Muscle synergists are frequently thought to share function at a common joint. However, whether a muscle synergist is monoarticular or multiarticular affects how the muscle transmits force and generates work between joints within the limb. Bi- or multiarticular muscles can function to transfer energy across joints within the limb, which may enhance running stability. Monoarticular muscles necessarily act locally at a particular joint. Nevertheless, muscle-tendon architecture affects contractile strain patterns within individual muscles, so that even monoarticular muscles may exhibit differential strain and net work output associated with a given motor task. Thus, functional diversification may occur within, as well as between, muscle synergists. Differences in motor function within and between muscle synergists likely reflect differences in fiber type and motor recruitment. Intrinsic force-velocity and force-length muscle properties also serve to provide immediate stabilizing responses to perturbations during fast movements, such as running, prior to neural feedback for control adjustments. These have been termed 'preflexes' but necessarily act throughout a muscle's contraction, providing a temporal buffer to reflex modulation of a muscle's neural activation.
Muscle contraction is driven by interactions between two sets of protein filaments: actin and myosin. 'Cross-bridges' on the myosin filaments bind to the actin filaments and undergo a structural change which causes the filaments to slide relative to one another. During each interaction, a cross-bridge generates forces of a few piconewtons and motions of a few nanometres. Much of our knowledge of muscle mechanics, however, is based on experiments performed on single muscle cells, where forces are measured in millinewtons and motions are measured in micrometres or millimetres. Based on such experiments, the mechanical properties of the cross-bridges and the rates of the biochemical reactions between actin and myosin have been inferred. In particular, the study of the response of muscle cells to step changes in force or length (e.g. Huxley and Simmons, 1971. Nature 233, 533-538.) has been extremely important in the development of current models of how muscles work. However, a typical muscle cell (5 mm long, 80 micrometres diameter) contains approximately 10^12 cross-bridges and is known to exhibit inhomogeneous behaviour along its length. The question that I would like to answer is this: is it truly possible to infer molecular mechanical properties from experiments on whole cells, or do the properties that we observe at the level of the cell, and have attributed to the cross-bridges, in fact arise at a higher level of organisation? I will present my thoughts on how to tackle this question and would appreciate suggestions as to the best approach and opinions as to what the answer is.
Humans and other animals require metabolic energy to walk. Minimizing this metabolic cost appears to determine many aspects of how we walk (e.g. our preferred speed). While the importance of metabolic cost in determining our preferred walking biomechanics was recognized long ago, an understanding of why walking exacts a metabolic cost has remained elusive. Our approach has been to use mathematical models inspired by passive dynamic walking to make quantitative predictions regarding the determinants of metabolic cost and test these predictions using empirical experiments on humans and other animals. In this manner, we have identified two major biomechanical determinants of the metabolic cost of healthy human walking, step-to-step transition and limb swing costs, that account for about 90% of total metabolic cost at moderate speeds. There is a tradeoff between these two costs-transition costs are minimized with short narrow steps while swing costs are minimized with long wide steps. Minimizing the sum of these two costs predicts measured preferred walking biomechanics remarkably well. Another intriguing determinant of the metabolic cost of walking is balance. Our theoretical and empirical results support the idea that lateral motion is passively unstable, is actively stabilized using medio-lateral foot placement, and this active stabilization exacts a modest metabolic cost (~10%). A new direction for our research is towards understanding the determinants of the metabolic cost of pathological gait. Dynamic walking models predict that symmetrical transitions, those during which the trailing leg performs an equal amount of positive work to replace the energy dissipated by the leading leg negative work, minimize the required mechanical work. Two common characteristics of pathological gait - be it from stroke, spinal cord injury or amputation-is left-right asymmetry and an elevation of the metabolic cost of gait. In subjects with hemiparesis due to stroke and in healthy subjects with simulated hemiparesis, we are currently testing the general hypothesis that the elevated metabolic cost of pathological gait is due to the increased muscle mechanical work required of asymmetrical step-to-step transitions. Our early empirical results are generally supportive of this hypothesis. It is our intention to use the results to guide the design of rehabilitation strategies, rehabilitation devices and assistive devices aimed at lowering metabolic cost and increasing patient mobility by improving transition symmetry.
The remarkable manipulative skills of the human hand are neither the result of rapid sensorimotor processes nor of fast or powerful effector mechanisms. Rather, the secret lies in the way manual tasks are organized and controlled by the nervous system. Successful manipulation requires the selection of motor commands tailored to the task at hand and the relevant physical properties of the manipulated objects. Because of long time delays associated with sensorimotor loops, feedback control cannot support the swift and skilled coordination of fingertip forces observed in most manipulation tasks. Instead, the brain relies on feedforward control mechanisms that take advantage of the stable and predictable physical properties of these objects. In this talk, I will provide evidence that the brain uses internal models of object and limb dynamics both to estimate the motor commands required to achieve desired sensory outputs and to predict the sensory consequences of motor commands [1-6]. I will also discuss results from recent experiments suggesting that predictions about object weight, used when lifting, result from the interaction between two sources of knowledge: slowly adapting statistical knowledge about the relationship between size and weight for families of objects and rapidly adapting knowledge about the weights of specific objects [7-8].
This work was supported by NSERC, CIHR, and HFSP.
Given sufficient sensory feedback, time to practice, and sometimes a bit of instruction, the (embodied) mind has an uncanny ability to harness the dynamics of implements to its purposes. Among other implements, the piano makes a valuable case study in the situated mind extending itself out beyond the reaches of the body. Our lab is involved in modeling the mind/limb/instrument as a multivariable control system and we are applying fundamental design limitations theory to produce statements that apply irrespective of control design or implementation. On the experimental side, we use haptic interface to covertly modify the dynamics of virtual objects such as balls to be thrown or spring-mass systems to be excited into resonance to investigate the role of power exchange in supporting the development of muscle memory and the expression of "physical intelligence". We have found that nonspecific practice improves performance and knowledge of results transfers across conditions in a manner that suggests the calculus of internal models is richly grounded in the calculus of the physical world. We propose that appropriate sensory feedback, including feedback associated with mechanical coupling is critical for the development of motor skills that encompass tools and implements.
Work done in collaboration with Art Kuo and Jim Freudenberg under NSF support.
Rats are nocturnal, burrowing animals that use their vibrissae (whiskers) to tactually explore the environment. Using only its whiskers, a rat can determine object size, shape, orientation, and texture. This makes the rat vibrissal system an excellent model to explore the structure of movements that subserve sensing. In this talk I will describe recent experiments in our laboratory that have aimed to understand neural processing in the vibrissal system from the outside-in. I will walk through our laboratory's suggested answers to the following questions:
Many neuroscientists, working across different sensory modalities, have posited that perception is based at least in part on comparing "actual" with "predicted" sensory information. Validation of our hypothesis would provide a quantitative, physiological basis for this comparison to occur. We will provide some suggestions for how our results from the rat whisker system may generalize across modalities, and look forward to discussing these issues with the workshop group.
Sarcomeres (or half sarcomeres) comprise the smallest contractile unit of striated muscles. Their mechanical properties have come to dominate the thinking of muscle physiologists, although mechanical properties of isolated sarcomeres have never been measured. For example, the force-length property of sarcomeres was first derived in the classical study by Gordon et al. (1966), but in their study sarcomere length was never measured, forces never reached steady-state and thus were extrapolated, and shortening known to decrease isometric steady-state forces was used prior to some measurements but not in others. As a result, textbooks show the sarcomere force-length relationship, arguably the most basic property of skeletal muscle, with a linear descending limb from approximately 2.2\mu m to 3.7\mu m (for frog skeletal muscle) which is associated with the linear loss of actin-myosin filament overlap and the number of potential cross-bridge attachment sites within a (half-) sarcomere. We have been able to make measurements in single sarcomeres and in isolated myofibrils, which are sub-cellular organelles containing sarcomeres in a strictly in series arrangement, thus forces measured at the end of the myofibril can be related to the individual (lengths of) sarcomeres. Two basic observations have been made in these preparations: (i) isometric sarcomeres on the descending limb of the force-length relationship can support the same force but reside at vastly different lengths (Herzog et al., 2006; Rassier et al., 2003); and (ii) when stretched or shortened prior to an isometric contraction, sarcomeres at a given length will show a series of different steady-state isometric forces (Joumaa et al. unpublished observations).
The basic conclusion arising from these results is that, in contrast to common believe and textbook knowledge, sarcomere lengths (or myofilament overlap) beyond optimal length are not uniquely associated with steady-state isometric force. Different steady-state forces for these situations have been observed previously in muscle and fibre preparations, but here we demonstrate that this also holds for isolated myofibrils and single sarcomeres. We hypothesize that the decrease in steady-state isometric forces (following active sarcomere shortening) is associated with a decrease in the proportion of attached cross-bridges (this result is supported by decreased sarcomere stiffness in these situations), while the increase in steady-state forces (following active stretching) is associated with an increase in stiffness of passive structural elements (we identified the molecular spring titin as a potential candidate for this hypothesis) and an increase in the average force per cross-bridge (which could be accomplished by an increase in the average attachment length for these cross-bridges).
Since these observations have been made for steady-state situations in which cross-bridges go through many attachment/detachment cycles, our results suggest that cross-bridges possess a memory for their contractile history. Since steady-state isometric forces are constant following deactivation and subsequent isometric reactivation (independent of contractile history), this memory is erased upon deactivation.
Work done in collaboration with V. Joumaa and T. Leonard
Despite the skill and dexterity of our motor behavior, human control of slow, continuous movements is remarkably inept. When asked to move the elbow in a smooth rhythm at low frequencies, motion profiles develop pronounced irregularity1. It appears to have no musculo-skeletal cause and persists in the absence of visual feedback even when the target cadence is displayed without explicit timing cues (either of the latter might evoke corrective submovements). Discrete movements (reaching in space2,3 and pointing with wrist rotation4) also exhibit pronounced kinematic irregularity when subjects are asked to move slowly. This appears to reflect a fundamental feature of the neural control architecture, a suggestion supported by the pronounced "chunking" of movements made by patients recovering after neurological injury5,6,7
What is the origin of this phenomenon? Might it be due to controlling through a neural architecture with nonlinear dynamics originally specialized for rapid rhythmic behavior? Alternatively, it might be an adaptation to simplify the complexity of motor planning and control - but is temporal "chunking" necessarily simpler than continuous control? What might this phenomenon imply about the form of internal models of external objects? Even highly flexible objects with prodigiously complicated dynamics (ropes, fly rods. . .) can be manipulated skillfully. Is a "calculus" of combining temporal chunks sufficiently rich to describe these objects? Alternatively, is it better suited than continuous dynamic models?
Robotics researchers have spent 30 years and millions of dollars in the attempt to create effective grasping and manipulation capabilities. This effort has largely followed an anthropomorphic approach and many multifingered robot hands with elaborate sensing and control systems have been constructed. By almost any measure, this endeavor has failed: robots today cannot autonomously perform even simple grasping tasks like picking up a newspaper or a coffee cup in a typical home setting.
We have followed a different approach to create a robot hand for unstructured environments. This hand combines carefully tuned passive joint stiffness with adaptive transmissions that equilibrate joint torques between fingers. The result is a hand with four fingers and eight joints but only a single motor. The hand uses no sensors of any kind and control is binary, so when the hand is activated the motor simply tightens the tendons to a preset force level. Despite its simplicity, experimental results demonstrate that the hand can successfully grasp objects roughly 3 to 25 cm in size despite visual targeting errors of 5 cm in object location.
This robotic system emulates key aspects of human grasping ability but its methods are fundamentally different from human hands. Human grasping and manipulation relies on many sophisticated feedforward and feedback control schemes, based on vast numbers of afferents in the skin, joints, and muscles. While the importance of passive impedance is well-established in biological motor control, there have been few studies of its role in grasping. In this talk we will consider the implications of the success of our robotic hand, and the relative roles of passive impedance and sensory feedback in human grasping.
Epimuscular myofascial force transmission is sideways transmission between muscles and / or between muscles and non-muscular connective tissues. As force is transmitted this way from the muscle belly, force exerted at the muscular origin and insertion are not equal. Very substantial proximo-distal force differences of up + or - 40 % have been found in several species. The size of this difference may indicate the potential of this force transmission mechanism. Simple reasoning, as well as physical and finite element modeling indicates that such effects should lead to enhanced serial (within muscle fibers), as well as parallel (between muscle fibers) distributions of sarcomere lengths. Such transmission between antagonistic muscles was indicated also by decreases of force (in rat between -15 and 30%) as a function of antagonistic muscle length, for muscles kept at t constant length. All this has substantial functional consequences for muscle function, adaptation etc.. (for reviews see Huijing, 2003; Huijing and Jaspers, 2005; Huijing et al.., 2007) So far, in physiological in situ experiments without dissection of the muscle belly, passive muscles or, muscle undergoing either maximal or sub-maximal activation have been studied under conditions full recruitment. Another limitation, of the above experiments, whose aims were to show the existence and extent of the mechanism, may be that in those experiments muscles are activated to an identical degree. It is clear that substantial work needs to be done in the future, using differential degrees of activation. An initial report of a start of such work was presented recently (Maas and Sandercock, 2007) indicating that at least prior to tenotomy, activation of a single muscle does not yield evidence for interaction between an active and passive synergistic muscles. Despite this finding (of which many details are still to be discussed), we are still working on the hypothesis that, also in health, epimuscular myofascial force transmission is potentially an important mechanism, as this result is only one out of a multidimensional problem. After tenotomy of Soleus muscle, this muscle still contributed to force exerted at the ankle joint. Previously, we had shown such effects also for human flexor carpi-ulnaris muscle (e.g. Kreulen et al., 2002).
Many people use Hill-type models to assess muscular function. It is clear that such models, are as long as they in principle constitute one giant sarcomere for each muscle (in other words pooled properties), cannot deal with phenomena of distributions of sarcomere lengths and other variables. Finite element models (FEM) are equipped to deal with those phenomena, but it is presently not realistic to make FEM of for example the muscle within a limb or even a segment of a limb. The problem could in principle be solved for Hill-type models, by positioning 3 models in series to deal with serial sarcomere length distributions. Of course this complicates the model calculations substantially; and one has to resort to FEM type strategies to calculate converging equilibrium conditions. This is a question that should be addressed in the future by multidisciplinary teamwork.
Dexterous object manipulation is a hallmark of human skill. This talk will focus on the sensorimotor control of fingertip actions with special emphasis on the role of tactile sensory mechanisms. It highlights the importance of sensory predictions, especially related to mechanical contact events around which manipulation tasks are organized, and discuss how such predictions are influenced by tactile afferent signals recorded in single neurons in awake humans. It is generally assumed that primary sensory neurons transmit information by their firing rates. However, during natural manipulation tasks, tactile information from the fingertips is used faster than can be readily explained by rate codes. Recent evidence indicate that the relative timing of the first impulses elicited in individual units of ensembles of afferents reliably conveys information about complex contact parameters important for planning an control of dexterous fingertip actions. The sequence in which different afferents initially discharge in response to mechanical fingertip events provides information about these events faster than the fastest possible rate code and fast enough to account for the use of tactile signals in natural manipulation.
I will present a view on motor learning and adaptation that starts with assumptions about the way the mechanical properties of the body and the world change over time. We use Bayesian algorithms to predict how people could adapt in an optimal fashion. Our research shows that many published adaptation phenomena can be understood as optimal adaptation in this framework.
How is walking best controlled? Human legs act like pendulums during walking, and their dynamics present both an opportunity and a problem for control. The opportunity is to take advantage of passive dynamics to produce a cyclic motion with little effort. The problem is that these same dynamics integrate all forces including unwanted disturbances, leading to imperfect motions. The motion can be controlled through a combination of feedforward trajectories and closed-loop feedback. But what combination is best? The prevailing view is that central pattern generators produce rhythmic commands that are modulated by feedback. But given the ability of passive dynamics to generate rhythmic motion on their own, the role of the central pattern generator may deserve reconsideration. We will use control systems methodology to examine the relative roles of dynamics, feedback, and feedforward in the control of locomotion. Analysis indicates that walking can be controlled by feedback with remarkably low gains, and with remarkably little need for periodic, centrally-generated command signals. The rhythmic signal may have more to do with sensing than commanding the resulting motion. Central pattern generators may therefore be state estimators in disguise.
Muscle has been modeled at two different scales, for two distinct purposes. When the task is to estimate muscle force from kinematics and external forces, lumped models of whole muscle can be simulated easily, and have state variables that may be comparable to data at hand. However, they are often not consistent with muscle structure, and lack predictive power, as they are essentially models of data. When the task is to understand how the mechanics, energetics and biochemistry of motor proteins integrate into mechanical, energetic and metabolic behavior of whole muscle, models of actomyosin interactions in single sarcomeres are more apt, may be tested against simpler data, and allow predictions, as they are models of process. However, they are more complicated to simulate, and may not be comparable to whole muscle data. Now is a time of opportunity to build on advances at the motor protein and single fiber scales to make models that can answer questions with significance at larger scales. I plan to show how recent approaches to modeling at the sarcomere level can answer larger questions, and to identify experimental and modeling problems that will need to be solved in order to achieve integration from molecules to muscle.
In a muscle which must both generate resistive force and undertake work on the centre of mass (e.g. a limb extensor), there is a trade off where short fibres reduce activation cost but long fibres reduce the sarcomere shortening velocity; both of which reduce the energetic cost of contraction. Elastic tissues in series with the muscle fibres, specifically the tendon and the aponeurosis, may aid in a muscle's versatility. Legs appear as springs during running and in some larger animals, most of the muscle tendon unit length change occurs in series elastic tissues (Biewener et al., 1998; Lichtwark and Wilson, 2006; Roberts et al., 1997). So it seems that elastic energy can contribute significantly to power generation during locomotion. But the question remains: does elastic energy enhance muscle efficiency during stretch shorten cycles?
We have attempted to answer this question based on the assumption that if series elastic compliance enhances muscle efficiency, then its architecture (muscle fibre length and tendinous compliance) will be tuned such that efficiency is nearly maximal for everyday tasks like locomotion. We hypothesise that there is an optimum combination of fascicle length and tendon compliance which maximises muscle efficiency for a particular stretch-shorten task. To test this hypothesis we have furthered the ideas of Alexander (1997) to estimate a muscles metabolic costs based on known muscle physiological properties. We developed a Hill-type muscle model which could reasonably predict the total energetic cost associated with sinusoidal work-loop experiments in Dogfish muscle (Lichtwark and Wilson, 2005). Such a model is useful because it allows us to understand the influence of changing the model parameters under a given set of conditions. We then applied the model to experimentally obtained data from the medial gastrocnemius (muscle force and muscle-tendon unit length changes) for walking and running at different speeds and varied the muscle fibre length and tendon compliance. Maximum isometric force of the muscle was adjusted for fibre length changes.
The results indicate that although running requires a stiffer Achilles tendon and longer medial gastrocnemius muscle fibres to maximise muscle efficiency, a relatively broad range of combinations can still maintain a high efficiency for both gaits and different speeds. These combinations agree well with real-life values, where muscle fibres are generally short compared to the whole muscle and the tendon is compliant relative to its maximum force producing capacity. Therefore it seems that having a compliant Achilles tendon and relatively short muscle fibres does indeed enable a muscle to enhance its efficiency during locomotion.
It is difficult to define the accuracy of such modelling approaches to muscle energetics. This is largely because much of the data on the energetics of mammalian muscle contraction comes from experiments on muscles from small rodents which have much stiffer series elastic tissues relative to their force producing capacity compared to humans or other larger animals. To overcome this we have begun artificially introducing series elasticity to small muscle preparations and using myothermic techniques to quantify energetic cost during stretch-shorten cycles. This will allow us to explore how both the power output and efficiency of a muscle is influenced by parameters like tendon compliance and amplitude and frequency of the cyclic contractions and also to further refine models to accurately predict muscle power and energetic cost. We are also interested in investigating the influence of sub-maximal activations on the series elastic compliance (Hof, 1997) and how this influences the power output, energetics and control of movement.
Work done in collaboration with Chris Barclay and Alan Wilson.
The control of limb and hand movement is complex. We have measured the intrinsic structural properties of forearm muscles and have revealed a wide range of architectures that cause these muscles to be highly specialized for production of force and/or excursion (1,2,3). Additionally, synergistic muscles were often seen to have diverging architectural design, which greatly decreases the total muscle mass required to accomplish a wide range of functional tasks (4). These findings have dramatic implications for surgical reconstruction of the upper extremity where muscles are moved in order to provide function for those muscles that have become paralyzed or traumatized (5,6). A specialized laser diffraction tool was developed to measure the sarcomere length operating range of upper extremity muscles (7). Such measurements also demonstrated that wrist muscles have a stereotypical operating range on the length-tension curve that produces a wrist with great stability throughout the normal range of motion (8,9). Taken together, these studies demonstrate the highly specialized nature of human upper extremity muscles which is, apparently, developmentally encoded and even regulated in adult tissue which easily remodels in the face of transfer or trauma.
This work was supported by the NIH and the Department of Veterans Affairs.
The design and control of a robot for physical collaboration with a human requires an understanding of the human motor control system, so that the coupled human-robot system can be directed by the human in an ergonomic, intuitive, or energy-efficient manner. One class of collaborative robots creates programmable kinematic constraints. The design of constraints for collaborative manipulation requires an understanding of how humans naturally take advantage of kinematic constraints. In this talk I will describe some simple experiments in single-arm interaction with a planar manipulandum enforcing a programmable rail. The results indicate that the forces subjects apply against the rail can be described by optimization of an objective function. We derive the objective function directly from the experimental data, without resorting to a particular hypothesis or biomechanical model.
This work is joint with Peng Pan, Ed Colgate, and Michael Peshkin.
Do the same, or different, mechanisms contribute to vestibular perception and action? We used motion paradigms that provided identical sinusoidal inter-aural otolith cues across a broad frequency range. We accomplished this by sinusoidally tilting (20 degrees, 0.005-0.7 Hz) subjects in roll about an earth-horizontal, head-centered, rotation axis ("Tilt") or sinusoidally accelerating (3.3 m/s2, 0.005-0.7 Hz) subjects along their inter-aural axis ("Translation"). While identical inter-aural otolith cues were provided by these motion paradigms, the canal cues were substantially different, since roll rotations were present during Tilt but not during Translation. We found that perception was dependent on canal cues, since the reported perceptions of both roll tilt and inter-aural translation were substantially different during Translation and Tilt. These findings match internal model predictions that rotational cues from the canals influence the neural processing of otolith cues. We also found horizontal translational VORs at frequencies above 0.2 Hz during both Translation and Tilt. These responses were dependent on otolith cues and match simple filtering predictions that translational VORs include contributions via simple high-pass filtering of otolith cues. More generally, these findings demonstrate that internal models govern human vestibular "perception" across a broad range of frequencies and that simple high-pass filters contribute to human horizontal translational VORs ("action") at frequencies above about 0.2 Hz.
A central topic in neuroscience research has been the manner in which the nervous system controls movement. The follow-up length servo hypothesis of Merton (1953), for example, posited that muscular length could be controlled by the central nervous system independently of any specified mechanical properties of muscle, save that the muscle can produce force, by the provision of a simple negative feedback circuit in the spinal cord. As a result of newer interdisciplinary approaches, however, it has been recognized that the peripheral motor apparatus possesses attributes that contribute to coordinated movement as well, and that the associated organization and mechanical properties of the musculoskeletal system constitute an integral part of the control system. Several examples from receptor physiology and limb biomechanics support this viewpoint. First, the mechanical behavior of a muscle in situ results in part from the intrinsic properties of the muscle and in part from feedback from the muscle spindle receptor. Under some conditions, the behavior is dominated by feedback, and under others by intrinsic properties. This interaction arises from the complementary mechanical properties of receptor and muscle, and not from attributes of the associated neural pathways. Therefore, feedback modulation results from mechanical sources as well as neuromodulation in the spinal cord and brainstem. Second, interjoint coordination is enhanced not only by neural pathways arising from Golgi tendon organs that span different joints and axes of rotation, but also by multi-articular muscles that anatomically span more than one joint or axis of rotation. Indeed, these two systems further illustrate the parallel nature of mechanical and neural mechanisms of coordination. Third, the much touted "degree of freedom problem" may be "solved" in part by selection and hierarchical organization in the central nervous system, but also in part by passive structures in the limb such as fascia. Fascia can provide important channels for force transmission and can constrain the degrees of freedom of joints depending on the motor task. During rapid locomotion and higher forces, the degrees of freedom of a limb are limited by both neural and mechanical mechanisms, including force feedback, multi-articular attachments of muscles, and passive structures such as fascia.
The cerebellum evidently plays an important role in motor agility. I have used a comparative-evolutionary approach and computational modeling to try to untangle disparate evidence about what that role is. Recent advances in passive dynamics and under-actuated robotics have drawn attention to the importance of mechanical design in trajectory and gait generation. It has become clear that understanding what the brain (particularly the cerebellum) does for motor agility also requires understanding the physics of the musculoskeletal system and its interactions with the environment. To develop such an understanding, and to provide a convenient computational environment for exploring neural and biomechanical mechanisms underlying agile movement, I have written a MATLAB toolbox. The Dyanimat toolbox has methods for constructing and simulating multi-link legged animals/robots on irregular terrain, with embedded neural controllers. I will use dyanimat demonstrations to argue in favor of my favorite hypothesis about what the cerebellum does (it's a Bayesian dynamical state estimator), and to present a model of spike-based computation in the cerebellum (it's a random-lattice bootstrap particle filter).
Occam's razor  encourages us to search for simple, parsimonious explanations for phenomena. But how can we ever expect to find a simple explanation of a complex biological system? Ultimately we can't. However, we still can be guided toward a complete, complex explanation by initially finding a relatively simple explanation that accounts for a large fraction of experimental observations, and then incrementally adding complexity to account for deviations that are not initially explained. When our explanation is quantitative, then we say that we are developing a mathematical model of the system, and that the current model represents our working hypothesis about the structure/function of the system. There are potential pitfalls, but if we do this right, we leave in our wake not just one complex explanation, but a useful set of explanations that can be interpreted and communicated at different levels depending on the needs of our audience.
Our work on understanding sensorimotor integration in the human postural control system illustrates the process described above. By the 1970s, a lot was known about the detailed properties of sensory receptors contributing to postural control [cf. 2,3]. An early model incorporated these details in order to explain the control of upright stance . However, this model did not motivate further research, possibly because the model simply quantified the prevailing view that postural control was determined by a set of reflexes (e.g., stretch reflexes, vestibulo-spinal reflexes). However, in our attempt to develop a model to explain the major features of a comprehensive data set obtained by applying various postural perturbations , it did not seem necessary to consider the details of sensory receptors, but rather we made the simpler assumption that the nervous system combines sensory information to obtain an internal estimate of body orientation, and then the motor system acts on this estimate to generate a corrective torque in proportion to the angular deviation and the derivative of the angular deviation from an internal set point (i.e. and PD controller). This model quantified the sensory reweighting phenomenon whereby the nervous system alters its reliance on orientation information as a function of environmental conditions and external perturbations, and the model explained postural changes caused by a loss of vestibular sensory function . The model has motivated addition experiments to investigate the sensory reweighting phenomenon , has provided an explanation for a very odd experimental result whereby a particular manipulation can evoke sustained 1 Hz body oscillations , and has helped to explain how orientation information from a balance prosthesis is incorporated into the postural control system . This model has since been modified to account for other features of the experimental data set that were not fully explained by the original model [6,9]. The model continues to motivate our research, and our new experimental results motivate further development of the model.
This work was supported by NIH grant AG-17960.
At the cellular (fiber) level, the mechanical behavior of muscle is determined by the interaction of motor proteins within muscle sarcomeres. The properties of sarcomeres are well understood, and have provided predictive power for the development of musculoskeletal models of movement as well as a framework for interpreting empirical data. In recent years it has become increasingly apparent that the force, velocity, and power output of muscles is also highly dependent on the extracellular structures through which muscle fiber forces are transmitted. These effects are apparent in our studies of the effect of tendon elasticity on the mechanical and energetic performance of muscles during steady-speed running and acceleration. Series elastic elements are important during both activities but have opposite effects on muscle work. During steady-speed running, elastic elements reduce muscle work (1,2), while during accelerations elastic elements increase muscle work (3,4); in both cases the effect results from tendon's ability to uncouple body movements from muscle movements (5). We have also investigated the influence of dynamic changes in muscle architecture on the velocity of muscle contraction. Changes in muscle shape appear to mediate a variable gearing of muscle fibers, such that muscle fiber velocities are amplified during fast, low-force contractions. The evolution of these non-sarcomeric modulators of muscle force and velocity has likely been driven by the need to overcome the relatively constrained performance of skeletal muscle motors.
How do we perform tasks requiring that multiple muscles to be coordinated across a multifunctional joint?
We can choose to activate different parts of a given multifunctional muscle, perhaps by recruiting particular sets of motor units that are tuned to generate forces in targeted directions - This seems feasible, but we do not know how broadly such units can change the net force direction that is being generated by the multifunctional muscle.
Alternatively we can pair two or more muscles to achieve the desired torque magnitude and direction.
This presentation will review the potential for spatial tuning of motor units within a multifunctional muscle, and assess the estimated range of forces that could be generated in this way.
Work done in collaboration with N.L., Suresh, A.D. Kuo, and J. Kutch.
In the past, computational motor control has been approached from at least two major frameworks: the dynamic systems approach and the viewpoint of optimal control. The dynamic systems approach emphasizes motor control as a process of self-organization between an animal and its environment -- passive dynamic walking, for instance, could be the outcome of such a self-organizing process. Nonlinear differential equations that can model entrainment and synchronization behavior are among the most favorable tools of dynamic systems modelers. In contrast, optimal control approaches view motor control as the evolutionary or development result of a nervous system that tries to optimize rather general organizational principles, e.g., energy consumption or accurate task achievement. Optimal control theory is usually employed to develop appropriate theories. Interestingly, there is rather little interaction between dynamic systems and optimal control modelers as the two approaches follow rather different philosophies and are often viewed as diametrically opposing. In this talk, we present a computational approach to motor control that offers a unifying modeling framework for both dynamic systems and optimal control approaches. In discussions of several behavioral experiments and some theoretical and robotics studies, we demonstrate how our computational ideas allow both the representation of self-organizing processes and the optimization of movement based on reward criteria. Our modeling framework is rather simple and general, and opens opportunities to revisit many previous modeling results from this novel unifying view. Finally, we will also critically review which obstacles research in motor control has to address in the future to overcome the shortcoming of current theories. Examples from biped locomotion and quadruped locomotion over rough terrain, including robotics implementations, will serve as guiding examples for this discussion.
Primary motor cortex (MI) is a key component of the volitional motor system providing the largest contribution to the corticospinal tract and receiving input from many cortical and subcortical structures. The most common approach for interpreting MI function has been based on the notion of sensorimotor transformations, focusing attention on experiments that identify which coordinate frame best describes neural activity in MI. However, myriad coordinate frames or neural representations have been observed illustrating correlations with spatial goals, hand motion, joint motion, muscular torque, muscular power and EMG activity. How all these 'representations' contribute or create coordinated motor behavior remains unclear. The focus of my talk will be to provide an alternate approach for interpreting MI function based on optimal feedback control. This approach re-emphasizes the importance of sensory feedback to MI function and the adaptive nature of long-latency reflexes that predominantly involve a transcortical pathway through MI. I will describe experiments that examine the highly adaptive nature of long-latency reflexes in humans either to maintain online control or to rapidly switch behavioral goals. I hope to present preliminary observations illustrating corresponding rapid context-dependent changes in the response of MI neurons of non-human primates for these same tasks.
During the last four years, a number of simplistic robotic platforms were developed at the Lauflabor Locomotion Laboratory at the University of Jena, Germany. The concept behind these systems can be understood in the framework of the conceptual models on gait stability. With the exception of the MarcoHopper - which addresses reflex mechanisms in the leg - all other robots do not require sensory feedback in order to achieve steady locomotion. However, sensory feedback can be used to further enhance gait stability.
In this talk the interplay of robot experiments, conceptual models and human trails will be demonstrated at selected system levels. This includes effects of leg segmentation, gait pattern generation, impact management, trunk stability and muscle-reflex dynamics during locomotion. We will address issues which are specific to technical systems and compare them to solutions found in animals and humans. Finally, some implications for future robot developments and prosthetic or orthotic devices will be given.
Among all the different ways that we can travel, using two legs, from one point to another, we choose to use relatively stereotypical gaits, walking when we are not in a hurry and running when we are rushed. We show that minimizing a simple work-based metabolic cost in the context of a simple bipedal animal discovers idealized versions of these stereotypical gaits, and the gait transition.
Next, we show that strategies that avoid muscle mechanical work are often indistinguishable from those that minimize other simple models of metabolic cost, both ad hoc and physiologically-based. Fo instance, we show that when there is a perfect compliant spring in series with the leg muscles, minimizing muscle work and letting the tendon perform the work required to redirect the center of mass velocity, also minimizes a variety of other costs, including the integrated quasisteady ATP cleavage model used by Minetti & Alexander (1997). We will comment on extensions to these models, toward a better-predictive but still-simple model of steady legged locomotion.
The calculations described above make predictions only about steady locomotion -- but they can be simply modified to make predictions about how a legged animal, with full current state information, will respond to discrete perturbations if it tried to reliably get back to its nominal gait trajectory and minimize the effort required to do so. We will comment on some specific examples of such optimal "feedback" control calculations.
In the late 20th century roboticists reinvigorated the quantitative study of motor learning by applying robotic theory to human motor control and by using robotic arms as tools to interrogate the central nervous system. These interrogations revealed important computational features of internal representations by testing how people learn and how motor memories transfer to novel movement spaces and new tasks. In this decade research into motor memories continues, but a new approach changes movement environments across individual movements. This variance enables identification of how the nervous systems transforms individual sensed movements into incremental updates of predictive control. Although this approach centers on only the fastest components of motor adaptation, a coupling of these behavioral experiments with theoretical models has revealed important, precise characterizations of motor adaptation. This research considers the generalization of incremental adaptation, the real valued transformation of sense into adaptation, and how the adaptive process itself can change in concert with environmental demands.
I will review the evolution of trial-by-trial adaptation research, view current accomplishments in the field, and preview our nascent application of trial-by-trial approaches to understand cognition and childhood development.
Recent research suggests that the nervous system controls muscles by activating flexible combinations of muscle synergies to produce a wide repertoire of movements. Muscle synergies are like building blocks, defining characteristic patterns of activation across multiple muscles that may be unique to each individual, but perform similar functions. The identification of muscle synergies has strong implications for the organization and structure of the nervous system, providing a mechanism by which task-level motor intentions are translated into detailed, low-level muscle activation patterns. Understanding the complex interplay between neural circuits and biomechanics that give rise to muscle synergies will be critical to advancing our understanding of neural control mechanisms for movement. We propose that muscle synergies emerge from the interacting constraints and features of the nervous and musculoskeletal systems. Our rationale is supported by computational studies of motor cortex topography demonstrating that functionally-organized regions of the cortex may arise from interactions between the biomechanical characteristics of the behavioral repertoire and the biases in the nervous system towards co-localizing neurons that process similar information. As an example relevant to muscle synergies, consider the energetic constraints on the musculoskeletal and nervous systems during locomotion. Movement patterns are energetically efficient in a mechanical sense when joint motions are functionally immobilized or linearly correlated (e.g. "inverted pendulum," or "spring-mass" dynamics in locomotion). Simultaneously, energetic efficiency in neural systems - limiting the number of neurons dedicated to encode task performance - may favor piecewise-linear representations of complex elements. Thus the combined neural and mechanical energetic pressures may give rise to a motor control strategy of activating linear combinations of muscle synergies that coordinate the musculoskeletal system to act in low-dimensional movement patterns. Because of the large solution space of muscle synergies sufficiently near the energetically optimal operating regions defined by simple biomechanical models, a cascade of ancillary factors may influence the specific muscle synergy patterns within each individual.
We will demonstrate through neuromechanical modeling that muscle synergies represent a balance between biomechanical optimality and neural parsimony. In a nominal postural configuration, the experimental forces generated by muscle synergies are close to that predicted from an energetic optimum. However, in non-standard postural configurations, the experimental forces become non-optimal, but can be predicted if the assumption that the same muscle synergies are used at all postural configurations. These results suggest that the use of muscle synergies comes at a cost to motor performance. We propose that there is an associated neural "cost" to generating additional muscle synergies that must be balanced against mechanical and energetic costs.
Work done in collaboration with J. Lucas McKay
If you are going to act, you might as well act in the best way possible. Optimal control is a way to formalize this intuitive idea. Here I will summarize evidence that motor behavior is optimal with respect to ecologically-valid cost functions, and describe surprising properties that emerge from optimality. I will also discuss the features of the sensorimotor system which may enable it to find near-optimal solutions to seemingly intractable control problems. Inspired in part by these observations we have developed efficient new methods for optimal control. They include hierarchical feedback control through sensorimotor synergies, and a new mathematical framework in which the problem of nonlinear stochastic optimal control is reduced to a linear eigenvalue problem. These new methods are useful both in terms of developing more realistic theories of brain function and in terms of solving hard problems in control engineering.
The control of movement is highly complex, involving the specification of a large number of variables and elements with complex dynamical properties. It has been hypothesized that the nervous system overcomes this complexity by utilizing a low dimensional control strategy of muscle synergies, with each such synergy specifying a particular balance of activation across a set of muscles. However, it has been unclear how the set of muscles within a synergy are specified and whether such a control strategy is capable of producing realistic behaviors. We propose here a principle for the specification of muscle synergies: that synergies are chosen to best control the significant dynamics of the limb. We further evaluate the ability of the synergies chosen according to this principle to produce behaviors. Our analyses can provide insights into the strategies that might simplify the control of behavior by the nervous system, as well as suggesting potential strategies for rehabilitation or the reanimation of paralyzed limbs.
Dexterous manipulation has a certain mystique. This is largely because it, together with cognition and language, is intimately associated with our identity as a species. Much of the work in our laboratory is dedicated to developing the scientific paradigm and conceptual/computational tools to properly phrase and address a broad question: whether and how sensorimotor control for manipulation differs from, or recapitulates, systems and principles evolved for other, older systems in humans. Are we dexterous because of, or in spite of, hand anatomy? How do the specialized neural structures associated with hand function enable specific functional features of manipulation? In this short presentation, I will discuss two specific hypotheses. The first, on the "peripheral" side, is that the anatomically complexity of the hand (traditionally oversimplified or ignored) may in fact be critical to understanding brain-body co-evolution and neuromuscular control. The second, on the "central" side, is that even common and simple functional features like grasp acquisition for precision pinch require exquisitely time-critical sensorimotor function-which begins to explain the evolution of specialized neural circuits for the human hand. I will also discuss the implications of this work to progress in clinical rehabilitation and development of innovative robotic systems.
The human neural and musculoskeletal system has the capability to execute a wide range of muscle and limb coordination patterns. Yet, we tend to execute movements in a stereotypical manner. Since the human locomotor system is not dominated by mechanism constraints or hardwired neural circuits, it is usually assumed that such movements are voluntarily chosen because they are optimal. In the specific case of human gait, minimal energy is generally thought to be the optimality criterion. There is considerable experimental evidence for this concept. For instance, when human subjects are asked to make changes in speed, step length, cadence, or joint kinematics, an increase in oxygen uptake is always observed.
If gait is indeed governed by a minimal energy principle, it should be possible to predict human gait with a computational model that has sufficiently realistic mathematical descriptions of system dynamics and energetics. The ability to do this is important for two reasons. First, it would confirm our fundamental understanding of form-function relationships in human locomotion. Second, it would allow effective computer-aided design of prosthetics and orthotics, surgical alterations of the musculoskeletal system, and sports techniques. In such clinical applications, a mechanistic model for prediction of clinical outcome is crucial because it is difficult to apply the usual biomedical research paradigm of animal models to select a small number of specific treatments which can then be tested in clinical trials.
In spite of the existence of sophisticated computational models for the human musculoskeletal system, convincing de novo predictions of human gait have not been generated. We will present a review of the state of the art in this field. Much can be learned from observing which features of gait are predicted correctly, and which are not. In our own work and in that of others, we have observed that a minimal energy principle tends to predict a gait pattern in which the knee remains fully extended in the stance phase and quadriceps activation is unrealistically small. Humans flex the knee to about 15 degrees in the stance phase and have significant quadriceps activation in early stance. Several explanations are possible for our current inability to predict this characteristic feature of human gait from basic principles. One possible explanation is that our computational models are still an incomplete representation of skeletal dynamics and the mechanical/energetic properties of muscle. Another possibility is that (in spite of experimental evidence) energy is not the only optimization principle for human gait. These two hypotheses suggest radically different strategies for future research.
Work done in collaboration with Marko Ackermann.
Acknowledgments: This work was supported by the U.S. National Science Foundation (BES0302259) and the U.S. National Institutes of Health (1R01EB006735).