Generalizability of foot placement control strategies during unperturbed and perturbed gait.

Control of foot placement is an essential strategy for maintaining balance during walking. During unperturbed, steady-state walking, foot placement can be accurately described as a linear function of the body's centre of mass (CoM) state at midstance. However, it is uncertain if this mapping from CoM state to foot placement generalizes to larger perturbations that could potentially cause falls. Recovery from these perturbations may require reactive control strategies not observed during unperturbed walking. Here, we used unpredictable changes in treadmill belt speed to assess the generalizability of foot placement mappings identified during unperturbed walking. We found that foot placement mappings generalized poorly from unperturbed to perturbed walking and differed for forward perturbation versus backward perturbation. We also used the singular value decomposition of the mapping matrix to reveal that people were more sensitive to backward versus forward perturbations. Together, these results indicate that a single linear mapping cannot describe the foot placement control during both forward and backward losses of balance induced by treadmill belt speed perturbations. Better characterization of human balance control strategies could improve our understanding of why different neuromotor disorders result in heightened fall risk and inform the design of controllers for balance-assisting devices.

NSF DARE-Transforming modeling in neurorehabilitation: Four threads for catalyzing progress.

We present an overview of the Conference on Transformative Opportunities for Modeling in Neurorehabilitation held in March 2023. It was supported by the Disability and Rehabilitation Engineering (DARE) program from the National Science Foundation's Engineering Biology and Health Cluster. The conference brought together experts and trainees from around the world to discuss critical questions, challenges, and opportunities at the intersection of computational modeling and neurorehabilitation to understand, optimize, and improve clinical translation of neurorehabilitation. We organized the conference around four key, relevant, and promising Focus Areas for modeling: Adaptation & Plasticity, Personalization, Human-Device Interactions, and Modeling 'In-the-Wild'. We identified four common threads across the Focus Areas that, if addressed, can catalyze progress in the short, medium, and long terms. These were: (i) the need to capture and curate appropriate and useful data necessary to develop, validate, and deploy useful computational models (ii) the need to create multi-scale models that span the personalization spectrum from individuals to populations, and from cellular to behavioral levels (iii) the need for algorithms that extract as much information from available data, while requiring as little data as possible from each client (iv) the insistence on leveraging readily available sensors and data systems to push model-driven treatments from the lab, and into the clinic, home, workplace, and community. The conference archive can be found at (dare2023.usc.edu). These topics are also extended by three perspective papers prepared by trainees and junior faculty, clinician researchers, and federal funding agency representatives who attended the conference.

Transforming modeling in neurorehabilitation: clinical insights for personalized rehabilitation.

Practicing clinicians in neurorehabilitation continue to lack a systematic evidence base to personalize rehabilitation therapies to individual patients and thereby maximize outcomes. Computational modeling- collecting, analyzing, and modeling neurorehabilitation data- holds great promise. A key question is how can computational modeling contribute to the evidence base for personalized rehabilitation? As representatives of the clinicians and clinician-scientists who attended the 2023 NSF DARE conference at USC, here we offer our perspectives and discussion on this topic. Our overarching thesis is that clinical insight should inform all steps of modeling, from construction to output, in neurorehabilitation and that this process requires close collaboration between researchers and the clinical community. We start with two clinical case examples focused on motor rehabilitation after stroke which provide context to the heterogeneity of neurologic injury, the complexity of post-acute neurologic care, the neuroscience of recovery, and the current state of outcome assessment in rehabilitation clinical care. Do we provide different therapies to these two different patients to maximize outcomes? Asking this question leads to a corollary: how do we build the evidence base to support the use of different therapies for individual patients? We discuss seven points critical to clinical translation of computational modeling research in neurorehabilitation- (i) clinical endpoints, (ii) hypothesis- versus data-driven models, (iii) biological processes, (iv) contextualizing outcome measures, (v) clinical collaboration for device translation, (vi) modeling in the real world and (vii) clinical touchpoints across all stages of research. We conclude with our views on key avenues for future investment (clinical-research collaboration, new educational pathways, interdisciplinary engagement) to enable maximal translational value of computational modeling research in neurorehabilitation.

An α-MN collateral to γ-MNs can mitigate velocity-dependent stretch reflexes during voluntary movement: A computational study.

The primary motor cortex does not uniquely or directly produce α-MN drive to muscles during voluntary movement. Rather, α-MN drive emerges from the synthesis and competition among excitatory and inhibitory inputs from multiple descending tracts, spinal interneurons, sensory inputs, and proprioceptive afferents. One such fundamental input is velocity-dependent stretch reflexes in lengthening (antagonist) muscles, which are thought to be inhibited by the shortening (agonist) muscles. It remains an open question, however, the extent to which velocity-dependent stretch reflexes disrupt voluntary movement, and whether and how they are inhibited in limbs with numerous monoand multi-articular muscles where agonist and antagonist roles become unclear and can switch during a movement. We used a computational model of a Rhesus Macaque arm to simulate movements with feedforward α-MN commands only, and with added velocity-dependent stretch reflex feedback. We found that velocity-dependent stretch reflex caused movement-specific, typically large and variable disruptions to the arm endpoint trajectories. In contrast, these disruptions became small when the velocity-dependent stretch reflexes were simply scaled by the α-MN drive to each muscle (equivalent to an α-MN excitatory collateral to its homologous γ-MNs, but distinct from α-γ co-activation. We argue this circuitry is more neuroanatomically tenable, generalizable, and scalable than α-γ co-activation or movement-specific reciprocal inhibition. We propose that this mechanism at the homologous propriospinal level, by locally and automatically regulating the highly nonlinear neuro-musculo-skeletal mechanics of the limb, could be a critical low-level enabler of learning, adaptation, and performance via cerebellar and cortical mechanisms.

Muscle redundancy is greatly reduced by the spatiotemporal nature of neuromuscular control.

Animals must control numerous muscles to produce forces and movements with their limbs. Current theories of motor optimization and synergistic control are predicated on the assumption that there are multiple highly diverse feasible activations for any motor task ("muscle redundancy"). Here, we demonstrate that the dimensionality of the neuromuscular control problem is greatly reduced when adding the temporal constraints inherent to any sequence of motor commands: the physiological time constants for muscle activation-contraction dynamics. We used a seven-muscle model of a human finger to fully characterize the seven-dimensional polytope of all possible motor commands that can produce fingertip force vector in any direction in 3D, in alignment with the core models of Feasibility Theory. For a given sequence of seven force vectors lasting 300 ms, a novel single-step extended linear program finds the 49-dimensional polytope of all possible motor commands that can produce the sequence of forces. We find that muscle redundancy is severely reduced when the temporal limits on muscle activation-contraction dynamics are added. For example, allowing a generous ±12% change in muscle activation within 50 ms allows visiting only ∼7% of the feasible activation space in the next time step. By considering that every motor command conditions future commands, we find that the motor-control landscape is much more highly structured and spatially constrained than previously recognized. We discuss how this challenges traditional computational and conceptual theories of motor control and neurorehabilitation for which muscle redundancy is a foundational assumption.

Dynamical Analyses Show That Professional Archers Exhibit Tighter, Finer and More Fluid Dynamical Control Than Neophytes.

Quantifying the dynamical features of discrete tasks is essential to understanding athletic performance for many sports that are not repetitive or cyclical. We compared three dynamical features of the (i) bow hand, (ii) drawing hand, and (iii) center of mass during a single bow-draw movement between professional and neophyte archers: dispersion (convex hull volume of their phase portraits), persistence (tendency to continue a trend as per Hurst exponents), and regularity (sample entropy). Although differences in the two groups are expected due to their differences in skill, our results demonstrate we can quantify these differences. The center of mass of professional athletes exhibits tighter movements compared to neophyte archers (6.3 < 11.2 convex hull volume), which are nevertheless less persistent (0.82 < 0.86 Hurst exponent) and less regular (0.035 > 0.025 sample entropy). In particular, the movements of the bow hand and center of mass differed more between groups in Hurst exponent analysis, and the drawing hand and center of mass were more different in sample entropy analysis. This suggests tighter neuromuscular control over the more fluid dynamics of the movement that exhibits more active corrections that are more individualized. Our work, therefore, provides proof of principle of how well-established dynamical analysis techniques can be used to quantify the nature and features of neuromuscular expertise for discrete movements in elite athletes.

Task-dependent alteration of beta-band intermuscular coherence is associated with ipsilateral corticospinal tract excitability.

Beta-band (15-30 Hz) synchronization between the EMG signals of active limb muscles can serve as a non-invasive assay of corticospinal tract integrity. Tasks engaging a single limb often primarily utilize one corticospinal pathway, although bilateral neural circuits can participate in goal-directed actions involving multi-muscle coordination and utilization of feedback. Suboptimal utilization of such circuits after CNS injury can result in unintended mirror movements and activation of pathological synergies. Accordingly, it is important to understand how the actions of one limb (e.g., a less-affected limb after strokes) influence the opposite corticospinal pathway for the rehabilitation target. Certain unimanual actions decrease the excitability of the "unengaged" corticospinal tract, presumably to prevent mirror movement, but there is no direct way to predict the extent to which this will occur. In this study, we tested the hypothesis that task-dependent changes in beta-band drives to muscles of one hand will inversely correlate with changes in the opposite corticospinal tract excitability. Ten participants completed spring pinching tasks known to induce differential 15-30 Hz drive to muscles. During compressions, transcranial magnetic stimulation single pulses to the ipsilateral M1 were delivered to generate motor-evoked potentials in the unengaged hand. The task-induced changes in ipsilateral corticospinal excitability were inversely correlated with associated changes in EMG-EMG coherence of the task hand. These results demonstrate a novel connection between intermuscular coherence and the excitability of the "unengaged" corticospinal tract and provide a springboard for further mechanistic studies of unimanual tasks of varying difficulty and their effects on neural pathways relevant to rehabilitation.

Generalizability of foot-placement control strategies during unperturbed and perturbed gait.

Control of foot placement is an essential strategy for maintaining balance during walking. During unperturbed, steady-state walking, foot placement can be accurately described as a linear function of the body's center of mass state at midstance. However, it is uncertain if this mapping from center of mass state to foot placement generalizes to larger perturbations that may be more likely to cause falls. These perturbations may cause balance disturbances and generate reactive control strategies not observed during unperturbed walking. Here, we used unpredictable changes in treadmill speed to assess the generalizability of foot placement mappings identified during unperturbed walking. We found that foot placement mappings generalized poorly from unperturbed to perturbed walking and differed for forward versus backward perturbations. We also used singular value decomposition of the mapping matrix to reveal that people were more sensitive to backward versus forward perturbations. Together, these results indicate that control of foot placement during losses of balance differs from the control strategies used during unperturbed walking. Better characterization of human balance control strategies could improve our understanding of why different neuromotor disorders result in heightened fall risk and inform the design of controllers for balance-assisting devices.

Edge Computing in Nature: Minimal pre-processing of multi-muscle ensembles of spindle signals improves discriminability of limb movements.

Multiple proprioceptive signals, like those from muscle spindles, are thought to enable robust estimates of body configuration. Yet, it remains unknown whether spindle signals suffice to discriminate limb movements. Here, a simulated 4-musculotendon, 2-joint planar limb model produced repeated cycles of five end-point trajectories in forward and reverse directions, which generated spindle Ia and II afferent signals (proprioceptors for velocity and length, respectively) from each musculotendon. We find that cross-correlation of the 8D time series of raw firing rates (four Ia, four II) cannot discriminate among most movement pairs (∼ 29% accuracy). However, projecting these signals onto their 1st and 2nd principal components greatly improves discriminability of movement pairs (82% accuracy). We conclude that high-dimensional ensembles of muscle proprioceptors can discriminate among limb movements-but only after dimensionality reduction. This may explain the pre-processing of some afferent signals before arriving at the somatosensory cortex, such as processing of cutaneous signals at the cat's cuneate nucleus.

Mountain Hiking: Prolonged Eccentric Muscle Contraction during Simulated Downhill Walking Perturbs Sensorimotor Control Loops Needed for Safe Dynamic Foot-Ground Interactions.

Safe mountain hiking requires precise control of dynamic foot-ground interactions. In addition to vision and vestibular afferents, limb proprioception, sensorimotor control loops, and reflex responses are used to adapt to the specific nature of the ground contact. Diminished leg dexterity and balance during downhill walking is usually attributed to fatigue. We investigated the supplementary hypothesis that the eccentric contractions inherent to downhill walking can also disrupt muscle proprioception, as well as the sensorimotor control loops and reflex responses that depend on it. In this study, we measured leg dexterity (LD), anterior-posterior (AP) and medio-lateral (ML) bipedal balance, and maximal voluntary leg extension strength in young and healthy participants before and after 30 min of simulated downhill walking at a natural pace on a treadmill at a 20° decline. Post-pre comparisons of LD (p < 0.001) and AP balance (p = 0.001) revealed significant reductions in dynamic foot-ground interactions after eccentric exercise without an accompanying reduction in leg extension strength. We conclude that eccentric contractions during downhill walking can disrupt the control of dynamic foot-ground interactions independently of fatigue. We speculate that mountaineering safety could be improved by increasing conscious attention to compensate for unadjusted proprioception weighting, especially in the descent.

"Fine synergies" describe motor adaptation in people with drop foot in a way that supplements traditional "coarse synergies".

Synergy analysis via dimensionality reduction is a standard approach in biomechanics to capture the dominant features of limb kinematics or muscle activation signals, which can be called "coarse synergies." Here we demonstrate that the less dominant features of these signals, which are often explicitly disregarded or considered noise, can nevertheless exhibit "fine synergies" that reveal subtle, yet functionally important, adaptations. To find the coarse synergies, we applied non-negative matrix factorization (NMF) to unilateral EMG data from eight muscles of the involved leg in ten people with drop-foot (DF), and of the right leg of 16 unimpaired (control) participants. We then extracted the fine synergies for each group by removing the coarse synergies (i.e., first two factors explaining ≥ 85% of variance) from the data and applying Principal Component Analysis (PCA) to those residuals. Surprisingly, the time histories and structure of the coarse EMG synergies showed few differences between DF and controls-even though the kinematics of drop-foot gait is evidently different from unimpaired gait. In contrast, the structure of the fine EMG synergies (as per their PCA loadings) showed significant differences between groups. In particular, loadings for Tibialis Anterior, Peroneus Longus, Gastrocnemius Lateralis, Biceps and Rectus Femoris, Vastus Medialis and Lateralis muscles differed between groups ( p < 0.05 ). We conclude that the multiple differences found in the structure of the fine synergies extracted from EMG in people with drop-foot vs. unimpaired controls-not visible in the coarse synergies-likely reflect differences in their motor strategies. Coarse synergies, in contrast, seem to mostly reflect the gross features of EMG in bipedal gait that must be met by all participants-and thus show few differences between groups. However, drawing insights into the clinical origin of these differences requires well-controlled clinical trials. We propose that fine synergies should not be disregarded in biomechanical analysis, as they may be more informative of the disruption and adaptation of muscle coordination strategies in participants due to drop-foot, age and/or other gait impairments.

Estimating Center of Pressure of a Bipedal Mechanism Using a Proprioceptive Artificial Skin around its Ankles.

Estimating the Center of Pressure (CoP) under legged robots is useful to control their posture and gait. This is traditionally done using contact sensors at the base of the foot or with sensors on distal joints, which are subject to wear and damage due to impulse forces. In vertebrates, skin and ligament deformation at the ankle is a particularly rich source of sensory information for locomotion. For our bipedal mechanism, afferent signals from sensors on synthetic skin wrapped around the ankles sufficed to estimate the location of the CoP with a mean accuracy >81.5%. For this we used K-Nearest Neighbors (KNN) algorithm trained on the same force magnitude applied at four and nine ground-truth CoP locations. For a single mechanical foot (i.e., single stance), signals from skin or ligaments (i.e., elastic rubber sheets and cables, respectively) also sufficed to calculate the CoP (Mean prediction accuracy >91.3%). Moreover, the visco-elasticity of these elements serves to passively stabilize the ankle. Importantly, training the single leg case with forces of different magnitudes also resulted in similarly accurate mean CoP prediction accuracy >84.5%. We show that using bio-inspired proprioceptive skins and/or ligament arrangements can provide reliable COP predictions, while permitting arbitrary postures of the ankle and no sensors on the sole of the foot prone to wear and damage. This novel approach to estimation of the CoP can be used to improve locomotion control in a new class of bio-inspired rigid, soft and hybrid (soft-rigid) legged robots.

Data-efficient Causal Decoding of Spiking Neural Activity using Weighted Voting.

Brain-Computer Interface systems can contribute to a vast set of applications such as overcoming physical disabilities in people with neural injuries or hands-free control of devices in healthy individuals. However, having systems that can accurately interpret intention online remains a challenge in this field. Robust and data-efficient decoding-despite the dynamical nature of cortical activity and causality requirements for physical function-is among the most important challenges that limit the widespread use of these devices for real-world applications. Here, we present a causal, data-efficient neural decoding pipeline that predicts intention by first classifying recordings in short sliding windows. Next, it performs weighted voting over initial predictions up to the current point in time to report a refined final prediction. We demonstrate its utility by classifying spiking neural activity collected from the human posterior parietal cortex for a cue, delay, imaginary motor task. This pipeline provides higher classification accuracy than state-of-the-art time windowed spiking activity based causal methods, and is robust to the choice of hyper-parameters.

insideOut: A Bio-Inspired Machine Learning Approach to Estimating Posture in Robots Driven by Compliant Tendons.

Estimates of limb posture are critical for controlling robotic systems. This is generally accomplished with angle sensors at individual joints that simplify control but can complicate mechanical design and robustness. Limb posture should be derivable from each joint's actuator shaft angle but this is problematic for compliant tendon-driven systems where (i) motors are not placed at the joints and (ii) nonlinear tendon stiffness decouples the relationship between motor and joint angles. Here we propose a novel machine learning algorithm to accurately estimate joint posture during dynamic tasks by limited training of an artificial neural network (ANN) receiving motor angles and tendon tensions, analogous to biological muscle and tendon mechanoreceptors. Simulating an inverted pendulum-antagonistically-driven by motors and nonlinearly-elastic tendons-we compare how accurately ANNs estimate joint angles when trained with different sets of non-collocated sensory information generated via random motor-babbling. Cross-validating with new movements, we find that ANNs trained with motor angles and tendon tension data predict joint angles more accurately than ANNs trained without tendon tension. Furthermore, these results are robust to changes in network/mechanical hyper-parameters. We conclude that regardless of the tendon properties, actuator behavior, or movement demands, tendon tension information invariably improves joint angle estimates from non-collocated sensory signals.

Temporal control of muscle synergies is linked with alpha-band neural drive.

KEY POINTS: It is theorized that the nervous system controls groups of muscles together as functional units, or 'synergies', resulting in correlated electromyographic (EMG) signals among muscles. However, such correlation does not necessarily imply group-level neural control. Oscillatory synchronization (coherence) among EMG signals implies neural coupling, but it is not clear how this relates to control of muscle synergies. EMG was recorded from seven arm muscles of 10 adult participants rotating an upper limb ergometer, and EMG-EMG coherence, EMG amplitude correlations and their relationship with each other were characterized. A novel method to derive multi-muscle synergies from EMG-EMG coherence is presented and these are compared with classically defined synergies. Coherent alpha-band (8-16 Hz) drive was strongest among muscles whose gross activity levels are well correlated within a given task. The cross-muscle distribution and temporal modulation of coherent alpha-band drive suggests a possible role in the neural coordination/monitoring of synergies. ABSTRACT: During movement, groups of muscles may be controlled together by the nervous system as an adaptable functional entity, or 'synergy'. The rules governing when (or if) this occurs during voluntary behaviour in humans are not well understood, at least in part because synergies are usually defined by correlated patterns of muscle activity without regard for the underlying structure of their neural control. In this study, we investigated the extent to which comodulation of muscle output (i.e. correlation of electromyographic (EMG) amplitudes) implies that muscles share intermuscular neural input (assessed via EMG-EMG coherence analysis). We first examined this relationship among pairs of upper limb muscles engaged in an arm cycling task. We then applied a novel multidimensional EMG-EMG coherence analysis allowing synergies to be characterized on the basis of shared neural drive. We found that alpha-band coherence (8-16 Hz) is related to the degree to which overall muscle activity levels correlate over time. The extension of this coherence analysis to describe the cross-muscle distribution and temporal modulation of alpha-band drive revealed a close match to the temporal and structural features of traditionally defined muscle synergies. Interestingly, the coherence-derived neural drive was inversely associated with, and preceded, changes in EMG amplitudes by ∼200 ms. Our novel characterization of how alpha-band neural drive is dynamically distributed among muscles is a fundamental step forward in understanding the neural origins and correlates of muscle synergies.

Force variability is mostly not motor noise: Theoretical implications for motor control.

Variability in muscle force is a hallmark of healthy and pathological human behavior. Predominant theories of sensorimotor control assume 'motor noise' leads to force variability and its 'signal dependence' (variability in muscle force whose amplitude increases with intensity of neural drive). Here, we demonstrate that the two proposed mechanisms for motor noise (i.e. the stochastic nature of motor unit discharge and unfused tetanic contraction) cannot account for the majority of force variability nor for its signal dependence. We do so by considering three previously underappreciated but physiologically important features of a population of motor units: 1) fusion of motor unit twitches, 2) coupling among motoneuron discharge rate, cross-bridge dynamics, and muscle mechanics, and 3) a series-elastic element to account for the aponeurosis and tendon. These results argue strongly against the idea that force variability and the resulting kinematic variability are generated primarily by 'motor noise.' Rather, they underscore the importance of variability arising from properties of control strategies embodied through distributed sensorimotor systems. As such, our study provides a critical path toward developing theories and models of sensorimotor control that provide a physiologically valid and clinically useful understanding of healthy and pathologic force variability.

Sampling-Based Nonlinear Stochastic Optimal Control for Neuromechanical Systems.

Determining how the nervous system controls tendon-driven bodies remains an open question. Stochastic optimal control (SOC) has been proposed as a plausible analogy in the neuroscience community. SOC relies on solving the Hamilton-Jacobi-Bellman equation, which seeks to minimize a desired cost function for a given task with noisy controls. We evaluate and compare three SOC methodologies to produce tapping by a simulated planar 3-joint human index finger: iterative Linear Quadratic Gaussian (iLQG), Model-Predictive Path Integral Control (MPPI), and Deep Forward-Backward Stochastic Differential Equations (FBSDE). We show that averaged over 128 repeats these methodologies can place the fingertip at the desired final joint angles but-because of kinematic redundancy and the presence of noise-they each have joint trajectories and final postures with different means and variances. iLQG in particular, had the largest kinematic variance and departure from the final desired joint angles. We demonstrate that MPPI and FBSDE have superior performance for such nonlinear, tendon-driven systems with noisy controls.Clinical relevance- The mathematical framework provided by MPPI and FBSDE may be best suited for tendon-driven anthropomorphic robots, exoskeletons, and prostheses for amputees.

Parkinson's Disease Exhibits Amplified Intermuscular Coherence During Dynamic Voluntary Action.

Parkinson's disease (PD) is typically diagnosed and evaluated on the basis of overt motor dysfunction, however, subtle changes in the frequency spectrum of neural drive to muscles have been reported as well. During dynamic actions, coactive muscles of healthy adults often share a common source of 6-15 Hz (alpha-band) neural drive, creating synchronous alpha-band activity in their EMG signals. Individuals with PD commonly exhibit kinetic action tremor at similar frequencies, but the potential relationship between the intermuscular alpha-band neural drive seen in healthy adults and the action tremor associated with PD is not well-understood. A close relationship is most tenable during voluntary dynamic tasks where alpha-band neural drive is strongest in healthy adults, and where neural circuits affected by PD are most engaged. In this study, we characterized the frequency spectrum of EMG synchronization (intermuscular coherence) in 16 participants with PD and 15 age-matched controls during two dynamic motor tasks: (1) rotation of a dial between the thumb and index finger, and (2) dynamic scaling of isometric precision pinch force. These tasks produce different profiles of coherence between the first dorsal interosseous and abductor pollicis brevis muscles. We sought to determine if alpha-band intermuscular coherence would be amplified in participants with PD relative to controls, if such differences would be task-specific, and if they would correlate with symptom severity. We found that relative to controls, the PD group displayed amplified, but similarly task-dependent, coherence in the alpha-band. The magnitude of coherence during the rotation task correlated with overall symptom severity as per the UPDRS rating scale. Finally, we explored the potential for our coherence measures, with no additional information, to discriminate individuals with PD from controls. The area under the Receiver Operating Characteristic curve (AUC) indicated a clear separation between groups (AUC = 0.96), even though participants with PD were on their typical medication and displayed only mild-moderate symptoms. We conclude that a task-dependent, intermuscular neural drive within the alpha-band is amplified in PD. Its quantification via intermuscular coherence analysis may provide a useful tool for detecting the presence of PD, or assessing its progression.

Autonomous Functional Movements in a Tendon-Driven Limb via Limited Experience.

Robots will become ubiquitously useful only when they can use few attempts to teach themselves to perform different tasks, even with complex bodies and in dynamical environments. Vertebrates, in fact, use sparse trial-and-error to learn multiple tasks despite their intricate tendon-driven anatomies-which are particularly hard to control because they are simultaneously nonlinear, under-determined, and over-determined. We demonstrate-for the first time in simulation and hardware-how a model-free, open-loop approach allows few-shot autonomous learning to produce effective movements in a 3-tendon 2-joint limb. We use a short period of motor babbling (to create an initial inverse map) followed by building functional habits by reinforcing high-reward behavior and refinements of the inverse map in a movement's neighborhood. This biologically-plausible algorithm, which we call G2P (General-to-Particular), can potentially enable quick, robust and versatile adaptation in robots as well as shed light on the foundations of the enviable functional versatility of organisms.