NeckSense: Neck-Worn Robotics & Sensing

Updated 3 July 2026

NeckSense is a multidisciplinary framework that leverages the neck’s unique anatomical properties for haptic perception, real-time sensing, and control.
It integrates sensor fusion—from IMUs, EMG, bio-impedance to force sensors—to enable applications in posture recognition, silent speech, and object classification.
The technology supports diverse applications including health monitoring, ergonomic optimization, and immersive navigation through advanced computational models.

NeckSense refers to a diverse set of wearable and robotic systems built around the neck, drawing on the region’s rich anatomical, physiological, and biomechanical properties. These systems span several research domains: haptic perception via flexible manipulators, multimodal sensing for biomechanics and health, activity recognition via sensor fusion, speech decoding through neck-sited EMG, bio-impedance-based head pose estimation, and ergonomic or comfort modeling in situational contexts. Across these applications, the neck serves as a versatile substrate for both sensing and actuation, enabling novel approaches to perception, control, interaction, and health monitoring.

1. Flexible-Body Haptic Perception and Morphological Computation

The "NeckSense" flexible manipulator system models haptic perception mechanisms observed in the avian neck, particularly the ostrich. The platform is constructed as a serial chain of 17 revolute joints (C2–C18) operating in the sagittal plane, with a free elastic head joint. A tendon driven from a base-mounted motor (routed through guide bushings at each joint and anchored at C6) enables pecking motions; slackened, the neck exhibits passive viscoelastic dynamics. Joint "cartilages" use silicone elastomers (Dragon Skin 20, Ecoflex 00-30), and damping is emulated via friction brakes or rotary dashpots. Metal springs span adjacent joints to mimic the Ligamentum elasticum interlaminare, providing gravity compensation. Sensing includes magnetic absolute encoders (AS5600) for joint angles (17 DOF) and a two-axis force sensor at the beak tip.

The system is formulated dynamically as a cascade of rigid links with joint-wise linear viscoelasticity:

$I_i \ddot{\theta}_i + D_i \dot{\theta}_i + S_i (\theta_i - \theta_i^0) + \tau_{\text{friction},i} = \tau_{\text{tendon},i} + \tau_{\text{contact},i} + \tau_{\text{gravity},i}$

in which $I_i$ , $D_i$ , and $S_i$ denote inertia, damping, and stiffness parameters, respectively. The state-space update is given by $\dot{x} = F(x,u)$ , with $x$ comprising joint angles and velocities.

The platform employs a physical reservoir computing framework: time-series of joint angles and beak-tip reaction forces, sampled at 700–750 Hz, are concatenated and fed to a softmax regression readout trained via cross-entropy loss. This enables real-time classification of object softness, leveraging the high-dimensional, nonlinear morphodynamic response of the structure as a computational substrate. Peak classification accuracy reaches ≈90% in both 9-class (simulation) and 3-class (robot) tasks, with haptic memory (pre-impact classification) maintained at ~70–80%. Crucially, the structural heterogeneity (cranial-caudal gradient in stiffness/damping, gravity-compensating ligaments) enlarges the performance envelope for both separability and memory, demonstrating "morphological computation": dynamic feature extraction and filtering by the body itself, relegating learning to a light softmax readout (Nakano et al., 12 Apr 2025).

2. Multimodal Sensing for Posture and Strain Monitoring

NeckSense systems in the health and ergonomics domain utilize the neck’s proximity to relevant muscles and posture determinants. The "NeckCare" and related designs leverage off-the-shelf IMUs, dry electrodes, or acoustic sensors in headbands, necklaces, and hearables.

IMU/acoustic fusion: Systems such as NeckCare integrate 6-axis IMUs and ultrasonic acoustic ranging (using stereo hearable microphones and a laptop-mounted speaker) to classify postures (Neutral, FHP, Slight Bend, Severe Bend, Hunch). Features include mean/std/min/max of pitch ( $\theta$ ), displacement ( $x$ ), spectral characteristics, and acoustic distance (via ToF-based cross-correlation). A Random Forest classifier, trained on ~70 hours of data, achieves up to 99% accuracy with full-modal fusion (96% IMU-only), and millimeter-level distance estimation. Real-time feedback (visual popup or haptic buzz) is triggered when poor posture or close viewing is sustained (Chhaglani et al., 2024).
Integration of kinetics and kinematics: A smart neck-band positioned over the Sternum Jugular Notch employs a MetaWear CPRO IMU (accelerometer), with option for complementary kinetic data synthesized via OpenSim musculoskeletal modeling. Feature vectors combining accelerations, positions, and (simulated) hyoid muscle tendon forces achieve 100% classification of 9 neck postures with a Random Forest (Kumar et al., 2020).
Head tracker + surface EMG: The NeckCheck system synchronizes high-rate IMU and surface EMG from the upper trapezius to predict muscle strain using Random Forest regression. Model $R^2$ reaches 0.97, with real-time latency under 20 ms. Principal feature importances are pitch (0.549), roll (0.286), and yaw (0.165). Recommendations for robust, generalizable NeckSense implementations include sensor fusion of multi-site EMG, individualized calibration, and LSTM/CNN-based temporal fatigue modeling (Chhaglani et al., 17 Mar 2025).

3. Activity and Behavior Recognition with Wearable Neck Devices

The original "NeckSense" multi-sensor necklace is designed for free-living, all-day eating/activity recognition. Core sensor suite comprises a proximity and ambient light sensor (Vishay VCNL4040), a tri-axial IMU (Bosch BMI-0080), and a microcontroller with BLE and on-board SD storage. The system samples all channels at 20 Hz and computes derived signals: proximity $p(t)$ , ambient light $I_i$ 0, Lean-Forward Angle $I_i$ 1 (from IMU quaternion $I_i$ 2), and tri-axial accelerometer "energy" $I_i$ 3.

Chewing sequence segmentation employs prominence-based peak detection and $I_i$ 4-periodic subsequence algorithms (dynamic programming-based), with chew periodicities in $I_i$ 5 s. Rich feature sets (257 dimensions) are extracted per candidate window, encompassing time/frequency-domain statistics and meta-context. XGBoost classifiers trained via LOSOCV yield F1-scores of 81.6% (exploratory) and 77.1% (free-living) per-episode, outperforming previous single-modality approaches. Power analysis confirms >15.8 hours operational time on a 350 mAh Li-ion cell. Noted limitations include degraded recall in reclined/supine eating, low-light environments, or soft foods (Zhang et al., 2019).

4. Neck-Mounted Electrophysiological Interfaces

Recent research exploits the neck as a substrate for wearable electrophysiology beyond the face. A necklace-form EMG-to-speech interface arranges ten gold-plated, dry electrodes in a 360° ring with two midline anchors. 13 differential channels (10 neck, 3 face) are sampled at 1 kHz, bandpass filtered (20–450 Hz), and processed into both time-domain statistics and STFT-based spectrogram features. A Random Forest classifier (100 trees, max depth 32) achieves 92.7% accuracy (neck only) for 11-word discrimination, with minimal decrement relative to additional facial electrodes or benchmark face-only configurations. Ablation shows ≥8–10 neck electrodes are needed for ≥92% accuracy. Speech-EMG linear mapping (to WavLM embeddings) reveals 33.1% of EMG spectrogram bins correlate with speech representations at $I_i$ 6 (Wu et al., 2024).

The EMG necklace platform is applicable for silent-speech communication, assistive devices for laryngectomy or dysarthria, and hands-free or privacy-preserving controls, contingent on integration of on-device DSP, real-time streaming, and deep network classifiers.

5. Bio-Impedance-Based Head and Neck Pose Tracking

A line-of-sight-free, compact head pose tracker utilizes bio-impedance changes from tissue displacement and muscle activity around the neck. NeckSense (bio-impedance variant) mounts five soft, textile-based electrodes on a collar. Four voltages are sensed (AD5941 AFE, 100 μA at 100 kHz), yielding complex impedance measurements offset-corrected against a neutral-pose baseline. The eight-dimensional feature vector (four magnitude, four phase) is low-pass filtered and windowed for model input.

The transformer-based Imp2Head architecture processes 90-frame (input) windows and generates 10-frame sequences of SMPL-compatible neck/head/jaw axis-angle rotations. Anatomical priors are enforced via range-constraint loss. Quantitative leave-one-person-out cross-validation demonstrates a mean per-vertex error (MPVE) of 25.9 mm (corrected for ground-truth estimator variance), matching current state-of-the-art vision methods and outperforming IMU-based alternatives. The approach does not pursue EIT but models the neck as a lumped impedance sensitive to major head kinematic axes (Liu et al., 17 Jul 2025).

6. Sensorimotor Control, Modeling, and Ergonomic Optimization

Computational models of neck-head dynamics and ergonomic strain have emerged for VR/AR and automated vehicle applications:

Ergonomics in VR/AR: Surface EMG from sternocleidomastoid and splenius is modeled alongside head pose to train CNN-MLP cascades (MCLNet), learning to map pose and derived accelerations to instantaneous muscle contraction levels (MCL). Pose-to-trajectory regressors (TrajectoryNet) provide pre-hoc strain predictions based only on planned endpoints. NRMSE for after-movement MCL estimation is $I_i$ 7, and user studies confirm the model’s predictive validity for discomfort, enabling ergonomic content layout with observable comfort gains (Zhang et al., 2023).
Head-neck postural stabilization: In automated vehicle contexts, the 2-segment, 5-DOF (lower/upper neck: roll, pitch, yaw) head-neck system is modeled in state-space with muscle torques ( $I_i$ 8) and sensory perturbation ( $I_i$ 9). A finite-horizon MPC controller minimizes muscle effort and angular-rate somatosensory weights, replicating the CNS’s partial feedback stabilization under unpredictable (e.g., lateral) perturbations. Integrators (HiS, HoT) are eschewed for dynamic realism; real-time factors reach 8–11 $D_i$ 0 on multicore CPUs. Validation confirms alignment with human time/frequency response (Messiou et al., 30 Jul 2025).

Hapbeat-driven NeckSense systems use mechanical vibration actuated by modulated musical waveforms across both sides of the neck for 2D navigation and affective experience. Two DC-motor/tensioned-ribbon actuators provide haptic cues: balance encodes horizontal direction, intensity encodes proximity to targets; all cues are achieved by amplitude modulation of the musical carrier. Experimental results show mean angular localization error ≈20° and >80% shortest-path achievement in VR environments, with no interference to music enjoyment and increased “groove” and immersion compared to audio-only or wrist/waist haptic cues (Yamazaki et al., 2022).

These approaches highlight the neck's ability to host haptic informational channels for navigation, feedback, and emotion, supporting immersive, eyes-free, and multi-modal scenarios without burdening the user’s auditory or visual bandwidth.

By synthesizing advances across haptic robotics, multi-modal sensing, electrophysiology, bio-impedance, and computational modeling, NeckSense establishes the neck as an effective anatomical locus for both perception and interaction. It enables robust real-time object classification, posture/strain feedback, activity tracking, silent speech decoding, and ergonomic modeling, with applications in assistive technology, human-machine interaction, nutrition science, and user-comfort optimization (Nakano et al., 12 Apr 2025, Chhaglani et al., 2024, Kumar et al., 2020, Chhaglani et al., 17 Mar 2025, Yamazaki et al., 2022, Wu et al., 2024, Zhang et al., 2023, Zhang et al., 2019, Liu et al., 17 Jul 2025, Messiou et al., 30 Jul 2025).