Low-Cost Sensor Glove for Hand Kinematics

• Low-Cost Sensor Glove is a wearable device that uses resistive flex sensors and microcontrollers to capture hand kinematics at an affordable cost.
• It employs per-user calibration and simple normalization methods to accurately infer finger postures and support object shape discrimination.
• The hardware design, costing under $100, enables rapid prototyping and research in teleoperation, gesture recognition, and machine learning applications.

A low-cost sensor glove is a wearable device designed to capture human hand kinematics, contact, or muscle activity using affordable components and simplified calibration, targeting applications in human–object interaction, teleoperation, and gesture recognition. The approach leverages commodity resistive flex sensors, basic analog electronics, and per-user normalization to infer finger postures or classify grasped objects while maintaining a total bill of materials well below that of commercial data gloves. Recent systems implement an array of finger-mounted flex sensors coupled to microcontroller-based signal acquisition and statistical or machine learning pipelines for object discrimination. The central aim is to achieve robust, repeatable sensing of multi-finger grasp patterns and shape-dependent response at less than \$100 per device and with minimal dependence on per-user reconfiguration (Le et al., 2022).

1. Hardware Architecture and Cost Structure

The canonical form consists of five Spectra-Symbol resistive flex sensors (length 112.5 mm, \$10–12 each) adhered to the dorsal side of each finger. Their outputs are routed to the analog inputs (A0–A4) of an Arduino UNO R3 microcontroller (\$20–25) via individual 100 kΩ resistor dividers, providing the necessary voltage attenuation for the 10-bit (0–1023) analog-to-digital converter. The physical assembly employs off-the-shelf Lycra or knit gloves (\$5–10) as the mounting base, with optional 3D-printed enclosures or simple breadboard headers for cable management (\$5–10 combined).

The following table summarizes the components and costs (2022 values for typical suppliers):

Component	Unit Cost (\$) | Qty | Subtotal (\$)
Spectra-Symbol flex sensor (112.5mm)	10–12
Arduino UNO R3	20–25
100 kΩ resistor	0.05
Jumper wires, breadboard headers	0.50
Knit/glove base	5–10
Optional: Enclosure/PCB	5–10
Total (per glove)

The sensors are fixed and aligned along the dorsal axis of each finger; channel-to-port mapping is standardized (Thumb→A4, Index→A0, Middle→A1, Ring→A2, Pinky→A3).

2. Calibration Protocol and Signal Processing

Per-user calibration is mandatory due to inter-individual hand size and baseline offset variability. The system employs min–max normalization for each sensor:

1. Each user grasps objects with the smallest (e.g., 6 cm) and largest (e.g., 16 cm) known diameters.
2. For each finger, record ADCmin and ADCmax—these correspond to maximal and minimal bending.
3. Normalize any subsequent reading ADCraw for each channel by

$$\mathrm{normalizedADC} = \frac{\mathrm{ADC_{raw}} - \mathrm{ADC_{min}}}{\mathrm{ADC_{max}} - \mathrm{ADC_{min}}}$$

so that $\mathrm{normalizedADC} \in [0,1]$.

Assuming a near-linear sensor response, conversion from raw ADC to flexion angle $\theta$ (for a known max angle $\theta_{\max}$) is given by

$$\theta = \theta_{\max} \cdot \frac{\mathrm{ADC_{raw}}-\mathrm{ADC_{min}}}{\mathrm{ADC_{max}}-\mathrm{ADC_{min}}}$$

or, in operational terms,

$$\theta = \theta_{\max} \cdot \mathrm{normalizedADC}$$

No additional digital filtering is applied in practice beyond simple trial-averaging, and signal drift over an 8-minute window is limited to ±1 ADC count.

3. Data Acquisition and Statistical Analysis

Each flex channel is sampled at 20 Hz (50 ms interval) over a 5-second window, yielding 100 samples per finger per grasp. The Arduino UNO provides a 10-bit ADC resolution (≈4.9 mV/count).

Statistical evaluation uses, per object/finger/size combination, the average of 100 samples across $n$ users ($n=11$ for spheres, $n=8$ for cylinders). The standard error of the mean (SEM) is

$\mathrm{SEM} = \frac{\sigma}{\sqrt{n}}$

where $\sigma$ is the sample standard deviation of the mean flex readings.

Group-wise comparisons utilize 95% confidence intervals via

$$\mathrm{CI} = \bar{x} \pm t_{\alpha/2,\,n-1} \cdot \frac{\sigma}{\sqrt{n}}$$

with $t_{\alpha/2,\,n-1}\approx2.1$ for $n=8$–11, or $z=1.96$ for larger $n$.

Per-finger responses exhibit a systematic slope in normalized ADC versus diameter, but the gradient and intercept are both finger- and shape-specific. Spheres and cylinders of identical radii yield statistically separable signals (non-overlapping confidence intervals) on at least one finger except near 13 cm diameter, and all five channels respond differently to shape as object radii are varied.

For linear regression of normalized ADC against object diameter, ring finger $R^2$ values can exceed 0.95 across the full test span, though above 10cm diameter channel responses begin to saturate, degrading fit ($R^2<0.5$ in some fingers).

4. Object Classification Algorithms and Metrics

While no explicit classifier is deployed, the system is constructed to support downstream classification by feeding five-dimensional normalized flexion vectors $[f_1,\dots,f_5]$ into a supervised model (such as SVM, decision tree, or small neural network) for online detection of object shape (sphere versus cylinder) and dimension.

Empirical separability is quantified by the non-overlap of 95% confidence intervals, which act as a proxy for classification performance approaching 100% at most tested diameters. However, neither confusion matrices nor cross-validated accuracy rates are reported. Such procedural extensions would follow after collecting labeled vectors under controlled conditions and subjecting them to standard machine learning workflows.

5. Limitations and Prospective Improvements

Identified challenges include:

• Sensor nonlinearity and signal saturation below curvature radii of 12 cm and at object diameters above 10 cm.
• Substantial inter-user variation necessitating individual calibration.
• Gaps in data collection (e.g., no measurement for 10 cm cylinder), limited $n$ per shape instance.
• Restriction to static grasps; dynamical or partial grip patterns are not addressed and may exhibit higher noise.

Future development priorities include:

• Richer, multi-pose, or polynomial calibration to extend the usable curvature range.
• Fusion with complementary sensing modalities (pressure pads, IMUs) to improve robustness and resilience to saturation and user variation.
• Automated, cross-validated classification with larger user panels and full confusion matrices.
• Exploration of embedded, on-board machine learning classifiers (e.g., Arduino-side inferencing) for low-latency wearable deployment.

6. Implementation and Application Scope

The low-cost sensor glove in this configuration is specifically optimized for tasks such as real-time object shape and size discrimination, exploratory research in multi-fingered tactile patterning, and as an enabling front-end for machine learning-based hand-object interaction analysis. Its hardware simplicity, cost-effectiveness ($75–100 per unit), and ease of assembly make it suited as a research or prototyping platform where individual per-user calibration is tolerable and absolute pose accuracy is not the primary constraint. The primary limitation is reliance on resistive flex sensing, which constrains performance in highly dynamic tasks or in situations where complex motion artifacts or finger pressures nonlinearly influence output. Nevertheless, the architecture provides a reproducible basis for the development of statistical and machine learning models for multi-class object identification—or for rapid prototyping of more elaborate hand-wearable sensor systems (Le et al., 2022).