GuideTouch: Wearable Obstacle Avoidance

• GuideTouch is a wearable tactile navigation aid that detects head and chest obstacles using dual Time-of-Flight sensors and a four-point vibrotactile interface.
• It integrates sensor fusion on an ESP32 with a self-cleaning optical cover and an auditory drop alarm to deliver rapid, sub-120 ms feedback.
• User studies reported over 92% recognition accuracy for simple directional cues, confirming its role as a practical enhancement to traditional mobility aids.

Searching arXiv for “GuideTouch” and related tactile guidance work to ground the article in current papers. I’ll look up the named paper and closely related “GuideTouch”/touch-guidance entries on arXiv. GuideTouch is a compact, scarf-style, standalone wearable system for visually impaired individuals that detects obstacles at head and chest level and conveys their location through intuitively mapped vibrotactile cues. Introduced as an autonomous obstacle avoidance device, it combines two vertically aligned Time-of-Flight (ToF) sensors for three-dimensional environmental perception, four vibrotactile actuators distributed across the shoulders and upper chest, a centrifugal self-cleaning optical cover, and an auditory “find-me” alarm. Its stated purpose is to address the head-level detection gap left by traditional white canes while preserving ambient hearing through a silent haptic channel; the reported system delivers sub-120 ms obstacle alerts and primary directional cue recognition above 92% in the reported studies (Kozlov et al., 20 Jan 2026).

1. Problem setting and design objective

Safe navigation for visually impaired individuals remains a critical challenge, especially concerning head-level obstacles, which traditional mobility aids often fail to detect. GuideTouch is positioned specifically against that failure mode. The device is described as compact, affordable, standalone, and wearable, with an explicit focus on autonomous obstacle avoidance at head and chest level rather than generic environmental sensing (Kozlov et al., 20 Jan 2026).

The authors emphasize a scarf-style form factor and a 4-point vibrotactile interface located across the user’s shoulders and upper chest. This places the system in the category of body-worn haptic navigation aids, but with a narrower operational target: rapid localization of nearby obstacles in a small forward volume. The reported emphasis on directional proximity cues, rather than full scene reconstruction or speech-based assistance, indicates a deliberately low-bandwidth but high-immediacy interaction model.

A common misconception would be to read GuideTouch as a replacement for existing mobility aids. The paper does not make that claim. Instead, it presents the device as a practical enhancement to existing mobility aids, aimed at improving safety and independence by filling the “head-level” detection gap (Kozlov et al., 20 Jan 2026).

2. Hardware architecture and physical implementation

GuideTouch comprises four functional blocks: a sensor module with two vertically stacked ToF sensors, a centrifugal self-cleaning optical cover, four vibrotactile actuators embedded in a fabric scarf, and an auditory “find-me” alarm. All electronics and mechanics are mounted in two 250 g enclosures joined by a fabric scarf; total mass is under 500 g (Kozlov et al., 20 Jan 2026).

Block	Specification	Role
Sensor module	Two VL53L5CX ToF sensors on custom PCBs	3D obstacle detection
Haptic interface	Four coin-type vibrotactile motors at L1, L2, R1, R2	Directional cue delivery
Optical protection	Circular IR-transparent glass disk with centrifugal cleaning	Real-world robustness
Recovery aid	Piezo buzzer with contact clip trigger	Device location if dropped

The two VL53L5CX sensors implement multizone ranging on an $8\times 8$ grid. Their measurement principle is the standard ToF relation

$d = \frac{c \cdot \Delta t}{2},$

where $c \approx 3 \cdot 10^8\ \mathrm{m/s}$. Each zone covers a square field of view with 65° diagonal, and the two sensors are tilted 30° relative to one another to yield a combined vertical FoV of 90°. Reliable detection is reported for obstacles of at least 4 cm at up to 1 m, with zone size at distance $d$ given by

$$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$

using $\mathrm{FoV}=60^\circ$ and $N=8$ zones per side (Kozlov et al., 20 Jan 2026).

The vibrotactile interface uses four coin-type motors of diameter approximately 10 mm and thickness approximately 3 mm, placed at lower left ($L1$), upper left ($L2$), lower right ($R1$), and upper right ($d = \frac{c \cdot \Delta t}{2},$0). Their nominal operating frequency is 200–250 Hz, peak acceleration is 0.8–1.2 g, and each is driven at 3.3 V with current approximately 80 mA. This actuator placement directly encodes the device’s spatial discretization: left versus right and high versus low.

The optical self-cleaning cover is a distinctive mechanical element. It is a circular IR-transparent glass disk driven by a small brushless DC motor and belt transmission at $d = \frac{c \cdot \Delta t}{2},$1, corresponding to $d = \frac{c \cdot \Delta t}{2},$2. With radius $d = \frac{c \cdot \Delta t}{2},$3, the centrifugal acceleration at the rim is approximately $d = \frac{c \cdot \Delta t}{2},$4. The design condition is that the radial force $d = \frac{c \cdot \Delta t}{2},$5 exceed the surface-tension force $d = \frac{c \cdot \Delta t}{2},$6, with $d = \frac{c \cdot \Delta t}{2},$7 for water. The motor produces approximately 5 mNm, which the authors state is ample to maintain 3000 rpm and fling off droplets in less than 1 s (Kozlov et al., 20 Jan 2026).

The alarm subsystem uses a passive piezo buzzer driven by an NE555 timer at approximately 3.5 kHz. A clothing-mounted contact clip closes the circuit if the device falls off, triggering the buzzer. Control and data acquisition are handled by an ESP32 microcontroller over I$d = \frac{c \cdot \Delta t}{2},$8C and PWM. Power is provided by a 7.4 V Li-ion pack with 12 h endurance, or 4 h with continuous cleaning engaged. The inclusion of both the cleaner and the drop alarm suggests an explicit concern with field robustness rather than laboratory-only demonstration.

3. Perception-to-haptic pipeline

The GuideTouch processing loop runs at 10 Hz, or every 0.1 s, on the ESP32. Two $d = \frac{c \cdot \Delta t}{2},$9 distance matrices, $c \approx 3 \cdot 10^8\ \mathrm{m/s}$0 and $c \approx 3 \cdot 10^8\ \mathrm{m/s}$1, are acquired from the upward- and downward-tilted sensors and fused into a single $c \approx 3 \cdot 10^8\ \mathrm{m/s}$2 matrix $c \approx 3 \cdot 10^8\ \mathrm{m/s}$3 by vertical stacking. Temporal outlier filtering is then applied cellwise using

$c \approx 3 \cdot 10^8\ \mathrm{m/s}$4

to suppress spurious spikes (Kozlov et al., 20 Jan 2026).

Obstacle detection is performed by thresholding: cells with $c \approx 3 \cdot 10^8\ \mathrm{m/s}$5, where $c \approx 3 \cdot 10^8\ \mathrm{m/s}$6, are marked as obstacle cells. The resulting $c \approx 3 \cdot 10^8\ \mathrm{m/s}$7 grid is partitioned into four coarse regions, Left/Right $c \approx 3 \cdot 10^8\ \mathrm{m/s}$8 High/Low, which map directly to $c \approx 3 \cdot 10^8\ \mathrm{m/s}$9, $d$0, $d$1, and $d$2. In this sense, the system intentionally compresses 3D perception into a four-channel body-centric haptic code.

The haptic command is represented as a 4-vector

$d$3

For each region $d$4, the controller sets $d$5 if any cell in that zone satisfies $d$6, and $d$7 otherwise. Single-motor patterns encode isolated directional cues, such as $d$8, while double-motor patterns encode composite cues, such as $d$9. The paper defines the pattern set as

$$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$0

This mapping is deliberately coarse. It does not attempt continuous angular rendering or graded distance display; instead, it prioritizes rapid categorical recognition.

The real-time performance figures are correspondingly tight. Acquisition, filtering, zone analysis, and motor update complete in less than 50 ms, total loop time is approximately 100 ms, and the vibromotors respond within approximately 20 ms to PWM drive changes. Reported end-to-end latency from obstacle entry into the FoV to final vibration is less than 120 ms, which the authors state ensures perceptually continuous feedback at human walking speeds (Kozlov et al., 20 Jan 2026).

4. Experimental evaluation and statistical results

The evaluation comprises a sighted-user study focused on vibrotactile pattern recognition and preliminary trials with visually impaired users. The sighted-user study included 22 participants, 17 male and 5 female, aged 21–48 years with mean $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$1 and standard deviation $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$2. These participants were divided into Group A, which received all 15 patterns spanning single-, double-, and triple-motor activations, and Group B, which received 10 patterns restricted to single and double activations. Each pattern was played five times per participant. The preliminary visually impaired trials involved 14 participants, 8 male and 6 female, aged 16–60, during workplace visits (Kozlov et al., 20 Jan 2026).

Recognition accuracy per participant was defined as

$$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$3

The analysis used one-way repeated-measures ANOVA, with

$$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$4

and post-hoc Bonferroni and Tukey HSD comparisons. Confidence intervals for a proportion $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$5 were computed as

$$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$6

The reported quantitative outcomes separate sharply by pattern complexity. Group B, which used only single and double patterns, achieved average accuracy of 92.9%, with no significant variance across patterns: $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$7, $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$8. Group A, which included all patterns, achieved average accuracy of 78.4%, with a significant difference between simple 1–2 motor patterns and complex 3–4 motor patterns: $$s = 2d \cdot \tan\!\left(\frac{(\mathrm{FoV}/2)\cdot(\pi/180)}{N}\right),$$9, $\mathrm{FoV}=60^\circ$0. In the visually impaired cohort, average recognition for single and double cues was 93.75% (Kozlov et al., 20 Jan 2026).

The confusion matrices are summarized qualitatively in the paper: most errors occur when users confuse multi-motor combinations with overlapping single or double patterns, while simple directional cues remain highly reliable, with greater than 95% recognition for single motors. This is an important boundary condition on the system’s interface design. A plausible implication is that the four-actuator layout is effective for primary spatial cueing precisely because the control vocabulary is intentionally limited. The study does not support the claim that arbitrarily rich combinatorial haptic alphabets are equally usable in this form factor.

5. Safety claims, deployment characteristics, and limitations

The authors state that GuideTouch successfully fills the “head-level” detection gap left by traditional white canes. They further argue that the high recognition accuracy of directional cues, reported as greater than 92%, enables users to instantaneously sense and avoid obstacles. In pilot blind-user trials, participants reported increased confidence when walking under overhangs and near street signage. The vibrotactile channel is also described as silent, preserving ambient hearing and thereby further enhancing safety (Kozlov et al., 20 Jan 2026).

Several deployment-oriented characteristics are explicit. The device is standalone, under 500 g, and approximately USD 100 in cost. It includes a self-cleaning optical cover for droplets and a buzzer that activates if the device falls off. These elements indicate that the design goal extends beyond sensing performance to practical daily use. The scarf-based placement also avoids occupying the hands, which is consistent with coexistence alongside a white cane or other mobility aid.

The limitations are also explicit and technically significant. Current tests were static and involved first-time users. Accuracy for complex patterns is lower, approximately 65–80%, and the authors hypothesize that multi-day training and improved physical contact will raise performance. Planned future work includes dynamic navigation trials in real environments and crowds, multi-day user training, integration of a higher-resolution ToF or hybrid CV module for pedestrian and crowd detection, and ergonomic refinements to ensure consistent motor-skin coupling across body types (Kozlov et al., 20 Jan 2026).

These limitations delimit what has and has not been demonstrated. The paper supports high recognition of simple directional cues and preliminary validation with visually impaired users; it does not yet establish long-term adaptation, crowd navigation performance, or robustness under continuous real-world locomotion.

6. Terminological ambiguity and relation to adjacent tactile-guidance research

The name “GuideTouch” is not unique across the broader touch-related literature. The obstacle-avoidance wearable introduced in “GuideTouch: An Obstacle Avoidance Device for Visually Impaired” is a specific scarf-style assistive system for head- and chest-level obstacle alerts (Kozlov et al., 20 Jan 2026). This should not be conflated with unrelated uses of similar terminology in other domains.

One nearby source of ambiguity is “TOUCH: Text-guided Controllable Generation of Free-Form Hand-Object Interactions,” whose implementation notes describe a “GuideTouch” as a three-stage text-driven framework for generating 3D hand-object interactions. That usage concerns explicit contact modeling, diffusion-based pose synthesis, and refinement for free-form HOI generation, not assistive obstacle avoidance (Han et al., 16 Oct 2025). Another is “Touch-Augmented Gaussian Splatting for Enhanced 3D Scene Reconstruction,” whose exposition refers to a GuideTouch-style tactile-guided online reconstruction pipeline that interleaves vision and touch for 3D Gaussian Splatting; again, this is a reconstruction framework rather than a wearable aid (Gao et al., 11 Aug 2025).

A broader neighboring line is robotic touch-guided control. “TouchGuide: Inference-Time Steering of Visuomotor Policies via Touch Guidance” steers pre-trained diffusion or flow-matching visuomotor policies using a Contact Physical Model at inference time, within a low-dimensional action space. Its problem setting is fine-grained contact-rich manipulation, not human navigation assistance (Zhang et al., 28 Jan 2026). Earlier haptic guidance work such as “Three-dimensional hand guidance by midair haptic display” demonstrated guidance of a participant’s hand to the goal in a 30 cm cube workspace with an error of 64.34 mm using a virtual cone rendered by ultrasonic midair stimulation; this establishes a different tactile-guidance modality, non-contact and waypoint-oriented, rather than body-worn obstacle localization (Hiura et al., 2023).

This suggests that GuideTouch, in the assistive-wearable sense, belongs to a broader family of touch-guidance systems but occupies a distinct niche: low-latency, body-centric obstacle cueing for visually impaired navigation. Its contribution is not large-scale scene understanding, robotic contact planning, or midair waypoint rendering. It is the integration of coarse 3D ToF sensing, simple directional vibrotactile coding, and field-oriented robustness features into a lightweight standalone mobility aid.