Characterizing Healthy & Post-Stroke Neuromotor Behavior During 6D Upper-Limb Isometric Gaming: Implications for Design of End-Effector Rehabilitation Robot Interfaces

Ajay Anand; Gabriel Parra; Chad A. Berghoff; Laura A. Hallock

cs.ROMar 10, 2026

Characterizing Healthy & Post-Stroke Neuromotor Behavior During 6D Upper-Limb Isometric Gaming: Implications for Design of End-Effector Rehabilitation Robot Interfaces

Authors: Ajay Anand, Gabriel Parra, Chad A. Berghoff, Laura A. Hallock

Organizations: The authors are with the Departments of Mechanical Engineering, Physical Medicine & Rehabilitation, and Mathematics, the Kahlert School of Computing, and the Robotics Center at the University of Utah, Salt Lake City, UT 84112, USA.

Abstract

Successful robot-mediated rehabilitation requires designing games and robot interventions that promote healthy motor practice. However, the interplay between a given user's neuromotor behavior, the gaming interface, and the physical robot makes designing system elements -- and even characterizing what behaviors are "healthy" or pathological -- challenging. We leverage our OpenRobotRehab 1.0 open access data set to assess the characteristics of 13 healthy and 2 post-stroke users' force output, muscle activations, and game performance while executing isometric trajectory tracking tasks using an end-effector rehabilitation robot. We present an assessment of how subtle aspects of interface design impact user behavior; an analysis of how pathological neuromotor behaviors are reflected in end-effector force dynamics; and a novel hidden Markov model (HMM)-based neuromotor behavior classification method based on surface electromyography (sEMG) signals during cyclic motions. We demonstrate that task specification (including which axes are constrained and how users interpret tracking instructions) shapes user behavior; that pathology-related features are detectable in 6D end-effector force data during isometric task execution (with significant differences between healthy and post-stroke profiles in force error and average force production at $p=0.05$ ); and that healthy neuromotor strategies are heterogeneous and inherently difficult to characterize. We also show that our HMM-based models discriminate healthy and post-stroke neuromotor dynamics where synergy-based decompositions reflect no such differentiation. Lastly, we discuss these results' implications for the design of adaptive end-effector rehabilitation robots capable of promoting healthier movement strategies across diverse user populations.

Figures & tables

Fig. 1 : End-effector robot–mediated rehabilitation — in which users are instructed to perform varied trajectory tracking tasks while assisted (or resisted) by a robot — consists of complex feedback loops ( blue ) between task specification and display, human neuromotor behavior, and robot intervention. In this work, we leverage multimodal biosensor data, collected from 13 healthy and 2 post-stroke participants during completion of 8 isometric gaming tasks under 2 pose conditions [ 1 ] , to address three questions key to effective therapy design ( gold ), toward development of rehabilitation robots that can induce targeted, healthy neuromotor practice.

Fig. 2 : Motor rehabilitation platform enabling measurement of muscle engagement and end effector forces during trajectory tracking tasks. Users exert forces and torques on a SensONE 6-axis load cell (a) (Bota Systems AG, Zürich, Switzerland) through the attached handle, which are then mapped to

x

–

y

coordinates of on-screen avatar (b) to allow users to follow red target ball (c) through different trajectories within a custom gamified rehabilitation environment developed in Unity (Unity Software Inc., San Francisco, CA, USA). During each trajectory tracking task, surface electromyography (sEMG) electrodes (d) — two Trigno Quattro 4-channel sensor motes, controlled through a Trigno Base Station (Delsys, Inc., Natick, MA, USA) — placed on key muscles of the arm (e) record muscle activations. The system currently supports isometric rehabilitation tasks at arbitrary poses — the LBR iiwa 14 R820 7-degree-of-freedom cobot (f) (KUKA AG, Augsburg, Germany) remains static — but will be expanded in the future to support a variety of robot controllers. Surface EMG electrode placements: anterior deltoid (AD), middle deltoid (MD), posterior deltoid (PD), and biceps brachii (BB), grounded at the acromion (G1); triceps brachii (long head, TR), brachioradialis (BR), wrist flexors (FL), and wrist extensors (EX), grounded at the olecranon (G2). Reprinted from [ 1 ] .

Fig. 3 : Experimental flow during collection of pilot data set. Participants were consented and surveyed, then completed trajectory tracking tasks at two ADL-inspired poses — Conditions A (shoulder adducted, elbow flexed $\sim 90\degree$ , forearm horizontal) and B (shoulder slightly flexed above horizontal with near full elbow extension) — in randomized order before providing final survey feedback. Survey data are not examined in this work. Reprinted from [ 1 ] .

Goal Trajectory	Input	Output (Game)	Repetitions
$x$ -axis	$x$ -axis force	$x$ coordinate	7
$y$ -axis	$y$ -axis force	$y$ coordinate	7
$z$ -axis	$z$ -axis force	$y$ coordinate	7
torque	$z$ -axis torque	$x$ coordinate	5
circle (CW 1 )	$(x,y)$ -axis force	$(x,y)$ coordinate	3
circle (CCW 2 )	$(x,y)$ -axis force	$(x,y)$ coordinate	3

TABLE I : Trajectory Tracking Tasks & Mappings

Fig. 4 : Illustration of three distinct observed target tracking behaviors, depicted during the spline 2 trajectory tracking task. Participants were instructed to follow target ball

t

with avatar ball

a

as it traversed the specified trajectory. Observed strategies included (1) pursuit of the target along the trajectory, (2) shortest-distance pursuit to

t

’s current position, and (3) movement toward an anticipated future position of

t

Fig. 5 : Exemplar time series force profiles of two healthy participants (02, top and 08, middle ) and one post-stroke participant (21, bottom ) performing the

x

-axis trajectory tracking task in pose Condition A. All participants exhibit not only “productive” force exertions (

x

, red ) contributing to task completion, but substantial “non-productive” output in orthogonal dimensions (

y

, green and

z

, blue ). The magnitude and temporal characteristics of these non-productive exertions vary within both healthy and post-stroke populations, illustrating the importance of careful specification when prescribing movement tasks. The post-stroke participant also exhibits markedly altered force production, with sustained off-axis force generation and reduced efficiency in producing task-relevant forces, reflecting underlying pathology.

Fig. 6 : Aggregate force output — measured as the mean and variance of each rectified force time series — for all 3 force dimensions during completion of the

x

-axis tracking task in pose Condition A for both healthy (02–20, left ) and post-stroke (21–22, right ) participants. While productive

x

-axis ( red ) force magnitudes are largely consistent across individuals, post-stroke participants exhibit higher variance and elevated levels of off-axis forces (

y

, green and

z

, blue ), indicating increased non-productive effort. Healthy participant behavior is, however, heterogeneous, with some healthy individuals exerting substantial non-productive forces (e.g., 08, 12), and others (e.g., 04, 14) exhibiting only minor off-axis force production and thus improved task efficiency. In this and other tasks, most individuals exhibited highest and most variable non-productive exertions in the

z

-axis (i.e., in/out). Box plots show the median (50th percentile), interquartile range (25th–75th percentiles), whiskers extending to 1.5

\times

IQR, and outliers beyond this range.

Fig. 7 : Exemplar time series data of both player avatar ( solid ) and target ball ( dashed )

x-y

position during completion of the spline 2 trajectory tracking task for two exemplar participants. While some participants, such as 02 ( top ), performed sufficiently tight tracking that their trajectories were largely indistinguishable from that of the target ball, most participants (including 21, bottom , when performing the displayed task), both healthy and post-stroke, exhibited multiple distinct target tracking behaviors to varied extents, which must be considered in analyses of task performance and comparisons of neuromotor behavior across individuals. Annotations (1), (2), and (3) indicate times at which each of the 3 pursuit strategies identified in Figure 4 can be observed within a single trial.

Panel of Figure 9 — Fig. 8 : “Productive” force RMSE (a), total force impulse (b), average force output (c), and peak force output (d), as calculated for each of the 7 force trajectory tracking tasks at pose Condition A, aggregated across all healthy (02–20, turquoise ) and post-stroke (21–22, pink ) participants. Post-stroke participants exhibit significantly higher force RMSE, consistent with existing literature, as well as significantly higher average force output. Total force impulse and peak force output appeared higher for impaired participants, but results were not statistically significant at $p=0.05$ . These findings were consistent across both pose conditions, with only marginal differences between them. Box plots show the median (50th percentile), interquartile range (25th–75th percentiles), whiskers extending to 1.5 $\times$ IQR, and outliers beyond this range.

Fig. 9 : Optimal synergy count (OSC) reflecting neuromotor behavior complexity, aggregated across all tracking tasks in pose Condition A, for all participants. Both healthy (02–20, turquoise ) and post-stroke (21–22, pink ) participants exhibit comparable OSCs of 3–5 (standard deviation displayed as black lines), challenging previous findings [ 31 ] documenting an inverse relationship between OSC and impairment severity. These findings hold across both pose conditions with only marginal differences.

Fig. 10 : HMM subtask classification error — defined in section III-E2 as the fraction of time points for which HMM-predicted action sequences and target direction subtasks were not the same — is higher in average magnitude and variance (as calculated across 25 independently trained HMMs for each trial) for post-stroke participants (21–22, pink ) for the

x

-axis trajectory tracking task as compared with most healthy participants (02–20, turquoise ), a finding that holds across the majority of examined tracking tasks in each pose condition (see appendix V), providing a metric for discriminating healthy and pathological neuromotor behavior. Some healthy participants (here, 12) exhibit behavior more consistent with post-stroke participants as quantified by this metric intermittently across tasks, illustrating the challenge of discriminating truly pathological motor behavior from uncommon but healthy motor strategies. Box plots show the median (50th percentile), interquartile range (25th–75th percentiles), whiskers extending to 1.5

\times

IQR, and outliers beyond this range.

Fig. 11 : Exemplar HMM-predicted action sequences ( bottom , solid black ), fit to 8-channel sEMG time series data ( top ) as detailed in section III-E2 , for the $x$ -axis trajectory tracking task in pose Condition A, displayed alongside target $x$ -axis game position ( bottom , dashed gray ) illustrating targeted cyclic horizontal direction changes. HMM-predicted action sequences consistently coincide with direction-specific subtasks for the healthy participant shown (04, left ) and for most healthy participants during most tasks, suggesting that directionality is the dominant factor determining underlying motor behavior. On the other hand, HMM-predicted action sequences substantially deviate from the target subtask sequence for the post-stroke participant shown (22, right ) and for both post-stroke participants during most tasks, indicating unmodeled complexity in neuromotor behaviors and/or additional time-varying factors impacting motor behavior. In this exemplar, HMM-identified actions correspond to temporal regions in which pathologically high wrist flexor activation was and was not present, indicating that HMM-based decompositions may be able to segment regions in which maladaptive compensatory behaviors are and are not expressed. Additional data on the distribution of HMM classification error across participants are available in appendix V.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 13 — Fig. 1: “Productive” force RMSE (a), total force impulse (b), average force output (c), and peak force output (d), as calculated for each of the 7 force trajectory tracking tasks at pose Condition B, aggregated across all healthy (02–20, turquoise ) and post-stroke (21–22, pink ) participants. Post-stroke participants exhibit significantly higher force RMSE, consistent with existing literature, as well as significantly higher average force output. Total force impulse and peak force output appeared higher for impaired participants, but results were not statistically significant. These findings were consistent across both pose conditions. Box plots show the median (50th percentile), interquartile range (25th–75th percentiles), whiskers extending to 1.5 $\times$ IQR, and outliers beyond this range.

Panel of Figure 14 — Fig. 2: Participants falling into non-normative ( orange ) — as contrasted with normative ( blue ) — loss score distribution. Post-stroke participants (21, 22) or participants who were observed to have atypical behavior (8) are the only participants who fell into the non-normative distribution on half or more of the tasks.

Task	Condition	Participants in Non-Normative Loss Distribution
$x$ -axis	Condition 1	8, 12, 13, 15, 21, 22
$x$ -axis	Condition 2	12, 21, 22
$y$ -axis	Condition 1	3, 5, 21, 22
$y$ -axis	Condition 2	3, 15
$z$ -axis	Condition 1	20, 22
$z$ -axis	Condition 2	7, 8, 13, 22

Appendix

TABLE I: Participants in Non-Normative Loss Distribution

Explore similar work

Mar 17, 2026

Real-Time Decoding of Movement Onset and Offset for Brain-Controlled Rehabilitation Exoskeleton

Robot-assisted therapy can deliver high-dose, task-specific training after neurologic injury, but most systems act primarily at the limb level-engaging the impaired neural circuits only indirectly-which remains a key barrier to truly contingent, neuroplasticity-targeted rehabilitation. We address this gap by implementing online, dual-state motor imagery control of an upper-limb exoskeleton, enabling goal-directed reaches to be both initiated and terminated directly from non-invasive EEG. Eight participants used EEG to initiate assistance and then volitionally halt the robot mid-trajectory. Across two online sessions, group-mean hit rates were 61.5% for onset and 64.5% for offset, demonstrating reliable start-stop command delivery despite instrumental noise and passive arm motion. Methodologically, we reveal a systematic, class-driven bias induced by common task-based recentering using an asymmetric margin diagnostic, and we introduce a class-agnostic fixation-based recentering method that tracks drift without sampling command classes while preserving class geometry. This substantially improves threshold-free separability (AUC gains: onset +56%, p = 0.0117; offset +34%, p = 0.0251) and reduces bias within and across days. Together, these results help bridge offline decoding and practical, intention-driven start-stop control of a rehabilitation exoskeleton, enabling precisely timed, contingent assistance aligned with neuroplasticity goals while supporting future clinical translation.

Kanishka Mitra, Satyam Kumar, Frigyes Samuel Racz +3

Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA. · Chandra Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, USA. · Department of Neurology, The University of Texas at Austin, Austin, TX, USA. +3

Brain-Computer Interfaces Assistive Robotics MI Classification Electroencephalography

Ask PDF

Mar 2, 2026

Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

Post-stroke rehabilitation is often necessary for patients to regain proper walking gait. However, the typical therapy process can be exhausting and physically demanding for therapists, potentially reducing therapy intensity, duration, and consistency over time. We propose a Patient-Therapist Force Field (PTFF) to visualize therapist responses to patient kinematics and a Synthetic Therapist (ST) machine learning model to support the therapist in dyadic robot-mediated physical interaction therapy. The first encodes patient and therapist stride kinematics into a shared low-dimensional latent manifold using a Variational Autoencoder (VAE) and models their interaction through a Gaussian Mixture Model (GMM), which learns a probabilistic vector field mapping patient latent states to therapist responses. This representation visualizes patient-therapist interaction dynamics to inform therapy strategies and robot controller design. The latter is implemented as a Long Short-Term Memory (LSTM) network trained on patient-therapist interaction data to predict therapist-applied joint torques from patient kinematics. Trained and validated using leave-one-out cross-validation across eight post-stroke patients, the model was integrated into a ROS-based exoskeleton controller to generate real-time torque assistance based on predicted therapist responses. Offline results and preliminary testing indicate the potential of their use as an alternative approach to post-stroke exoskeleton therapy. The PTFF provides understanding of the therapist's actions while the ST frees the human therapist from the exoskeleton, allowing them to continuously monitor the patient's nuanced condition.

Grayson Snyder, Lorenzo Vianello, Levi Hargrove +2

Center for Robotics and Biosystems, Northwestern University, Evanston, IL, USA. · Legs and Walking Lab, Shirley Ryan AbilityLab, Chicago, IL, USA.

Assistive Robotics Human-Robot Interaction Policy Learning Robot Imitation Learning Ankle Exoskeleton Control

Ask PDF

Mar 15, 2026

See, Learn, Assist: Safe and Self-Paced Robotic Rehabilitation via Video-Based Learning from Demonstration

In this paper, we propose a novel framework that allows therapists to teach robot-assisted rehabilitation exercises remotely via RGB-D video. Our system encodes demonstrations as 6-DoF body-centric trajectories using Cartesian Dynamic Movement Primitives (DMPs), ensuring accurate posture-independent spatial generalisation across diverse patient anatomies. Crucially, we execute these trajectories through a decoupled hybrid control architecture that constructs a spatially compliant virtual tunnel, paired with an effort-based temporal dilation mechanism. This architecture is applied to three distinct rehabilitation modalities: Passive, Active-Assisted, and Active-Resistive, by dynamically linking the exercise's execution phase to the patient's tangential force contribution. To guarantee safety, a Gaussian Mixture Regression (GMR) model is learned on-the-fly from the patient's own limb. This allows the detection of abnormal interaction forces and, if necessary, reverses the trajectory to prevent injury. Experimental validation demonstrates the system's precision, achieving an average trajectory reproduction error of 3.7cm and a range of motion (ROM) error of 5.5 degrees. Furthermore, dynamic interaction trials confirm that the controller successfully enforces effort-based progression while maintaining strict spatial path adherence against human disturbances.

Ali Alabbas, Camillo Murgia, Joanne Regan +1

Faculty of Mechanical Engineering, Atlantic Technological University, Galway, Ireland. · Department of Health and Nutritional Science, Atlantic Technological University, Sligo, Ireland.

Assistive Robotics Robotic Control Robot Motion Generation Robot Imitation Learning

Ask PDF