Robotic Perception with a Large Tactile-Vision-Language Model for Physical Property Inference

Gan Ma; Hengxiang Chen; Nutan Chen; Qiang Li; Qiusang Qiu; Xinheng Mai; Zexiang Guo; Zhanat Kappassov

REVIEW 4 major objections 7 minor 23 references

Robotic Perception with a Large Tactile-Vision-Language Model for Physical Property Inference

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fusing touch and vision ranks object properties better than one sense

desk verdict A plausible tactile-vision-language pipeline with real instrument measurements, but the central fusion claim rests on an unmatched, domain-shifted tactile baseline. read the letter →

arxiv 2506.19303 v1 pith:RAZIUMDX submitted 2025-06-24 cs.RO

Zexiang Guo , Hengxiang Chen , Xinheng Mai , Qiusang Qiu , Gan Ma , Zhanat Kappassov , Qiang Li , Nutan Chen This is my paper · ORCID

classification cs.RO

keywords tactileperceptionvisual-tactilefusionphysicalpropertyinferencemultimodalintegrationvision-languagemodelrobothardnessestimationroughness

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that a robot can estimate an object's hardness, elasticity, and roughness before committing to a grasp by feeding one camera image and one tactile image into the same large vision-language model. On 35 household objects, the fused model ranks all three properties in closer agreement with instrument measurements than models using only vision or only touch. The clearest gain is for roughness, where the model's scores reach a Spearman correlation of $\rho = 0.64$; hardness and elasticity reach $0.50$ and $0.53$. If these correlations hold, robots could choose safer grasps and adapt to unfamiliar materials from a single multimodal observation.

What carries the argument

The load-bearing mechanism is a feature-concatenation fusion path built around three pretrained components: CLIP (a contrastively trained image-text embedding model) provides the visual encoder; the Octopi tactile encoder (a tactile-language encoder pretrained on annotated tactile videos, also CLIP-based) turns a sequence of tactile images into temporal features; and Vicuna-7B (a 7-billion-parameter instruction-tuned language model) is the reasoning backbone. A LLaVA-trained linear projection maps the CLIP visual and tactile features into the language model's word-embedding space, where they are concatenated with text tokens between special boundary markers. Around this, a structured two-phase prompt first asks the model to identify the object from vision and then to score hardness, elasticity, and roughness on a 10-point scale with a material rationale. The paper credits the correlation gains to this unification of visual and tactile features in one embedding space plus property-specific prompting, since ablations that remove either modality, or switch to the baseline's original coarse three-level prompt, lose most of the correlation.

What would settle it

Re-measure the same 35 objects with a second independent protocol—for example, operator-administered Shore hardness tests, a different tensile sample geometry, and human perceptual ratings—and recompute the three Spearman correlations. If the fused model's advantage over vision-only shrinks or disappears when the analysis is restricted to textiles, foam, and paper, where the original instruments are least reliable, the reported correlations would reflect label noise rather than the benefit of combining vision and touch.

Watch

Extended reading notes

Core claim

The central claim is that cross-modal fusion—an object image, a tactile image, and a structured text prompt fed together through one language model—lets the model assign 1–10 scores for hardness, elasticity, and roughness that rank-order objects significantly better than either modality alone. In the paper's zero-shot evaluation, the full model obtains Spearman $\rho = 0.501$ for hardness ($p = 0.005$), $0.530$ for elasticity ($p = 0.003$), and $0.643$ for roughness ($p = 0.0001$) against instrument ground truth, while the vision-only version gives $0.307$, $0.452$, and $0.413$, and the tactile-only baseline stays near zero. Because elasticity is inversely related to elastic modulus, the paper reports the absolute value of the correlation. The paper reads these numbers as evidence that visual priors compensate for the limited contact area of tactile sensors and that property-specific prompting activates the language model's physical reasoning.

Load-bearing premise

The paper assumes the durometer, tensile tester, and surface-roughness meter give trustworthy hardness, elasticity, and roughness readings for all 35 objects, including soft textiles, foam, and paper; if those instruments are unreliable on compliant or fibrous materials, the reported correlations would not measure true physical properties.

Editorial extensions

If this is right

A robot can produce usable 1–10 scores for hardness, elasticity, and roughness from a single camera frame and one tactile image, before committing to a grasp force or strategy.
Roughness benefits most from combining vision and touch, which suggests surface texture is where the two modalities carry complementary information; hardness and elasticity gains are smaller but still statistically significant.
In this setup vision alone beats tactile alone, so adding a camera to an existing tactile-language model may be a cheaper path to better physical reasoning than collecting more tactile data.
Prompt design matters but is not sufficient: the tactile-only baseline with the refined prompt still scores near-zero correlations for elasticity and roughness, and only the fusion with vision produces large, significant correlations.
The model transfers zero-shot to a new sensor and a new object set, since the reported scores come from the pretrained model applied to the lab's 35 objects without per-object retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the instrument labels are trustworthy, the same recipe—concatenate a CLIP image embedding, a tactile encoder, and a structured scoring prompt—could be applied to other physical properties such as friction, thermal feel, or fragility by changing only the prompt and the ground-truth instrument.
The comparison in the paper mixes two changes at once: adding vision and moving to a new tactile sensor domain. A cleaner test of the fusion claim would retrain or adapt the tactile-only baseline on GelSight Mini images, so the tactile model is not handicapped by domain shift, and then re-run the three comparisons.
Because the rating scales are material-anchored (for example, cotton and sponge at 1–2 hardness), the model's scores are interpretable by a human operator, which may make the outputs usable as grasp-planning priors or as training labels for downstream policies.
A direct extension would be to record the model's predicted scores over repeated contacts on the same object and check whether the residual errors track measurement noise at different locations; if they do, the framework is capturing per-location material variation rather than just object identity.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know: this is a modest engineering contribution with a real data collection effort, but the paper's central claim—that fusing vision and touch beats either alone—is not actually supported by the experiments as reported.

What is new: the architecture combines Octopi's tactile encoder, CLIP ViT-L/14 for vision, LLaVA's projection, and Vicuna-7B with structured scoring prompts. The evaluation uses 35 physical objects with instrument ground truth (Shore hardness, elastic modulus, Ra roughness), which is more rigorous than the typical self-reported labels. They report Spearman correlations and p-values, and they explicitly acknowledge that the zero-shot Octopi baseline fails due to sensor domain shift—that honesty is to their credit.

The main problem is the comparison. The tactile-only baselines are used zero-shot with weights trained on a different GelSight sensor; the paper itself says this causes failure. So the near-zero Octopi correlations do not measure what a tactile-only model could do on this dataset after adaptation or fine-tuning. There is no matched ablation: they never run their own architecture with the tactile branch removed but fine-tuned the same way. The vision-only baseline is their own model, so the observed gain could come from fine-tuning, the visual branch, or the prompt, not from fusion. Also, the difference between full model and vision-only correlations is not tested for significance; the reported p-values are within-method, not between-method. Minor issues: single run, no confidence intervals, post hoc absolute value for elasticity, no code or data release, and the 35-object set spans nine material categories but with probably very few per category.

That said, the result that a model can achieve rho ~0.5-0.64 against instrument measurements is still useful evidence that VLMs with tactile images can rank physical properties. The paper is not misleading in its data, just overclaiming in its attribution.

Who is this for: robotics researchers working on multimodal perception and material property estimation. A serious referee could ask for a matched ablation (same architecture, tactile-only and vision-only, fine-tuned) and for a test of correlation difference, plus code/data. I'd want to see that before relying on the fusion claim.

Recommendation: send to peer review with requests for additional experiments; it's a worthwhile dataset and architecture but not yet proved as claimed.

Referee Report

4 major / 7 minor

Summary. The paper proposes Octopi-ViTaL, a multimodal vision-tactile-language model that concatenates CLIP visual features, Octopi-based tactile features, and text embeddings into a Vicuna-7B LLM, and uses a structured 10-point prompting protocol (Table 1) to score hardness, elasticity, and roughness from a single RGB image and a short tactile video. The authors evaluate on 35 objects across nine material categories, comparing Spearman correlations between model scores and instrument measurements (Shore hardness, elastic modulus, Ra roughness) against two Octopi variants and a vision-only version of their own model. The main reported result is that the full model achieves statistically significant correlations (rho = 0.501, 0.530, 0.643) that exceed the unimodal baselines.

Significance. If the fusion benefit were established, this would be a useful step toward applying vision-language models to robotic physical property inference, and the paper's introduction of a structured rating-scale prompt (Table 1) is a practically relevant contribution. The paper reports explicit correlation coefficients and p-values for all conditions, which make the evaluation transparent, and the use of real instrument measurements on a multi-material object set is a strength. However, the claimed advantage of multimodal fusion is currently not supported by a matched ablation, and the training and ground-truth validity issues would need to be resolved before the empirical claim can be credited.

major comments (4)

[Section 4.3, Table 2] The central claim that fusing vision and touch outperforms unimodal encoders is not supported by the reported comparisons because the tactile-only baseline (Octopi) is zero-shot and the paper itself attributes its failure to domain shift in sensor resolution, lighting, and calibration (Section 4.3). A matched ablation is missing: the proposed architecture or an equivalent fine-tuned model evaluated with tactile input only would be needed to rule out that the gain comes from adaptation, fine-tuning, or the visual branch alone. Please add such an ablation or soften the fusion attribution.
[Section 4.3, elasticity analysis] The reported elasticity correlation uses the absolute value of Spearman's rho (rho = 0.530, p = 0.003) on the grounds that elastic modulus is inversely proportional to perceived elasticity. Because this sign conversion is introduced post hoc, please report the signed correlation coefficient and the number of objects whose ranks moved in the expected versus opposite direction, so the reader can verify that the negative relationship actually holds. Without this, the absolute value could conceal inconsistency.
[Sections 4.1 and 4.2] The ground-truth instruments may not be valid for all nine material categories: Shore hardness (PosiTector SHD) is designed for rubber and plastics, the C610H tensile tester assumes samples with particular geometry, and the RUGOSURF 20 contact profilometer may be problematic for textiles and foam. If some ground-truth labels are unreliable or not commensurable across categories, the correlations in Table 2 do not measure the stated physical properties. Please report material-by-material measurements or justify the instruments for each category.
[Sections 3 and 4.2] The training procedure for Octopi-ViTaL is not described: there is no dataset, fine-tuning protocol, number of epochs, learning rate, or data split, despite Section 3.1 mentioning 'frozen during fine-tuning.' The paper calls the evaluation zero-shot, but the model is clearly not zero-shot with respect to its own training data, and the reader cannot determine whether the rating scale in Table 1 was learned or supplied. Please include full training details and clarify what 'zero-shot' means here.

minor comments (7)

[Section 4.3, Table 2] The p-values are not adjusted for multiple comparisons across the three attributes and four methods; if the authors wish to emphasize statistical significance, a multiple-comparison correction or an explicit statement that findings are exploratory should be added.
[Section 3.1] The sentence 'fed into the LLM (see Fig. 2)' repeats the earlier 'As shown in Fig. 2' in the same paragraph; remove the duplicate reference for clarity.
[Section 3.3] Equation (1) is a simple concatenation, yet the text says this 'retains distinguishing features from each modality while enabling cross-modal interaction'; concatenation alone does not guarantee interaction. Please clarify how the LLM's attention layers are expected to perform the cross-modal interaction.
[Section 4.2] The tactile videos are sampled at 250 ms intervals, but the number of frames per object and the spatial resolution of the tactile encoder input are not given; specify these for reproducibility.
[References] Reference [17] has an unintended 'bibliography' prefix in the author list; clean up the BibTeX entry.
[Section 4.3] The sentence 'Octopi uses our prompt to score the physical properties of objects' is interjected parenthetically; it should be moved to the experimental setup to avoid confusing the fine-grained baseline definition.
[Abstract and Conclusion] The claim of 'strong zero-shot generalization' is stronger than the evidence, which evaluates a single 35-object set without a held-out generalization experiment; rephrase to 'promising zero-shot performance on the current object set.'

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: model outputs are validated against independent instrument measurements; self-citations appear only in related work and are not load-bearing.

full rationale

The paper's central claim is empirical: the proposed Octopi-ViTaL model produces physical-property scores that correlate with instrument-measured values. The ground truth is obtained externally with a PosiTector SHD durometer, a C610H tensile tester, and a RUGOSURF 20 profilometer, none of which are derived from the model. The hand-designed 10-point scoring rubric (Table 1) defines a qualitative mapping from material examples to score ranges; it may shape model outputs, but it is not fitted to the ground-truth measurements and does not make the prediction equal to the label by construction. The comparison against Octopi is a zero-shot external baseline, and the paper explicitly acknowledges the sensor domain shift that handicaps the baseline; that is an experimental-design concern, not circularity. Self-citations (Refs. 5 and 6) appear only in the related-work section and do not carry the argument. No equation reduces the prediction to the input, and no fitted parameter is relabeled as a prediction. The reported correlations are therefore an independent, externally grounded empirical result, with at most minor concerns about label validity for compliant materials that belong under correctness risk rather than circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of heterogeneous physical measurements across material types, the transferability of the Octopi tactile encoder despite known domain shift, a post hoc inversion of elasticity correlation, and the zero-shot reasoning capability of the LLM with the custom prompt. None of these are established independently in the paper.

assumptions (4)

domain assumption Shore hardness, elastic modulus (via tensile test), and Ra roughness are valid, comparable ground-truth measures across all 35 objects.
Section 4.1. Diverse materials (foam, textile, paper, glass) may not be accurately measured by the same instruments; no exclusion or correction is described.
domain assumption The pretrained Octopi tactile encoder provides useful features for the GelSight Mini even though Octopi itself fails zero-shot on this sensor due to domain shift.
Section 3.2 and Section 4.3. The paper reports Octopi's near-zero correlations on the same data, yet uses its encoder in the full model without describing any adaptation.
ad hoc to paper Perceived elasticity is inversely proportional to elastic modulus, justifying the use of the absolute value of Spearman's rho.
Section 4.3. This assumption is introduced post hoc to interpret the correlation sign and is not independently validated.
domain assumption Vicuna-7B, prompted with the structured protocol, can map visual-tactile features to consistent Likert-scale scores in a zero-shot manner.
The entire method depends on the reasoning capability of the LLM with the custom prompt; no validation of prompt reliability beyond the 35-object study is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robotic Perception with a Large Tactile-Vision-Language Model for Physical Property Inference." pith.science (2026). https://pith.science/paper/RAZIUMDX

@misc{pith2026250619303,
  author       = {Pith},
  title        = {Pith review of: Robotic Perception with a Large Tactile-Vision-Language Model for Physical Property Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RAZIUMDX}},
  note         = {Machine review of arXiv:2506.19303}
}

read the original abstract

Inferring physical properties can significantly enhance robotic manipulation by enabling robots to handle objects safely and efficiently through adaptive grasping strategies. Previous approaches have typically relied on either tactile or visual data, limiting their ability to fully capture properties. We introduce a novel cross-modal perception framework that integrates visual observations with tactile representations within a multimodal vision-language model. Our physical reasoning framework, which employs a hierarchical feature alignment mechanism and a refined prompting strategy, enables our model to make property-specific predictions that strongly correlate with ground-truth measurements. Evaluated on 35 diverse objects, our approach outperforms existing baselines and demonstrates strong zero-shot generalization. Keywords: tactile perception, visual-tactile fusion, physical property inference, multimodal integration, robot perception

Figures

Figures reproduced from arXiv: 2506.19303 by the authors.

**Figure 2.** The architecture of a multimodal large model. After embedding and to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Vision processing pipeline. The image is partitioned into multiple regions [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: A sequence of tactile images is first processed by the tactile encoder to [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Object set of 35 common household items, spanning nine major mate [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: During the experiment, we selected 35 objects in the laboratory, measured [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 20 canonical work pages

[1]

Brain Res.134, 107–125 (2000)

Jeka, J., Oie, K.S., Kiemel, T.: Multisensory information for human postural con- trol: integrating touch and vision.Exp. Brain Res.134, 107–125 (2000)

work page 2000
[2]

94, 62–75 (2014)

Fleming, R.W.: Visual perception of materials and their properties.Vision Res. 94, 62–75 (2014)

work page 2014
[3]

Sensors18(4), 948 (2018)

Chi, C., Sun, X., Xue, N., et al.: Recent progress in technologies for tactile sensors. Sensors18(4), 948 (2018)

work page 2018
[4]

arXiv preprintarXiv:2311.07226 (2023)

Zeng, F., Gan, W., Wang, Y., et al.: Large language models for robotics: A survey. arXiv preprintarXiv:2311.07226 (2023)

arXiv 2023
[5]

IEEE/RSJ Int

Liu, H., Huang, B., Li, Q., et al.: Multi-fingered tactile servoing for grasping ad- justment under partial observation.Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS)2022, 7781–7788 (2022) 12 Zexiang Guo et al

work page 2022
[6]

IEEE/RSJ Int

Van Hoof, H., Chen, N., Karl, M., et al.: Stable reinforcement learning with au- toencoders for tactile and visual data.Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS)2016, 3928–3934 (2016)

work page 2016
[7]

Yuan, W., Dong, S., Adelson, E.H.: Gelsight: High-resolution robot tactile sensors for estimating geometry and force.Sensors17(12), 2762 (2017)

work page 2017
[8]

In:2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Donlon, E., Dong, S., Liu, M., et al.: Gelslim: A high-resolution, compact, robust, and calibrated tactile-sensing finger. In:2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1927–1934. IEEE (2018)

work page 2018

Show all 23 references

[9]

Neural Inf

Smith, E., Calandra, R., Romero, A., et al.: 3D shape reconstruction from vision and touch.Adv. Neural Inf. Process. Syst.33, 14193–14206 (2020)

2020
[10]

In:2024 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp

Dave, V., Lygerakis, F., Rueckert, E.: Multimodal visual-tactile representation learning through self-supervised contrastive pre-training. In:2024 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp. 8013–8020. IEEE (2024)

2024
[11]

In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

Xu, W., Yu, Z., Xue, H., et al.: Visual-tactile sensing for in-hand object recon- struction. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8803–8812 (2023)

2023
[12]

Taunyazov, T., Sng, W., See, H.H., et al.: Event-driven visual-tactile sensing and learning for robots.arXiv preprintarXiv:2009.07083 (2020)

2020 arXiv
[13]

Robot.39(5), 3838–3856 (2023)

Li, S., Yu, H., Ding, W., et al.: Visual–tactile fusion for transparent object grasping in complex backgrounds.IEEE Trans. Robot.39(5), 3838–3856 (2023)

2023
[14]

Li, S., Tang, H.: Multimodal Alignment and Fusion: A Survey.arXiv preprint arXiv:2411.17040 (2024)

2024
[15]

Liu, H., Yu, Y., Sun, F., et al.: Visual–tactile fusion for object recognition.IEEE Trans. Autom. Sci. Eng.14(2), 996–1008 (2016)

2016
[16]

In:2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Meyer, J., Eitel, A., Brox, T., et al.: Improving unimodal object recognition with multimodal contrastive learning. In:2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5656–5663. IEEE (2020)

2020
[17]

bibliography Wong, C.Y., Vergez, L., Suleiman, W.: Vision-and tactile-based con- tinuous multimodal intention and attention recognition for safer physical human– robot interaction.IEEE Trans. Autom. Sci. Eng.(2023)

2023
[18]

In:2024 IEEE International Conference on Robotics and Automation (ICRA), pp

Gao, J., Sarkar, B., Xia, F., et al.: Physically grounded vision-language models for robotic manipulation. In:2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 12462–12469. IEEE (2024)

2024
[19]

In:2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Lai, W., Zhang, T., Lam, T.L., et al.: Vision-language model-based physical rea- soning for robot liquid perception. In:2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 9652–9659. IEEE (2024)

2024
[20]

Yu, S., Lin, K., Xiao, A., et al.: Octopi: Object property reasoning with large tactile-language models.arXiv preprintarXiv:2405.02794 (2024)

2024 arXiv
[21]

In:International Conference on Machine Learning (ICML), pp

Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual mod- els from natural language supervision. In:International Conference on Machine Learning (ICML), pp. 8748–8763. PMLR (2021)

2021
[22]

Chiang, W.L., Li, Z., Lin, Z., et al.: Vicuna: An open-source chatbot impressing GPT-4 with 90%* ChatGPT quality.See https://vicuna.lmsys.org (accessed 14 April 2023)2(3), 6 (2023)

2023
[23]

Neural Inf

Liu, H., Li, C., Wu, Q., et al.: Visual instruction tuning.Adv. Neural Inf. Process. Syst.36, 34892–34916 (2023)

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Brain Res.134, 107–125 (2000)

Jeka, J., Oie, K.S., Kiemel, T.: Multisensory information for human postural con- trol: integrating touch and vision.Exp. Brain Res.134, 107–125 (2000)

work page 2000

[2] [2]

94, 62–75 (2014)

Fleming, R.W.: Visual perception of materials and their properties.Vision Res. 94, 62–75 (2014)

work page 2014

[3] [3]

Sensors18(4), 948 (2018)

Chi, C., Sun, X., Xue, N., et al.: Recent progress in technologies for tactile sensors. Sensors18(4), 948 (2018)

work page 2018

[4] [4]

arXiv preprintarXiv:2311.07226 (2023)

Zeng, F., Gan, W., Wang, Y., et al.: Large language models for robotics: A survey. arXiv preprintarXiv:2311.07226 (2023)

arXiv 2023

[5] [5]

IEEE/RSJ Int

Liu, H., Huang, B., Li, Q., et al.: Multi-fingered tactile servoing for grasping ad- justment under partial observation.Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS)2022, 7781–7788 (2022) 12 Zexiang Guo et al

work page 2022

[6] [6]

IEEE/RSJ Int

Van Hoof, H., Chen, N., Karl, M., et al.: Stable reinforcement learning with au- toencoders for tactile and visual data.Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS)2016, 3928–3934 (2016)

work page 2016

[7] [7]

Yuan, W., Dong, S., Adelson, E.H.: Gelsight: High-resolution robot tactile sensors for estimating geometry and force.Sensors17(12), 2762 (2017)

work page 2017

[8] [8]

In:2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Donlon, E., Dong, S., Liu, M., et al.: Gelslim: A high-resolution, compact, robust, and calibrated tactile-sensing finger. In:2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1927–1934. IEEE (2018)

work page 2018