A 3D-Printable Dataset for Fair Testing and Comparisons of Tactile Sensors

Abstract: Existing texture datasets for tactile sensing primarily consist of sensor readings from a specific sensor interacting with available surfaces/objects rather than describing the textures themselves, limiting fair comparison between tactile sensors and hindering reproducible research. In this work, we introduce a 3D-printable dataset of mathematically defined textures designed to be fabricated reliably across different printers and filament types. The dataset consists of six parametrically generated surface patterns derived from combinations of sine-wave and Fourier-based functions, giving controlled variation in spatial frequency, amplitude, and directional structure. We evaluate the reproducibility of these textures across three popular 3D printers and multiple filament types by measuring variance in images captured using an optical TacTip sensor under controlled contact conditions. Our results show that print quality, particularly peak sharpness and stringing, affects tactile variance, with higher-end printers producing significantly more consistent signatures. Classification experiments using neural networks and PCA-based models further demonstrate that high-quality prints support strong within-printer generalisation, while cross-printer generalisation remains challenging due to geometric inconsistencies. This work establishes the first openly available, physically reproducible 3D-printed texture benchmark, providing a foundation for fair comparison of tactile sensors.

• The paper introduces six mathematically defined, openly available 3D-printable textures that isolate physical stimuli from sensor-specific data for reproducible tactile benchmarking.
• The paper finds that resin and high-quality Bambu FDM prints produce the lowest variability, while printer quality affects consistency more than filament choice and sensor noise creates a 2.6-count baseline.
• The paper shows that classifiers trained on high-fidelity resin prints reach up to 70.66% accuracy on unseen prints, compared with about 31% after training on high-variance Ender-3 output, revealing the impact of manufacturing artifacts on generalization.

Motivation and contribution

Texture classification is a common benchmark task for tactile sensors, yet comparisons between sensors are frequently confounded by the stimuli themselves. Most existing texture datasets consist of sensor readings collected from a particular sensor interacting with whatever surfaces were available, rather than of the textures as independently specified physical objects (2606.25886). This means a classifier trained on one sensor's data cannot be meaningfully compared against another sensor's results, because the underlying stimulus conditions are neither identical nor reproducible across laboratories.

This paper addresses that gap by releasing an openly available dataset of six mathematically defined textures designed for 3D printing, together with an empirical characterization of how reliably those textures can be fabricated across different printers and filaments. The authors' central claim is that this constitutes the first openly available, physically reproducible 3D-printed texture benchmark suitable for fair comparison of tactile sensors. The design philosophy follows the USAF calibration test tradition: rather than scanning real-world surfaces (which the authors rejected after finding non-uniformity within single samples made contact location a confound), each texture is generated from parametric equations so that every point on the surface is known and uniform.

Texture design

The six textures are all derived from sums of sinusoids, parameterized by amplitude and frequency along two spatial axes plus a phase offset. Textures 1 and 2 use a two-term sine sum $z_1 = A_1 \sin(f_1 x) + A_2 \sin(f_2 y + \phi)$, with texture 1 being a unidirectional ripple ($A_2 = 0$) and texture 2 adding variation in both dimensions. Textures 3–6 use Fourier-series-like summations over odd harmonics up to $N = 25$, differing in coefficient decay and dimensionality: textures 3 and 4 are symmetric 2D patterns with alternating-sign ($8/\pi^2 i^2$) and non-alternating ($4/\pi i$) coefficients respectively, the latter producing sharper ridges; textures 5 and 6 are extruded 1D profiles using the same two coefficient families.

The deliberate overlap between the governing equations is presented as a feature: it allows new, progressively harder textures to be generated by small parameter changes, supporting difficulty scaling in future benchmarks. All blocks share a maximum height of 7 mm so the sensor can be lowered to a fixed position, and blocks are deliberately non-square (23.3 mm × 30 mm) to prevent orientation errors during mounting. STL resolution was set at a step size of 0.1 (0.53 mm per step), which the authors argue exceeds what any 0.4 mm nozzle can resolve, so higher mesh resolution would yield no benefit.

Measuring print reproducibility

Reproducibility was quantified by pressing an optical TacTip sensor (133 markers of 1 mm diameter) onto each texture at 25 positions on a 5×5 grid, twice per position, at three contact forces (approximately 0.08 N, 1.91 N, and 3.01 N), yielding 150 images per block. Pixel-wise variance across repeated prints served as the variability metric. Four printers were characterized—Ender-3, Ender-3 V3 SE, Bambu Lab P1P, and Formlabs Form 3 resin printer—with three FDM filaments (PETG, PLA+, PLA-Lite) plus Grey V4 resin.

Three findings stand out:

• Resin printing dominates. The Form 3 produced the lowest variance on nearly every texture (e.g., 2.11 average pixel variance for texture 2 versus 10.36 for the Ender-3 V3 SE), consistent with visual inspection showing no stringing artifacts.
• Printer tier matters more than filament choice within tiers. The Bambu P1P with cheaper PLA-Lite achieved variances of roughly 2.6–3.7 across textures, comparable to resin, while the Ender-3 V3 SE with PLA+ showed variances up to 12.12.
• Sensor noise sets a floor. Baseline recordings with a stationary TacTip showed an average pixel variation of 2.6 counts attributable mainly to LED power fluctuations. This is an important caveat: much of the measured variance for high-quality prints reflects sensor noise rather than geometric inconsistency, meaning the true manufacturing variance of resin and Bambu prints may be even lower than reported.

Visually, stringing between peaks was common on lower-cost printers but did not affect the top surface; inconsistent peak sharpness and layer heights were identified as the properties most likely to influence tactile readings.

Effect of manufacturing variability on classification

To connect print quality to downstream task performance, the authors trained a random forest classifier (max depth 50) and a feedforward ANN (one hidden layer of 50 units, Adam optimizer, cross-entropy loss) on TacTip images from all printer-filament combinations—8,100 images total, preprocessed via greyscale conversion, downsampling to 40% resolution, Gaussian noise augmentation, brightness perturbation, and PCA to 25 components.

Within-printer performance is uniformly strong: both classifiers reach approximately 99–100% train and test accuracy regardless of training batch. The decisive result concerns cross-manufacturing generalization:

Training set	Test accuracy	Unseen print accuracy
RFC, Ender-3 V3 SE	99.53%	31.4%
ANN, Ender-3 V3 SE	99.3%	33.21%
RFC, Bambu PLA-	99.7%	63.98%
ANN, Bambu PLA-	99.8%	53.73%
RFC, Resin	99.3%	70.66%
ANN, Resin	99.9%	61.41%

A model trained on the highest-variance printer transfers poorly (roughly chance-plus performance against a 16.7% chance level), while models trained on resin or Bambu prints transfer substantially better. The implication is direct: classifiers trained on low-fidelity prints learn print-specific artifacts rather than texture identity, so manufacturing quality directly bounds cross-lab generalization of tactile classification models. Confusion analysis shows residual errors concentrate on perceptually similar pairs—textures 2/3 and 1/5—which differ only subtly by design, suggesting the remaining errors reflect genuine stimulus similarity rather than random noise.

Limitations and open questions

The paper is candid about several constraints. Only one physical scale has been validated; because the texture equations carry no units, other sizes are printable but untested. The evaluation covers four printers and three filaments—a limited sample—though the variance-based protocol itself is offered as a repeatable method others can apply to their own equipment. Slicer-level optimization and printer calibration were not explored, leaving open whether tuned settings could close the gap for low-cost FDM systems. The TacTip's own ~2.6-count noise floor means the reported variance metric conflates sensor noise with manufacturing variance; separating these would strengthen the benchmark. Finally, the claim that digital copies enable sim-to-real transfer is asserted but not demonstrated experimentally here.

Conclusion

This work provides a parametrically defined, openly available set of 3D-printable textures together with a quantitative protocol for assessing fabrication fidelity via optical tactile variance. Its empirical results establish that printer quality—not filament cost—is the dominant factor in print consistency, that resin printing yields the most reproducible textures, and that classifiers trained on high-fidelity prints generalize far better across manufacturing conditions than those trained on low-cost FDM output. By decoupling the stimulus specification from any particular sensor, the dataset enables the fair, reproducible cross-sensor comparisons that existing reading-based datasets cannot support.