ABSTRACT ENGINEERING SYNTHETIC FEEDBACK TO PROMOTE RECOVERY OF SELF-FEEDING SKILLS IN PEOPLE WITH SENSORY DEFICITS DUE TO STROKE Alexis Krueger Marquette University, 2016 Kinesthesia ref
INTRODUCTION
Rationale and Specific Aims
Kinesthesia refers to sensations of limb position and movement (Bastian 1887) derived predominantly from information encoded by muscle spindle afferents (c.f., Proske and Gandevia
Kinesthetic feedback depends on muscle length and the rate of length change, as described by Edin and Valbo (1990) After stroke, deficits in kinesthetic feedback are common, with about 50% of survivors experiencing impaired limb position sense in their contralesional arm (Carey and Matyas, 2011; Dukelow et al., 2010; Connell et al., 2008) This loss of kinesthetic sensation contributes to impaired control of reaching and stabilization movements (Scheidt and Stoeckmann, 2007; Zackowski et al.).
Independence in daily living and basic tasks like self-feeding depend on intact kinesthetic feedback, yet when this input is compromised, people often rely on visual cues to guide movement; the visual system’s processing delays of about 100–200 ms produce movements that are slow, poorly coordinated, and require high concentration, with visually guided corrections arriving too late and resulting in jerky, unstable movements.
Stroke rehabilitation often focuses on motor retraining, yet many survivors stop using their contralesional limb due to sensorimotor deficits, a phenomenon described by Taub et al (1993) This disengagement can significantly reduce quality of life, as shown in studies by Abela et al (2012) and Tyson et al (2008) The critical interactions between sensory and motor systems are frequently overlooked in current therapies Developing techniques that specifically address impaired kinesthesia could yield integrated interventions that combine sensory and motor rehabilitation to improve the overall sensorimotor function of stroke survivors.
The long-term objective is to mitigate post-stroke kinesthesia deficits by developing sensory substitution technologies that provide real-time feedback of the hypoesthetic contralesional arm’s state—its position and velocity—to a body site with preserved somatosensation, such as the ipsilesional arm Providing supplemental feedback to mitigate sensory deficits has been explored for decades, with vibrotactile feedback among the successful modalities reported since White et al (1970) However, existing systems often neglect explicit limb kinesthesia or interfere with daily activities like eating and speaking when using electrotactile stimulation of the tongue (e.g., BrainPort, WiCab, Inc.) To address these limitations, we propose vibrotactile stimulation on one arm to deliver kinesthetic feedback about the motion of the other arm, offering meaningful information while avoiding disruption to speech or sight.
Numerous strategies exist for encoding information about a moving limb within a vibrotactile feedback stream, but it remains unclear which approach most effectively supports closed-loop control of goal-directed stabilization and reaching movements One possibility is encoding the limb's state, such as its position and velocity A second distinct approach focuses on representing error signals or performance metrics to guide corrective actions.
Goal-aware feedback, described by Tzorakoleftherakis et al (2016), encodes information about the current task's objectives For example, error feedback—a simple form of goal-aware vibrotactile feedback—indicates the instantaneous error between the hand's current position and the position of a visual target Each encoding scheme might offer distinct advantages in terms of user performance and practicality.
To advance the goal of reducing post-stroke kinesthesia deficits, this study first evaluated whether people with no neuromotor impairments could control goal-directed arm actions when provided with real-time vibrotactile feedback about the moving arm from the opposite arm The focus was on reaching, stabilizing, and tracking—fundamental building blocks of daily living activities Drawing on motor control literature that shows healthy participants make systematic performance errors, such as proprioceptive drift, in the absence of ongoing visual feedback, the study compared the extent of improvement in proprioceptive drift when participants received supplemental vibrotactile state feedback versus error feedback In a second stage, the researchers conducted case studies to examine to what extent stroke survivors could learn to use supplemental vibrotactile feedback to improve control of the contralesional arm.
We hypothesize that task-relevant vibration signals, deployed as a sensory substitution method, can augment or restore closed-loop kinesthetic feedback in stroke survivors with impaired kinesthesia while preserving residual motor capacity in the contralesional arm and hand By reducing the impact of kinesthetic deficits, this approach aims to improve functional outcomes for individuals with this pattern of sensorimotor impairments As an initial step, the study investigates synthetic vibratory feedback as a sensory substitution mechanism to compensate for the inherent limitations of kinesthetic feedback in healthy participants, with the goal of identifying what information types are most useful for encoding supplemental kinesthetic cues.
We will also investigate the performance benefits conferred by vibrotactile feedback in a small cohort of stroke survivors Three specific aims are addressed
1 Determine if the performance of healthy participants during reaching and stabilization tasks varies systematically with the amount of limb position or velocity information encoded in synthetic vibrotactile feedback, in order to identify the optimal combination of limb position and velocity information
2 Compare the performance of healthy participants conducting reaching and stabilization tasks in the absence of vision with three different types of vibration encodings – optimal limb state feedback (containing both position and velocity information), goal-aware hand position error feedback, or task-irrelevant random sham feedback
3 Perform a series of case studies, wherein stroke survivors attempt to use supplemental kinesthetic feedback to improve performance of reaching and stabilization behaviors performed with the contralesional arm
In the course of conducting these three aims, this study seeks to provide guidance for the future development of vibrotactile sensory substitution devices for stroke survivors.
Outline of the Thesis
Chapter 1 presents the study’s rationale and its specific aims Chapter 2 presents a review of pertinent literature Chapter 3 presents the primary experimental study, which has been submitted to the Journal of Neuroengineering.
Rehabilitation research is currently under revision Chapter 4 offers an initial investigation into the extent to which supplemental kinesthetic feedback can improve the arm's control bandwidth when concurrent visual feedback of performance is absent.
5 presents results from five case studies in stroke survivors Finally, Chapter 6 presents overall conclusions and suggestions for future work.
BACKGROUND
Proprioception and Arm Control in Neurologically-Intact Individuals
To justify studying technologies designed for stroke survivors in neurologically intact individuals, we point out that the vast majority of people—even healthy ones—exhibit imperfect somatosensory control of the arm and hand when visual feedback is absent The most conspicuous and ubiquitous sign of this imperfect somatosensation is proprioceptive drift, a phenomenon described by Wann and Ibrahim (1992) and reinforced in later work such as that of Smeets et al.
Marked errors in the perceived position of the unseen hand emerge within a 12–15 second window Proprioceptive error is likely due to a progressive drift between visual and proprioceptive maps of body configuration when vision of the relative positions of the body and the visual target is precluded (Paillard and Brouchon 1968; Jeannerod 1989).
Wann and Ibrahim (1992) conducted an experiment in which participants tracked a moving visual target with the fingertip of their visually occluded dominant arm for 150 seconds, and every 15 seconds they used their non-dominant hand to indicate the dominant fingertip’s position, with some trials offering brief glimpses of the tracking arm and others providing no visual feedback; a 1.6 cm error emerged within the first 15 seconds in both conditions, but in trials with intermittent visual access the error stopped growing after the initial glimpse, whereas in trials without any visual feedback the error continued to accumulate linearly throughout the 150-second trial The authors argue that these findings are best explained by drift between proprioceptive and visual spatial maps when visuomotor information is not activated simultaneously, i.e., proprioceptive drift occurs naturally in neurologically intact individuals who do not frequently see their arm, providing a model for impaired kinesthesia in stroke survivors and informing our understanding of proprioception and visuomotor integration in rehabilitation.
Control Actions in Neurologically-Intact Individuals
Goal-directed actions like reaching and keeping the hand steady against environmental perturbations rely on two independent control processes: trajectory control, which shapes the movement path, and end-point stabilization, which preserves the final hand position (Scheidt and Ghez, 2007) Depending on the task, these control processes are differentially engaged, with stabilization tasks—where the hand must remain at a location despite perturbations—primarily relying on end-point stabilization.
Tracking tasks require primarily trajectory control to follow a moving target, while reaching tasks require both trajectory control and end-point stabilization to reach and end on a target Proprioceptive drift influences these control actions and succinctly predicts the pattern of performance errors observed during goal-directed reaching and stabilizing actions performed with the hand in the absence of visual feedback, as shown by Scheidt et al (2005) and Suminski et al (2007).
Vibrotactile Sensation in Neurologically-Intact Individuals
Hairy skin vibration is thought to be detected primarily by two receptor classes: shallow rapidly adapting (RA) mechanoreceptors and deeper Pacinian corpuscles (PC) RA receptors encode dynamic, lower-frequency skin deformations near the surface, whereas Pacinian corpuscles respond to high-frequency vibrations detected in deeper tissue layers This dual-receptor arrangement provides complementary temporal and spatial information essential for tactile perception, as described in studies by Bensmaia and Hollins (1999), Mahns et al (2005), Mountcastle et al (1972), and Bolanowski et al.
Dating back to 1988, the RA channel is described as lying superficially in the skin and can be compromised by local skin anesthesia, whereas the PC channel is located deeper, extending into subcutaneous tissues—particularly in hairy skin—and is not affected by local anesthetic.
Rapidly adapting (RA) receptors respond to low-frequency vibrations (30 Hz) The sensation in the mid-frequency overlap area appears to arise from a combination of RA and PC input, and this overlap sensation is somewhat affected by anesthesia Pattern discrimination and the minimum detectable change in vibration intensity are thought to be better encoded by RA channels than PC channels (Bensmaia and Hollins, 1999; Mountcastle et al., 1972) Tannan et al (2006) report that adaptation in RA channels within the first few seconds of exposure to low-frequency vibration (25 Hz) enhances localization ability.
Multisensory Integration in Neurologically-Intact Individuals
During everyday tasks like answering a ringing phone, the brain integrates multiple sensory streams—vision for the phone and surrounding objects and the hand, proprioception for the arm’s speed and position in joint space, and auditory-vestibular cues for the phone’s location based on head orientation and balance—each in its own reference frame to precisely locate the phone and the moving limb The sensory inputs are weighted by their reliability in the current context, and a Bayesian framework yields optimal estimates rather than simple averaging Importantly, one sense can map onto another’s reference frame, so auditory information can compensate for weak visual cues and proprioception can fill in for missing visual input of hand motion; this cross-modal mapping explains how people can still pick up the phone in the dark, or with impaired proprioception or hearing, such as in stroke or noisy environments.
Because one sense can fill in and substitute for another, this scheme allows for supplemental sensory substitution as a compensatory technique.
Vibrotactile Feedback for Sensory Substitution
Supplemental feedback to mitigate sensory deficits has been explored for decades, with early work such as White et al (1970) Notable successes include cochlear implants (Loeb 1990) and non-invasive systems that encode video images into vibratory or electrical signals delivered to the skin at several sites—including the abdomen, back, thigh, fingertip, forehead, and tongue (Kaczmarek et al 1991) Vibrotactile approaches for enhancing postural stabilization in individuals with vestibular disorders have also been proposed (Sienko et al 2008; Lee et al.).
Synthesized vibrotactile feedback that encompasses all task-relevant states shows promise, aligning with prior findings that full state encoding enhances performance (Peterka et al., 2006; Lee et al., 2011) In stroke rehabilitation, vibration tolerability studies indicate comfort and acceptance of the feedback, with no anxiety or adverse complications reported among participants (Bento et al., 2012).
Vibrotactile systems also show promise for providing information about grasp force and hand aperture to users of myoelectric forearm prostheses (Witteveen et al 2014)
An example of vibrotactile feedback’s potential is shown by a study describing a vibrotactile armband that reduces arm-angle error while teaching rehabilitative motor skills The device, designed to aid physical therapists in stroke rehabilitation, provides feedback to patients during repetitive practice of arm movements Eight vibrotactile motors are mounted around the arm—four near the bicep and four near the wrist—and arm motion is tracked with a Microsoft Kinect 360 to compute error in each arm segment and deliver repulsive vibrotactile cues akin to light touch corrections, with only one vibrator active at a time Visual feedback appears on a computer monitor as two avatars: one mirrors the user’s motion and the other shows the target motion In a study with 26 participants across four sessions, participants practiced six arm motions with either visual feedback alone (V) or visual feedback with vibrotactile feedback (VT), and the tested movements had 1–3 degrees of freedom.
Bark et al found that arm-angle errors for 1-DOF movements were significantly lower when participants used combined vibrotactile and visual feedback (VT) compared with visual feedback alone (V), while vibrotactile feedback did not affect 2- or 3-DOF arm motions Post-session workload surveys showed participants experienced higher workload with VT than with V, though workload decreased across sessions in both conditions Eighteen of the 26 participants preferred practicing the arm motions with VT rather than with V Overall, the study indicates vibrotactile feedback can help neurologically intact individuals learn simple rehabilitative arm motions, and despite the higher initial workload, most participants chose VT practice over V.
Previous studies reveal the brain's remarkable ability to integrate synthetic feedback for both perception and motor control Yet these works have notable gaps: they often neglect limb kinesthesia, emphasize laboratory conditions over real-world use, and employ feedback systems that can interfere with daily tasks such as eating and speaking For instance, devices that use electrotactile tongue stimulation may disrupt normal mouth movements and conversation Addressing these limitations is essential to translate neuroprosthetic feedback into practical, everyday applications.
Drawing on BrainPort, WiCab, Inc., we propose using vibrotactile stimulation on one arm to deliver supplemental kinesthetic feedback about the motion of the other arm This approach rests on the idea that, aside from the palms and fingers, tactile feedback from the surface of the arm is not typically critical for completing most daily living activities, so the vibrotactile display on the stimulated arm is unlikely to impede its use for other tasks.
Encoding Information in Vibrotactile Feedback
Vibrotactile feedback streams can encode information about a moving limb in several ways, and it remains unclear which encoding best supports closed-loop control of goal-directed stabilization and reaching One option conveys limb state, such as hand position and velocity, and the hardware and software needed to build standalone wearables that detect, synthesize, and deliver these limb-state cues in unconstrained environments are readily available A second approach, called "goal-aware" feedback, adds information about the current task objectives A simple form of goal-aware vibrotactile feedback might indicate the instantaneous error between the elbow’s current position and its target position; in Lieberman et al (2015), participants received joint-angle error feedback via eight tactors—four around the wrist and four around the elbow—while a computer screen displayed several discrete arm postures that composed the target motion The vibrotactile signal was proportional to the degrees of error at each step, and participants reduced joint-angle error by 27% using this feedback compared with performing the task without vibratory feedback.
Goal-aware feedback can be implemented by encoding the required arm movement direction through a computational model that optimizes the trade-off between kinematic performance and energetic cost In the study by Tzorakoleftherakis and colleagues, participants completed a single session where they attempted to balance a one-degree-of-freedom virtual inverted pendulum mounted on a cart Two vibrotactile actuators were placed on the right thumb and little finger, and moving the right hand drove the cart in the same direction The vibrotactile cues provided a real-time teaching signal indicating the exact hand—and thus cart—motion needed to stabilize the pendulum Participants performed the task under three feedback modalities: visual only (V), vibrotactile only (T), and combined visual and vibrotactile feedback (VT) Without participant input, the pendulum would fall, highlighting the importance of timely, direction-specific feedback for balance control.
In a study comparing feedback modalities for maintaining an inverted pendulum, visual feedback alone yielded an average time to failure of 13 seconds, while vibrotactile feedback alone produced 9 seconds, indicating that visual cues were more effective than vibration alone When both visual and vibrotactile feedback were combined, participants maintained the inverted pendulum for an average duration longer than either single-modality feedback, highlighting the advantage of multimodal feedback for balance control.
37 seconds In addition, a large standard deviation for this condition suggests that participants were still learning how to best use the dual-feedback, and could improve with further practice
Goal-aware feedback can potentially outperform limb-state feedback by incorporating extra task-specific information, but it faces distinctive technical challenges that state feedback avoids Specifically, inferring a user’s motor intentions and movement goals from moment to moment in dynamic or uncontrolled environments is a daunting task Any inaccuracies in estimating intent can lead to unreliable feedback, thereby undermining the usability of the vibrotactile display.
Human physiology provides no clear guidance on how kinesthetic information might be encoded within supplemental vibrotactile feedback to optimize augmented closed-loop control of stabilization and reaching behaviors For example, muscle spindle primary endings encode muscle length and the rate of length change in a joint-based coordinate reference frame whereas muscle spindle secondary endings encode primarily muscle length information (Proske and Gandevia 2012) Simulated vibrotactile limb state feedback could readily emulate these types of native feedback By contrast, Golgi tendon organs encode muscle tension (Proske and
Estimating and emulating the gamma-motor system is difficult because gamma-motor neurons modulate muscle spindle sensitivity in ways that can resemble error encoding; in some cases this modulation appears to carry error-related signals, as suggested by Houk and Rymer (1981), though the functional relationship between spindle feedback and gamma-drive remains intricate, as Grillner (1969) and later reviews (Gandevia 2012) emphasize the complexity of how spindle feedback depends on gamma-motor activity, underscoring challenges for accurate modeling and interpretation.
OPTIMIZING VIBROTACTILE FEEDBACK TO ENHANCE REAL-TIME CONTROL OF THE
Background
Kinesthesia—the sense of limb position and movement—derives predominantly from muscle spindle afferents that are sensitive to muscle length and the rate of length change (Bastian 1887; Proske and Gandevia 2012) After stroke, deficits in kinesthetic feedback are common, with almost 50% of survivors experiencing impaired limb position sense in the contralesional arm (Carey and Matyas 2011; Dukelow et al 2010).
Loss of kinesthetic sensation impairs the control of reaching and stabilization behaviors that underpin independent living When kinesthetic feedback is absent, individuals may compensate by relying on vision of their own limbs; however, the visual system’s processing delays of about 100–200 ms tend to produce movements that are slow, poorly coordinated, and require sustained concentration.
Studies by Sainburg et al (1993) and Ghez et al (1995), along with Sarlegna (2006), show that visually guided corrections arrive too late and produce jerky, unstable movements In stroke survivors, this timing mismatch can leave patients with persistent motor challenges, prompting many to abandon using their contralesional limb due to sensorimotor deficits (Taub et al 1993) Consequently, avoiding the affected limb is associated with a reduced quality of life (Abela et al 2012; Tyson et al 2008).
Long-term goal: mitigate the negative impact of post-stroke kinesthesia deficits by creating sensory substitution technologies that provide real-time feedback of contralesional arm state (e.g., position and velocity of the insensate limb) to a somatosensation-preserving site on the body, such as the ipsilesional arm As an initial step, the study tested whether individuals with no neuromotor deficits can perform goal-directed actions using supplementary vibrotactile stimuli that convey real-time information about the moving limb to a body region not involved in movement or essential functions like speaking or eating This approach is justified because the vast majority of people—even neurologically intact individuals—exhibit imperfect somatosensory control of the arm and hand without ongoing visual feedback A conspicuous manifestation of this imperfect somatosensation is proprioceptive drift, where marked errors in perceiving the position of the unseen hand develop within 12 to 15 seconds.
Proprioceptive error likely results from a progressive drift between visual and proprioceptive maps of body configuration when vision of the body's relative position to a target is precluded (Jeannerod 1989) This proprioceptive drift reliably predicts the pattern of performance errors observed during goal-directed reaching and stabilizing actions (Scheidt et al 2005).
(Suminski et al 2007) performed with the hand in the absence of visual feedback
Decades of research have explored supplemental feedback as a way to mitigate sensory deficits, tracing back to White et al (1970) Notable successes include cochlear implants (Loeb, 1990) and non-invasive systems that translate video images into vibratory or electrical signals applied to the skin at multiple sites—abdomen, back, thigh, fingertip, forehead, and tongue (Kaczmarek et al., 1991) Vibrotactile approaches to enhance postural stabilization in individuals with vestibular impairment have also been proposed, underscoring the potential of tactile feedback to augment balance and sensory perception.
Prior work indicates promise for synthesized feedback that encodes all task-relevant states, with studies showing improved perception and control when feedback conveys complete task information (Peterka et al 2006; Lee et al 2011) Vibrotactile systems have demonstrated the ability to convey grasp force and hand aperture to users of myoelectric forearm prostheses (Witteveen et al 2014) Although these results highlight the brain’s capacity to integrate artificial feedback for perceptual and motor control, they often overlook limb kinesthesia and include feedback approaches that can degrade quality of life, such as systems that interfere with verbal communication (e.g., BrainPort, WiCab) To address these gaps, we propose delivering supplemental kinesthetic feedback via vibrotactile stimulation on one arm that reflects the motion of the opposite arm This design is based on the rationale that tactile feedback from the arm surface, aside from the palms and fingers, is not typically critical for most daily activities, thereby reducing the likelihood that the vibrotactile display would impede use of the stimulated limb for other tasks.
Multiple viable strategies exist for encoding information about a moving limb within a vibrotactile feedback stream, and it remains to be seen which approach best supports closed‑loop control of goal‑directed stabilization and reaching One option is to encode limb state—such as hand position and velocity—into the vibrotactile signal, while a distinct, goal‑aware feedback approach (Tzorakoleftherakis et al 2016) also conveys information about the task’s objectives From a technical standpoint, the hardware and software needed to develop standalone wearable technologies that detect, synthesize, and deliver limb‑state information in unconstrained environments are readily available A simple form of goal‑aware vibrotactile feedback might indicate the instantaneous error between the hand’s current position and a visual target.
A more complex approach could encode the arm's intended direction based on the output of a computational model that optimizes the trade-off between kinematic performance and energetic efficiency Although goal-aware feedback may yield better performance than limb-state feedback because it incorporates additional task-specific information, this approach introduces several unique technological challenges that state-feedback methods avoid In particular, inferring the user’s motor intentions and movement goals from moment to moment in a dynamically changing environment is a daunting task, and errors in estimating intent can lead to unreliable feedback, undermining the usability of the vibrotactile display.
Human physiology provides limited guidance on how kinesthetic information should be encoded within supplemental vibrotactile feedback to optimize augmented closed-loop control of stabilization and reaching behaviors Muscle spindle primary endings encode joint-based length and the rate of length change, while secondary endings primarily convey muscle length information, and simulated vibrotactile limb-state feedback could readily emulate these native signals By contrast, Golgi tendon organs encode muscle tension, which is harder to estimate and reproduce Additionally, gamma-motor neurons modulate spindle sensitivity in ways that can resemble error encoding, though the functional dependence of spindle feedback on gamma activity is complex.
Developing sensory substitution technologies to enhance closed-loop control of goal-directed actions in people with impaired somatosensation is the ultimate objective The approach bypasses damaged feedback pathways by encoding motion information from a moving limb, such as the dominant arm, into a vibrotactile feedback stream delivered to a non-moving body part, such as the non-dominant arm, to improve performance when continuous visual feedback is not available.
To illustrate how this could work, we examine a simplified, single-joint, human-in-the-loop state-feedback control model (Figure 3.1A) While not intended to replicate the full complexity of human sensorimotor control (Heenan et al 2014), the model includes key elements: a viscoelastic limb, feedback delay from sensory transduction to processing, and a central mechanism that translates performance errors into corrective motor commands, implemented here with the simplest control law—proportional control This choice is justified because the goal is to design a sensory substitution system that imposes minimal information processing demands on the stroke-injured brain.
Model feedback can simulate visual or proprioceptive feedback by adjusting the sensor function to emulate physiological systems Emulating visual feedback relies on position feedback in the sensor function, with θ_f = θ_a, to represent the information that drives saccadic eye movements and encodes object location in space (Biguer et al.).
Within this model, visual feedback encodes position-based saccadic movements because the target position is static, while velocity-based smooth pursuit is employed for tracking moving targets Proprioception is emulated by combining position and velocity feedback in the sensor function, θ_f = θ_a + 0.15 dθ_a/dt, to reflect physiological input from muscle spindle primaries and secondaries This integrated approach captures both rapid, position-driven corrections and slower, velocity-driven tracking to produce realistic motor control.
From a feedback-control perspective, adding velocity information to the feedback path produces a change in the overall transfer function that is analogous to the addition of derivative action in the controller Specifically, we compare three transfer functions: the proportional controller with only position feedback, H1(S); the proportional controller with position plus velocity (derivative) feedback, H2(S); and the proportional‑plus‑derivative controller with only position feedback, H3(S), with K as a constant (Equations 1).
Methods
Twenty-six healthy adults (13 female) were recruited from the University of Genoa community, and all provided written informed consent to participate in the study Procedures were approved by the local Institutional Review Boards serving the University of Genoa (ASL3 Genovese) and Marquette University, in accordance with the 1964 Declaration of Helsinki None of the participants had known neurological disorders, and ages ranged from 22 to 32 years (mean approximately 26 years).
3) years All of the participants self-reported to be right handed All participants had normal or corrected-to-normal vision and all were nạve to the purposes of the study
Participants took part in up to three experimental sessions on separate days to assess whether encoding state information or error information about a moving limb into a vibrotactile feedback stream delivered to a non-moving body part best enhances stabilization and reaching when visual feedback is withheld Experiment 1 aimed to identify the optimal combination of limb state information—hand position and velocity feedback—to encode within the vibrotactile feedback applied to the contralateral arm The second and third sessions were designed to compare the effects of encoding the optimal state feedback against encoding an objective measure of hand position error.
Participants sat comfortably in a high-backed, adjustable-height chair in front of a horizontal planar robotic manipulandum, as described in Casadio et al (2006) and illustrated in Fig 3.2A They were positioned about 25 cm from the workspace center, with the right arm strapped to the robotic handle and its integrated arm support Seat height was adjusted so that the right shoulder abduction angle ranged from 75° to 85° The left arm rested on a horizontal planar armrest below the robot’s motion plane, with the forearm and hand pointed forward, as shown in Fig 3.2A.
An opaque shield was placed over the workspace to block the participant's view of the moving arm and the robotic apparatus, while the stationary arm remained visible A vertical computer monitor was mounted in direct view, 0.7 m in front of the participant and just above the shield to minimize neck strain; this display provided visual cues of hand and target position and motion when appropriate, with the timing of visual feedback described below.
Figure 3.2 depicts the experimental setup and protocol: a participant at a robot holds the end effector of a planar manipulandum under a visual occlusion shield, while the left arm shows the standard placement of four tactors (red dots) The figure also outlines the tasks (B) and the sequence of events in each experiment (C), with E1 indicating the first experiment.
Experiments 1 and 2 used a counterbalanced design in which baseline 2 and test 2 conditions were rotated across participants to balance order effects Visual feedback (V) and vibrotactile feedback (T) were implemented in three modes: continuous (+), absent (-), or provided only at the end of each task as knowledge of results (KR) The same sequence was applied across two sessions, with the only difference being that the vibrotactile feedback encoded either error information or state information.
Supplemental kinesthetic feedback was provided using a two-channel (four-tactor) vibrotactile display attached to the non-moving arm Each tactor consisted of a micro-motor with an integrated eccentric rotating mass, such as Pico Vibe 10 mm vibration motors from Precision Microdrives.
Microdrives Inc., Model 310-117 vibrotactile actuators operated over a frequency range of 50–250 Hz deliver a peak vibrational amplitude of 0.97 N, corresponding to an expected maximal forearm-plus-hand acceleration between 0.53 and 0.77 m/s^2, depending on participant anthropometrics The tactors were driven with a pulse-step control scheme in which increments in activation are realized by first fully activating the tactor for a brief moment (0.8 milliseconds) before reducing to the commanded level, a method that overcomes static friction and improves repeatability of tactor excitation In the standard configuration, tactors were arranged with one on the back of the hand, two on the forearm, and one on the upper arm (Fig 3.2A; default locations indicated by red spheres) The hand tactor is located approximately 1 cm proximal to the first and second metacarpophalangeal joints; forearm tactors are positioned about 3 cm distal to the cubital fossa, one on each side of the forearm (+x right, −x left); the upper-arm tactor lies about 5 cm proximal to the cubital fossa on the biceps muscle belly Elastic fabric bands secure the tactors.
We conducted a verification procedure to refine tactor locations so that each participant could reliably indicate which tactor or pair of tactors was active, using low, middle, and high vibration intensities at approximately 10%, 40%, and 90% of the full-scale range The procedure began by asking participants to place the hand’s cursor at each of the four screen corners and report which tactors were vibrating at ~90% FSR, repeated twice—once near the center (~10% FSR) and again midway between the center and the edge (~40% FSR) Next, participants placed the hand cursor at the screen center and moved away from and back toward the center to feel the changing intensity If a participant could not clearly indicate the active tactor and the direction of activation change, personalized tests were used to isolate and resolve the problem This setup identified well-discriminated stimulation sites in all participants and typically took about 10 minutes, though for 16 healthy participants finding well-discriminated sites required repeated adjustments (see 3.4 DISCUSSION).
The vibrotactile display was calibrated to the robot’s workspace so that rightward motion of the robot handle activates the +X tactor, whereas motion of the handle away from the participant toward the monitor activates the +Y tactor The subsequent sections describe the various mappings of hand kinematics onto vibratory stimuli and explain how different movements are translated into tactile feedback.
During three days of testing, participants completed two distinct experimental tasks (Fig 3.2B): first, stabilizing the hand at a fixed point in space against robotic perturbations; and second, reaching to 16 spatial targets that sampled 16 movement directions and two movement extents.
During the stabilization task, participants aimed to hold the robot's handle steady at a central, comfortable "home" position within the workspace In each one-minute stabilization trial, the robot generated spatially complex perturbations that were modeled as a sum of sinusoidal forces, combining a predictable low-frequency component with an unpredictable high-frequency component (Equations 2).
Hand motion in the study was described by the vertical component F_Y, modeled as F_Y = 0.75 sin(2π·1.65 t) + 0.75 sin(2π·1.1 t) + 6 sin(2π·0.25 t) During pilot testing, some participants stabilized the hand by stiffening the arm and ignored vibrotactile feedback, so they were instructed: “Without stiffening your arm, keep your hand as steady as possible using the vibration feedback.” The stabilization task was performed under three visual feedback conditions: continuous visual feedback (V+), where a 0.5 cm radius cursor continuously tracked hand motion; no visual feedback (V-), where the cursor was never shown and participants stabilized at the home position without cursor feedback; and Knowledge of Results (VKR), where cursor feedback of hand position was provided only after the trial Reminders to avoid arm stiffening and to focus on the vibration were repeated throughout the stabilization trials.
Reaching: In this task, participants performed out-and-back reaches to 16 targets For each target, participants reached to the target, paused briefly, and executed a return-to-home movement, totaling 32 discrete goal-directed reaches Each target-capture movement began from the same comfortable “home” position used in the stabilization task.
Results
Hand position drift was a conspicuous feature of kinematic performance during stabilization in neurologically intact participants In a representative trial without visual feedback, the hand drifted steadily to the left Drift was well modeled as a linear function of time for both the X- and Y-axis projections, enabling decomposition of the raw stabilization kinematics (RMSEtotal) into drift-related (RMSEdrift) and residual (RMSEresidual) components, with the latter reflecting the participant’s ability to compensate for moment-to-moment changes in the imposed robotic forces Across participants, some degree of hand position drift was observed in all cases. -**Support Pollinations.AI:**🌸 **Ad** 🌸 Unlock advanced kinematic analysis for your research with [Pollinations.AI free text APIs](https://pollinations.ai/redirect/kofi)—the perfect tool for studying hand position drift and stabilization!
Figure 3 shows Experiment 1: Selected subject performance in the stabilization task at λ = 1.0 The cursor trajectory exhibits drift over time, illustrated by line shading, with the drift modeled from 5 seconds into the trial to its end at t′ seconds The figure also displays the time courses of the endpoint’s x (black) and y (blue) components from 5 seconds to t′ seconds, and, after drift removal, the residuals of these components over the same interval, highlighting the stabilization dynamics once drift is accounted for.
Across the study population, RMSEtotal varied with the state weighting variable l (Fig
3.4A), with approximately equal contributions of RMSEdrift (Fig 3.4B) and RMSEresidual (Fig 3.4C) at low l values and an increasing drift contribution at higher l values We fit a third-order polynomial to the pooled population RMSETOTAL data and found this relationship to be minimized when l was approximately 0.2 A similar result was obtained upon fitting a third-order polynomial to the pooled RMSEdrift data (Fig 3.4B) By contrast, variation acrossl values in the RMSEresidual data appeared to be minimal (Fig 3.4C)
Figure 4 summarizes Experiment 1, detailing population performance in the stabilization task as a function of the state mixture parameter lambda, with a third-order polynomial population fit and 95% function bounds Panel A reports the RMSE of the end-effector trajectory, Panel B the RMSE of the drift component of the end-effector trajectory, and Panel C the RMSE of the residuals after the drift is removed.
These observations were confirmed using repeated measures MANOVA to compare stabilization performance {RMSEtotal, RMSEdrift, RMSEresidual} across l values MANOVA found significant variation across l values [Wilk's F(15,188) = 2.536; p = 0.002] Post-hoc ANOVA found significant variation in RMSEtotal values [F(5,70) = 6.02, p < 0.0005] such that Dunnett multiple comparison tests (referenced to control level: l = 0.2) revealed significant increases in RMSEtotal when l = 0.6 (p = 0.022),l = 0.8 (p = 0.001) and l = 1.0 (p = 0.001) Similarly, post-hoc
ANOVA [F(5,70) = 5.52, p < 0.0005] indicates a significant effect, with Dunnett's multiple comparison tests revealing substantial increases in RMSEdrift from the control level l = 0.2 to l = 0.8 (p = 0.002) and to l = 1.0 (p = 0.005) By contrast, post-hoc ANOVA shows no significant variation in RMSEresidual across the levels of l [F(5,70) = 1.22, p = 0.311].
Kinematic performance of reaching movements varied with l, despite only brief exposure to each new l encoding scheme prior to structured reach training This variation is most evident in target capture variability, quantified from final hand positions recorded during return-to-home movements Figure 3.5 shows the final hand positions for a single participant as they reached without ongoing visual feedback but with l-weighted vibrotactile feedback The yellow ellipses denote the 95% confidence intervals on the distributions of return-to-home movement endpoints under each l feedback condition Precision of return-to-home reaches declines substantially as l increases l-dependent performance differences were harder to detect for reaches toward near and far targets.
Figure 5 from Experiment 1 shows selected subject performance in the reaching task across each lambda value under the VKR visual condition The yellow ellipses represent the two-dimensional 95% confidence intervals of the return-to-home reach endpoints, illustrating the spatial variability of endpoint positions.
Across the study population, target capture variability for return-to-home reaches varied significantly across l values (ANOVA: F(5,70) = 6.01, p < 0.0005) Dunnett's multiple comparison tests revealed significant increases in target capture variability at l = 0.6 (p = 0.034), l = 0.8 (p = 0.028), and l = 1.0 (p = 0.006) when compared with the baseline l value These results indicate that higher l values are associated with greater variability in target capture during return-to-home tasks.
0.2 (Fig 3.6A) This outcome suggests that low l values enhance spatial localization of the hand about the central reference location
Figure 6 summarizes Experiment 1's population statistics for a reaching task as a function of the state mixture parameter lambda, with error bars representing ±1 SEM Panel A quantifies the variability of raw reach endpoints about the home target by fitting an ellipse to the endpoints and measuring its area Panel B shows the variability of reach endpoints at the central target after collapsing across movement directions Panel C reports the mean absolute error (RMSE) at the central target, with red lines indicating results that are significant at p < 0.05.
Target capture variability and mean absolute error were computed at the center, near, and far targets after collapsing movement directions by counter-rotating the reach endpoints to align with the intended movement direction around the home target Across the six datasets (two performance measures × three target sets), MANOVA revealed a significant variation across l values (Wilk’s F(35,271) = 1.986; p = 0.001) Post-hoc ANOVAs showed a significant main effect of l for both performance measures at the center target (F(5,70) > 3.60; p < 0.006 in each case; Figures 3.6B, 3.6C), but no main effect of l for either measure at the near and far target sets (F(5,70) < 1.31; p not significant).
> 0.270 in all cases] (data not shown) At the center target, Dunnett multiple comparison tests revealed significant increases in target capture variability when l = 0.8 (p = 0.011) and l = 1.0
(p = 0.018) compared to when l = 0.2 (Fig 3.6B) Dunnett multiple comparison tests also revealed significant increases in mean absolute error magnitude when = 0.6 (p = 0.025), l =
0.8 (p < 0.001) and l = 1.0 (p < 0.001) compared to when l = 0.2 (Fig 3.6C)
3.3.2 Experiment 2 - Comparison of Optimal State vs Error Feedback
Figure 3.7 demonstrates that one participant’s stabilization performance in Experiment 2 varied with the type of vibrotactile feedback: during task familiarization with ongoing visual feedback but no tactor feedback (V+T-), the hand remained centered on the home target with virtually no drift; in baseline testing, after visual feedback was removed and vibrotactile stimulation had not yet been introduced (V-T-), the hand gradually drifted away from the home target with drift directions differing across trials; after about 45 minutes of training with vibrotactile feedback during reaching and stabilizing, post-training baseline assessments still showed drift, though its magnitude decreased somewhat; by contrast, drift was eliminated when vibrotactile feedback conveyed either state or error information (V-T+), and this improvement arose from the informational content of the feedback rather than the mere presence of vibrotactile stimulation, since drift was at least as large during sham stimulation trials (V-TSHAM) as during baseline.
Figure 7 presents Experiment 2, examining selected subjects’ performance in the stabilization task The cursor trajectory shows drift over time, with line shading indicating how drift varies with the presence and type of vibration feedback Drift is modeled from t = 5 seconds to the end of the trial at t' seconds, and the red values denote the RMSE Drift for each trial.
Based on these observations, we focused our analysis of population behavior on
RMSEdrift during the post-training baseline (V - T -), the V - T+ test phase, and the sham feedback (V - TSHAM) phase varied significantly across phases when stabilizing about the home target, as shown by a two-way repeated measures ANOVA [F(2,70) = 23.76, p < 0.0005], but did not vary with feedback conditions [F(1,70) = 0.56, p = 0.457], and there was no interaction between phase and feedback [F(2,70) = 2.74, p = 0.071] (Fig 3.8) Dunnett multiple comparison tests indicated that the significant main effect arose from a decrease in RMSEdrift during the V - T+ test phase relative to both the V - T - post-training baseline phase (p < 0.0005) and the sham stimulation phase (V - TSHAM: p < 0.0005) The difference between the post-training baseline and test phases was not due to an order effect, as the presentation order of these two phases was counterbalanced across participants.
Vibrotactile feedback specifically improved RMSEdrift, while neither feedback type yielded a systematic improvement in RMSEresidual when comparing the test phase to post-training baseline performance; both ANOVA and Dunnett's tests produced p-values greater than 0.211, indicating no significant effect in either case.
Figure 8: Experiment 2: Population statistics in the stabilization task for error and state feedback Red lines: p < 0.05
Discussion
This line of research aims to develop sensory-substitution technologies that improve closed-loop control of goal-directed actions for people with impaired somatosensation In this initial study, we evaluated the objective and subjective utility of two vibrotactile feedback schemes—one encoding the limb state and the other encoding hand-position error—in enhancing real-time arm stabilization and reaching in neurologically healthy participants To reflect practical constraints faced by stroke survivors who often have diminished somatosensation in the affected arm, the feedback was delivered to the opposite arm, a body part not directly involved in the task This cross-arm approach is a sensible first step because even neurologically normal individuals show drift in their internal hand-position estimates during localization tasks.
Two experiments show that vibrotactile encoding schemes can effectively eliminate drift when stabilizing the hand at the origin of the feedback encoding space (the "home target"), even with only 1 minute of prior exposure Both schemes immediately improve the accuracy and precision of reaching movements directed toward the encoding-space origin However, they differ in their ability to enhance reach accuracy and precision at other locations in the arm’s workspace In Experiment 1, 1 minute of training was not enough for participants to use any tested state-feedback scheme to successfully reach near and far targets The best scheme—80% hand position information plus 20% hand velocity information—along with error feedback, improved reach accuracy and precision across spatial targets throughout the reachable workspace.
Across roughly 45 minutes of training in Experiment 2, the study found that the benefits of state feedback and error feedback depended on the informational content encoded by vibrotactile cues, as non-informative sham stimulation produced no meaningful performance gains This finding has implications for sensory substitution technologies in neurorehabilitation, showing that individuals can learn to use simple state feedback to boost goal-directed stabilization and reaching when visual feedback is unavailable Notably, during reaching tasks, error feedback surpassed optimal state feedback not only on objective outcomes such as target capture accuracy and precision, but also on a subjective assessment of usefulness.
3.4.1 Importance of information content within supplemental vibrotactile feedback
Model simulations shown in Figure 3.1 indicate that for systems dominated by second‑order dynamics with delayed feedback, incorporating a modest amount of velocity information into position feedback—around a 20% velocity to 80% position mix—can enhance control, since velocity feedback is predictive of position feedback, with velocity leading position by 90° in phase and potentially improving the perception of limb state changes Experiment 1 tested these predictions and found evidence that a vibrotactile feedback mix of 20% velocity and 80% position reduced hand position drift during stabilization and improved reach accuracy and precision at the center target, but varying the relative composition of state feedback did not produce significant changes in the ability to reject moment‑by‑moment force perturbations, likely due to limited training with each feedback combination Across near and far targets and across both experiments, results suggest that participants required up to about 45 minutes of training with the optimal state encoding to begin learning to use that feedback form, and training‑dependent gains were limited to drift reduction and improved point‑to‑point targeting, with no significant improvement in RMSE residual This implies that full integration of supplemental kinesthetic feedback into moment‑by‑moment control may demand substantially more training than the brief sessions provided in the current studies, and might also benefit from task‑specific stabilization training rather than the reaching emphasis used in these experiments.
Subjective feedback after Experiment 1 revealed that participants often solved the stabilization task by stiffening the arm rather than using the vibrotactile feedback as intended, which led us to develop targeted task instructions and repeatedly remind participants to avoid stiffness and rely on vibration to complete each task Across pilot testing and the survey period after each trial, participants reported that they generally used the vibration as instructed, but when the vibration conveyed more velocity information than position information, some temporarily stiffened their arm to discern position within the signal before using it effectively Consequently, the value of l produced markedly different state-feedback encodings and subjective experiences, which in turn could drive distinct strategies for integrating the supplementary vibrotactile feedback into real-time arm control.
Qualitative differences between state and error feedback encodings account for performance differences in reaching near and far targets in Experiment 2 We defined the origin of the arm’s workspace at the center of the home target, making error and state encodings equivalent when the target is centered and the parameter l is small (e.g., 0.2) but independent when l = 1.0 However, the origin of the error encoding jumps to the current goal location anywhere within the workspace, while the origin of the state encoding remains fixed As a result, when driven by error feedback, test-phase target capture performance at near and far targets closely reproduces home-target performance in the same phase and resembles performance with visual feedback observed during the familiarization phase By contrast, optimal state feedback yields much better performance than baseline and sham phases, yet target capture variability and RMS error across all target sets are roughly twice as large with optimal state encoding versus error encoding.
Qualitative differences between state and error feedback encoding are critical for designing practical, supplemental feedback systems, and they become especially important when moving from lab-based tasks to real-world use In highly constrained experiments with a robotic manipulandum, implementing an error-encoding scheme is straightforward because we can measure instantaneous hand position relative to well-defined spatial targets; however, defining error feedback for wearable technology in unstructured environments is much more challenging, since the system must predict the user’s intent and movement goals on a moment-by-moment basis Although no current technology reliably performs intent and goal prediction in uncertain environments, low-cost wearables that integrate MEMS accelerometers, gyroscopes, and magnetometers can still estimate limb state These hurdles aren’t insurmountable, because real-time computing systems already support goal-aware feedback encodings that enhance human performance on difficult but well-defined tasks, such as balancing an inverted pendulum while minimizing both kinematic error and control effort.
3.4.2 Exposure to vibrotactile feedback of limb state induces spatial learning
Participants learned to map hand position in the horizontal plane to target location in the display’s vertical plane, effectively acquiring an inverse kinematic transformation that converts desired visual cursor changes into corresponding hand movements While performing tasks with vibrotactile cues, they also learned two aspects of an additional interposed transformation: how hand motion drives vibrotactile display activity and how changes in visual target location modulate vibrotactile patterns Experiment 1 tested the first aspect and found that a state encoding with l = 0.2 yielded better stabilization and return-to-home reaches than several alternative encoding schemes Experiment 2 tested the second aspect and found that, within a single session, error feedback outperformed optimal state feedback in guiding reaches, particularly to near and far targets.
Two experimental observations indicate that state feedback encourages learning of an additional spatial map, providing a rationale for state feedback-based supplemental vibrotactile systems First, after about 45 minutes of practice with optimal state feedback, participants showed performance improvements at near and far targets during test phases with vibrotactile feedback; because the order of V-T+ test and V-T- baseline phases was counter-balanced across subjects, the learning effect cannot be attributed to longer practice in the V-T+ condition, suggesting that optimal state feedback enhanced reach accuracy and precision Near and far targets mapped to non-zero activation patterns in the vibrotactile display, so the observed improvements were not confounded by mere similarity to error encoding at the center target Second, post-training baseline reaches performed without ongoing vibrotactile stimulation showed less overshoot than pre-training baselines, especially at the near targets, implying that training with optimal state vibrotactile feedback facilitated learning of an internal spatial map that could be recalled to guide reaches in the absence of stimulation This improvement was evident across all three target sets, particularly for target capture error (RMSE), indicating that optimal state feedback can be as effective as error feedback in teaching spatial relations among target location, hand position, and vibrotactile input Future studies should directly investigate how participants learn to use supplemental vibrotactile feedback to shape internal representations of body configuration.
3.4.3 Potential applications of supplemental vibrotactile stimulation
Wearable technologies that deliver subsensory, random vibrotactile stimulation to the soles of the feet or to the tendons of finger flexor muscles have potential to augment human motor performance, with applications in standing balance and grip force production This augmentation is often explained by stochastic resonance, a nonlinear cooperative interaction in which a weak cutaneous input, when combined with injected noise, enhances the sensitivity of threshold detectors In our study, however, the observed benefits were not attributable to stochastic resonance, because performance improvements dissipated in the presence of sham stimulation.
Recent experimental work shows that artificial activation of sensory afferents in distal arm muscles via 70 Hz wrist tendon vibration can improve proximal-arm control (shoulder and elbow) during reaching and tracking tasks in the paretic arm of stroke survivors by providing excitatory drive to central and peripheral sensorimotor circuits Compared with baseline, vibrotactile stimulation enhances end-point hand stability at the end of reach, reduces muscle activity throughout the arm, and lowers grip pressure during movement, with effects observed during stimulation and for a brief period after it ceases Potential mechanisms include improved cortical sensorimotor integration within spared neural circuits mediating the hemiparetic arm or heightened spinal reflex thresholds through modulation of spinal reflex activity, reducing spastic hypertonia Unlike the vibrotactile feedback used in some prior work, which did not encode meaningful information, the present stimulation conveys information encoded in the tactile data stream, yielding effects specific to the conveyed sensory content.
Vibrotactile displays offer a way to inject useful information into the human nervous system, with applications ranging from tactile navigation displays for aircraft pilots seeking to fly toward a target (Van Erp 2005) and to hover a helicopter (Raj et al 2000), to indicating direction in low-visibility conditions (Sklar and Sarter 1999), and even enabling vibrotactile transmission of spoken language (Novich and Eagleman 2015) In our study, injecting informative vibrotactile feedback to the non-moving arm recruits alternate sensorimotor control pathways to guide the moving arm The results show that for neurologically intact individuals, this approach can improve the accuracy of goal-directed reaching in the absence of ongoing visual feedback and eliminate limb-position drift during limb stabilization without vision For stroke survivors who retain some proximal arm strength and the capacity to produce flexion and extension torques, vibrotactile feedback could promote greater use of the affected arm by restoring a sense of movement, thereby enhancing arm control when visual feedback is unavailable.
Although the current study doesn’t delve into neural mechanisms in depth, both vibration and joint-position sense follow the dorsal column–medial lemniscus pathway, projecting through the ventral posterior lateral nucleus (VPL) of the thalamus to the primary somatosensory cortex Functional neuroimaging shows that these regions contribute to real-time, closed-loop control of the distal upper extremity and are notably vulnerable to injury from the most common form of stroke Recent research also identifies networks of neurons that interconnect the two sides of the brainstem and spinal gray matter, along with intrahemispheric transcallosal connections that may form bilateral coordination networks for sensorimotor integration.
Building on the concept of "detour circuits" for recovery of function, as reviewed by Jankowska and Edgley (2006), we speculate that detour circuits may provide a route for the supplemental kinesthetic feedback described here to tap into residual cerebello-thalamo-cortical circuits that participate in the real-time, closed-loop control of the contralesional arm and hand.
Participants in the current study found error feedback more useful than state feedback during the reaching task, yet showed no clear preference for either form while stabilizing at the home target, likely because state and error feedback were virtually identical there; in Experiment 2, every participant reported preferring informative vibrotactile feedback during both training and testing over no vibration, and many expressed dismay at having to repeat the task without vibration after practicing with it; these subjective responses highlight a positive user experience for wearable vibrotactile encoding of state and/or error feedback aimed at enhancing motor performance of goal-directed arm actions.
Conclusions
This study demonstrates the immediate utility and relative merits of two vibrotactile kinesthetic feedback methods for stabilizing and guiding reaching movements of the arm and hand in neurotypical individuals In the first experiments, a specific pairing of hand position and velocity information was identified as optimal for state-feedback control of both stabilization and reaching after very limited practice In the second set, however, error feedback—a simple, goal-aware form—yielded superior performance across the entire reachable workspace compared with optimized state feedback These results show that the intact human brain can integrate vibrotactile kinesthetic feedback into real-time control of moving the arm and hand, even when feedback is applied to a body part not directly involved in the action (e.g., the opposite arm) Together, the findings provide strong empirical support for developing sensory substitution technologies aimed at counteracting impaired proprioceptive sensation in stroke survivors who retain motor capacity in the more involved arm.