Neurorehabilitation robotics: how much control should therapists have?

Robotic technologies for rehabilitating motor impairments from neurological injuries have been the focus of intensive research and capital investment for more than 30 years. However, these devices have failed to convincingly demonstrate greater restoration of patient function compared to conventional therapy. Nevertheless, robots have value in reducing the manual effort required for physical therapists to provide high-intensity, high-dose interventions. In most robotic systems, therapists remain outside the control loop to act as high-level supervisors, selecting and initiating robot control algorithms to achieve a therapeutic goal. The low-level physical interactions between the robot and the patient are handled by adaptive algorithms that can provide progressive therapy. In this perspective, we examine the physical therapist's role in the control of rehabilitation robotics and whether embedding therapists in lower-level robot control loops could enhance rehabilitation outcomes. We discuss how the features of many automated robotic systems, which can provide repeatable patterns of physical interaction, may work against the goal of driving neuroplastic changes that promote retention and generalization of sensorimotor learning in patients. We highlight the benefits and limitations of letting therapists physically interact with patients through online control of robotic rehabilitation systems, and explore the concept of trust in human-robot interaction as it applies to patient-robot-therapist relationships. We conclude by highlighting several open questions to guide the future of therapist-in-the-loop rehabilitation robotics, including how much control to give therapists and possible approaches for having the robotic system learn from therapist-patient interactions.

Recurrence Quantification Analysis of Ankle Kinematics During Gait in Individuals With Chronic Ankle Instability.

Purpose: An investigation of the ankle dynamics in a motor task may generate insights into the etiology of chronic ankle instability (CAI). This study presents a novel application of recurrence quantification analysis (RQA) to examine the ankle dynamics during walking. We hypothesized that CAI is associated with changes in the ankle dynamics as assessed by measures of determinism and laminarity using RQA. Methods: We recorded and analyzed the ankle position trajectories in the frontal and sagittal planes from 12 participants with CAI and 12 healthy controls during treadmill walking. We used time-delay embedding to reconstruct the position trajectories to a phase space that represents the states of the ankle dynamics. Based on the phase space trajectory, a recurrence plot was constructed and two RQA variables, the percent determinism (%DET) and the percent laminarity (%LAM), were derived from the recurrence plot to quantify the ankle dynamics. Results: In the frontal plane, the %LAM in the CAI group was significantly lower than that in the control group (p < 0.05. effect size = 0.86). This indicated that the ankle dynamics in individuals with CAI is less likely to remain in the same state. No significant results were found in the %DET or in the sagittal plane. Conclusion: A lower frontal-plane %LAM may reflect more frequent switching between different patterns of neuromuscular control states due to the instabilities associated with CAI. With further study and development, %LAM may have the potential to become a useful biomarker for CAI.

Visual programming for accessible interactive musculoskeletal models.

OBJECTIVE: Musculoskeletal modeling and simulation are powerful research and education tools in engineering, neuroscience, and rehabilitation. Interactive musculoskeletal models (IMMs) can be controlled by muscle activity recorded with electromyography (EMG). IMMs are typically coded using textual programming languages that present barriers to understanding for non-experts. The goal of this project was to use a visual programming language (Simulink) to create and test an IMM that is accessible to non-specialists for research and educational purposes. RESULTS: The developed IMM allows users to practice a goal-directed task with different control modes (keyboard, mouse, and EMG) and actuator types (muscle model, force generator, and torque generator). Example data were collected using both keyboard and EMG control. One male participant in his early 40's performed a goal-directed task for four sequential trials using each control mode. For EMG control, the participant used a low-cost EMG system with consumer-grade EMG sensors and an Arduino microprocessor. The participant successfully performed the task with both control modes, but the inability to grade muscle model excitation and co-activate antagonist muscles limited performance with keyboard control. The IMM developed for this project serves as a foundation that can be further tailored to specific research and education needs.

Exploiting telerobotics for sensorimotor rehabilitation: a locomotor embodiment.

BACKGROUND: Manual treadmill training is used for rehabilitating locomotor impairments but can be physically demanding for trainers. This has been addressed by enlisting robots, but in doing so, the ability of trainers to use their experience and judgment to modulate locomotor assistance on the fly has been lost. This paper explores the feasibility of a telerobotics approach for locomotor training that allows patients to receive remote physical assistance from trainers. METHODS: In the approach, a trainer holds a small robotic manipulandum that shadows the motion of a large robotic arm magnetically attached to a locomoting patient's leg. When the trainer deflects the manipulandum, the robotic arm applies a proportional force to the patient. An initial evaluation of the telerobotic system's transparency (ability to follow the leg during unassisted locomotion) was performed with two unimpaired participants. Transparency was quantified by the magnitude of unwanted robot interaction forces. In a small six-session feasibility study, six individuals who had prior strokes telerobotically interacted with two trainers (separately), who assisted in altering a targeted gait feature: an increase in the affected leg's swing length. RESULTS: During unassisted walking, unwanted robot interaction forces averaged 3-4 N (swing-stance) for unimpaired individuals and 2-3 N for the patients who survived strokes. Transients averaging about 10 N were sometimes present at heel-strike/toe-off. For five of six patients, these forces increased with treadmill speed during stance (R2 = .99; p < 0.001) and increased with patient height during swing (R2 = .71; p = 0.073). During assisted walking, the trainers applied 3.0 ± 2.8 N (mean ± standard deviation across patients) and 14.1 ± 3.4 N of force anteriorly and upwards, respectively. The patients exhibited a 20 ± 21% increase in unassisted swing length between Days 1-6 (p = 0.058). CONCLUSIONS: The results support the feasibility of locomotor assistance with a telerobotics approach. Simultaneous measurement of trainer manipulative actions, patient motor responses, and the forces associated with these interactions may prove useful for testing sensorimotor rehabilitation hypotheses. Further research with clinicians as operators and randomized controlled trials are needed before conclusions regarding efficacy can be made.

Visual dynamics cues in learning complex physical interactions.

This study investigated the role of visual dynamics cues (VDCs) in learning to interact with a complex physical system. Manual gait training was used as an exemplary case, as it requires therapists to control the non-trivial locomotor dynamics of patients. A virtual analog was developed that allowed naïve subjects to manipulate the leg of a virtual stroke survivor (a virtual patient; VP) walking on a treadmill using a small robotic manipulandum. The task was to make the VP's leg pass through early, mid, and late swing gait targets. One group of subjects (n = 17) started practice seeing the VP's affected thigh and shank (i.e., VDCs); a second control group (n = 16) only saw the point-of-contact (VP ankle). It was hypothesized that, if seeing the VP's leg provides beneficial dynamics information, the VDC group would have better task performance and generalization than controls. Results were not supportive. Both groups had similar task performance, and for the late swing gait target, a decrement in manipulative accuracy was observed when VDCs were removed in a generalization task. This suggests that when learning to manipulate complex dynamics, VDCs can create a dependency that negatively affects generalization if the visual context is changed.

Learning to shape virtual patient locomotor patterns: internal representations adapt to exploit interactive dynamics.

This work aimed to understand the sensorimotor processes used by humans when learning how to manipulate a virtual model of locomotor dynamics. Prior research shows that when interacting with novel dynamics humans develop internal models that map neural commands to limb motion and vice versa. Whether this can be extrapolated to locomotor rehabilitation, a continuous and rhythmic activity that involves dynamically complex interactions, is unknown. In this case, humans could default to model-free strategies. These competing hypotheses were tested with a novel interactive locomotor simulator that reproduced the dynamics of hemiparetic gait. A group of 16 healthy subjects practiced using a small robotic manipulandum to alter the gait of a virtual patient (VP) that had an asymmetric locomotor pattern modeled after stroke survivors. The point of interaction was the ankle of the VP's affected leg, and the goal was to make the VP's gait symmetric. Internal model formation was probed with unexpected force channels and null force fields. Generalization was assessed by changing the target locomotor pattern and comparing outcomes with a second group of 10 naive subjects who did not practice the initial symmetric target pattern. Results supported the internal model hypothesis with aftereffects and generalization of manipulation skill. Internal models demonstrated refinements that capitalized on the natural pendular dynamics of human locomotion. This work shows that despite the complex interactive dynamics involved in shaping locomotor patterns, humans nevertheless develop and use internal models that are refined with experience. NEW & NOTEWORTHY This study aimed to understand how humans manipulate the physics of locomotion, a common task for physical therapists during locomotor rehabilitation. To achieve this aim, a novel locomotor simulator was developed that allowed participants to feel like they were manipulating the leg of a miniature virtual stroke survivor walking on a treadmill. As participants practiced improving the simulated patient's gait, they developed generalizable internal models that capitalized on the natural pendular dynamics of locomotion.

Magnetic resonance images and measurements of the volume, proportion, and longitudinal distribution of contractile and non-contractile tissue in the dorsi- and plantar flexor muscles of healthy young and older adults.

OBJECTIVE: This paper presents magnetic resonance images of the dorsi- and plantar flexor muscles for individual young and older healthy adults. Also included are measurements of the volume, proportion, and longitudinal distribution of contractile and non-contractile tissue. This dataset was previously used to quantify age-related differences in these measures, constrain subject- and muscle-specific estimates of dorsi- and plantar flexor maximal isometric force capability, and quantify the degree to which maximal isometric force capability explains the age-related variance in postural control. DATA DESCRIPTION: The data include contiguous axial magnetic resonance images of the lower leg for 12 young (21-31 years) and 12 older (66-79 years) healthy adults. The data are in the form of MATLAB binary files with a freely distributable custom MATLAB analysis program that allows image viewing and navigation in two and three dimensions, muscle outlining, tissue segmentation, and cross-sectional area calculation. The latter measurements are contained in a set of companion MATLAB binary files, which are included with the image data files. If desired, the magnetic resonance images could be used to identify other anatomical structures, or the MATLAB programs could be used to analyze other image sets.

An Interactive Simulator for Imposing Virtual Musculoskeletal Dynamics.

OBJECTIVE: disease processes are often marked by both neural and muscular changes that alter movement control and execution, but these adaptations are difficult to tease apart because they occur simultaneously. This is addressed by swapping an individual's limb dynamics with a neurally controlled facsimile using an interactive musculoskeletal simulator (IMS) that allows controlled modifications of musculoskeletal dynamics. This paper details the design and operation of the IMS, quantifies and describes human adaptation to the IMS, and determines whether the IMS allows users to move naturally, a prerequisite for manipulation experiments. METHODS: healthy volunteers (n = 4) practiced a swift goal-directed task (back-and-forth elbow flexion/extension) for 90 trials with the IMS off (normal dynamics) and 240 trials with the IMS on, i.e., the actions of a user's personalized electromyography-driven musculoskeletal model are robotically imposed back onto the user. RESULTS: after practicing with the IMS on, subjects could complete the task with end-point errors of 1.56°, close to the speed-matched IMS-off error of 0.57°. Muscle activity, joint torque, and arm kinematics for IMS-on and -off conditions were well matched for three subjects (root-mean-squared error [RMSE] = 0.16 N·m), but the error was higher for one subject with a small stature (RMSE = 0.25 N·m). CONCLUSION: a well-matched musculoskeletal model allowed IMS users to perform a goal-directed task nearly as well as when the IMS was not active. SIGNIFICANCE: this advancement permits real-time manipulations of musculoskeletal dynamics, which could increase our understanding of muscular and neural co-adaptations to injury, disease, disuse, and aging.

Elucidating Sensorimotor Control Principles with Myoelectric Musculoskeletal Models.

There is an old saying that you must walk a mile in someone's shoes to truly understand them. This mini-review will synthesize and discuss recent research that attempts to make humans "walk a mile" in an artificial musculoskeletal system to gain insight into the principles governing human movement control. In this approach, electromyography (EMG) is used to sample human motor commands; these commands serve as inputs to mathematical models of muscular dynamics, which in turn act on a model of skeletal dynamics to produce a simulated motor action in real-time (i.e., the model's state is updated fast enough produce smooth motion without noticeable transitions; Manal et al., 2002). In this mini-review, these are termed myoelectric musculoskeletal models (MMMs). After a brief overview of typical MMM design and operation principles, the review will highlight how MMMs have been used for understanding human sensorimotor control and learning by evoking apparent alterations in a user's biomechanics, neural control, and sensory feedback experiences.

Predictability and Robustness in the Manipulation of Dynamically Complex Objects.

Manipulation of complex objects and tools is a hallmark of many activities of daily living, but how the human neuromotor control system interacts with such objects is not well understood. Even the seemingly simple task of transporting a cup of coffee without spilling creates complex interaction forces that humans need to compensate for. Predicting the behavior of an underactuated object with nonlinear fluid dynamics based on an internal model appears daunting. Hence, this research tests the hypothesis that humans learn strategies that make interactions predictable and robust to inaccuracies in neural representations of object dynamics. The task of moving a cup of coffee is modeled with a cart-and-pendulum system that is rendered in a virtual environment, where subjects interact with a virtual cup with a rolling ball inside using a robotic manipulandum. To gain insight into human control strategies, we operationalize predictability and robustness to permit quantitative theory-based assessment. Predictability is quantified by the mutual information between the applied force and the object dynamics; robustness is quantified by the energy margin away from failure. Three studies are reviewed that show how with practice subjects develop movement strategies that are predictable and robust. Alternative criteria, common for free movement, such as maximization of smoothness and minimization of force, do not account for the observed data. As manual dexterity is compromised in many individuals with neurological disorders, the experimental paradigm and its analyses are a promising platform to gain insights into neurological diseases, such as dystonia and multiple sclerosis, as well as healthy aging.

Neuromotor Noise Is Malleable by Amplifying Perceived Errors.

Variability in motor performance results from the interplay of error correction and neuromotor noise. This study examined whether visual amplification of error, previously shown to improve performance, affects not only error correction, but also neuromotor noise, typically regarded as inaccessible to intervention. Seven groups of healthy individuals, with six participants in each group, practiced a virtual throwing task for three days until reaching a performance plateau. Over three more days of practice, six of the groups received different magnitudes of visual error amplification; three of these groups also had noise added. An additional control group was not subjected to any manipulations for all six practice days. The results showed that the control group did not improve further after the first three practice days, but the error amplification groups continued to decrease their error under the manipulations. Analysis of the temporal structure of participants' corrective actions based on stochastic learning models revealed that these performance gains were attained by reducing neuromotor noise and, to a considerably lesser degree, by increasing the size of corrective actions. Based on these results, error amplification presents a promising intervention to improve motor function by decreasing neuromotor noise after performance has reached an asymptote. These results are relevant for patients with neurological disorders and the elderly. More fundamentally, these results suggest that neuromotor noise may be accessible to practice interventions.

Neural Control Adaptation to Motor Noise Manipulation.

Antagonistic muscular co-activation can compensate for movement variability induced by motor noise at the expense of increased energetic costs. Greater antagonistic co-activation is commonly observed in older adults, which could be an adaptation to increased motor noise. The present study tested this hypothesis by manipulating motor noise in 12 young subjects while they practiced a goal-directed task using a myoelectric virtual arm, which was controlled by their biceps and triceps muscle activity. Motor noise was increased by increasing the coefficient of variation (CV) of the myoelectric signals. As hypothesized, subjects adapted by increasing antagonistic co-activation, and this was associated with reduced noise-induced performance decrements. A second hypothesis was that a virtual decrease in motor noise, achieved by smoothing the myoelectric signals, would have the opposite effect: co-activation would decrease and motor performance would improve. However, the results showed that a decrease in noise made performance worse instead of better, with no change in co-activation. Overall, these findings suggest that the nervous system adapts to virtual increases in motor noise by increasing antagonistic co-activation, and this preserves motor performance. Reducing noise may have failed to benefit performance due to characteristics of the filtering process itself, e.g., delays are introduced and muscle activity bursts are attenuated. The observed adaptations to increased noise may explain in part why older adults and many patient populations have greater antagonistic co-activation, which could represent an adaptation to increased motor noise, along with a desire for increased joint stability.

A reinforcement learning approach to gait training improves retention.

Many gait training programs are based on supervised learning principles: an individual is guided towards a desired gait pattern with directional error feedback. While this results in rapid adaptation, improvements quickly disappear. This study tested the hypothesis that a reinforcement learning approach improves retention and transfer of a new gait pattern. The results of a pilot study and larger experiment are presented. Healthy subjects were randomly assigned to either a supervised group, who received explicit instructions and directional error feedback while they learned a new gait pattern on a treadmill, or a reinforcement group, who was only shown whether they were close to or far from the desired gait. Subjects practiced for 10 min, followed by immediate and overnight retention and over-ground transfer tests. The pilot study showed that subjects could learn a new gait pattern under a reinforcement learning paradigm. The larger experiment, which had twice as many subjects (16 in each group) showed that the reinforcement group had better overnight retention than the supervised group (a 32% vs. 120% error increase, respectively), but there were no differences for over-ground transfer. These results suggest that encouraging participants to find rewarding actions through self-guided exploration is beneficial for retention.

Effects of kinematic vibrotactile feedback on learning to control a virtual prosthetic arm.

BACKGROUND: After a limb is lost a prosthesis can restore function. For maximum utility, prosthetic limbs should accept movement commands and provide force and motion feedback, which can be conveyed with vibrotactile feedback (VIBF). While prior studies have shown that force-based VIBF benefits control, the merits of motion-based VIBF are unclear. Our goal was to clarify the effectiveness of position- and velocity-based VIBF for prosthetic arm control. METHODS: Healthy adults with normal limb function practiced a goal-directed task with a virtual myoelectric prosthetic arm. A linear resonant actuator on the wrist provided VIBF. Two groups with nine subjects each received amplitude modulated VIBF in addition to visual feedback while practicing the task. In one group, the VIBF was proportional to the virtual arm's position, and in the other group, velocity. A control group of nine subjects received only visual feedback. Subjects practiced for 240 trials, followed by 180 trials with feedback manipulations for the VIBF groups. Performance was characterized by end-point error, movement time, and a composite skill measure that combined these quantities. A second experiment with a new group of five subjects assessed discrimination capabilities between different position- and velocity-based VIBF profiles. RESULTS: With practice all groups improved their skill in controlling the virtual prosthetic arm. Subjects who received additional position- and velocity-based VIBF learned at the same rate as the control group, who received only visual feedback (learning rate time constant: about 40 trials). When visual feedback was subsequently removed leaving only VIBF, performance was no better than with no feedback at all. When VIBF was removed leaving only visual feedback, about half of the participants performed better, instead of worse. The VIBF discrimination tests showed that subjects could detect virtual arm angular position and velocity differences of about 5 deg and 20 deg/s, respectively. CONCLUSIONS: Kinematic VIBF did not increase the rate of skill acquisition or improve performance when controlling a virtual myoelectric prosthetic arm, whether provided in isolation or coupled with visual feedback. VIBF had a deleterious effect on performance for some individuals, who may have had difficulty integrating kinematic VIBF information into their control strategies.

An arm for a leg: Adapting a robotic arm for gait rehabilitation.

The purpose of this study was to adapt a multipurpose robotic arm for gait rehabilitation. An advantage of this approach is versatility: a robotic arm can be attached to almost any point on the body to assist with lower- and upper-extremity rehabilitation. This may be more cost-effective than purchasing and training rehabilitation staff to use several specialized rehabilitation robots. Robotic arms also have a more human-like morphology, which may make them less intimidating or alien to patients. In this study a mechanical interface was developed that allows a fast, secure, and safe attachment between a robotic arm and a human limb. The effectiveness of this interface was assessed by having two healthy subjects walk on a treadmill with and without a robotic arm attached to their legs. The robot's ability to follow the subjects' swinging legs was evaluated at slow and fast walking speeds. Two different control schemes were evaluated: one using the standard manufacturer-provided control algorithm, and another using a custom algorithm that actively compensated for robot-human interaction forces. The results showed that both robot control schemes performed well for slow walking. There were negligible differences between subjects' gait kinematics with and without the robot. During fast walking with the robot, similar results were obtained for one subject; however, the second subject demonstrated noticeable gait modifications. Together, these results show the feasibility of adapting a multipurpose robotic arm for gait rehabilitation.

Rhythmic manipulation of objects with complex dynamics: predictability over chaos.

The study of object manipulation has been largely confined to discrete tasks, where accuracy, mechanical effort, or smoothness were examined to explain subjects' preferred movements. This study investigated a rhythmic manipulation task, which involved continuous interaction with a nonlinear object that led to unpredictable object behavior. Using a simplified virtual version of the task of carrying a cup of coffee, we studied how this unpredictable object behavior affected the selected strategies. The experiment was conducted in a virtual set-up, where subjects moved a cup with a ball inside, modeled by cart-and-pendulum dynamics. Inverse dynamics calculations of the system showed that performing the task with different amplitudes and relative phases required different force profiles and rendered the object's dynamics with different degrees of predictability (quantified by Mutual Information between the applied force and the cup kinematics and its sensitivity). Subjects (n = 8) oscillated the virtual cup between two targets via a robotic manipulandum, paced by a metronome at 1 Hz for 50 trials, each lasting 45 s. They were free to choose their movement amplitude and relative phase between the ball and cup. Experimental results showed that subjects increased their movement amplitudes, which rendered the interactions with the object more predictable and with lower sensitivity to the execution variables. These solutions were associated with higher average exerted force and lower object smoothness, contradicting common expectations from studies on discrete object manipulation and unrestrained movements. Instead, the findings showed that humans selected strategies with higher predictability of interaction dynamics. This finding expressed that humans seek movement strategies where force and kinematics synchronize to repeatable patterns that may require less sensorimotor information processing.

Safety margins in older adults increase with improved control of a dynamic object.

Older adults face decreasing motor capabilities due to pervasive neuromuscular degradations. As a consequence, errors in movement control increase. Thus, older individuals should maintain larger safety margins than younger adults. While this has been shown for object manipulation tasks, several reports on whole-body activities, such as posture and locomotion, demonstrate age-related reductions in safety margins. This is despite increased costs for control errors, such as a fall. We posit that this paradox could be explained by the dynamic challenge presented by the body or also an external object, and that age-related reductions in safety margins are in part due to a decreased ability to control dynamics. To test this conjecture we used a virtual ball-in-cup task that had challenging dynamics, yet afforded an explicit rendering of the physics and safety margin. The hypotheses were: (1) When manipulating an object with challenging dynamics, older adults have smaller safety margins than younger adults. (2) Older adults increase their safety margins with practice. Nine young and 10 healthy older adults practiced moving the virtual ball-in-cup to a target location in exactly 2 s. The accuracy and precision of the timing error quantified skill, and the ball energy relative to an escape threshold quantified the safety margin. Compared to the young adults, older adults had increased timing errors, greater variability, and decreased safety margins. With practice, both young and older adults improved their ability to control the object with decreased timing errors and variability, and increased their safety margins. These results suggest that safety margins are related to the ability to control dynamics, and may explain why in tasks with simple dynamics older adults use adequate safety margins, but in more complex tasks, safety margins may be inadequate. Further, the results indicate that task-specific training may improve safety margins in older adults.

Balance decrements are associated with age-related muscle property changes.

In this study, a comprehensive evaluation of static and dynamic balance abilities was performed in young and older adults and regression analysis was used to test whether age-related variations in individual ankle muscle mechanical properties could explain differences in balance performance. The mechanical properties included estimates of the maximal isometric force capability, force-length, force-velocity, and series elastic properties of the dorsiflexors and individual plantarflexor muscles (gastrocnemius and soleus). As expected, the older adults performed more poorly on most balance tasks. Muscular maximal isometric force, optimal fiber length, tendon slack length, and velocity-dependent force capabilities accounted for up to 60% of the age-related variation in performance on the static and dynamic balance tests. In general, the plantarflexors had a stronger predictive role than the dorsiflexors. Plantarflexor stiffness was strongly related to general balance performance, particularly in quiet stance; but this effect did not depend on age. Together, these results suggest that age-related differences in balance performance are explained in part by alterations in muscular mechanical properties.

Neural representation of muscle dynamics in voluntary movement control.

Several theories of motor control posit that the nervous system has access to a neural representation of muscle dynamics. Yet, this has not been tested experimentally. Should such a representation exist, it was hypothesized that subjects who learned to control a virtual limb using virtual muscles would improve performance faster and show greater generalization than those who learned with a less dynamically complex virtual force generator. Healthy adults practiced using their biceps brachii activity to move a myoelectrically controlled virtual limb from rest to a standard target position with maximum speed and accuracy. Throughout practice, generalization was assessed with untrained target trials and sensitivity to actuator dynamics was probed by unexpected actuator model switches. In a muscle model subject group (n = 10), the biceps electromyographic signal activated a virtual muscle that pulled on the virtual limb with a force governed by muscle dynamics, defined by a nonlinear force-length-velocity relation and series elastic stiffness. A force generator group (n = 10) performed the same task, but the actuation force was a linear function of the biceps activation signal. Both groups made significant errors with unexpected actuator dynamics switches, supporting task sensitivity to actuator dynamics. The muscle model group improved performance as fast as the force generator group and showed greater generalization in early practice, despite using an actuator with more complex dynamics. These results are consistent with a preexisting neural representation of muscle dynamics, which may have offset any learning challenges associated with the more dynamically complex virtual muscle model.

Energy margins in dynamic object manipulation.

Many tasks require humans to manipulate dynamically complex objects and maintain appropriate safety margins, such as placing a cup of coffee on a coaster without spilling. This study examined how humans learn such safety margins and how they are shaped by task constraints and changing variability with improved skill. Eighteen subjects used a manipulandum to transport a shallow virtual cup containing a ball to a target without losing the ball. Half were to complete the cup transit in a comfortable target time of 2 s (a redundant task with infinitely many equivalent solutions), and the other half in minimum time (a nonredundant task with one explicit cost to optimize). The safety margin was defined as the ball energy relative to escape, i.e., as an energy margin. The first hypothesis, that subjects converge to a single strategy in the minimum-time task but choose different strategies in the less constrained target-time task, was not supported. Both groups developed individualized strategies with practice. The second hypothesis, that subjects decrease safety margins in the minimum-time task but increase them in the target-time task, was supported. The third hypothesis, that in both tasks subjects modulate energy margins according to their execution variability, was partially supported. In the target-time group, changes in energy margins correlated positively with changes in execution variability; in the minimum-time group, such a relation was observed only at the end of practice, not across practice. These results show that when learning a redundant object manipulation task, most subjects increase their safety margins and shape their movement strategies in accordance with their changing variability.