A 3D-Printable Dataset for Fair Testing and Comparisons of Tactile Sensors Dexter R. Shepherd1,∗ , Nicolas Herzig2 , Phil Husbands1 , Andrew Philippides1 , Chris Johnson1 , William Kimbell1 1
arXiv:2606.25886v1 [cs.RO] 24 Jun 2026
2
University of Sussex, Department of Informatics University of Sussex, Department of Engineering 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.
1
Introduction
While texture classification is an important topic in robotics (1; 3; 4; 5) and used as a benchmark task (2), fairly comparing texture classification performance between different tactile sensors is difficult due to an inability to reliably reproduce the textures themselves. Existing approaches typically rely on self-collected datasets from available objects (6; 7), which limits the accuracy and fairness of comparisons with newly developed sensors. Indeed, most available texture datasets consist of sensor readings from specific tactile sensors (8; 9; 10), rather than the textures themselves. This limits their use to model optimisation or transfer learning, rather than allowing a direct comparison with other sensors. This paper fills this gap by designing a set of textures that can be 3D printed by researchers, and ensures it’s physical reporducinility by assessing the variability introduced by different printers and filament types. Standardised datasets have been used in other fields. 3D printed textures have been employed by Tymms et al. to measure human tactile perception (18). However, this task is aimed at investigating the resolution of human sensing and consists [say what it consists of]. Although it could make a good benchmark for determining the minimal resolution of tactile sensors, it does not evaluate different textures so we require something different for texture classification. In object classification 3D datasets of objects exist, which allows the objects to be imported into simulations (11). However, little work has been done with physical instantiations aside from random objects (12; 2), and Braille (13; 14; 17). The latter is a good example of a standardized and reproducible dataset, which is the goal of this paper. Following on from the ideas presented in (18; 13; 14) we propose a 3D printed dataset of textures that can be used to compare the performance of various tactile sensors. To do this we use equation generated texture blocks which can be 3D printed. We then use a TacTip to gather a visual dataset across different printer-filament combinations to determine print variance. Importantly, 3D printing quality varies between different printers and filaments, as well as being affected by other factors that are not always easy to control, such as humidity (15). We therefore
1
investigate how reproducible a dataset is across various popular printers and a range of filament types. Finally, we train texture classifiers on the resultant data to show the impact of dataset variation on classification. In so doing, this work paves the way for fair and reproducible tactile sensor comparisons. Furthermore, having digital copies of the real world textures could be used for sim-to-real transfer using matched digital and physical assets.
2
Methods
2.1
Creation of Textures
Various approaches were considered for generating textures in this dataset. Initially, we explored 3D scanning real-world surfaces to better reflect natural textures. However, this introduced significant variability, including non-uniformity within a single sample. As a result, different regions of the same texture could produce inconsistent sensor responses depending on the contact location, reducing reproducibility and making controlled comparisons difficult. To address this, we instead draw inspiration from the USAF calibration test (16), which employs well-defined geometric patterns that can be systematically scaled without loss of fidelity. By using mathematically generated patterns parameterised by spatial frequency, we ensure that each texture is both uniform and precisely controllable. This allows for consistent sampling across the surface, improves reproducibility, and enables more rigorous benchmarking of sensor performance across well-defined stimulus conditions. For this project, we produce 6 textures, shown in Figure 1. Each of the 2D surfaces are made from a summation of sin waves. As other repetitive shapes can be obtained from a sum of sine functions, finding the smallest variation in sine patterns that the sensor can detect may help benchmark the sensor for more advanced shapes. Each 2D surface is governed by 5 parameters which specify amplitude and frequency along each of the 2 axes as well as a phase shift between the directions. Note that units are not specified at this point which allows the datasets to be printed at different sizes. However, we have only validated one size (specified in section 2.2).
Textures One and Two In equation 1, A1 and A2 represent the amplitudes of the sine waves in the x- and y-directions, respectively. The parameters f1 and f2 denote the frequencies of the sine components along the x- and y-axes. The variable ϕ is a phase offset applied to the sine term involving y. The variables x and y are spatial inputs, and the output z1 is the sum of two sine functions modulated by their respective amplitudes, frequencies, and phase and gives the height of the surface. z1 = A1 sin(f1 x) + A2 sin(f2 y + ϕ) (1) We use A1 = 1, A2 = 0, f1 = 1, f2 = 1, ϕ = π/2 with x ∈ [0, 6π + 0.5] and y ∈ [0, 6π + 1.2]. ϕ makes no difference here, as the second part of the equation is multiplied by zero. Texture two also uses equation 1, with the same parameters for A1 , f1 , and f2 . However A2 = 1 introducing sinusoidal variation in both dimensions.
Texture Three We define the function z3 over a grid using the following summation: z3 =
N X
(Ai · sin(ix) + Ai · sin(iy))
(2)
i=1 i odd
In this expression, N is the number of terms in the summation, taken over odd integers from 1 to N . For our texture we used N = 25. Each term in the series uses the coefficient i−1
8 (−1) 2 · (3) π2 i2 which controls the amplitude of the sine wave. The sine functions sin(ix) and sin(iy) vary with spatial coordinates x and y, respectively, and share the same frequency f = i and amplitude Ai . The phase offset ϕ is set to zero and is therefore omitted in the formula. As z3 is built-up by adding sine wave components in both the x- and y-directions, the result is a 2D Fourier-series-like symmetrical pattern. Ai =
2
Texture Four Texture four uses the same equation (equation 2) as texture three, but with a modified coefficient 4 (4) πi The decay and lack of sign alternation lead to sharper features, with more pronounced ridges and edges. Ai =
Texture Five Texture five used the same coefficient as texture three (equation 3) for the x component but with the y-component coefficient set to 0 which leads to no variation in one axis, making the shape an extruded profile: z5 =
N X
(Ai · sin(ix))
(5)
i=1 i odd
Texture Six Our final texture uses the same equation as equation 5 but with the same coefficient as texture four (equation 4).
3
(a) Texture 1
(b) Texture 2
(c) Texture 3
(d) Texture 4
(e) Texture 5
(f) Texture 6
Figure 1: The 3D printable textures generated from the equations listed. We can see that there is a lot of overlap between these textures equations, making the generation of new patterns that are not widely different very possible.
2.2
Conversion to STL
All the models were converted to file type STL as blocks to be 3D printed. These models were all constrained to be the same maximum height of 7mm as we wanted to ensure the sensor can be lowered to the same position for each experiment. When generating the patterns listed in equations 1 - 5, we used a step size of 0.1 with converts to 0.53mm per step, with 189 points in the x-direction and 201 points in the y-direction so that the blocks are not perfectly square, giving blocks of dimension 23.3mm × 30mm. This was done to prevent human error and ensure the blocks only fit in the testing rig one way. These points form the vertices in the STL generation. The smaller the step size, the larger the STL file and higher resolution. As the physical resolution is determined by the nozzle aperture (listed in table 1), having a higher resolution in the STL yielded no benefits. Any slicer software will convert any minor inaccuracies to flat layers. The same resolution constraints do not apply to the resin printer. However, from prototype prints, a step size of 0.1 showed smooth accuracy upon visual inspection, so we did not need to increase resolution.
4
Figure 2: Example wireframe of the STL file for texture 1 generated by equation 1. We used step size 0.1, which gave more than high enough resolution given the constraints of nozzle size of the printers.
2.3
Evaluation of Materials and Print Quality
Several 3D print filament types were used to make a protruded sine-wave block. We first wanted to evaluate how different filament brands and types differ in terms of material properties. We additionally evaluated the textures produced by different 3D printers. If a cheaper filament, or printer, is significantly worse in quality than others then we need to advise against using this, or find combinations of print settings mitigating these issues. To assess variation, we used the variance of images from a TacTip sensor as pressing the sensor on the same part of the block should result in similar sensor readings. Visual inspection is also used to judge the quality of a print, particularly for ”stringing”. Stringing is when the printer’s nozzle leaks a small amount of molten filament while moving between separate parts or sections, leaving thin strands of plastic connecting them. If from a human perspective quality varies little, this will result in lower variance. If there is insignificant difference in variance between datasets then we can assume that the print quality makes no difference to the model.
2.4
Experimental Setup
A texture housing unit (seen in figure 4) was 3D printed to hold each of the texture blocks seen in figure 5. We used a TacTip tactile sensor (1) with 133 1mm optical markers. The testing rig pressed the sensor over each texture by pressing the sensor in a rectangular grid taking snapshots of the tactile sensation. The centre of the TacTip was placed at xy positions. The x position moves to 5 equally spaced positions between 4mm and 9mm, and the y axis moves to 5 equally-spaced 1mm positions between 18mm and 23mm, giving 25 points across the texture. This was repeated twice giving a dataset of 50 images. We further repeated this experiment with varying contact forces (similar to prior work (7)) at 0.08045±0.00195 N, 1.91295±0.00985 N, and 3.0137±0.0098 N. We specify a range as despite setting the pressure to the same values each time, the measured forces ranged through a margin of error from the soft body and measuring device. Therefore, each texture block had 150 images corresponding to it.
5
Figure 3: A diagram to show an example block, and the positions that the readings have been taken from in the 5 × 5 grid. Each red point is where the summit of the TacTip soft body will press.
Figure 4: Rig with new texture housing unit attached. Each texture slots into the unit, allowing for the sensor to be pressed. We used four printers, chosen for popularity and access: An Ender 3 printer; a Creality ender V3-SE printer,; a BambuLab P1P printer; a Formlabs Form 3 resin printer. The printer specifications are outlined in table 1. Resin printers are generally more expensive but yield better quality for small detail. Labs will have different printers to these, so replicating our methods will allow researchers to see whether their printer can produce a high enough standard.
6
Spec
Ender-3
Ender-3 SE
V3
Bambu P1P
Build volume
220×220×250 mm Cartesian Bed slingers ≤20,000 mm/s2 up to 300◦ C ≤100◦ C Direct drive
220×220×250 mm Cartesian Bed slingers ∼2,500 mm/s2
256×256×256 mm CoreXY
Formlabs Form 3 (Resin) 145×145× 185 mm na
20,000 mm/s2
na
up to 260◦ C ≤100◦ C Sprite direct drive 0.4 mm CR Touch + strain Touchscreen, SD/USB-C
up to 300◦ C ≤100◦ C Direct drive, all-metal 0.4 mm Auto (built-in)
na na na
App, slicer
USB
Motion system Acceleration Hotend temp Bed temp Extruder Extruder aperture Leveling Connectivity
0.4 mm Auto (varies by model) Touchscreen, SD/USB
cloud,
na na
Table 1: Comparison of Ender-3 V3, Creality Ender-3 V3 SE, Bambu Lab P1P printers and Formlabs Form 3 resin printer. It is worth noting that the printers all used 0.4 nozzles, though this can be reduced further. We used three different 1.75 mm filaments: Rapid PETG (Elegoo), PLA+ (eSUN), and PLALite (eSUN). For the resin printer we used Grey V4. According to manufacturer specifications, these filaments have a dimensional tolerance of ± 0.02 mm. We chose the PLA filaments as PLA is an economic standard printed across many 3D-printer users. The PLA-Lite (we refer to as PLA-) was cheaper than the PLA+, though the PLA+ was supposed to be higher quality so we chose to explore whether this had an impact. We used PETG as it provides durability but can be trickier to print with, potentially introducing more variance unless tuned properly.
7
3
Results
(a) Texture 1
(b) Texture 2
(c) Texture 3
(d) Texture 4
(e) Texture 5
(f) Texture 6
Figure 5: The 3D printable textures printed in PLA- on a Bambu printer.
8
3.1
Variability Across Printer and Filaments
(a) Resin
(b) Ender-3 pla+
(d) Ender-3 V3 SE PLA-
(c) Bambu PLA-
(e) eNDER-3 PLA+
3D textures printed on two printers 1
Texture
2
3
4
5
6
1
1
2
2
3
3
4
4
5
5
6
6
Test batch
(f) All the filaments
Figure 6: Examples of the prints. The black filament (close up shown in b & d ) shows the Ender printer using PLA+, whereas c & e shows shows the Bambu prints with PLA- which is grey. The cheaper printer (Ender) had a lot more stringing, and ridges were more visible. This undoubtedly led to higher variance. It is clear that there are no stringing effects on the resin printer (a), making it a high quality print. Finally we show an example of the Bambu and Eder V3 SE printed 6 times. When we inspect the varying qualities of the prints, some have more plastic webbing/string between blocks, especially with on prints produced with lower-cost PLA (seen in figure 6b&d where there
9
are thin plastic lines between the peaks), but these did not affect the surface of the block. More importantly, some of the peaks are more defined than other peaks on the same print. The sensor for dataset gathering typically comes from the top and presses down so stringing between peaks is unlikely to be picked up. However differences in peak heights and well defined summits will be. This was not a problem with the resin printer, which visually looked and felt the same. Upon visual inspection we saw that the Ender-3 was the poorest performing. It suffered the most from webbing, and inconsistent layer heights. We next selected a subset of printers for large-scale evaluation that spans the print-quality spectrum (as judged visuallY): the highest-performing filament printer (Bambu Lab) paired with the lower-cost PLA- filament, the lowest-performing filament printer (Ender-3 V3) paired with PLA+, and the resin printer as a consistently low-variance baseline. This selection enables detailed comparison of print variability across different levels of printer capability and material quality. To do this, we made 6 prints of each of 6 textures for each of the three selected printer-filament combinations. To assess the variability, we collected TacTip readings from the 25 positions shown in Figure 3, again taking 2 readings per position. For each position, texture and printer/filament, we calculate the variance of each pixel across the 6 x 2 samples. The average variance, across all pixels, positions, is shown for the selected printers and filaments and per texture in table 2. From the filament printers, Bambu printer led to the least variance between prints, making it the most suited out of the three chosen filament fed printers for this task. However, the resin printer had a lower variance than all the printers tested. These results backed up what we say visually. While differences did vary according to texture, apart from texture 1 the pattern was consistent. Note that there is also some variance from the TacTip. Baseline video recordings showed an average pixel variation of 2.6 counts even when the sensor was stationary on the same texture block, attributable mainly to small LED power fluctuations. Thus, some apparent differences between repeated measurements of the same block reflected background sensor noise rather than true texture differences.
Printer Class Ender-3 V3 SE PLA+ Bambu PLAResin
Texture 1
Texture 2
Texture 3
Texture 4
Texture 5
Texture 6
2.60
10.36
12.12
10.75
7.026
8.28
2.80
2.64
3.61
3.67
3.63
3.51
2.31
2.11
2.8
3.06
3.32
3.14
Table 2: Printer variability for three different qualities of printer-filament combination. Each value shows the average pixel variance rounded to 2 decimal places, averaged from 300 images for each texture.
10
4
Effect of printer variability on texture classification
Map of variance per texture averaged over trials and blocks 14
Texture 1
12
Texture 2
10 Texture 3
8
Texture 4
6
Texture 5
4 2
Texture 6
re 6
Tex tu
re 5
Tex tu
re 4
Tex tu
re 3
Tex tu
re 2
Tex tu
Tex tu
re 1
0
Figure 7: The variance between pixels for each texture, based on the Ender PLA- dataset, with averaged pixel variance across 6 prints of each texture. In the previous section we saw that different printers different in print variability. We next wanted to know whether this variance has an effect on classification? To get a first intuition, we on this from looking at the between texture variation for one of the worst performing printer to evaluate where similarities arise, Figure 7 displays the variance between each texture. As expected, texture 2 and 3 have low variance (as they are very similar in design). Texture 2 and 1 have some similarities, but not as close to texture 2 and 3. We then wanted a more direct test of the influence of variability on classification. To do this we trained classifiers to classify textures from different printer and filament combinations. To make sure our results were not due to a particular classifier, we trained both a random forest classifier and a feedforward artificial neural network (ANN) (models chosen following our previous work on tactile texture classification (7)). We used a random forest classifier with a max depth of 50 as they have been shown to work well with noisy data and on tactile tasks (7; 20; 21). The ANN using the same setup as existing TacTip classification (7) namely a single hidden layer of 50 and 6 outputs. To train the ANN we used cross entropy loss with the Adam optimizer at a learning rate of 0.001. Training took place over 1500 epochs. The dataset consisted of all six textures produced using all printer–filament combinations, resulting in a total of 8,100 images. To reduce memory requirements, all images were converted to greyscale and downsampled to 40% of their original resolution, a preprocessing step previously shown not to affect classification performance (7). We applied augmentation to the training datasets to account for LED noise and global lighting variation. Specifically, for each image we added Gaussian noise with mean 0 and standard deviation 5 to each pixel, making one new image. We also added/subtracted 10 to each pixels to ‘vary’ the lighting, resulting in 2 new images. This effectively quadrupled the data to 1200 images per filament-printer combo. After some preliminary investigation, we also decided to use PCA to convert the dataset to a smaller dimensionality thus removing noise and highlighting useful features for the classifiers. PCA was applied to the whole dataset to convert the images to 25 principal components (chosen based on preliminary results). The principal components were then used as input to both Random Forest Classifier and ANN. To assess the impact of manufacturing quality, each classifier was trained using data from a single printer-filament combination and then evaluated on data from all unseen printer-filament combinations. This allowed us to determine how well classifiers generalised to textures produced
11
under different manufacturing conditions. We trained the model on one of the filament printer types to begin with (Ender-3 V3 SE PLA-) as it was shown to be the configuration with the most variance from table 2. Each classifier had a 20% test split. Results are displayed in table 3. The random forest classifier does best with unseen print/filament configurations. Beyond the test dataset made up of the same print-filament batch, we then tested the trained models with the rest of the dataset to evaluate performance across each unseen filament and printer combination. Each classifier was trained on a specific batch of printer-filament combination and averaged over 20 trials, with this average accuracy shown in table 3. We found a model trained on one texture-printer combo did not transfer to the whole dataset. This is backed up by our previous results that show there is a higher variance between some of the filaments and textures (seen in figure 7). Next we trained a model only on the Bambu prints of PLA-, as it was the best performing filament-printer from the shortlisting. This gave us a dataset of size 2,050 images across the 6 textures. Accuracy increased as reflected in table 3. Finally we trained the resin prints, as the last shortlisting sample, and tested it across the same methods as previously, to find it leads to a much higher accuracy on unseen print types.
Model RFC Ender-3 V3 SE ANN Ender-3 V3 SE RFC Bambu only ANN Bambu only RFC Resin ANN Resin
Train accuracy 100.0% 99.7% 100% 100% 100% 100%
Test accuracy 99.53% 99.3% 99.7% 99.8% 99.3% 99.9 %
Unseen print accuracy 31.4% 33.21% 63.98% 53.73 % 70.66% 61.41 %
Table 3: The models, train accuracy, the average test accuracy for the textures/printers selected within the training, and average accuracy across untrained textures/printers objects. This was achieved by training on the one combination and testing on the unseen, and doing this for every single filament (and if not stated otherwise printer). We then average the results of the performance. When we look at confusion matrices of the random forest classifier trained on the Bambu dataset (which gave representative results, figure 8), we see much is being correctly classified. The pattern in incorrect classifications tend to revolve around certain textures. Texture 2 and 3 get misclassified, potentially because they are very similar. The same is true for textures 1 and 5 which are also very similar. See figure 5 for visual inspection of the texture patterns. Not only do textures 2 and 3 look similar, but they feel very similar from a human perspective. What distinguishes them could be errors in the print where one of the peaks is not as pointed as the others (as can be seen looking closely at figure 6). This would explain why the models do very well on the same print type, even on unseen images. Classifiers start to recognise the slight differences of various prints which aid in the classification task.
5
Conclusions
This paper introduces a method for generating geometric patterned 3D printed textures that can be reliably reproduced across different research teams, enabling fair and consistent comparisons between sensors in texture classification tasks. Our investigations highlight that both filament type and printer quality significantly influence the resulting texture fidelity. From the filament-fed printer models, PLA+ consistently produced higher-quality prints across tests, while lower-cost filaments exhibited increased stringing and reduced sharpness of fine features. Resin was best overall showing a lower variance between prints. To assess how varaiance affects classification accuracy, we evaluated calssification perofrmance across multiple printer–filament combinations. The resin printer consistently produced the lowest variance and most stable texture representations, leading to improved classification performance on unseen data. Among filament-based systems, higher-end printers such as the Bambu Lab performed competitively, while lower-cost systems such as the Ender-3 V3 exhibited greater variability. As a result, we recommend the resin printer as the preferred option for high-fidelity dataset generation, while acknowledging that the Bambu Lab paired with PLA provides a strong lower-cost alternative. Importantly, while we have only tested a limited number of printers, by introducing a variance-based evaluation method by which other prints can be compared to our dataset, we
12
Confusion Matrix Texture 1
62
13
0
0
24
1
Texture 2
0
20
74
2
1
3
70 60
True label
50 Texture 3
0
19
53
22
5
1 40
Texture 4
0
10
0
77
4
9
Texture 5
0
21
0
0
67
12
Texture 6
19
30
0
1
0
50
Texture 1 Texture 2 Texture 3 Texture 4 Texture 5 Texture 6 Predicted label
30 20 10 0
Figure 8: Confusion matrix for the random forest classifier trained on the Bambu dataset and evaluated on unseen printer–filament combinations, achieving an overall accuracy of 53%. Although performance is reduced by manufacturing variability, the model remains substantially above chance level (16.7% for six classes) and correctly identifies most texture classes. This suggests that the benchmark dataset contains texture features that are consistent across manufacturing conditions while remaining sufficiently challenging to assess classifier robustness and generalisation.
have introduced a repeatable method that can be applied to assess other printer and filament combinations. While this work did not explore slicer-level optimisation or printer calibration, such directions offer clear opportunities to further reduce variability, particularly for lower-cost systems. In addition, beyond the results presented here, the dataset can be faithfully reproduced in simulation, supporting broader benchmarking efforts. In summary, this study provides both a reproducible dataset and a structured evaluation framework for quantifying and comparing the effects of 3D printing parameters on texture consistency, which can support future work in both physical and simulated texture analysis. Acknowledgment. DS was funded by a be.AI Leverhulme Trust Doctoral Scholarship. We thank Nashil Sowaruth for his help with 3D printing models for us.
References [1] N. F. Lepora, Soft biomimetic optical tactile sensing with the TacTip: A review, IEEE Sensors Journal, vol. 21, no. 19, pp. 21131–21143, 2021. [2] P. Cavalin and L. S. Oliveira, A Review of Texture Classification Methods and Databases, in Proc. 30th SIBGRAPI Conf. Graphics, Patterns and Images Tutorials (SIBGRAPI-T), 2017, pp. 1–8, doi: 10.1109/SIBGRAPI-T.2017.10. [3] Ward-Cherrier, Benjamin and Pestell, Nicholas and Cramphorn, Luke and Winstone, Ben-
13
jamin and Giannaccini, Maria Elena and Rossiter, Jonathan and Lepora, Nathan F. 2018. The tactip family: Soft optical tactile sensors with 3d-printed biomimetic morphologies. Soft robotics. [4] Winstone, Benjamin and Griffiths, Gareth and Melhuish, Chris and Pipe, Tony and Rossiter, Jonathan. 2012. TACTIP—Tactile fingertip device, challenges in reduction of size to ready for robot hand integration. 2012 IEEE International Conference on Robotics and Biomimetics ROBIO. [5] James, Jasper Wollaston and Pestell, Nicholas and Lepora, Nathan F. 2018. Slip detection with a biomimetic tactile sensor. IEEE Robotics and Automation Letters. [6] M. Meribout, N. A. Takele, O. Derege, N. Rifiki, M. El Khalil, V. Tiwari, and J. Zhong, ”Tactile sensors: A review,” Measurement, p. 115332, 2024. [7] D. R. Shepherd, P. Husbands, A. Philippides, and C. Johnson, Texture and Friction Classification: Optical TacTip vs. Vibrational Piezoeletric and Accelerometer Tactile Sensors, Sensors, vol. 25, no. 16, art. 4971, 2025, doi: 10.3390/s25164971. [8] B. M. R. Lima, V. N. S. S. Danyamraju, T. E. A. de Oliveira, and V. P. da Fonseca, A multimodal tactile dataset for dynamic texture classification, Data in Brief, vol. 50, p. 109590, 2023. [9] M. Marzani, S. Khatibi, R. Masinjila, V. P. da Fonseca, and T. E. A. de Oliveira, A dataset for tactile textures on uneven surfaces collected using a BioIn-Tacto sensing module, Data in Brief, p. 111312, 2025. [10] S. Zhong, A. Albini, P. Maiolino, and I. Posner, TactGen: Tactile Sensory Data Generation via Zero-Shot Sim-to-Real Transfer, IEEE Transactions on Robotics, 2024. [11] R. Gao, Y.-Y. Chang, S. Mall, L. Fei-Fei, and J. Wu, Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations, arXiv preprint arXiv:2109.07991, 2021. [12] J. Cepriá-Bernal and A. Pérez-González, Dataset of tactile signatures of the human right hand in twenty-one activities of daily living using a high spatial resolution pressure sensor, Sensors, vol. 21, no. 8, p. 2594, 2021. [13] Z. Gao, L. Chang, B. Ren, J. Han, and J. Li, Enhanced braille recognition based on piezoresistive and piezoelectric dual-mode tactile sensors, Sensors and Actuators A: Physical, vol. 366, p. 115000, 2024. [14] X. Xu, N. F. Lepora, and B. Ward-Cherrier, A Neuromorphic Tactile System for Reliable Braille Reading in Noisy Environments, IEEE Robotics and Automation Letters, 2025. [15] L. Lendvai, I. Fekete, S. K. Jakab, G. Szarka, K. Verebélyi, and B. Iván, Influence of environmental humidity during filament storage on the structural and mechanical properties of material extrusion 3D-printed poly (lactic acid) parts, Results in Engineering, vol. 24, p. 103013, 2024. [16] Document Index for MIL-STD-150, archived Mar. 27, 2019, retrieved Aug. 10, 2019. Accessed https://web.archive.org/web/20190327102710/https://quicksearch.dla.mil/ qsDocDetails.aspx?ident_number=35531. [17] B. Ward-Cherrier, N. Pestell, and N. F. Lepora, NeuroTac: A neuromorphic optical tactile sensor applied to texture recognition, in Proc. IEEE Int. Conf. Robotics and Automation (ICRA), 2020, pp. 2654–2660. [18] C. Tymms, D. Zorin, and E. P. Gardner, Tactile perception of the roughness of 3D-printed textures, Journal of Neurophysiology, vol. 119, no. 3, pp. 862–876, 2018. [19] P. Cavalin and L. S. Oliveira, A Review of Texture Classification Methods and Databases, in Proc. 30th SIBGRAPI Conf. Graphics, Patterns and Images Tutorials (SIBGRAPI-T), 2017, pp. 1–8, doi: 10.1109/SIBGRAPI-T.2017.10. [20] B. M. R. Lima, V. P. da Fonseca, T. E. A. de Oliveira, Q. Zhu, and E. M. Petriu, Dynamic tactile exploration for texture classification using a miniaturized multi-modal tactile sensor and machine learning, in Proc. IEEE Int. Systems Conf. (SysCon), 2020, pp. 1–7.
14
[21] S. Pratap, J. Narayan, Y. Hatta, K. Ito, and S. M. Hazarika, From tactile signals to grasp classification: Exploring patterns with machine learning, in Proc. IEEE Int. Conf. Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI), vol. 2, 2024, pp. 1–6. [22] D. R. Shepherd, 3D Printable Tactile Dataset, University of Sussex, 2025. Available: https://doi.org/10.25377/sussex.30256453
15