acceptodds
Under review as a conference paper at ICLR 2027

LifelongTouch: Lifecycle-Invariant Tactile Representation

Abstract

Vision-based tactile sensors (VBTSs) provide rich contact information, yet their observations are susceptible to variations across sensors and state changes during long-term use. Existing tactile representation learning methods primarily focus on cross-sensor generalization, typically treating each sensor as a static domain and thus struggling with evolving sensor states during long-term deployment. To address this limitation, we propose LifelongTouch, a lifecycle-invariant tactile representation learning framework for generalization across sensors and lifecycle states. LifelongTouch jointly models contact observations and no-contact references under the current sensor state, and combines reference-conditioned physical canonicalization, physical geometry supervision, and sensor-state physical alignment to learn consistent physical representations across sensors and lifecycle states. Furthermore, we develop a lifecycle-aligned tactile simulation engine that provides cross-sensor, cross-state tactile observations with geometric supervision for representation learning. Experiments on object classification, contact pose estimation, force distribution estimation, and surface reconstruction demonstrate that LifelongTouch generalizes robustly to unseen sensors and lifecycle states, providing an effective approach toward reliable lifelong tactile perception. The data and code will be publicly released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Tactile sensing, representation learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Ub3nOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

LifelongTouch addresses generalization of vision-based tactile representations across sensor identities and changing sensor conditions, using a fixed model and a current no-contact reference. The authors also introduce a lifecycle simulation engine producing 870,400 observations across 100 virtual sensors. The experiment uses frozen representations for object classification, relative contact-position estimation, and dense force estimation, alongside qualitative surface reconstruction, demonstrating improved illumination robustness, strong cross-sensor pose performance under common linear probes, and promising force transfer.

Critical strengths

1.The main method innovation of this article lies in explicitly incorporating the sensor lifecycle state into the generalization problem, using the current reference graph for learning style feature modulation and interaction, and training a more robust representation of state changes through lifecycle simulation and local feature alignment. 2.The experimental results indicates that geometric information is transferable. The component ablation also supports the usefulness of reference adjustment. These results confirm that the proposed design goes beyond the scope of qualitative examples.

Critical weaknesses

None

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Appreciation for Your Recognition

We sincerely thank you for your high regard for our work and your positive assessment of its motivation, methodological contributions, and experimental results.

We particularly appreciate your understanding of the challenges associated with real-world evaluation in vision-based tactile sensing. Unlike conventional visual tasks, ground-truth contact geometry, surface deformation, and force distributions are often difficult to measure directly with reliable pixel-level annotations. This challenge becomes even more pronounced as sensors undergo prolonged use, wear, and aging, making it difficult to continuously obtain accurate geometric and physical ground truth under real interaction conditions. Addressing these challenges is essential for establishing reliable tactile benchmarks.

These practical difficulties also motivate our approach of combining simulation-based training with real-world evaluation. Specifically, physical simulation provides controllable and diverse training data with accurate ground truth, while real tactile datasets allow us to validate generalization across different sensors and lifecycle states.

We are also grateful for your recognition of our core methodological contributions, particularly the explicit incorporation of sensor lifecycle states into the generalization problem, dynamic-reference-based feature modulation, and the combination of geometric supervision and local feature alignment for learning robust representations. Your recognition of the transferability of tactile geometric information and the effectiveness of reference adjustment is especially encouraging to us.

Once again, we sincerely appreciate your positive assessment and expert perspective. Your recognition not only affirms the methodological design and experimental efforts behind LifelongTouch but also further motivates our continued efforts toward developing reliable tactile perception systems with long-term adaptability and generalization.

Reviewer_Bs9JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LifelongTouch learns a tactile representation space for vision-based tactile sensors that remains consistent across the sensors' lifecycle, which is invariant to the sensor lighting drift, gel wear and stains. The model takes a contact reading together with a no-contact reference reading (which shows the original reading of the sensor's current state). The encoder is trained on a simulated dataset with synthetic degradations to the sensor, and outperforms existing tactile representation methods on real tactile benchmark tasks, including object classification, contact pose estimation and force estimation.

Critical strengths
  1. Meaningful problem setup: Existing tactile representation method mainly treat each sensor as a sensor with consistent readings across the time. However in practice the sensor would change significantly, including the drift in lighting conditions, the aging/damage of the gel and the wearing of the surface. Despite only using synthetic data, this paper formulates this problem and build a dataset to simulate that, which has practical contribution to the long-lasting problem of tactile sensing.

  2. Sensor-invariant representation: According to the experiments, the proposed method is achieving the lower pose estimation error and force estimation erorr when compared to all baseline methods, even if the testing is performed on unseen sensors with degraded gels. This indicates that the learned representation is capturing the tactile information that is invariant to sensor-specific appearance.

Critical weaknesses
  1. The evaluation is mostly based on synthetic degradations: All the light-drift, wear, stain in LifetimeSITR are all synthetic degradations, and they are also similar to the degradation factors used in the pre-training simulation engine data. The result is that LifelongTouch has seen the same types of distribution shift during training while the baseline methods have not. Moreover, the wear and stain degradations barely affect the methods, and most of the improvement comes from light-drift. It would be interesting to evaluate on the real data collected from the same sensor over a long period of use time.

  2. The comparison is determined by different downstream heads: In table 1 and 2, LifelongTouch and SITR use the predicted normal map with a ImageNet-pretrained ResNet, while some of the baselines (e.g. Sparsh, AnyTouch and CTTP) only uses a linear layer. Under the unified linear probe in table 7, the pretrained ViT outperforms LifelongTouch on classification. This makes it hard to conclude how much of the gain comes from the representation itself. It would be interesting to report all methods' performance with the same heads to ablate on the representation itself.

Questions for authors

Corresponding to the weakness points above, I have two major questions:

  1. Would it be possible to evaluate on real tactile data collected from the same sensor at different points of its usage (e.g., before and after long-term use, or with a worn or stained gel)? If LifelongTouch still outperforms the baselines in these settings, it would support the claim that the representation generalizes to unseen lifecycle states.

  2. Would it be possible to report the main results of all methods with the same downstream heads? If LifelongTouch still outperforms baseline methods under the same heads, it would support the claim that the learned representation is the major improvement.

2 comments
AuthorsOpenReview ↗

R1. Synthetic Degradation Evaluation and Validation on Real Physical Aging (Q1)

(1) Validity of Simulated Degradation in LifetimeSITR.

We acknowledge that synthetic image degradation in LifetimeSITR cannot fully reproduce physical sensor aging. To reduce overfitting to specific degradation patterns, we use different generation mechanisms: LifelongTouch is pretrained on Blender-based physical simulations by varying sensor parameters, whereas LifetimeSITR applies post-processing perturbations to real tactile images. Thus, this evaluation primarily assesses generalization to unseen degradation patterns.

Moreover, we adopt a New-only training protocol, where downstream task heads are trained exclusively on New-state data and evaluated directly on degraded states without fine-tuning or recalibration. This simulates practical deployment, where a sensor is initially calibrated and must continue operating as its physical properties change. Therefore, LifetimeSITR evaluates representation robustness and cross-state transfer rather than replacing real physical aging evaluation.

(2) Evaluation on Real Tactile Sensors.

Collecting long-term data across multiple degradation stages of the same physical sensor is challenging. Therefore, LifetimeFEATS currently includes two real lifecycle states: New and Aged. We provide further analysis of performance across different usage stages of the same sensor.

LifetimeFEATS contains five physical tactile sensors (sensor00–sensor04). Both the training and NoT test sets involve sensor01–sensor04, whereas US-New and US-Aged both come from sensor00, which is unseen during training. Thus, the results in Table 3 of the main paper allow direct comparison of different lifecycle states while keeping the sensor identity fixed.

To further isolate sensor identity, we conduct a fine-grained evaluation on the UO-Press subset, which contains real contact data from the seen sensors sensor01 and sensor04 across different objects and both New and Aged conditions. This enables direct evaluation of physical gel aging under fixed sensor identities. Results are reported as RMSE/MAE (N↓).

MethodSensor01 NewSensor01 AgedSensor04 NewSensor04 Aged
FEATS0.01202/1.1230.02312/2.6430.02309/2.0020.01010/0.749
T30.00652/1.1620.01141/1.1040.01106/0.7890.00540/0.880
SITR0.00784/0.7040.01435/0.9560.01336/0.8560.00647/0.643
ViT0.00773/0.7150.01403/1.0970.01443/1.2570.00644/0.544
AnyTouch0.00640/0.7000.01191/0.8500.01153/0.6830.00546/0.598
Sparsh0.00793/0.6390.01405/1.5480.01371/0.7930.00697/0.613
CTSRL0.00768/1.0170.01324/1.1110.01255/0.7110.00595/0.671
CTTP0.01046/0.9750.01900/1.5150.01795/1.2630.00875/1.323
LifelongTouch0.00557/0.4390.01022/0.5190.01011/0.3860.00460/0.264

LifelongTouch achieves the lowest RMSE and MAE across all four conditions, demonstrating its ability to maintain accurate force perception under tactile observation changes caused by real gel aging.

In summary, we do not consider synthetic degradation evaluation a substitute for real physical aging. Instead, we combine controlled synthetic degradation experiments on LifetimeSITR with real aging experiments on LifetimeFEATS to validate LifelongTouch's generalization across both synthetic appearance variations and real physical lifecycle changes.

AuthorsOpenReview ↗

R2. Fair Comparison of Downstream Task Heads (Q2)

We thank the reviewer for pointing out this issue. Indeed, during this work, we also recognized the challenges of comparing tactile representation methods that employ different prediction heads for downstream evaluation. Existing methods do not share a unified downstream architecture, meaning that their final performance depends not only on the pretrained encoder but also on the task head architecture and its pretrained priors. Therefore, original downstream results alone may not accurately reflect the representation capability and generalization of the encoders themselves.

(1) Original downstream evaluation.

In Tables 1 and 2 of the main paper, to ensure a fair comparison with existing methods, we follow each baseline's original downstream evaluation protocol rather than forcing all methods to use an identical task head. For example, Sparsh, AnyTouch, and CTTP use global features with linear prediction layers, whereas SITR, CTSRL, and T3 leverage predicted surface normal maps combined with an ImageNet-pretrained ResNet for downstream prediction. We retain these native task heads to preserve each method's original feature interface and task adaptation strategy, allowing evaluation under their respective designed downstream paradigms.

However, we agree with the reviewer that this evaluation compares the complete pretrained representation and downstream task head, rather than strictly measuring the representation capability of the encoder itself.

(2) Unified Linear Probe Evaluation.

To eliminate the influence of downstream task head designs and directly evaluate pretrained encoder representations, we provide additional experiments using a unified Linear Probe Head on LifetimeSITR in Tables 7 and 8 of Appendix A.4.1.

For object classification, ImageNet-pretrained ViTs, particularly ViT-Large, exhibit strong linear separability and achieve the best classification performance, demonstrating the rich object appearance and semantic information learned through large-scale visual pretraining. Nevertheless, among tactile-pretrained encoders, LifelongTouch achieves the best overall classification performance.

For pose estimation, which relies more heavily on fine-grained contact geometry, the advantage of ImageNet-pretrained ViTs becomes less pronounced. In contrast, LifelongTouch achieves the lowest cross-sensor pose estimation RMSE across all lifecycle conditions, outperforming the other tactile encoders. This suggests that generic visual representations do not necessarily preserve fine-grained physical contact geometry. By incorporating contact geometry supervision and lifecycle alignment, LifelongTouch learns representations that are better suited to physical tactile perception and exhibit stronger cross-sensor and cross-lifecycle generalization.

(3) Unified Linear Probe Evaluation on LifetimeFEATS.

To further evaluate the ability of different pretrained encoders to represent fine-grained tactile force information, we conduct additional unified Linear Probe experiments on LifetimeFEATS. Specifically, all tactile encoders are frozen, and only a unified linear prediction head is trained. The encoder features are mapped to dense force fields of size 3×24×323\times24\times32 using Linear(D,2304), where D denotes the feature dimension of each encoder. Results are reported as RMSE (N).

MethodDNoT ↓US-New ↓US-Aged ↓UO-Press ↓
T37680.018130.016720.021380.01628
ViT-Base ImageNet7680.070600.098780.106550.09578
ViT-Large ImageNet10240.059930.088210.110860.09433
AnyTouch10240.030030.058830.071580.02552
SITR7680.019800.018690.019420.01686
CTSRL20480.027030.034700.035400.02274
Sparsh7680.019220.035980.031730.01751
CTTP20480.038680.036930.036600.04020
LifelongTouch7680.017230.015280.018350.01448

The results show that LifelongTouch achieves the best performance across all four evaluation settings. This further demonstrates that its learned representations preserve fine-grained tactile force information while providing strong generalization across sensor identities and lifecycle states.

Reviewer_YrXtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents LifelongTouch, a tactile representation learning framework designed to improve generalization across different vision-based tactile sensors and their changing states over time. The main motivation is that existing tactile representation methods largely focus on cross-sensor transfer while overlooking changes within the same sensor caused by illumination drift, gel aging, surface wear, and contamination.

The proposed method consists of three components: reference-conditioned physical canonicalization, physical geometry supervision, and sensor-state physical alignment. It uses a no-contact reference image to condition the representation on the current sensor state, while geometric supervision and feature alignment encourage consistent representations across sensors and lifecycle conditions. The authors also develop a simulation engine that generates aligned tactile observations under different sensor configurations and degradation states.

The method is evaluated on object classification, contact-pose estimation, force-distribution estimation, and surface reconstruction. The results show improvements over several existing tactile representation methods, particularly under simulated lifecycle variations and in force-distribution estimation on physically aged gels.

Critical strengths

1. The paper addresses a relevant and practical problem.

Most existing tactile representation learning methods focus on generalization across different sensors, while the performance degradation caused by changes within the same sensor over time has received less attention. I think this is an interesting and practically relevant direction, especially for robotic systems that rely on tactile sensing over extended periods. Extending sensor-invariant representations to account for lifecycle changes is a meaningful contribution.

2. The proposed framework is well motivated and supported by reasonably comprehensive experiments.

Using a no-contact reference to capture the current sensor state is intuitive, and combining it with geometric supervision and feature alignment is a reasonable design. The paper evaluates the method across several downstream tasks rather than relying on a single benchmark. The ablation studies are also useful and show that the reference conditioning and alignment components contribute to the overall performance. I particularly appreciate the inclusion of force-distribution estimation using real tactile data with different gel conditions.

Critical weaknesses

1. The real-world evaluation does not fully support the central claim of lifelong tactile perception.

My main concern is that the paper emphasizes generalization across the sensor lifecycle, but most of the experimental evidence comes from simulated degradation rather than actual long-term sensor usage. In LifetimeSITR, the lifecycle conditions are generated by applying artificial perturbations to existing tactile images. Although these perturbations are designed to simulate realistic changes, they may not fully capture the complex effects of physical aging and degradation.

The LifetimeFEATS benchmark does include data collected with physically aged gels, which is valuable. However, the US-Aged setting also involves unseen sensors, so the effects of sensor variation and gel aging are not clearly separated. It is therefore difficult to determine how much of the improvement comes specifically from handling lifecycle changes rather than general cross-sensor robustness.

I think the paper would be more convincing with a controlled evaluation of the same physical sensor at different stages of its lifecycle, or at least a clearer analysis that isolates the effects of sensor identity and aging.

2. The assumptions about no-contact reference images need further investigation.

The proposed method relies on a no-contact reference image representing the current sensor state. While this is reasonable in principle, I am not fully convinced that such references can always be obtained reliably in practical deployment. For example, a robot may operate under continuous contact, and the sensor state may change during operation.

I also have some concerns about how the reference images are constructed for LifetimeFEATS. According to Appendix A.2.2, the authors use contact masks derived from force-distribution labels and image similarity matching to select appropriate no-contact references. It is unclear whether this procedure can be reproduced during inference without access to ground-truth force information.

Since reference conditioning is one of the main contributions, I think it is important to clarify this assumption and evaluate how sensitive the method is to imperfect or outdated references.

Questions for authors

I am actually leaning towards positive for this work and below are my questions:

  1. In LifetimeFEATS, the US-Aged setting involves both unseen sensors and aged gel pads. Is it possible to separately evaluate the effect of gel aging while keeping the sensor identity fixed? This would help clarify whether the improvement comes from lifecycle robustness or general cross-sensor transfer.

  2. Could the authors clarify how the no-contact references are obtained during inference, particularly for LifetimeFEATS? Does the reference selection process require ground-truth force labels or contact masks? If so, how would this work in a real deployment setting?

  3. How sensitive is LifelongTouch to the quality of the reference image? For example, what happens if the reference is collected before significant sensor degradation, contains residual contact, or does not perfectly match the current sensor state?

  4. Table 2 shows that SITR achieves better average inter-sensor contact-pose estimation than LifelongTouch under aged conditions. Could the authors discuss why the proposed method does not consistently improve this task, despite the explicit geometric supervision and alignment objectives?

  5. Could the authors provide a comparison with simpler reference-based approaches, such as direct contact-reference subtraction or concatenating the two images, using the same backbone and training data? This would help clarify the contribution of the proposed reference-conditioning architecture.

6 comments
AuthorsOpenReview ↗

R1. Definition of Lifelong Tactile Perception and Real-World Evaluation (Q1)

We agree that simulated degradation cannot fully capture the physical aging of real sensors. We therefore clarify our definition of lifelong tactile perception, justify the simulated degradation evaluation, and provide additional experiments on physically aged sensors.

(1) Definition and motivation.

Inspired by human tactile perception, which remains relatively stable despite individual differences in skin properties and changes over time, artificial tactile systems should similarly adapt to variations across sensors and remain robust to changes caused by prolonged use, wear, and aging. Accordingly, we define lifelong tactile perception as the ability to generalize across both sensor identities and lifecycle states.

(2) Validity of simulated degradation in LifetimeSITR.

We acknowledge that synthetic image degradation cannot fully reproduce physical sensor aging. To reduce overfitting to specific degradation patterns, we deliberately use different degradation mechanisms: LifelongTouch is pretrained on Blender-based physical simulations with varying sensor parameters, whereas LifetimeSITR applies post-processing perturbations to real tactile images. This evaluation therefore primarily measures generalization to unseen degradation patterns.

Moreover, we adopt a New-only training protocol, where downstream task heads are trained exclusively on New-state data and evaluated directly on degraded states without additional fine-tuning or recalibration. This reflects practical deployment, where a sensor is initially calibrated and must remain functional as its properties evolve. Thus, LifetimeSITR evaluates representation robustness and cross-state transfer rather than replacing real-world aging evaluation.

(3) Further analysis of physical aging in LifetimeFEATS.

We thank the reviewer for highlighting the potential confounding between sensor identity and gel aging in the US-Aged setting. To address this concern, we clarify the dataset splits and provide a finer-grained evaluation under real sensor aging.

LifetimeFEATS contains five physical sensors (sensor00–sensor04). Both training and NoT testing involve sensor01–sensor04, while US-New and US-Aged exclusively use the unseen sensor00. Therefore, the results in Table 3 already enable comparisons across lifecycle states of the same unseen sensor, jointly evaluating cross-sensor and cross-lifecycle generalization.

To further isolate the effect of aging, we analyze the UO-Press subset, which contains real contact measurements from two seen sensors (sensor01 and sensor04) under both New and Aged conditions. This allows direct comparison before and after physical aging while keeping sensor identity fixed. Results are reported as RMSE/MAE (N↓).

MethodSensor01 NewSensor01 AgedSensor04 NewSensor04 Aged
FEATS0.01202/1.1230.02312/2.6430.02309/2.0020.01010/0.749
T30.00652/1.1620.01141/1.1040.01106/0.7890.00540/0.880
SITR0.00784/0.7040.01435/0.9560.01336/0.8560.00647/0.643
ViT0.00773/0.7150.01403/1.0970.01443/1.2570.00644/0.544
AnyTouch0.00640/0.7000.01191/0.8500.01153/0.6830.00546/0.598
Sparsh0.00793/0.6390.01405/1.5480.01371/0.7930.00697/0.613
CTSRL0.00768/1.0170.01324/1.1110.01255/0.7110.00595/0.671
CTTP0.01046/0.9750.01900/1.5150.01795/1.2630.00875/1.323
LifelongTouch0.00557/0.4390.01022/0.5190.01011/0.3860.00460/0.264

LifelongTouch achieves the lowest dense force prediction RMSE and integrated force MAE across all four conditions, demonstrating strong force perception accuracy despite changes induced by real gel aging.

In summary, we do not regard simulated degradation as a complete substitute for physical aging. Instead, we validate lifelong tactile perception through three complementary evaluations: (i) generalization to unseen degradation mechanisms, (ii) cross-state transfer without fine-tuning or recalibration under the New-only protocol, and (iii) robustness to real physical aging on LifetimeFEATS. Together, these results provide evidence of LifelongTouch's ability to generalize across sensor identities and lifecycle states.

AuthorsOpenReview ↗

R2. Further Investigation of Non-contact Reference Mechanisms (Q2-Q5)

(1) Acquisition of Non-contact References in Inference and LifetimeFEATS (Q2).

During deployment, LifelongTouch directly captures a tactile image whenever the sensor is not in contact, without requiring force labels or additional information.

Since FEATS provides no paired non-contact references, we construct them offline for LifetimeFEATS. Specifically, force labels are used to mask contact regions, and image similarity is computed over non-contact regions. For each contact image, we select the most similar candidate with a similarity score above 0.9 within the same subset. Unmatched samples are discarded, and matching quality is manually inspected.

Therefore, force labels are used only for offline dataset construction, not model inference. During deployment, reference images are directly captured without mask extraction or image matching.

(2) Sensitivity of LifelongTouch to reference image quality (Q3).

LifelongTouch adopts a dynamic non-contact reference that reflects the sensor's current physical state, rather than the fixed background image captured during initial calibration, as used by methods such as Sparsh, SITR, and T3. During normal operation, the reference frame can be updated during the non-contact interval before each interaction, preventing outdated references. However, residual contact in the reference image or sensor damage occurring during continuous contact may still introduce reference discrepancies. In such cases, the most recent valid reference must be used until the next non-contact observation becomes available.

Despite these limitations, unlike conventional approaches relying on fixed initial references, LifelongTouch can adapt to sensor aging simply by capturing a new non-contact image, without requiring retraining or large-scale data collection and recalibration. This dynamic adaptability helps reduce long-term maintenance costs.

(3) Ablation study on non-contact reference mechanisms (Q5).

Let I and R denote contact and non-contact images, respectively. Existing methods use three strategies: contact images only (I, T3 and AnyTouch), pixel-wise subtraction (I-R, SITR, CTSRL, and CTTP), and channel-wise concatenation ([I,R], Sparsh).

Intuitively, I-R is an effective reference cancellation strategy that captures the major geometric changes induced by contact. However, vision-based tactile sensing relies on gel deformation, whose visual response is nonlinearly related to physical contact. Consequently, although direct pixel-wise subtraction highlights the overall contact region, it may also suppress fine-grained tactile variations. To address this limitation, we introduce a learnable latent-level difference, computing F(I)-F(R) in the feature space. This allows the model to extract fine-grained contact-related information through a learned feature representation.

We compare these strategies under the same LifelongTouch architecture, changing only the reference fusion mechanism. Results on LifetimeFEATS are reported as RMSE/MAE (N↓).

ReferenceNoTUS-NewUS-AgedUO-Press
I0.00960/0.8460.00997/0.8920.00985/0.8020.00918/0.674
I−R0.00941/0.7140.01006/1.1490.00940/0.7410.00917/0.652
[I,R]0.01091/1.0330.01178/1.1230.01137/1.1470.01012/0.875
F(I)−F(R) (Ours)0.00735/0.3290.00900/0.7930.00871/0.7170.00677/0.350

Our latent-level difference achieves the lowest RMSE and MAE across all four settings, outperforming the other strategies under identical training conditions. Note that this comparison uses a unified architecture, which may not be optimal for every alternative strategy.

(4) Comparison with SITR (Q4).

As discussed in (3), SITR's pixel-wise subtraction is an effective reference cancellation strategy. However, since sensor aging in LifetimeSITR is simulated through synthetic image perturbations, the I-R operation largely cancels out these appearance changes, making it difficult to fully evaluate representation robustness to lifecycle degradation. Furthermore, SITR employs a ResNet18 decoder based on predicted surface normals, effectively exploiting tactile geometry and offering advantages in object classification and pose estimation. However, its performance drops noticeably on dense force field estimation in LifetimeFEATS (Table 3), which requires finer-grained tactile information.

It is worth noting that Tables 2 and 3 in the main paper use the original task heads of each method (Appendix A.3), reflecting the performance of the complete methods rather than the generalization capability of the encoders themselves. Therefore, we further evaluate the representations using a unified linear probe head in Appendix A.4.1. Tables 7 and 8 show that LifelongTouch achieves the best performance among all tactile-pretrained encoders on both object classification and pose estimation, further demonstrating the generalization capability of its learned representations.

AuthorsOpenReview ↗

Follow-up on Author Response Before Discussion Ends

Dear Reviewer YrXt,

We hope this message finds you well. As the discussion period is approaching its end, we would like to follow up with you regarding our author response.

We have carefully addressed the concerns you raised in your review, including the comparative experiments across different lifecycle states of the same sensor, and the ablation study on reference images. In addition, we have made corresponding improvements in the revised paper. We believe these points have now been clarified.

If you find that your concerns have been adequately addressed, we would be very grateful if you could consider updating your evaluation accordingly. Of course, if you still have any questions or concerns, please let us know and we will do our best to address them promptly.

Thank you very much for your time and for your constructive comments on our work.

Sincerely,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.