Toward Perceiving Robots as Humans: Three Handshake Models Face the Turing-Like Handshake Test.

In the Turing test a computer model is deemed to "think intelligently" if it can generate answers that are indistinguishable from those of a human. We developed an analogous Turing-like handshake test to determine if a machine can produce similarly indistinguishable movements. The test is administered through a telerobotic system in which an interrogator holds a robotic stylus and interacts with another party - artificial or human with varying levels of noise. The interrogator is asked which party seems to be more human. Here, we compare the human-likeness levels of three different models for handshake: (1) Tit-for-Tat model, (2) λ model, and (3) Machine Learning model. The Tit-for-Tat and the Machine Learning models generated handshakes that were perceived as the most human-like among the three models that were tested. Combining the best aspects of each of the three models into a single robotic handshake algorithm might allow us to advance our understanding of the way the nervous system controls sensorimotor interactions and further improve the human-likeness of robotic handshakes.

Force estimation from ensembles of Golgi tendon organs.

Golgi tendon organs (GTOs) located in the skeletal muscles provide the central nervous system with information about muscle tension. The ensemble firing of all GTO receptors in the muscle has been hypothesized to represent a reliable measure of the whole muscle force but the precision and accuracy of that information are largely unknown because it is impossible to record activity simultaneously from all GTOs in a muscle. In this study, we combined a new mathematical model of force sampling and transduction in individual GTOs with various models of motor unit (MU) organization and recruitment simulating various normal, pathological and neural prosthetic conditions. Our study suggests that in the intact muscle the ensemble GTO activity accurately encodes force information according to a nonlinear, monotonic relationship that has its steepest slope for low force levels and tends to saturate at the highest force levels. The relationship between the aggregate GTO activity and whole muscle tension under some pathological conditions is similar to one seen in the intact muscle during rapidly modulated, phasic excitation of the motor pool (typical for many natural movements) but quite different when the muscle is activated slowly or held at a given force level. Substantial deviations were also observed during simulated functional electrical stimulation.

Computationally efficient models of neuromuscular recruitment and mechanics.

We have improved the stability and computational efficiency of a physiologically realistic, virtual muscle (VM 3.*) model (Cheng et al 2000 J. Neurosci. Methods 101 117-30) by a simpler structure of lumped fiber types and a novel recruitment algorithm. In the new version (VM 4.0), the mathematical equations are reformulated into state-space representation and structured into a CMEX S-function in SIMULINK. A continuous recruitment scheme approximates the discrete recruitment of slow and fast motor units under physiological conditions. This makes it possible to predict force output during smooth recruitment and derecruitment without having to simulate explicitly a large number of independently recruited units. We removed the intermediate state variable, effective length (Leff), which had been introduced to model the delayed length dependency of the activation-frequency relationship, but which had little effect and could introduce instability under physiological conditions of use. Both of these changes greatly reduce the number of state variables with little loss of accuracy compared to the original VM. The performance of VM 4.0 was validated by comparison with VM 3.1.5 for both single-muscle force production and a multi-joint task. The improved VM 4.0 model is more suitable for the analysis of neural control of movements and for design of prosthetic systems to restore lost or impaired motor functions. VM 4.0 is available via the internet and includes options to use the original VM model, which remains useful for detailed simulations of single motor unit behavior.

Model-based sensorimotor integration for multi-joint control: development of a virtual arm model.

An integrated, sensorimotor virtual arm (VA) model has been developed and validated for simulation studies of control of human arm movements. Realistic anatomical features of shoulder, elbow and forearm joints were captured with a graphic modeling environment, SIMM. The model included 15 musculotendon elements acting at the shoulder, elbow and forearm. Muscle actions on joints were evaluated by SIMM generated moment arms that were matched to experimentally measured profiles. The Virtual Muscle (VM) model contained appropriate admixture of slow and fast twitch fibers with realistic physiological properties for force production. A realistic spindle model was embedded in each VM with inputs of fascicle length, gamma static (gamma(stat)) and dynamic (gamma(dyn)) controls and outputs of primary (I(a)) and secondary (II) afferents. A piecewise linear model of Golgi Tendon Organ (GTO) represented the ensemble sampling (I(b)) of the total muscle force at the tendon. All model components were integrated into a Simulink block using a special software tool. The complete VA model was validated with open-loop simulation at discrete hand positions within the full range of alpha and gamma drives to extrafusal and intrafusal muscle fibers. The model behaviors were consistent with a wide variety of physiological phenomena. Spindle afferents were effectively modulated by fusimotor drives and hand positions of the arm. These simulations validated the VA model as a computational tool for studying arm movement control. The VA model is available to researchers at website http://pt.usc.edu/cel .

Flexible communication and control protocol for injectable neuromuscular interfaces.

BION2 is a system based on injectable neuromuscular implants whose main goal is to restore the functional movement of paralyzed limbs. To achieve this objective, the functional requirements of the implanted interfaces include not only stimulation but also integrated sensors in order to detect patient intention, to provide servocontrol of muscle activation and to sense posture to inform more global motor planning and coordination. The technical constraints for managing the system include the efficient use of forward and reverse telemetry channels with limited capacity, minimization of adverse consequences from errors in data transmission or intermittent loss of power to the implants, and ability to adjust stimulation rates and phases to achieve efficient fine control of muscle force while minimizing fatigue. This paper describes a communication and control architecture with several novel features that address these requirements.

Functional electrical stimulation using microstimulators to correct foot drop: a case study.

This paper presents a case study that tested the feasibility and efficacy of using injectable microstimulators (BIONs) in a functional electrical stimulation (FES) device to correct foot drop. Compared with surface stimulation of the common peroneal nerve, stimulation with BIONs provides more selective activation of specific muscles. For example, stimulation of the tibialis anterior (TA) and extensor digitorum longus (EDL) muscles with BIONs produces ankle flexion without excessive inversion or eversion of the foot (i.e., balanced flexion). Efficacy was assessed using a 3-dimensional motion analysis of the ankle and foot trajectories during walking with and without stimulation. Without stimulation, the toe on the affected leg drags across the ground. BION stimulation of the TA muscle and deep peroneal nerve (which innervates TA and EDL) elevates the foot such that the toe clears the ground by 3 cm, which is equivalent to the toe clearance in the less affected leg. The physiological cost index (PCI) measured effort during walking. The PCI equals the change in heart rate (from rest to activity) divided by the walking speed; units are beats per metre. The PCI is high without stimulation (2.29 +/- 0.37, mean +/- SD) and greatly reduced with surface (1.29 +/- 0.10) and BIONic stimulation (1.46 +/- 0.24). Also, walking speed increased from 9.4 +/- 0.4 m/min without stimulation to 19.6 +/- 2.0 m/min with surface and 17.8 +/- 0.7 m/min with BIONic stimulation. These results suggest that FES delivered by a BION is an alternative to surface stimulation and provides selective control of muscle activation.

RF-powered BIONs for stimulation and sensing.

Virtually all bodily functions are controlled by electrical signals in nerves and muscles. Electrical stimulation can restore missing signals but this has been difficult to achieve practically because of limitations in the bioelectric interfaces. Wireless, injectable microdevices are versatile, robust and relatively inexpensive to implant in a variety of sites and applications. Several variants are now in clinical use or under development to perform stimulation and/or sensing functions and to operate autonomously or with continuous coordination and feedback control.

BIONic WalkAide for correcting foot drop.

The goal of this study was to test the feasibility and efficacy of using microstimulators (BIONs) to correct foot drop, the first human application of BIONs in functional electrical stimulation (FES). A prototype BIONic foot drop stimulator was developed by modifying a WalkAide2 stimulator to control BION stimulation of the ankle dorsiflexor muscles. BION stimulation was compared with surface stimulation of the common peroneal nerve provided by a normal WalkAide2 foot drop stimulator. Compared to surface stimulation, we found that BION stimulation of the deep peroneal nerve produces a more balanced ankle flexion movement without everting the foot. A 3-D motion analysis was performed to measure the ankle and foot kinematics with and without stimulation. Without stimulation, the toe on the affected leg drags across the ground. The BIONic WalkAide elevates the foot such that the toe clears the ground by 3 cm, which is equivalent to the toe clearance in the unaffected leg. The physiological cost index (PCI) was used to measure effort during walking. The PCI is high without stimulation (2.29 +/- 0.37; mean +/- S.D.) and greatly reduced with surface (1.29 +/- 0.10) and BION stimulation (1.46 +/- 0.24). Also, walking speed is increased from 9.4 +/- 0.4 m/min. without stimulation to 19.6 +/- 2.0 m/min. with surface and 17.8 +/- 0.7 m/min. with BION stimulation. We conclude that functional electrical stimulation with BIONs is a practical alternative to surface stimulation and provides more selective control of muscle activation.

Development of clinician-friendly software for musculoskeletal modeling and control.

Research and development in various fields dealing with human movement has been hampered by the lack of adequate software tools. We have formed a core development team to organize a collective effort by the research community to develop musculoskeletal modeling software that satisfies the requirements of both researchers and clinicians. We have identified initial requirements and have developed some of the basic components. We are developing common standards to facilitate sharing and reuse of musculoskeletal models and their component parts. Free distribution of the software and its source code will allow users to contribute to further development of the software as new models and data become available in the future.

Advanced modeling environment for developing and testing FES control systems.

Realistic models of neuromusculoskeletal systems can provide a safe and convenient environment for the design and evaluation of controllers for functional electrical stimulation (FES) prior to clinical trials. We have developed a set of integrated musculoskeletal modeling tools to facilitate the model building process. Simulink models of musculoskeletal systems are created using two software packages developed in our laboratory, Musculoskeletal Modeling in Simulink (MMS) and virtual muscle, in addition to one software package available commercially, SIMM (Musculographics Inc., USA). MMS converts anatomically accurate musculoskeletal models generated by SIMM into Simulink(R) blocks. It also removes run-time constraints on kinetic simulations in SIMM, and allows the development of complex musculoskeletal models without writing a line of code. Virtual muscle builds realistic Simulink models of muscles responding to either natural recruitment or FES. Models of sensorimotor control systems can be developed using various Matlab (Mathworks Inc., USA) toolboxes and integrated easily with these musculoskeletal blocks in the graphical environment of Simulink.

Neck muscles in the rhesus monkey. II. Electromyographic patterns of activation underlying postures and movements.

Electromyographic (EMG) activity was recorded in < or = 12 neck muscles in four alert monkeys whose heads were unrestrained to describe the spatial and temporal patterns of neck muscle activation accompanying a large range of head postures and movements. Some head postures and movements were elicited by training animals to generate gaze shifts to visual targets. Other spontaneous head movements were made during orienting, tracking, feeding, expressive, and head-shaking behaviors. These latter movements exhibited a wider range of kinematic patterns. Stable postures and small head movements of only a few degrees were associated with activation of a small number of muscles in a reproducible synergy. Additional muscles were recruited for more eccentric postures and larger movements. For head movements during trained gaze shifts, movement amplitude, velocity, and acceleration were correlated linearly and agonist muscles were recruited without antagonist muscles. Complex sequences of reciprocal bursts in agonist and antagonist muscles were observed during very brisk movements. Turning movements of similar amplitudes that began from different initial head positions were associated with systematic variations in the activities of different muscles and in the relative timings of these activities. Unique recruitment synergies were observed during feeding and head-shaking behaviors. Our results emphasize that the recruitment of a given muscle was generally ordered and consistent but that strategies for coordination among various neck muscles were often complex and appeared to depend on the specifics of musculoskeletal architecture, posture, and movement kinematics that differ substantially among species.

BION system for distributed neural prosthetic interfaces.

We have developed the first in a planned series of neural prosthetic interfaces that allow multichannel systems to be assembled from single-channel micromodules called BIONs (BIOnic Neurons). Multiple BION implants can be injected directly into the sites requiring stimulating or sensing channels, where they receive power and digital commands by inductive coupling to an externally generated radio-frequency magnetic field. This article describes some of the novel technology required to achieve the required microminiaturization, hermeticity, power efficiency and clinical performance. The BION1 implants are now being used to electrically exercise paralyzed and weak muscles to prevent or reverse disuse atrophy. This modular, wireless approach to interfacing with the peripheral nervous system should facilitate the development of progressively more complex systems required to address a growing range of clinical applications, leading ultimately to synthesizing complete voluntary functions such as reach and grasp.

Learning from the spinal cord.

The graceful control of multiarticulated limbs equipped with slow, non-linear actuators (muscles) is a difficult problem for which robotic engineering affords no general solution. The vertebrate spinal cord provides an existence proof that such control is, indeed, possible. The biological solution is complex and incompletely known, despite a century of meticulous neurophysiological research, celebrated in part by this symposium. This is frustrating for those who would reanimate paralysed limbs either through promoting regeneration of the injured spinal cord or by functional electrical stimulation. The importance of and general role played by the spinal cord might be more easily recognized by analogy to marionette puppets, another system in which a brain (the puppeteer's) must cope with a large number of partially redundant actuators (strings) moving a mechanical linkage with complex intrinsic properties.

Real-time sonography to estimate muscle thickness: comparison with MRI and CT.

PURPOSE: We investigated the feasibility of using real-time sonography to measure muscle thickness. Clinically, this technique would be used to measure the thickness of human muscles in which intramuscular microstimulators have been implanted to treat or prevent disuse atrophy. METHODS: Porcine muscles were implanted with microstimulators and imaged with sonography, MRI, and CT to assess image artifacts created by the microstimulators and to design protocols for image alignment between methods. Sonography and MRI were then used to image the deltoid and supraspinatus muscles of 6 healthy human subjects. RESULTS: Microstimulators could be imaged with all 3 methods, producing only small imaging artifacts. Muscle-thickness measurements agreed well between methods, particularly when external markers were used to precisely align the imaging planes. The correlation coefficients for sonographic and MRI measurements were 0.96 for the supraspinatus and 0.97 for the deltoid muscle. Repeated sonographic measurements had a low coefficient of variation: 2.3% for the supraspinatus and 3.1% for the deltoid muscle. CONCLUSIONS: Real-time sonography is a relatively simple and inexpensive method of accurately measuring muscle thickness as long as the operator adheres to a strict imaging protocol and avoids excessive pressure with the transducer.

Recruitment properties of intramuscular and nerve-trunk stimulating electrodes.

Functionally useful reanimation of paralyzed limbs generally requires reliable, finely graded control of muscle recruitment and force with minimal fatigue. We used force and electromyographic (EMG) recordings in combination with myofibrillar adenosine triphosphatase activity and glycogen depletion analysis to investigate the recruitment properties of intramuscular (IM) and nerve cuff (NC) stimulating electrodes implanted acutely or chronically in cat hindlimbs. Overall, 32 muscles were submaximally stimulated with current intensities producing approximately 20% of maximal twitch force using 330 ms trains of pulses at 20 and 40 pps. Both the glycogen-depletion and fatigue-test results were found to be difficult to interpret because NC stimulation resulted in surprisingly unstable recruitment during such trains. Fluctuations of force and M-waves within trains of identical stimuli were significantly greater for NC than for IM stimulation. NC stimulation produced much steeper recruitment curves and a reduced tetanus/twitch ratio compared to IM stimulation. IM stimulation produced more reliable and less fatigable recruitment of a mix of motor unit types that tended to be localized in neuromuscular compartments containing, or adjacent to, the IM electrode. We hypothesize that trains of submaximal stimulation applied through NC electrodes resulted in fluctuating recruitment because this electrode configuration magnifies the effects of refractoriness and small changes in axonal excitability during pulse trains.

Virtual muscle: a computational approach to understanding the effects of muscle properties on motor control.

This paper describes a computational approach to modeling the complex mechanical properties of muscles and tendons under physiological conditions of recruitment and kinematics. It is embodied as a software package for use with Matlab and Simulink that allows the creation of realistic musculotendon elements for use in motor control simulations. The software employs graphic user interfaces (GUI) and dynamic data exchange (DDE) to facilitate building custom muscle model blocks and linking them to kinetic analyses of complete musculoskeletal systems. It is scalable in complexity and accuracy. The model is based on recently published data on muscle and tendon properties measured in feline slow- and fast-twitch muscle, and incorporates a novel approach to simulating recruitment and frequency modulation of different fiber-types in mixed muscles. This software is distributed freely over the Internet at http://ami.usc.edu/mddf/virtualmuscle.

Measured and modeled properties of mammalian skeletal muscle: III. the effects of stimulus frequency on stretch-induced force enhancement and shortening-induced force depression.

Stretch-induced force enhancement and shortening-induced force depression were examined in fast-twitch feline caudofemoralis muscle at 37 degrees C. These phenomena were induced by applying ramp length changes during the first 100--200 ms of an otherwise isometric contraction. The effects of various stimulus frequencies ranging from 30 to 120 pps were investigated over lengths ranging from 0.85 to 1.15 L0. Distributed asynchronous stimulation of bundles of ventral roots was employed to produce smooth contractions at sub-tetanic stimulus frequencies in whole muscle. Of the two components of force enhancement identified by Noble (1992) we observed only the transient component that decays with time; we did not observe residual force enhancement. The force depression that we observed was symmetrical in almost all respects to the transient force enhancement, and was unlike the shortening-induced de-activation and residual force depression identified by Edman (Edman. 1975; Edman et al., 1993). Both transient force enhancement and depression were independent of work, load and activation. Reversals in the direction of ramp length changes following either an initial stretch or initial shortening were shown to cancel the effects of both transient force enhancement and transient force depression. The distances over which these cancellations could be achieved were different for the lengthening and shortening effects. This asymmetry can be reconciled with the predictions of Huxley's original cross-bridge mechanism by incorporating the recent suggestion that myosin heads can interact with multiple actin binding sites during a single 'working' stroke. We conclude that the types of force enhancement/ depression that are most likely to be encountered under physiological conditions are the transient effects observed here, but that even these will have relatively little effect on force production during most natural behaviors.

Measured and modeled properties of mammalian skeletal muscle: IV. dynamics of activation and deactivation.

The interactive effects of length and stimulus frequency on rise and fall times and on sag were investigated in fast-twitch feline caudofemoralis at normal body temperature. The length and stimulus frequency ranges studied were 0.8 1.2 L0 and 15 60 pps. Isometric rise times were shortest under two sets of conditions: short lengths + low stimulus frequencies and long lengths + high stimulus frequencies. In contrast the isometric fall time relationship showed a single minimum at short lengths + low stimulus frequencies. Velocity was shown to have an additional effect on fall time, but only at higher stimulus frequencies (40 60 pps): fall times were shorter during movement in either direction as compared to isometric. The effects of sag were greatest at shorter lengths and lower stimulus frequencies during isometric stimulus trains. Potential mechanisms underlying this last effect were investigated by comparing isometric twitches elicited prior to and immediately following a sag-inducing stimulus train. Post-sag twitches produced less force, reached peak force earlier and initially decayed more quickly compared to pre-sag twitches. However, the final rate of force decay and the initial rate of force rise (during the first 15 ms) were unaffected by sag. We construct a logical argument based on these findings to hypothesize that the predominant mechanism underlying sag is an increase in the rate of sarcoplasmic calcium ion removal. All of the above findings were used to construct a model of activation dynamics for fast-twitch muscle, which was then extrapolated to slow-twitch muscle. When coupled with a previous model of kinematic dynamics, the complete model produced accurate predictions of the forces actually recorded during experiments in which we applied concurrent dynamic changes in length. velocity and stimulus frequency.

What do reflex and voluntary mean? Modern views on an ancient debate.

Are the words reflex and voluntary useful scientific concepts, or are they prescientific terms that should be discarded? Physiologists use these words routinely in their publications, in laboratory experiments and, indeed, like most lay people, in their daily lives. The tacit assumption is that we all know, more or less, what they mean. However, the issue has a rich history of philosophical and scientific debate; and, as this article demonstrates, present-day researchers still cannot reach a consensus on the meaning of the words and on whether it is possible to draw a scientific distinction between them. The five authors present five quite different analyses. In broad terms, they split into two camps: those who equate voluntary behaviours with consciousness and suppressibility and those who view all behaviours as sensorimotor interactions, the complexity of which determines whether they are reflexive or voluntary. According to the first view, most movements of daily life are neither purely reflex nor purely voluntary. They fall into the middle ground of automatic motor programs. According to the second view, as neuroscience advances the class of reflex behaviours will grow and the class of voluntary behaviours will shrink.

Overcomplete musculature or underspecified tasks?

The number of muscles in the body is actually fairly close to the number required to control completely all its degrees of freedom. The apparent need for a coordinating principle arises from the experimental practice of asking subjects to perform simple movements and assuming that they make no implicit assumptions about other constraints. Natural activities include implicit constraints that differ greatly for different tasks and circumstances and that would be met best by a nervous system free of a priori principles.