] J. R. Flanagan, M. A. Beltzner, Independence of perceptual and
sensorimotor predictions in the size–weight illusion, Nature
neuroscience 3 (2000) 737–741.
[163] J. P. Warren, M. Santello, S. I. H. Tillery, Eﬀects of fusion
between tactile and proprioceptive inputs on tactile perception,
PloS one 6 (2011) e18073.
[164] C. Craj´e, M. Santello, A. M. Gordon, Eﬀects of visual cues of
object density on perception and anticipatory control of dexter-
ous manipulation, PloS one 8 (2013) e76855.
[165] A. E. Bryson, Applied optimal control: optimization, estima-
tion and control, CRC Press, 1975.
[166] R. Kalman, On the general theory of control systems, IRE
Transactions on Automatic Control 4 (1959) 110–110.
[167] R. E. Kalman,
Mathematical description of linear dynami-
cal systems, Journal of the Society for Industrial and Applied
Mathematics, Series A: Control 1 (1963) 152–192.
[168] D. Simon, Optimal state estimation: Kalman, H inﬁnity, and
nonlinear approaches, John Wiley & Sons, 2006.
[169] A. De Rugy, G. E. Loeb, T. J. Carroll, Muscle coordination
is habitual rather than optimal,
Journal of Neuroscience 32
(2012) 7384–7391.
[170] G. E. Loeb, Optimal isn’t good enough, Biological cybernetics
106 (2012) 757–765.
[171] Z. Shiller, S. Dubowsky, Robot path planning with obstacles,
actuator, gripper, and payload constraints, The International
Journal of Robotics Research 8 (1989) 3–18.
[172] G. Strang, Introduction to linear algebra, Wellesley Cambridge
Pr, 2003.
[173] L. T. Aschepkov, D. V. Dolgy, T. Kim, R. P. Agarwal, Optimal
Control, Springer, 2016.
[174] M. G. Safonov, Origins of robust control: Early history and
future speculations, IFAC Proceedings Volumes 45 (2012) 1–
8.
[175] M. Morari, C. Garcia, J. Lee, D. Prett, Model predictive con-
trol, Prentice Hall Englewood Cliﬀs, NJ, 1993.
[176] E. Theodorou, J. Buchli, S. Schaal, A generalized path inte-
gral control approach to reinforcement learning.,
Journal of
Machine Learning Research (2010) 3137–3181.
[177] E. Y. Chao, K. N. An, Graphical interpretation of the solution
to the redundant problem in biomechanics, Journal of Biome-
chanical Engineering 100 (1978) 159–67.
[178] R. Crowninshield, R. Brand, A physiologically based criterion
of muscle force prediction in locomotion, Journal of Biome-
chanics 14 (1981) 793–801.
[179] L. Righetti, M. Kalakrishnan, P. Pastor, J. Binney, J. Kelly,
R. C. Voorhies, G. S. Sukhatme, S. Schaal, An autonomous
manipulation system based on force control and optimization,
Autonomous Robots 36 (2014) 11–30.
[180] C. G. Cifuentes, J. Issac, M. Wuthrich, S. Schaal, J. Bohg,
Probabilistic articulated real-time tracking for robot manipu-
lation, IEEE Robotics and Automation Letters (2016).
[181] V. Kumar, E. Todorov, Mujoco haptix: A virtual reality system
for hand manipulation, in: Humanoid Robots (Humanoids),
2015 IEEE-RAS 15th International Conference on, IEEE, pp.
657–663.
[182] F. J. Valero-Cuevas, F. E. Zajac, C. G. Burgar, Large index-
ﬁngertip forces are produced by subject-independent patterns
of muscle excitation, Journal of Biomechanics 31 (1998) 693–
704.
[183] F. Valero-Cuevas, B. Cohn, H. Yngvason, E. Lawrence, Ex-
ploring the high-dimensional structure of muscle redundancy
via subject-speciﬁc and generic musculoskeletal models, Jour-
nal of biomechanics 48 (2015) 2887–2896.
[184] J. J. Kutch, F. J. Valero-Cuevas,
Challenges and new ap-
proaches to proving the existence of muscle synergies of neural
origin, PLoS computational biology 8 (2012) e1002434.
[185] J. M. Inouye, F. J. Valero-Cuevas, Muscle synergies heavily in-
ﬂuence the neural control of arm endpoint stiﬀness and energy
consumption, PLoS Comput Biol 12 (2016) e1004737.
[186] K. P. K¨ording, D. M. Wolpert,
Bayesian decision theory in
sensorimotor control, Trends in cognitive sciences 10 (2006)
319–326.
[187] M. A. Peters, W. J. Ma, L. Shams, The size-weight illusion is
not anti-bayesian after all: a unifying bayesian account, PeerJ
4 (2016) e2124.
20
[188] T. D. Sanger, Learning Visually Guided Risk-Aware Reach-
ing on a Robot Controlled by a GPU Spiking Neural Network,
Springer International Publishing, Cham, pp. 282–289.
[189] A. Dunning, A. Ghoreyshi, M. Bertucco, T. D. Sanger, The
tuning of human motor response to risk in a dynamic environ-
ment task, PloS one 10 (2015) e0125461.
[190] E. Theodorou, E. Todorov, F. J. Valero-Cuevas, Neuromuscu-
lar stochastic optimal control of a tendon driven index ﬁnger
model, in: American Control Conference (ACC), 2011, IEEE,
pp. 348–355.
[191] J. Rieﬀel, F. Valero-Cuevas, H. Lipson, Morphological Com-
munication: Exploiting Coupled Dynamics in a Complex Me-
chanical Structure to Achieve Locomotion,
Journal of the
Royal Society Interface In Press (2009).
[192] N. A. Bernstein, The co-ordination and regulation of move-
ment, Pergamon Press, Oxford, UK, 1967.
[193] A. T. Miller, P. K. Allen, Graspit! a versatile simulator for
robotic grasping, IEEE Robotics and Automation Magazine
11 (2004) 110–122.
[194] S. Giszter, V. Patil, C. Hart, Primitives, premotor drives, and
pattern generation: a combined computational and neuroetho-
logical perspective, Progress in brain research 165 (2007) 323–
346.
[195] J. Ewart, Neuroethology–an introduction to the neurophysio-
logical fundamentals of behaviour, 1980.
[196] J. J. Kutch, F. J. Valero-Cuevas, Muscle redundancy does not
imply robustness to muscle dysfunction,
Journal of Biome-
chanics 44 (2011) 1264–1270.
[197] D. A. Hagen, F. J. Valero-Cuevas, Similar movements are asso-
ciated with drastically diﬀerent muscle contraction velocities,
Journal of Biomechanics (2017).
[198] P. Brand, A. Hollister, Clinical mechanics of the hand, St.
Louis: Mosby-Year Book, Inc, 1993.
[199] F. J. Valero-Cuevas, N. Smaby, M. Venkadesan, M. Peterson,
T. Wright, The strength-dexterity test as a measure of dynamic
pinch performance, J Biomech 36 (2003) 265–270.
[200] E. L. Lawrence, I. Fassola, I. Werner, C. Leclercq, F. J. Valero-
Cuevas, Quantiﬁcation of dexterity as the dynamical regulation
of instabilities: comparisons across gender, age, and disease
(2014).
[201] N.-h. Ko, C. M. Laine, B. E. Fisher, F. J. Valero-Cuevas, Force
variability during dexterous manipulation in individuals with
mild to moderate parkinson’s disease, Frontiers in aging neu-
roscience 7 (2015).
[202] E. L. Lawrence,
S. Dayanidhi,
I. Fassola,
P. Requejo,
C. Leclercq, C. J. Winstein, F. J. Valero-Cuevas, Outcome mea-
sures for hand function naturally reveal three latent domains in
older adults: strength, coordinated upper extremity function,
and sensorimotor processing, Frontiers in aging neuroscience
7 (2015) 108.
[203] E. Pavlova, Å. Hedberg, E. Ponten, S. Gantelius, F. J. Valero-
Cuevas, H. Forssberg, Activity in the brain network for dy-
namic manipulation of unstable objects is robust to acute tac-
tile nerve block: an fmri study, brain research 1620 (2015)
98–106.
[204] Q. Fu, M. Santello, Retention and interference of learned dex-
terous manipulation: interaction between multiple sensorimo-
tor processes., Journal of Neurophysiology 113 (2015) 144–55.
[205] T. McGeer, et al., Passive dynamic walking, I. J. Robotic Res.
9 (1990) 62–82.
[206] S. Collins, A. Ruina, R. Tedrake, M. Wisse, Eﬃcient bipedal
robots based on passive-dynamic walkers, Science 307 (2005)
1082–1085.
[207] M. H. Schieber, M. Santello, Hand function: peripheral and
central constraints on performance, Journal of Applied Physi-
ology 96 (2004) 2293–2300.
[208] M. H. Schieber, Constraints on somatotopic organization in the
primary motor cortex, Journal of neurophysiology 86 (2001)
2125–2143.
[209] J. N. Sanes, M. H. Schieber, Orderly somatotopy in primary
motor cortex: does it exist?, 2001.
[210] K. Nazarpour, A. Barnard, A. Jackson, Flexible cortical control
of task-speciﬁc muscle synergies, The Journal of neuroscience
: the oﬃcial journal of the Society for Neuroscience 32 (2012)
12349–12360.
[211] S. F. Farmer, Rhythmicity, synchronization and binding in hu-
man and primate motor systems, The Journal of Physiology
509 (1998) 3.
[212] I. E. J. de Vries, A. Daﬀertshofer, D. F. Stegeman, T. W. Boon-
stra, Functional connectivity in the neuromuscular system un-
derlying bimanual coordination, Journal of Neurophysiology
116 (2016) 2576–2585.
[213] J.-A. Rathelot, P. L. Strick,
Subdivisions of primary motor
cortex based on cortico-motoneuronal cells,
Proceedings of
the National Academy of Sciences 106 (2009) 918–923.
[214] R. Shadmehr,
J. W. Krakauer,
A computational neu-
roanatomy for motor control,
Experimental brain research.
Experimentelle Hirnforschung. Experimentation cerebrale 185
(2008) 359–381.
[215] D. Nozaki, A. Yokoi, T. Kimura, M. Hirashima, J.-J. O.
de Xivry,
Tagging motor memories with transcranial direct
current stimulation allows later artiﬁcially-controlled retrieval,
Elife 5 (2016) e15378.
[216] E. Brown, N. Rodenberg, J. Amend, A. Mozeika, E. Steltz,
M. R. Zakin, H. Lipson, H. M. Jaeger, Universal robotic grip-
per based on the jamming of granular material, Proceedings of
the National Academy of Sciences 107 (2010) 18809–18814.
[217] T. Aﬂalo, S. Kellis, C. Klaes, B. Lee, Y. Shi, K. Pejsa, K. Shan-
ﬁeld, S. Hayes-Jackson, M. Aisen, C. Heck, et al.,
De-
coding motor imagery from the posterior parietal cortex of a
tetraplegic human, Science 348 (2015) 906–910.
[218] J. R. Flanagan, P. Vetter, R. S. Johansson, D. M. Wolpert, Pre-
diction precedes control in motor learning., Current Biology
13 (2003) 146–50.
[219] M. Haruno, D. M. Wolpert, M. Kawato, MOSAIC Model for
Sensorimotor Learning and Control, Neural Computation 13
(2001) 2201–20.
[220] K. P. Kording, J. B. Tenenbaum, R. Shadmehr, The dynamics
of memory as a consequence of optimal adaptation to a chang-
ing body., Nature Neuroscience 10 (2007) 779–86.
[221] J. W. Krakauer, C. Ghez, M. F. Ghilardi, Adaptation to visuo-
motor transformations: consolidation, interference, and forget-
ting., Journal of Neuroscience 25 (2005) 473–8.
[222] K. A. Thoroughman, R. Shadmehr, Learning of action through
adaptive combination of motor primitives, Nature 407 (2000)
742–7.
[223] J. Classen, J. Liepert, S. P. Wise, M. Hallett, L. G. Cohen,
Rapid plasticity of human cortical movement representation
induced by practice,
Journal of Neurophysiology 79 (1998)
1117–23.
[224] J. Diedrichsen, O. White, D. Newman, N. Lally,
Use-
dependent and error-based learning of motor behaviors., Jour-
nal of Neuroscience 30 (2010) 5159–66.
[225] T. Verstynen, P. N. Sabes, How each movement changes the
next: an experimental and theoretical study of fast adaptive
priors in reaching, Journal of Neuroscience 31 (2011) 10050–
9.
[226] V. S. Huang, A. M. Haith, P. Mazzoni, J. W. Krakauer, Re-
thinking motor learning and savings in adaptation paradigms:
model-free memory for successful actions combines with in-
21
ternal models., Neuron 70 (2011) 787–801.
[227] R. S. Johansson, K. J. Cole, Sensory-motor coordination dur-
ing grasping and manipulative actions,
Current Opinion in
Neurobiology 2 (1992) 815–23.
[228] B. M. Quaney, D. L. Rotella, C. Peterson, K. J. Cole, Sen-
sorimotor memory for ﬁngertip forces: evidence for a task-
independent motor memory.,
Journal of Neuroscience 23
(2003) 1981–86.
[229] B. M. Quaney, R. J. Nudo, K. J. Cole, Can internal models of
objects be utilized for diﬀerent prehension tasks?, Journal of
Neurophysiology 93 (2005) 2021–7.
[230] J. N. Ingram, I. S. Howard, J. R. Flanagan, D. M. Wolpert,
Multiple grasp-speciﬁc representations of tool dynamics medi-
ate skillful manipulation, Current Biology 20 (2010) 618–623.
[231] J.-Y. Lee, N. Schweighofer, Dual adaptation supports a parallel
architecture of motor memory,
Journal of Neuroscience 29
(2009) 10396–404.
[232] Q. Fu, M. Santello, Sensorimotor learning of dexterous manip-
ulation., Elsevier.
[233] C. C. Kemp, A. Edsinger, E. Torres-Jara,
Challenges for
robot manipulation in human environments [grand challenges
of robotics],
IEEE Robotics & Automation Magazine 14
(2007) 20–29.
[234] A. Saxena, J. Driemeyer, J. Kearns, A. Y. Ng, Robotic grasping
of novel objects, in: Advances in neural information process-
ing systems, pp. 1209–1216.
[235] K. Mojtahedi, Q. Fu, M. Santello, Extraction of Time and Fre-
quency Features From Grip Force Rates During Dexterous Ma-
nipulation, IEEE Transactions on Biomedical Engineering 62
(2015) 1363–1375.
[236] K. E. Adolph, W. G. Cole, M. Komati, J. S. Garciaguirre,
D. Badaly, J. M. Lingeman, G. L. Chan, R. B. Sotsky, How
do you learn to walk? thousands of steps and dozens of falls
per day, Psychological science 23 (2012) 1387–1394.
[237] M. Gladwell, Outliers: The story of success, Hachette UK,
2008.
[238] K. R. Lohse, C. E. Lang, L. A. Boyd,
Is more better? us-
ing metadata to explore dose–response relationships in stroke
rehabilitation, Stroke 45 (2014) 2053–2058.
[239] J. Bongard, V. Zykov, H. Lipson, Resilient machines through
continuous self-modeling, Science 314 (2006) 1118–1121.
[240] M. Kalakrishnan, J. Buchli, P. Pastor, M. Mistry, S. Schaal,
Learning, planning, and control for quadruped locomotion over
challenging terrain, The International Journal of Robotics Re-
search 30 (2011) 236–258.
[241] D. A. Bristow, M. Tharayil, A. G. Alleyne, A survey of itera-
tive learning control, IEEE Control Systems 26 (2006) 96–114.
[242] F. J. Valero-Cuevas, M. Venkadesan, E. Todorov, Structured
variability of muscle activations supports the minimal interven-
tion principle of motor control, J. Neurophysiol. 102 (2009)
59–68.
[243] K. R´acz, F. Valero-Cuevas, Spatio-temporal analysis reveals
active control of both task-relevant and task-irrelevant vari-
ables, Frontiers in computational neuroscience 7 (2013) 155.
[244] M. C. Tresch, A. Jarc, The case for and against muscle syner-
gies, Current Opinion in Neurobiology 19 (2009) 601–607.
[245] S. F. Giszter, J. McIntyre, E. Bizzi,
Kinematic strategies
and sensorimotor transformations in the wiping movements of
frogs, Journal of Neurophysiology 62 (1989) 750–767.
[246] J. P. Scholz, G. Sch¨oner, The uncontrolled manifold concept:
identifying control variables for a functional task, Experimen-
tal brain research 126 (1999) 289–306.
[247] L. H. Ting, J. L. McKay, Neuromechanics of muscle synergies
for posture and movement, Current Opinion in Neurobiology
17 (2007) 622–628.
[248] E. Bizzi, V. C. K. Cheung, A. D’Avella, P. Saltiel, M. C.
Tresch, Combining modules for movement, Brain Research
Reviews 57 (2008) 125–133.
[249] F. Lacquaniti, Y. P. Ivanenko, A. D’Avella, K. E. Zelik,
M. Zago, Evolutionary and developmental modules., Fron-
tiers in computational neuroscience 7 (2013) 61.
[250] V. C. Cheung, A. Turolla, M. Agostini, S. Silvoni, C. Ben-
nis, P. Kasi, S. Paganoni, P. Bonato, E. Bizzi, Muscle synergy
patterns as physiological markers of motor cortical damage,
Proceedings of the National Academy of Sciences 109 (2012)
14652–14656.
[251] M. Santello, C. E. Lang, Are movement disorders and sensori-
motor injuries pathologic synergies? when normal multi-joint
movement synergies become pathologic, Frontiers in human
neuroscience 8 (2015) 1050.
[252] M. Santello, G. Baud-Bovy, H. J¨orntell, Neural bases of hand
synergies., Frontiers in Computational Neuroscience 7 (2013)
23.
[253] M. Santello, M. Flanders, J. F. Soechting,
Patterns of hand
motion during grasping and the inﬂuence of sensory guidance.,
Journal of Neuroscience 22 (2002) 1426–1435.
[254] K. T. Reilly, G. R. Hammond, Independence of force produc-
tion by digits of the human hand, Neuroscience Letters 290
(2000) 53–56.
[255] V. M. Zatsiorsky, Z.-M. Li, M. L. Latash, Enslaving eﬀects in
multi-ﬁnger force production., Experimental Brain Research
131 (2000) 187–195.
[256] M. H. Schieber, L. S. Hibbard,
How somatotopic is the
motor cortex hand area?,
SCIENCE-NEW YORK THEN
WASHINGTON- (1993) 489–489.
[257] S. A. Overduin, A. D’Avella, J. M. Carmena, E. Bizzi, Micros-
timulation Activates a Handful of Muscle Synergies, Neuron
76 (2012) 1071–1077.
[258] A. Leo, G. Handjaras, M. Bianchi, H. Marino, M. Gabiccini,
A. Guidi, E. P. Scilingo, P. Pietrini, A. Bicchi, M. Santello,
E. Ricciardi, A synergy-based hand control is encoded in hu-
man motor cortical areas, eLife 5 (2016) e13420.
[259] N. Ejaz, M. Hamada, J. Diedrichsen, Hand use predicts the
structure of representations in sensorimotor cortex,
Nature
Neuroscience 103 (2015).
[260] S. Babikian, E. Kanso, J. J. Kutch, Cortical activity predicts
good variation in human motor output,
Experimental Brain
Research (2017) 1–9.
[261] S. F. Giszter, C. B. Hart, Motor primitives and synergies in
spinal cord and after injury - the current state of play, Ann N
Y Acad Sci. (2013) 114–126.
[262] S. A. Winges, M. Santello, Common input to motor units of
digit ﬂexors during multi-digit grasping., Journal of Neuro-
physiology 92 (2004) 3210–20.
[263] M. Santello, A. J. Fuglevand, Role of across-muscle motor unit
synchrony for the coordination of forces, Experimental Brain
Research 159 (2004) 501–508.
[264] S. a. Winges, K. W. Kornatz, M. Santello, Common input to
motor units of intrinsic and extrinsic hand muscles during two-
digit object hold., Journal of neurophysiology 99 (2008) 1119–
1126.
[265] M. Flanders, J. F. Soechting,
Kinematics of typing: paral-
lel control of the two hands, Journal of Neurophysiology 67
(1992) 1264–1274.
[266] M. Santello, J. F. Soechting, Force synergies for multiﬁngered
grasping, Experimental Brain Research 133 (2000) 457–467.
[267] M. P. Rearick, A. Casares, M. Santello, Task-dependent mod-
ulation of multi-digit force coordination patterns., Journal of
Neurophysiology 89 (2003) 1317–1326.
[268] V. Zatsiorsky, F. Gao, M. Latash, Prehension synergies: eﬀects
22
of object geometry and prescribed torques, Experimental Brain
Research 148 (2003) 77–87.
[269] M. Santello, M. Bianchi, M. Gabiccini, E. Ricciardi, G. Salvi-
etti, D. Prattichizzo, M. Ernst, A. Moscatelli, H. J????rntell,
A. M. L. Kappers, K. Kyriakopoulos, A. Albu-Sch????ﬀer,
C. Castellini, A. Bicchi,
Hand synergies:
Integration of
robotics and neuroscience for understanding the control of bio-
logical and artiﬁcial hands, Physics of Life Reviews 17 (2016)
1–23.
[270] J. P. Scholz, N. Kang, D. Patterson, M. L. Latash, Uncontrolled
manifold analysis of single trials during multi-ﬁnger force pro-
duction by persons with and without down syndrome, Experi-
mental Brain Research 153 (2003) 45–58.
[271] N. Kang, M. Shinohara, V. M. Zatsiorsky, M. L. Latash, Learn-
ing multi-ﬁnger synergies: an uncontrolled manifold analysis.,
Experimental Brain Research 157 (2004) 336–350.
[272] M. Xiloyannis, L. Cappello, D. B. Khanh, S.-C. Yen, L. Ma-
sia, Modelling and design of a synergy-based actuator for a
tendon-driven soft robotic glove, in: Biomedical Robotics and
Biomechatronics (BioRob), 2016 6th IEEE International Con-
ference on, IEEE, pp. 1213–1219.
[273] K. Zhao, R. Breighner, A. Theuer, S. Godfrey, M. Bianchi,
M. Catalano, et al., Application of a novel robotic hand as a
myoelectric prosthetic prototype: proof of concept in a single
patient, in: ISPO World Congress 2015.
[274] A. Bicchi, M. Gabiccini, M. Santello, Modelling natural and
artiﬁcial hands with synergies,
Philosophical Transactions
of the Royal Society of London B: Biological Sciences 366
(2011) 3153–3161.
[275] V. Kumar, Y. Tassa, T. Erez, E. Todorov, Real-time behaviour
synthesis for dynamic hand-manipulation,
in: Robotics and
Automation (ICRA), 2014 IEEE International Conference on,
IEEE, pp. 6808–6815.
[276] Q. Fu, A. Ushani, L. Jentoft, R. D. Howe, M. Santella, Human
reach-to-grasp compensation with object pose uncertainty, in:
Engineering in Medicine and Biology Society (EMBC), 2013
35th Annual International Conference of the IEEE, IEEE, pp.
6893–6896.
23