Under-determined vs. Over-determined mechanics
One of the central tenets 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 location in space [
185], or one of multiple types of grasps to hold the object just as well [
186]. From the muscle control perspective, vertebrates have multiple tendons crossing each joint, then there are multiple individual muscle forces that can produce a given net joint torque [
170]. In contrast, roboticists have emphasized design architectures to reduce kinematic and actuation redundancy and typically build robots with as few kinematic degrees of freedom or tendons to be controllable.
This begs the question why the evolutionary process has tended to converged on such so-called under-determined mechanical systems for vertebrates. As reviewed in [
54], thinking of biological systems as under-determined is paradoxical with respect to the evolutionary process and clinical reality. For example, why would organisms evolve, encode, grow, maintain, repair, 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 dysfunction even after injury to a few muscles, or mild neuropathology? Which muscle would you donate to improve 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 freedom for control. Behavior in the real world
7 consists of satisfying multiple—at times competing—demands. Therefore, a mathematical argument can be made [
54] that behavior in the real world, by virtue of needing to satisfy multiple demands or constraints, requires multiple muscles [
178]. And, by extension, that dysfunction of even a few muscles will make the limb less versatile [
187].
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 tendons cross each joint, then there is redundancy in the sense that multiple individual muscle forces at those tendons that can produce a given torque. However, the same is not true when we consider movement. The rotation of that single joint defines the lengths of all muscles that cross it [
54,
176,
188]. While in principle muscles can go slack, muscles with tone will shorten appropriately. However, muscles that lengthen must do so by a prescribed amount. Thus, the relationship 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 incorrect neural command or its stretch reflex fails to be appropriately modulated in time), the movement will be disrupted [
54,
176,
188]. Therefore, while multiple limb kinematics may be equivalent from a task perspective (i.e., reaching a cup or throwing a ball), they are far from equivalent from the perspective of neurophysiological control, robustness to sensorimotor noise, or time-critical modulation of activation and reflexes [
188].
Grasp vs. Manipulation
Another recent evolution in the biological side has been the explicit distinction between grasp and manipulation. Although these terms are often used interchangeably in the biological literature, there is a long tradition of creating clear taxonomies and descriptions of hand actions that clearly distinguish between the two [
27,
29,
149,
189]. 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 [
27,
29,
48]. Largely because studying precision or dexterous manipulation has important practical difficulties with motion capture of individual finger motions, and measurement of individual fingertip forces. Similarly, in spite of the mathematics of dexterous manipulation being well developed [
29], robotic hands and prostheses tend to focus on grasp because of the difficulties 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 animation [
174]. Similarly, some experimental methods have been developed to quantify dynamic dexterous manipulation, which has revealed novel aspects about the neuromechanical control of dexterous manipulation in development, adulthood, healthy aging and neurological conditions (for overviews see [
42,
190‐
194]), as well as how the interaction between cognitive and biomechanical factors affect dexterous manipulation performance (e.g., [
132,
133,
195]).
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 [
196,
197]. While effective mechanical function can be found in many, primates (and humans) are the beneficiaries of highly specialized neuroanatomical coevolution of brain and hand (e.g., [
2,
41,
198,
199]). 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 somatotopic [
200] and cortico-centric view of manipulation, especially in the cases of dynamic dexterous manipulation where time delays preclude active involvement [
191].
After all, the current concept of cortical control is not the exclusive micromanagement of individual muscle activations, but rather includes the ‘binding’ of motor neurons into flexible, context-dependent functional groups [
201‐
203], the utilization of primitive ‘synergies’ prepared by networks of spinal interneurons [
204], adjustment of sensory feedback gains [
205], and the formation/recall of motor memories [
206], to name just a few. Synergies are discussed in more detail in subsequent sections.
Nevertheless, as in the case of passive dynamic walkers, robotics provide counterexamples to such micromanagement 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 to control every degree of freedom independently). Such hands can display multiple versatile grasp functions, without requiring a controller [
53,
107] or even fingers [
207]. Such developments are alternatives that promise to develop multiple designs along the spectrum between under- and over-actuated robotic hands. This is especially useful in cases of brain-machine interfaces for hand prostheses, where only a few degrees of freedom of control can be extracted from the human pilot’s nervous system (e.g., [
208]).
Learning vs. Implementation vs. Adaptation
Human sensorimotor learning has been extensively studied [
80,
130]. One view posits that humans’ ability to perform skilled motor behaviors relies on learning both control and prediction through inverse and forward internal models (implicit, explicit, probabilistic or otherwise). Specifically, a given control strategy generates motor commands needed to create desired consequences (e.g., a given reach trajectory or grasping an object at specific locations), whereas prediction maps motor commands (i.e., efference copy) into expected sensory consequences (e.g., object contact or onset of acceleration at object lift off [
209,
210]). 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 predicted sensory outcome occur, i.e., error-based learning (e.g., [
211‐
213]) and use-dependent learning (e.g., [
214‐
217]). However, most of what we know about human sensorimotor learning for reach has been derived from studies of reaching movements over distances 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 [
81,
82,
218]. More recently, it has been shown that humans may acquire and retain multiple internal representations of manipulation [
219,
220]. Later studies provided further evidence supporting the concept of multiple sensorimotor mechanisms and how their different time scales may interfere with generalization or retrieval of previously learned manipulation [
221]—even when the object being manipulated is the same. A recent study has provided evidence for the co-existence of context-dependent and independent learning processes [
195], which would operate similarly to those described for adaptation of reaching movements [
222]. The advantage of context-dependent representations of manipulation is that they can be recalled when the object has strong contextual cues (i.e., object geometry and perhaps 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 upcoming 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 manipulation 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 [
223], multiple 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 manipulation context, and then transferring it to another context can be theoretically minimized or bypassed when designing an artificial controller. Some examples of successful robotic learning for grasp and manipulation show that this is possible [
172,
173,
224,
225]. Of course, these theoretical considerations assume that building multiple representations of learned manipulations allows them to operate independently, something that—as described above—clearly also challenges biological 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 object is grasped [
132,
133]. This phenomenon, which has been confirmed by several studies [
134‐
136,
226], 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 ([
132] and [
133], 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. Additionally, 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 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 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 biological 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 elucidate why learning to produce accurate, smooth and repeatable movements takes immense amounts of practice even in typically developing children [
227], why so few of us can become elite musicians or athletes [
228], and why rehabilitation requires very intensive practice [
229]. 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 learning. 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 cockpits, left- and right-handed scissors, frets in string instruments, the key system in clarinets, and touch screens are but a few examples.
Thus, biological hands in particular have an unfair advantage 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 performance 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 algorithms, motor habits, and statistically useful anticipatory strategies defined by their intended use or—in the case of prosthetics—the environment, job and preferences 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., [
46,
107]), as well as grippers with no fingers at all [
207]. Some salient examples of such learning (and re-learning) come from [
230‐
232].
Prescriptive vs. Descriptive synergies
What are the debates in the study of synergies in biological systems? A root cause of the debates is the nature of scientific inference based on experimental observations. The fact that experimental recordings detect dimensionality reduction is not surprising because sensorimotor control must, by definition, select motor actions from the low-dimensional subspace of feasible actions [
177]. Therefore, disambiguating
prescriptive synergies of neural origin (those that are prescribed by the nervous system as a control strategy) from
descriptive synergies (those that describe the expected dimensionality reduction) is difficult [
48,
177]. Thus, the main question is not
whether the nervous system inhabits a low-dimensional solution space to perform tasks, but rather
how it does so [
177,
233‐
235]. Moreover, 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 maybe critical to their performance [
48].
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 coordination and adaptability to task demands. The refinements afforded by such evolution can be appreciated when examining the efficacy of the neural control of several complex motor behaviors, including but not limited to speech production, locomotion, and manipulation. 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 fingers) in reality evolve in a lower-dimensional space (i.e., can be well approximated by roughly 5 variables) [
47]. Such motor
synergies have also been observed in the phase-locked coordination (or correlated action) of multiple muscles that produce complex behaviors [
236‐
238]
The theoretical framework of synergies has been extensively used and tested to account for the nervous system’s ability to control multiple muscles and multi-joint movements (for reviews see [
239‐
241]). 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 synergy framework. Synergies have also been used as a framework to understand pathological coordination of movement (for reviews see [
242,
243]). However, longstanding 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 sensorimotor learning in tasks that may benefit from, or be penalized by a synergy-like control structure [
150] (for a review see [
244]).
When considering the biomechanics of hand muscles, the existence of anatomical constraints would support synergistic actions of the fingers. These constraints can come from, for example, finger muscle-tendon complexes spanning several joints and passive linkages among tendons [
199]. Such synergistic actions have been described as subject-independent finger kinematic patterns for grasping [
47,
245] (for review see [
244]), 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 [
246,
247] (for review see [
198]). Early attempts to define the control of individuated finger forces in cortical neuron activity revealed a much more complex picture characterized by broadly distributed activity [
248]. More recent work in non-human primates, however, supports an organization of cortical activity that is compatible with the synergy framework [
249]. When searching for neural correlates of synergies in humans, a recent study revealed that the cortical representation of hand postures can be better accounted for by using a synergy-based network than somatotopic or muscle-based models [
250], which is compatible with the view of cortical organization of finger movement being shaped by habitual use [
251], and even goal equivalence in finger actions being implemented at a cortical level [
252].
Many studies have attempted to identify synergies at different levels of biological systems and species, including primary motor cortex [
250], spinal cord [
253], motor units [
254‐
256], motion [
47,
257], and forces [
258‐
260]. 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, including 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 [
244]; see also [
261]). 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 proposed by Bernstein [
185], or ensuring attainment of task goals by minimizing the variance that would be detrimental to performance [
237,
262,
263]. 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 [
235].
Robotics, in contrast, synthetically designs, assembles and operates engineered systems where synergies can be prescribed. Over the past two decades, roboticists have exploited the concept of (prescriptive) synergies to design robotic hands (for review see [
261]). Examples of these designs include the Pisa/IIT SoftHand [
46], whose design was based on the kinematic synergies extracted from grasping a set of imagined objects [
47], as well as devices to constrain motion of human fingers for rehabilitation of sensorimotor function [
264]. Here, the underlying design motivation is to capture human-like kinematic features, i.e, simultaneous motion of all fingers, by using a significantly smaller number of actuators than joints. Preliminary clinical applications of this approach for prosthetic applications have shown that individuals with upper limb loss can quickly adopt such synergy-based design with minimal training [
265]. A major goal and challenge for robotic grasping and manipulation is the implementation of force control using kinematic synergies. The results of computational modeling suggest that the first few hand postural synergies may play an important role for attaining force closure [
266]. Nevertheless, it remains to be investigated the extent to which robotic motion-to-force transition can fully leverage a synergy-based motion-to-force coordination. Experimental evidence and theoretical frameworks developed by studies of human multi-finger synergies might potentially be used to inspire a hierarchical control of high- and low-level grasp variables (i.e., task goal and distribution of individual fingertip forces, respectively) [
133], as well as ‘default’ vs. task-dependent modulation of fingertip force distributions [
259]. Nevertheless, a major challenge, both from neuroscientific and robotic perspective, is evaluating the role of sensory feedback elicited by object contact and force production on the coordination of multiple (human and robotic) fingers.