The motivation comes from a long-standing challenge in motor rehabilitation. Children with Developmental Coordination Disorder, a neurodevelopmental condition affecting roughly six percent of children, often struggle to convert sensory input into accurate motor responses, and conventional interventions combining occupational therapy, physiotherapy, and psychological support can be difficult to individualize. The authors argue that continuous, graded feedback—rather than discrete right-or-wrong signals—can support exploratory behavior and incremental self-correction, because learners can interpret the strength of a cue to adjust their direction and distance iteratively. OmniTrack operationalizes that idea with compact IoT modules that can be worn or attached to objects, each capable of delivering light, sound, and vibration.

Positioning is the technical heart of the system. Each mobile tag carries a Qorvo DW3000-series ultra-wideband (UWB) module that performs two-way time-of-flight ranging against fixed anchors placed at known coordinates, and a least-squares multilateration algorithm, solved iteratively with the Levenberg–Marquardt method, converts the distance estimates into a position. A BNO055 nine-axis inertial measurement unit supplies orientation, an ESP32-S3 microcontroller manages the electronics, and data flows over Bluetooth Low Energy into a ROS 2 processing pipeline that decides what feedback to send back to the actuators. A programmable Feedback Control System modulates the intensity and modality of the cues according to the user’s proximity to the target and to therapist-defined settings, so that vibration strength, LED brightness, or audio frequency change as the user moves through distance-based feedback zones.

Before any of that could work reliably, the team had to confront a sobering property of the raw ranging data. In a one-dimensional calibration experiment conducted in an indoor room under line-of-sight conditions, with one anchor and one tag held at fixed height and orientation, the researchers collected one thousand samples at each of a series of ground-truth separations. The raw UWB measurements showed a systematic, distance-dependent bias that grew with range—for example, a target at one metre was reported at nearly 1.9 metres, and a target at ten metres at more than eleven metres—making uncorrected ranging unusable for localization.

To correct this bias without overfitting, the team compared polynomial calibration models of degree one through five using leave-one-out cross-validation, in which each averaged calibration point is excluded from the fit and then predicted. The third-degree polynomial achieved the lowest cross-validated mean absolute error, 0.266 metres, and the lowest cross-validated root mean squared error, 0.373 metres, and also had the best Akaike information criterion, while the linear model had the lowest Bayesian information criterion but worse predictive performance. The third-degree model was therefore selected as the calibration mapping applied to raw ranges before localization. A first-order exponential smoothing filter was then applied to stabilize short-term fluctuations, with a smoothing factor chosen as a compromise between noise attenuation and responsiveness; the equivalent time constant of that filter depends directly on the Bluetooth update rate at which the system operates.

The second validation stage scaled up to a full localization pipeline. Three calibration strategies were compared: the one-dimensional third-degree correction carried over from the first experiment, a single global polynomial fitted to all anchors together, and an independent per-anchor polynomial selected by grouped cross-validation. The global model settled on a second-degree polynomial, while the per-anchor strategy selected degree one for five anchors, degree two for one anchor, and degree three for two anchors, reflecting genuinely different ranging behavior among anchors installed at different positions and heights.

The comparison showed that calibration strategy matters. Room-specific and per-anchor calibration improved multilateration performance relative to the general one-dimensional correction, and the per-anchor approach improved the consistency of the pipeline overall. But the residual error was not spatially uniform: the central region of the monitored area generally showed horizontal positioning errors below roughly 0.10 metres, while errors grew near the boundaries, where the tag is constrained by fewer surrounding anchors. The largest local error appeared near one boundary position, and the signed error components indicated a position-dependent directional bias rather than purely random noise. The authors attribute the pattern to a combination of factors—unfavorable anchor geometry at the edges, the tag’s low height relative to the anchors, floor and wall reflections, residual anchor-specific bias, and the physical footprint of the device itself—while cautioning that because these variables were not independently varied, no single cause can be isolated from the present data.

Timing received equally careful treatment, because delayed feedback can undermine motor learning, particularly for users with atypical sensory processing such as those with autism spectrum conditions or DCD. The team decomposed the sensing-to-actuation pipeline into UWB acquisition, BLE transmission, ROS 2 processing, command transmission, and actuator activation. When the exponential smoothing dynamics are included, the total timing window stretches to roughly 218 milliseconds or more at the chosen smoothing factor. The authors argue that for the hot–cold task this is acceptable, since feedback acts as a progressive proximity cue rather than an instantaneous corrective signal, but they note that faster movements, narrow feedback zones, or future discrete interaction modes would demand tighter timing and dedicated bench tests.

The physical prototype is designed for realistic sessions rather than laboratory demonstrations. It integrates the ESP32-S3 and UWB module, an addressable RGB LED ring, a piezo buzzer, the IMU, and a DRV2605L-driven vibration motor, all powered by a 1200 mAh lithium-polymer battery.

The team is explicit about what the study does and does not establish. This is a technical validation under controlled indoor conditions—fixed tag height, predominantly line-of-sight propagation, fixed device orientation—not a user study, and no claims are made yet about clinical efficacy. Future work will extend localization to full three dimensions, investigate orientation-aware calibration using IMU data, adapt the smoothing factor dynamically to the tag’s motion state, evaluate scalability with multiple moving tags, and, crucially, conduct structured usability tests with target users to assess the acceptability of the feedback modalities and to define effective hot–cold thresholds. As the authors conclude, OmniTrack validates its main technical components and identifies the calibration complexity, orientation sensitivity, latency, scalability, and usability questions that must be resolved before any clinical or educational deployment.

Subject of Research: Design and technical validation of a multimodal IoT platform using UWB indoor localization and proximity-based multimodal feedback for perceptual–motor training

Article Title: OmniTrack: design and preliminary validation of a multimodal IoT platform for indoor localization and proximity feedback

Article References: Tauro, G., Nicora, M. L., Vertechy, R., & Malosio, M. (2026). OmniTrack: design and preliminary validation of a multimodal IoT platform for indoor localization and proximity feedback. Frontiers in Communications and Networks, 7, Article 1829998. https://doi.org/10.3389/frcmn.2026.1829998

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1829998

Keywords: OmniTrack, UWB indoor localization, multimodal feedback, perceptual-motor training, Developmental Coordination Disorder, hot-cold paradigm, IoT wearable, multilateration, polynomial calibration, exponential smoothing, haptic feedback, ROS 2