A Biomimetic Gaze Control to Restore Head-Neck Movements

Colin Rubow; Haohan Zhang

cs.ROOct 2, 2026

A Biomimetic Gaze Control to Restore Head-Neck Movements

Authors: Colin Rubow, Haohan Zhang

Organizations: Department of Mechanical Engineering and Robotics Center, University of Utah, Salt Lake City, UT 84112, USA. · Rocky Mountain Center for Occupational and Environmental Health, University of Utah, Salt Lake City, UT 84112, USA.

Abstract

Powered neck exoskeletons are developed to restore head-neck motions for those with neck muscle weakness. Current methods using hand-held input devices to control these robotic systems are unintuitive or non-viable for these populations. Alternatively, the eyes coordinate with the head to accomplish many daily tasks. Therefore, movements of the eyes can be a useful signal to control the neck exoskeletons without needing to use the hands. However, previous control models lack biobehavioral basis as well as means to personalize the control behavior, causing unnatural head-neck movements. This paper presents a novel eye-tracking controller design leveraging the vestibulo-ocular reflex to emulate natural eyehead coordination. Personalization tools were also designed, consisting of sliders to allow users to adjust the parameters of this controller. Here, we present the technical design and validation of this system in human users. Compared to the previous eye-tracking model, the new controller provides natural head motions personalized to the user through more natural velocity profiles and a 24 percent average increase of accurate motion intent detection. This system provides a powerful platform for studying eye-tracking controllers in the context of personalization and preference, and ultimately restoring the lost head-neck mobility in patient populations.

Figures & tables

Fig. 1: Deadzone and gaze diagram with plot of rotational speed related to gaze shift distance.

\Delta\theta_{\text{des}}

is the desired angle of the rotation of the head.

\dot{\theta}_{h}

is the velocity of the head which is broken into its pitch and yaw rotational components.

d_{x}

and

d_{y}

are the lengths of the deadzone elliptical axis where the deadzone is centered in the world camera view. While the gaze is inside the deadzone, the head velocity is 0, but once outside the deadzone, the velocity increases linearly at the rate of

v_{p}

, the velocity proportion, in the direction of the gaze.

Fig. 2: Trajectory comparison results between the vector and region-based controllers in 100 computer simulated gaze shifts. (left) A typical gaze shift performance. The vector controller moves straight towards the target, while the region-based controller moves first to the right, then zigzags towards the target (enlarged in the zoomed view). Both controllers stop when the gaze target hits their respective deadzones, the dashed lines of the same color. The 100 gaze targets are distributed from directly rightward to upwards with a slope of 45 ∘ . (right) Box plot of the difference of path lengths between both controllers. Higher values correspond to a higher path length performed by the region-based controller as compared to the vector controller. The median value is

13.7^{\circ}

Fig. 3: The robotic system. The arrows denote the directional interaction subsystems have with each other. The android phone runs a program that allows gaze data streaming.

Fig. 4: The powered neck exoskeleton with eye-tracking integration. An inertial measurement unit (IMU) was included for the study as an independent measurement of head motions, but it was not used in the robot control. The headband can be easily detached from the rest of the robot by removing a set of fasteners.

Fig. 5: Mechanical sliders for manual parameter tuning. Stops restrict linear potentiometers to their linear region. The sliders consist of rails and low-friction rollers.

Fig. 6: Setup for the eye-head coordination model justification experiment. The IMU and eye-trackers were attached to the participant with a form fitting headband. The reflection blocking screen wrapped around the experiment space to block distracting motions and reflections.

Were the sliders smooth and easy to manipulate?

Do you feel like you could easily explore the various possibilities of the parameters?

Was having three parameters to change too confusing?

Do you feel like you were able to find an acceptable tuning of parameters?

TABLE I: survey questions after the third part of the experiment

	Velocity Proportion Regression		Deadzone Best Fit Ellipse
Participant	Slope ( $\frac{\text{rad}/\text{s}}{\text{rad}}$ )	R-Value	Horizontal Diameter ( ∘ )	Vertical Diameter ( ∘ )	Classification Error (%)
1	$2.298$	$0.946$	$53.16$	$55.53$	$26.50$
2	$2.752$	$0.928$	$94.47$	$75.12$	$9.50$
3	$1.941$	$0.926$	$7.10$	$77.27$	$16.00$
4	$2.346$	$0.877$	$20.88$	$92.34$	$15.00$
5	$2.221$	$0.767$	$19.76$	$49.00$	$13.00$

TABLE II: natural eye-head characteristics measured in the deadzone model justification experiment

Fig. 7: Natural characteristics for participant 4. (left) Deadzone best-fit ellipse. The squares and dots respectively represent the non head moving and head moving targets evenly distributed across the screen. Target locations relative to the center of the screen are measured in inches. The red ellipse is the best-fit ellipse projected onto the screen. Its axis lengths are reported along with its gaze target classification error. (right) Velocity proportion regression. Each sample represents a single head moving gaze shift. The red line is the regression result whose slope is the velocity proportion.

Fig. 8: Difference of classification errors for the vector and region-based deadzones with mean errors of 16.81 and 40.44 percent respectively. A Welch’s t-test gives a p-value of 0.014 indicating the difference is significant.

	Deadzone Ellipse	Velocity Proportion Regression		Absolute Mean(std) Difference End Orientation
Participant	Performance Error (%)	Slope ( $\frac{\text{rad}/\text{s}}{\text{rad}}$ )	R-Value	Bending ( ∘ )		Flexion ( ∘ )		Rotation ( ∘ )
1	$4.0$	$1.801$	$0.958$	$3.963$	(3.757)	$2.484$	(2.764)	$7.451$	(6.926)
2	$14.5$	$1.980$	$0.971$	$4.869$	(4.317)	$1.626$	(1.534)	$4.762$	(4.831)
3	$3.5$	$2.045$	$0.926$	$6.501$	(4.065)	$5.408$	(4.655)	$11.183$	(13.426)
4	$4.0$	$2.355$	$0.965$	$6.726$	(4.482)	$3.987$	(3.937)	$8.750$	(7.627)
5	$5.5$	$2.005$	$0.924$	$5.544$	(3.920)	$6.258$	(5.910)	$9.098$	(10.685)

TABLE III: performance measures of the exoskeleton from the second part of the experiment

Fig. 9: Gaze target classification and robot actuation in the world view camera for participant 4 during the second part of the experiment. Gaze targets are no longer in a grid due to errors such as camera distortion. Target locations in the world camera are measured in pixels. The red ellipse is the natural deadzone projected onto the world camera view. Its axis lengths are reported along with its respective error. (left) Controller classification of gaze targets based on relative position of the deadzone in software. Note that classification is perfect. (right) Actuation performance of the robot in correctly moving or not moving according to controller decisions. Note the robot incorrectly didn’t actuated for some gaze targets.

Fig. 10: Convergent tuning of parameters for participant 2. The participant explored till the 4th trial, then converged the parameters until the 12th trial. After convergence, the participant explored further. The decorators on the y-axis indicate the maximum tunable values. The deadzone parameters are tuned by the two degree-of-freedom Cartesian system and the velocity proportion parameter is tuned by the one degree-of-freedom system. There are 16 trials.

Fig. 11: Controller parameter values obtained from the sliders subtracted by the values obtained by the measurements from the first part of the experiment (Table II ). The difference of velocity proportion are shown on the left (

v_{p}

). The deadzone horizontal and vertical diameters are shown on the right (

\theta_{dx},\theta_{dy}

) . The mean absolute values from left to right are 0.9 radians per second per radian, 44 ∘ , and 46 ∘ .

Assistive Robotics

Assistive robotics develops robots and wearable devices that support mobility, daily activities, rehabilitation, and care.

See all 199 papers

A Multi-Layer Sim-to-Real Framework for Gaze-Driven Assistive Neck Exoskeletons
Mar 6, 2026Colin Rubow, Eric Brewer, Ian Bales +2Assistive Robotics Robotic Control
Dept. of Mechanical Engineering, University of Utah, UT 84112, USA. · Robotics Center, University of Utah, UT 84112, USA. · Kahlert School of Computing, University of Utah, UT 84112, USA.
Fabric Pneumatic Artificial Muscle-Based Head-Neck Exosuit: Design, Modeling, and Evaluation
Mar 13, 2026Katalin Schäffer, Ian Bales, Haohan Zhang +1Soft Pneumatic Actuators Assistive Robotics
Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN 46556, USA. · Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, 1083 Budapest, Hungary. · Department of Mechanical Engineering and Robotics Center, University of Utah, Salt Lake City, UT 84112 USA.
Predictive Gaze Is Preserved but Reorganized toward Monitoring during Robot-Mediated Manipulation
Jun 20, 2026Manuela Uliano, Silvia Fattorini, Marco ControzziRobot Teleoperation Human-Robot Interaction
The BioRobotics Institute, Scuola Superiore Sant’Anna; Pisa, Italy · Department of Excellence in Robotics and AI, Scuola Superiore Sant’Anna; Pisa, Italy

Mar 6, 2026J/K move · Enter open · S save

A Multi-Layer Sim-to-Real Framework for Gaze-Driven Assistive Neck Exoskeletons

Colin Rubow, Eric Brewer, Ian Bales, Haohan Zhang+1

Dept. of Mechanical Engineering, University of Utah, UT 84112, USA. · Robotics Center, University of Utah, UT 84112, USA. · Kahlert School of Computing, University of Utah, UT 84112, USA.

Assistive Robotics Robotic Control Sim-to-Real Transfer Human Motion Prediction

Open paper AskPDF arXiv

Dropped head syndrome, caused by neck muscle weakness from neurological diseases, severely impairs an individual's ability to support and move their head, causing pain and making everyday tasks challenging. Our long-term goal is to develop an assistive powered neck exoskeleton that restores natural movement. However, predicting a user's intended head movement remains a key challenge. We leverage virtual reality (VR) to collect coupled eye and head movement data from healthy individuals to train models capable of predicting head movement based solely on eye gaze. We also propose a novel multi-layer controller selection framework, where head control strategies are evaluated across decreasing levels of abstraction -- from simulation and VR to a physical neck exoskeleton. This pipeline effectively rejects poor-performing controllers early, identifying two novel gaze-driven models that achieve strong performance when deployed on the physical exoskeleton. Our results reveal that no single controller is universally preferred, highlighting the necessity for personalization in gaze-driven assistive control. Our work demonstrates the utility of VR-based evaluation for accelerating the development of intuitive, safe, and personalized assistive robots.

Abstract

Assistive Robotics

Explore similar work

A Multi-Layer Sim-to-Real Framework for Gaze-Driven Assistive Neck Exoskeletons

Fabric Pneumatic Artificial Muscle-Based Head-Neck Exosuit: Design, Modeling, and Evaluation

Predictive Gaze Is Preserved but Reorganized toward Monitoring during Robot-Mediated Manipulation