acceptodds
Under review as a conference paper at ICLR 2027

Touch Has Memory: Learning Shared Response Models for Few-Shot Calibration of Soft Tactile Sensors

Abstract

Maintaining accurate force estimation continuously after calibration is challenging for soft tactile sensors, whose responses vary across sensors and depend on loading history. For large-scale manufacturing, the challenge is to capture these history-dependent responses with expressive models while keeping local calibration effort low. We present a few-shot calibration framework that separates shared response modeling from local output calibration. We use four response models: static, path-dependent, temporal relaxation, and hysteresis. Each predicts normal contact force through a compact neural readout. We train the models independently on source recordings, using initial segments for calibration and later segments to optimize force prediction. At each new sensing location, we freeze the models and use a few initial force references to calibrate their output combination. After calibration, all model and calibration parameters remain fixed, while local memory states update causally from the incoming electrical data stream, enabling force estimation without further force references. We evaluate the framework at fourteen sensing locations across three tactile sensor arrays excluded from training, each containing 48 sensing elements. With three fitting references selected from dense initial force recordings, our framework achieves an overall RMSE of 0.375 N under full-window calibration, outperforming the evaluated external baselines. These results suggest that shared response models can support accurate force estimation under changing loading conditions while limiting local learning to output calibration.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Sensor calibration, Tactile sensing, Transfer learning, Hysteresis modeling, Force estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dtinOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to calibrate soft tactile sensors when their electrical response depends on how they have been loaded. The authors learn four shared models covering static behavior, path dependence, temporal relaxation, and hysteresis; these models can then be reused at a new sensing location by fitting only their output weights and an offset from a few selected force references. Once calibrated, the system estimates force from incoming electrical readings, while its memory states continue to evolve. The evaluation covers fourteen sensing locations on three arrays excluded from training.

Critical strengths
  1. Only five parameters are fitted at each sensing location, alongside 1,270 shared parameters, and I see practical value in keeping this local fitting step small. The experts cover both static behavior and effects of loading history, so the compact model still accounts for several aspects of the sensor response.

  2. Testing on fourteen locations across three arrays excluded from training shows that the shared models can be reused on new arrays. The method achieves lower force errors using the same calibration references as the other methods that require force labels, with RMSE of 0.375 N compared with 0.652 N for CAVIA under the full window. The advantage over the comparison methods also holds with the shorter calibration window.

  3. The regularized calibration gives a unique solution even with only three force references for five fitted parameters. This provides a sound basis for the local adaptation step, where the limited measurements alone would leave the expert weights and offset underdetermined.

Critical weaknesses
  1. The method uses only three force values for fitting, but these are selected from a continuous force recording. Even when the reference times are fixed in advance, the experiment still uses that recording to synchronize and interpolate the measurements, so I’m not sure how much force data would need to be collected in practice, or how much calibration effort the method would save.

  2. I’m not sure when fitting separate expert weights offers a practical advantage over the simpler alternatives. Under the full calibration window, equal weights with a fitted gain and offset already achieve 0.383 N RMSE, close to the 0.375 N obtained with local weight fitting, while proportions learned on the source data do slightly better.

  3. Since the evaluation uses signals sampled at 5 Hz and defines the Rapid subset through force changes over one second, it is difficult to judge how well the method handles faster changes in loading. More context on the force ranges and loading speeds would help clarify what conditions the experiments cover and how large the reported errors are relative to the forces being measured.

Questions for authors
  1. What would calibration require at a new location in practice? Could it work with only three acquired force measurements, or is a continuous reference trace still needed for synchronization or preprocessing?

  2. Equal weights are already close to the fitted mixture (see Weakness 2). Are the small gains from fitting separate weights consistent across sensing locations, and are there particular loading conditions where this helps most?

  3. What are the typical force ranges and loading speeds in the recordings? How large is the 0.375 N error relative to those forces, and what dynamic behavior is covered by the 5 Hz evaluation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dtin

Thank you for the review. Your question on what calibration requires in practice was especially useful. Answering it led us to test how accurate the three references need to be, and they do not need to be very accurate. We answer your three questions below, and the corresponding changes are in the revised paper.

Q1 (W1). What would calibration require at a new location in practice?

The short answer is that calibration needs three force readings, each with a time stamp accurate to about a second. A continuous force recording is not needed.

In our experiments we did record force continuously, and this may have caused the confusion. The recording was used for two other things. It is the ground truth against which we measure the error, and with force-based selection it is also used to detect contact and to choose the three references. Two results show that the calibration itself does not need it.

Result 1. This result is already in Table 1 in the paper, under time-based selection. We took the three references at fixed times, without looking at the force at all. Our method still works. Its RMSE is 0.415 N, compared with 0.597 N for CAVIA, the best baseline, with the same references. So a dense force recording is not needed to choose the references.

Result 2. In a new test, we deliberately made the three references inaccurate. A timing error of ±2 s increases the RMSE by at most 0.016 N, and force noise of 0.2 N by at most 0.007 N.

In practice, a hand-held force gauge read at three moments is enough. This saves a great deal of effort. A full force sweep takes about 14 minutes per taxel, which is about 11 hours per array. Three references take 30 s to 2 min per taxel, which is 24 minutes to 1.6 hours per array. This could be shortened further by pressing faster. The details are in Point 3 of the general response and in the new Appendix D.3.

Q2 (W2). Are the gains from fitting separate weights consistent, and where do they help most?

Yes, the gains are small but consistent. Fitted weights are better than equal weights at 11 of 14 locations under both windows. The gain in RMSE is 0.007 N with the full window and 0.004 N with the short window, and both confidence intervals exclude zero.

You also asked where fitted weights help most. With the full window they help most during unloading (0.012 N, at 11 of 14 locations) and less during loading (0.006 N, at 9 of 14 locations). With the short window they help most while the load is held (0.008 N, at 10 of 14 locations). On the final irregular segment with the full window, equal weights are slightly better, by 0.005 N.

You are right that equal weights are a reasonable simpler choice, and we looked into why they come so close. The four experts are mixed in almost the same proportions at every location, while the overall gain ranges from 1.3 to 3.5. In other words, locations differ mainly in how sensitive they are, and this sensitivity is exactly what calibration is meant to measure at each sensing location. The revised paper now explains this in Section 5.5. We keep the fitted weights because they cost nothing extra and give a small but real gain. Among the three options (fitted weights, equal weights, and weights learned on the source arrays), fitted weights are the only one that is never the worst under either window.

Q3 (W3). Force ranges, loading speeds, and what 5 Hz covers.

The new Appendix D.2 and the new Figure 6 in the revised paper describe the recordings. We summarize the main numbers here.

  • Force range. The peak force in the evaluated part of the recordings ranges from 1.6 to 6.5 N, with a median of 3.8 N.
  • Loading speed. The loading is slow to moderate. The force change over one second has a median of 0.11 N/s, a 95th percentile of 0.46 N/s, and a maximum of 1.50 N/s.
  • Size of the error. Our RMSE of 0.375 N is about 10% of the peak force (6.7% to 21.6% across recordings). For CAVIA, the best baseline, it is 17.5%.
  • Faster loading. On the samples where the force changes by more than 0.5 N within one second, our RMSE is 0.586 N, compared with 0.981 N for the path expert alone.
  • What 5 Hz covers. At 5 Hz and at most 1.5 N/s, consecutive samples differ by at most 0.3 N, so the sampling follows the loading used here. The 5 Hz rate is set in preprocessing and is not a limit of the hardware. We chose it because the reference force is measured by an Instron machine, which records at 10 Hz. Our customized PCB supports about 1,200 frames per second for one tactile sensor, and the target recordings were logged at about 60 frames per second. Faster loading, such as impacts, is beyond what this test machine can apply and measure.
Reviewer_dtinOpenReview ↗

Post Rebuttal Comment

Thanks for the clarifications and additional analysis. The explanation of the three-reference calibration and the timing tests help address my main concern. Please make sure the practical calibration requirments and the tested loading conditions are clear in the final version, including the limits of the 5 Hz evaluation. My overall assessment of the work and my score remains unchanged.

Thank you, glad it helped.

All three pts are in the revised paper.

  • What calibration needs is in the experimental setup and Appendix D.3.
  • The force ranges and loading speeds are in Appendix D.2 and Figure 6.
  • What the 5 Hz evaluation covers is in Appendix D.2.

We have also made them easier to find from the main text, and this will be in the final version.

Thanks again for practical questions during the disscussion process.

Reviewer_TPUBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method for calibrating tactile sensors by reweighing the outputs of four individual expert models. Each expert model maps the sensor reading (and its derived, history-dependent features) to a force estimate. The experts are trained in a calibration-aware fashion. The calibration of the weight of each expert and the shared offset is then obtained from three ground-truth force readings.

Critical strengths
  • This paper addresses a relevant problem, the few-shot calibration of tactile sensors.
  • The paper compares against several relevant baselines, and runs ablations on its expert design.
  • The paper also admits to its shortcomings, e.g., the fact that more references during calibration in fact hurt short-window force accuracy.
Critical weaknesses
  • Writing: Personally, I found the paper hard to parse. First, the terminology is often vague and loose, and leaves the reader unsure of what the authors are referring to. For example, the authors use the term sensing location throughout the paper without a clear explanation or definition. Is it just an individual taxel? Or a physically defined location on the sensor surface that isn’t necessarily aligned with a single taxel? How does it differ from a sensing element which itself is defined as a taxel? Similarly, the word calibration is used in both 4.2 and 4.3 to refer to two different things: one used during training of an individual expert (4.2), and one doing the actual calibration of the mixture of the experts (4.3). Second, many details are left out, rendering reproducibility and interpretability difficult. As an example: the physical tactile sensor (dimensions, sensing principle, etc.), and what the digital reading refers to (current? capacitance?) are never explained. Third, the paper is mostly descriptive, and little of the text is dedicated to providing insights or interpretations the authors may have that go beyond the method, equations, and numbers themselves. For example, in Section 4.1, the input features for three of the four experts are described in a single sentence each, accompanied by equations. More intuition and explanation of how each feature captures relevant information for the given expert’s purpose would be appreciated, and would also motivate more on why exactly these four experts are needed. Another example is sections 5.2 to 5.5, where many numbers from the tables and figures are simply restated. The fact that simply concatenating the features and training a single network leads to much worse performance warrants some discussion. Further, the fact that more references during calibration in fact hurt short-window accuracy is not explained or discussed. The fact that removing certain experts can in some situations improve accuracy is also simply stated as a fact and not discussed.
  • Results: No uncertainty measures are stated throughout the experiments and ablations for the RMSE numbers, and the statistical significance of the results is thus hard to interpret. Further, the ablations imply that the presented design may not be optimal. Removing certain experts improves full-window accuracy, and equal weighting almost matches per-expert weighting.
  • Fit: This paper in its current form would, in my opinion, be more suitable at a different venue, as its contribution to machine learning and representation learning, as well as the generality of its findings, is limited. The input features are hand-engineered (and mostly established in the literature), and results are shown on a single sensor type only.
Questions for authors
  • How is a sensing location exactly defined?
  • Do you have any interpretation of why the concatenated features with the MLP have worse performance?
  • Can you provide insight as to why the removal of the static or temporal relaxation experts improves full-window accuracy?
  • Is there an explanation for the fact that more references hurt short-window calibration performance?
9 comments
AuthorsOpenReview ↗

Response to Reviewer TPUB (Part 1 of 4)

(Part 1/4)

Thank you for the detailed review. Your question on why more references hurt short-window accuracy was especially useful. Answering it helped us understand that what matters is how spread out the reference forces are, not how many there are. We answer each point below, and the corresponding changes are in the revised paper.

Writing

The sensor and the meaning of the digital reading.

Each tactile array is a piezoresistive sensor made of a commonly used conductive fabric placed between two electrode layers, a construction that is well suited to production at scale. The sensing elements (taxels) lie on a 7 by 8 grid with a pitch of 3.5 mm. The two upper corners, each a 2 by 2 block, are left out, which gives 48 taxels per array. The array is about 4 cm by 3 cm and covers the whole finger, including the fingertip and its sides. The digital reading is the ADC readout. It follows the resistance of the taxel and increases with force. To answer your question, it is a resistance-based reading and not a current or a capacitance. Force is applied to one taxel at a time with a plastic indenter that has a flat tip of 3 mm diameter. This description, with photographs of the setup, is now in the new Appendix D.1 of the revised paper (new Figure 5).

Q1. How is a sensing location defined?

Yes, a sensing location is an individual taxel. It is one of the 48 taxels of an array, the one pressed by the indenter. "Taxel" and "sensing element" mean the same thing, and a "sensing location" is the taxel that is being calibrated and tested. Different sensing locations on an array are different taxels. An array has many taxels, and the question this paper studies is how to calibrate all of them quickly once the tactile array has been assembled, before it leaves the factory. The revised paper gives this definition where the term first appears, in Section 3.

"Calibration" used for two things (Sections 4.2 and 4.3).

The word was indeed used for two different things. In the revised paper, Section 4.2 gives the gain-and-offset fit used during training its own name, the episode-level affine fit, and states that it is different from the calibration at a new location in Section 4.3.

Intuition for the four experts (Section 4.1).

You asked what each expert captures and why these four are needed. We added an explanation of each expert to Section 4.1, which we summarize here.

  • Static. It uses only the current reading and represents the memoryless response of the sensor. The other three experts add memory on top of it.
  • Path-dependent. It tracks whether the reading has recently been rising or falling, which indicates whether the sensor is on the loading or the unloading branch. The same reading corresponds to different forces on the two branches, as is typical of soft materials. The memory is measured in reading change and not in elapsed time, so it is independent of the loading rate.
  • Temporal relaxation. It compares the current reading with its recent average. The difference is large right after a load change and decays while the load is held, which corresponds to creep and stress relaxation, both typical of soft materials. It is the only expert whose memory depends on elapsed time.
  • Hysteresis. It behaves like a gear with backlash. It follows the reading only after the reading has moved far enough, so it records past loading reversals, both large and small. Such hysteresis is typical of soft materials.

Together the four experts cover the response without memory, the memory of loading direction, the memory of time, and the memory of past reversals.

"The paper is mostly descriptive." You noted that the paper mostly describes the method and restates numbers, with little interpretation. Sections 5.2 to 5.5 in the revised paper now explain the results. The explanations follow our answers below to Q2, Q3, and Q4 that you raised.

AuthorsOpenReview ↗

Response to Reviewer TPUB (Part 2 of 4)

(Part 2/4)

Results

Statistical uncertainty and equal weighting.

Every result now comes with a 95% confidence interval, the variation across training seeds, and a location-by-location comparison. Against each external baseline, the confidence interval of the difference excludes zero under both windows, and our method is better at 13 or 14 of the 14 locations. Our RMSE changes by only 0.012 N across training seeds, compared with 0.084 N for CAVIA. The full tables are in Point 1 of the general response and in the new Appendix E.

You also noted that equal weighting almost matches per-expert weighting. This is correct. Fitted weights are better by a small but real margin, at 11 of the 14 locations under both windows. Equal weights come close because locations differ mainly in how sensitive they are, and this sensitivity is exactly what the calibration measures. Point 2 of the general response gives the details. The effect of removing an expert is covered in Q3 below.

Questions

Q2. Why do concatenated features with a single MLP perform worse?

The short answer is that the single network fits the training arrays better but transfers worse to new arrays.

The comparison is fair. The concatenated-feature models receive the same 36 features as our experts, are trained with the same objective (Eq. 9), and are calibrated in comparable ways. So the difference does not come from the features, the training objective, or the calibration. It comes from how the features are combined. Two results show this.

Result 1. On the source arrays, with one array held out for validation, the single MLP reaches an RMSE of 0.183 N and our model 0.272 N, so the MLP is better there. On the three target arrays the order reverses. With 16 references the RMSE is 0.551 N for the MLP and 0.374 N for ours. With three references it is 0.769 N for the MLP and 0.375 N for ours.

Result 2. The MLP is also much less stable. Its RMSE changes by 0.112 N across training seeds, compared with 0.012 N for ours, which is nearly 90% lower.

Our explanation is as follows. A large network is flexible, so it picks up patterns that hold only in the training data. During training it saw only regular, repeated presses. The target recordings also contain large single presses and an irregular segment (new Figure 6), which it had never seen, so those patterns stop working. On top of this, the MLP has to set 65 output parameters from only three references, which is not enough. In our model each expert handles one kind of memory and sees only its own 12 features, so it cannot pick up such patterns across features. The four experts also complement each other, so their combination is more accurate than any single one. Finally, a new location needs only five parameters, and three references are enough for that. This structure is also modular. For another sensor type, such as a capacitive one, the physical characteristics of that sensor could in principle be added as a further expert, without changing the rest. We have not tested this yet, and we name it as future work. This discussion is now in Section 5.2 and Appendix F.5 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer TPUB (Part 3 of 4)

(Part 3/4)

Q3. Why does removing the static or temporal relaxation expert improve full-window accuracy?

The short answer is that it does not really improve. The change is within the noise.

With the full window, removing the static expert changes the RMSE by 0.003 N, and removing the temporal relaxation expert changes it by 0.008 N. Both confidence intervals include zero, and the full mixture is still better at 9 of 14 locations. With the short window, removing either expert makes the error larger at 12 of 14 locations.

One expert can be removed with little effect because the four experts are trained to predict the same force, so their outputs are strongly correlated (0.86 to 0.97). They still complement each other, which is why the mixture (0.375 N) is more accurate than the best single expert (0.487 N). When one expert is removed, the other three cover most of what it contributed. The hysteresis expert is the exception. Removing it makes the error clearly larger under both windows. All four experts are therefore worth keeping. This explanation is now in Section 5.5 of the revised paper.

Q4. Why do more references hurt short-window accuracy?

In the short window, more references do not add new information. With a fixed prior, they instead pull the fitted gain down.

How the calibration works. At a new location, five parameters (four weights and one offset) are fitted from K references. With K=3 the references alone cannot determine all five, so the calibration starts from a prior. The prior says to weight the four experts equally, with a gain of one and no offset, unless the references say otherwise. The strength of this prior is a single value. It was chosen on the source arrays and is used for every K, both windows, and all target locations.

What the references decide. The four experts give similar outputs (see Q3), so the references mainly decide one thing, the overall gain, which is how sensitive that location is. The more the reference forces differ from each other, the more reliably they set the gain. References at similar forces leave the gain close to the prior.

What happens in the short window. The short window ends 30 s after contact, which is typically before the first press reaches its peak (8 s before, on average). The window therefore covers only the initial loading ramp and contains no unloading. With K=3 the references are taken at the start, the middle, and the end of the ramp, which is the most spread-out choice. A larger K adds references in between, so the reference forces become less spread out. The prior then has more influence, the fitted gain moves towards one, and the estimates become too low.

KShort: SD of reference forces (N)Short: fitted gainShort: bias (N)Short: RMSE (N)Full: fitted gainFull: RMSE (N)
31.301.71-0.0120.3972.000.375
81.041.55-0.0550.4201.820.376
160.971.51-0.0650.4331.820.374
320.941.50-0.0670.4371.840.380

The table shows the averages, and the same pattern holds at the individual locations. From K=3 to K=16, the fitted gain decreases at all 14 locations, and the RMSE increases at 12 of them. The full window contains the complete press, including unloading. There the references stay spread out, the gain stays near 1.8, and the RMSE stays flat (0.371 to 0.382 N for K from 3 to 32).

What this means in practice. First, three references already use what a 30 s window can offer. The short-window RMSE with K=3 (0.397 N) is within 0.022 N of the 2-minute window (0.375 N). Second, what matters is how spread out the reference forces are, not how many there are. Third, the trend comes from the fixed strength of the prior and not from the experts. As a check, a weaker prior (0.1 in place of 1) reverses the short-window trend (0.457 N at K=3 and 0.407 N at K=16). This is only a check. All results in the paper keep the value chosen on the source arrays. This analysis is in Section 5.3 and Appendix F.1 of the revised paper (new Figure 8).

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

This paper presents a few-shot calibration framework for soft tactile sensors. It learns four response models from source sensor readings each with a small neural readout, and at a new sensing location only fits their output combination using a few initial force references. After calibration, all parameters stay fixed while the memory states keep updating from the readings, and the method outperforms the evaluated baselines on force estimation at sensing locations from three unseen sensor arrays.

Critical strengths
  1. Practical and low-cost calibration: Calibrating tactile sensors is a real bottleneck for practical usage of the tactile sensors, since each sensing element responds differently and depends on its loading history. The proposed framework only needs three force references at each new sensing location, and the whole model is light-weight, with all parameters fixed after calibration. This makes it realistic to deploy the system in practice.

  2. Consistent improvement over baselines: The method outperforms all evaluated baselines, reducing RMSE by around 40% compared to the baseline method, and it stays better on all three unseen sensor arrays. The performance is also consistent as the number of references increases, and it still outperforms the baselines when the references are selected at fixed times. This suggests that modeling the loading history with shared models transfers well to new sensors.

Critical weaknesses
  1. The ablations do not clearly support the key design choices: The main idea of the paper is to calibrate the combination of four shared experts at each new sensing location. However, most of the improvement comes from averaging the experts with a simple gain and offset, while fitting the combination weights locally only reduces the error slightly, and weights learned on the source arrays even perform better under the full calibration window. Removing the static or temporal relaxation expert also slightly improves the full-window error, and on irregular loading both equal weights and removing the hysteresis expert outperform the full model. It would be interesting to report the results with confidence intervals across sensing locations and training seeds, to show which components are actually necessary.

  2. The evaluation is limited compared with the claims: The paper motivates the method with large-scale manufacturing and accurate force estimation throughout operation, but the evaluation only covers 14 sensing locations on three arrays of one sensor type, with short recordings (around 15 minutes in Figure 2) and only normal force. Long-term drift and other sensor types are not evaluated, and the sensor hardware itself is not described in the main paper. Moreover, the three references are selected from a dense initial force recording, which still requires the force measurement setup during calibration. It would be interesting to evaluate on more sensing locations and sensor types, and over longer periods of use, to support the claim that the shared models transfer to new sensors.

Questions for authors

Would it be possible to evaluate the proposed method on more sensing locations, other sensor types, or recordings over longer periods of use? If the performance remains similar in these settings, it would support the claim that the framework is practical for large-scale manufacturing and long-term force estimation.

2 comments
AuthorsOpenReview ↗

Response to Reviewer srNx (Part 1 of 2)

(Part 1/2)

Thank you for the review and for the concrete suggestions. Your request for confidence intervals was especially useful. Adding them showed which differences in the ablations are real and which are within the noise. We answer each point below, and the corresponding changes are in the revised paper.

W1. Do the ablations support the design choices?

You asked for confidence intervals across sensing locations and training seeds, to show which components are needed. Every result now has them. We summarize the findings here, so this response can be read on its own. The full tables are in Points 1 and 2 of the general response and in the new Appendix E.

  • Against the baselines. Our method is better at 13 or 14 of the 14 locations under both windows, and the confidence interval of every difference excludes zero. Its RMSE changes by only 0.012 N across training seeds, compared with 0.084 N for CAVIA.
  • Fitted weights versus equal weights. Fitted weights are better. The gain in RMSE is small (0.007 N with the full window, 0.004 N with the short window), but it shows up at 11 of 14 locations under both windows.
  • Fitted weights versus weights learned on the source arrays. With the full window the two perform the same (the confidence interval includes zero). With the short window fitted weights are clearly better, at 12 of 14 locations.
  • Removing an expert. All four experts are worth keeping. With the short window, removing any one of them makes the error larger at 12 or 13 of 14 locations. With the full window, removing the static or temporal relaxation expert makes no real difference (0.003 and 0.008 N, within the confidence intervals). Removing the hysteresis expert makes the error clearly larger under both windows.

You observed that most of the improvement comes from averaging the experts with a simple gain and offset. This is correct, and looking into it helped us describe the calibration step more precisely. The four experts are mixed in almost the same proportions at every location, while the overall gain ranges from 1.3 to 3.5. In other words, locations differ mainly in how sensitive they are, and this sensitivity is exactly what calibration is meant to measure at each sensing location. The revised paper says this in Section 5.5 and in the conclusion.

Irregular loading. You also noted that on irregular loading both equal weights and removing the hysteresis expert do better than the full model. This is true with the full window, and we looked into the reason. The cause is the level of the estimate, not a harmful expert. With the full window the full model reads slightly too high on this segment, by 0.15 N on average. Equal weights are lower by only 0.005 N. The three removals that help are exactly the ones that reduce this offset, and the one removal that does not reduce it does not help. With the short window the offset is only 0.03 N, and no removal gives a real improvement. Removing the hysteresis or temporal relaxation expert then makes the RMSE on this segment larger, by 0.042 and 0.033 N. Section 5.4 and Appendix F.3 of the revised paper now explain this.

AuthorsOpenReview ↗

Response to Reviewer srNx (Part 2 of 2)

(Part 2/2)

W2 and Question. More arrays, longer recordings, and the calibration setup.

More arrays. We evaluated the submitted model, with nothing changed or tuned, on a further tactile array with 48 sensing elements. This array was not used for training or for selecting the model. It has two recordings at one sensing location, with 50 and 35 loading cycles. We also ran the baselines of Table 1 in the paper on it, with the same three references. D_Former was not rerun for this array.

MethodFull window RMSE (N)Short window RMSE (N)
Classical PI0.2370.405
Raw + GRU0.4320.477
Concat features + MLP0.3090.528
Concat features + GRU0.4100.422
TDNN + OT0.7700.491
CAVIA0.2270.337
Ours0.1330.178

Our method has the lowest RMSE in both recordings and under both windows. Compared with CAVIA, which is again the best baseline, the RMSE is 41% lower with the full window and 47% lower with the short window. This is in line with the three target arrays in the paper (42% and 35%). The RMSE is 6% and 9% of the peak force in the two recordings.

Longer recordings. The longer recording on this array runs for 20 minutes after calibration. Its RMSE in successive intervals (0 to 3, 3 to 6, 6 to 9, 9 to 13, and 13 to 20 minutes) is 0.18, 0.13, 0.12, 0.12, and 0.16 N, and the average error stays within 0.02 N. The error does not grow over these 20 minutes. If the response of a sensor changes over longer use, a location can be calibrated again with three new references, which takes less than two minutes. You raise an important point about long-term use, and we fully agree with it. Unlike vision systems, whose calibration stays valid once it is done, soft materials are hard to keep calibrated, because their physical properties change over time, and the calibration may then have to be repeated. This has to be solved before soft sensors can be used at scale in real applications, and not only in experimental settings. This is what motivates our work, which has two aims. The first is to shorten the calibration time. The second is to follow how the calibration changes by modeling the physical characteristics of the sensor. This paper addresses the first aim, and the shared experts are our starting point for the second.

Force measurement during calibration. Calibration at a new location needs only three force readings, each with a rough time stamp. It does not need a continuous force recording. Two results show this. First, we took the three references at fixed times, without looking at the force at all. Our method still works. Its RMSE is 0.415 N, compared with 0.597 N for CAVIA, the best baseline, with the same references. So a dense force recording is not needed to choose the references. Second, we deliberately made the three references inaccurate. A timing error of ±2 s increases the RMSE by at most 0.016 N, and force noise of 0.2 N by at most 0.007 N. In practice, a hand-held force gauge read at three moments is enough. The details are in Point 3 of the general response and in the new Appendix D.3.

Sensor hardware. Each tactile array is a piezoresistive sensor made of a commonly used conductive fabric placed between two electrode layers. It has 48 taxels on a 7 by 8 grid with a pitch of 3.5 mm, with a 2 by 2 block left out at each upper corner. The array is about 4 cm by 3 cm and covers the whole finger. The digital reading is the ADC readout and follows the resistance of the taxel. Force is applied to one taxel at a time with a plastic indenter that has a flat tip of 3 mm diameter. This description, with photographs of the setup, is now in the new Appendix D.1 (new Figure 5). An array has 48 sensing elements of the same type, yet each one has to be pressed and calibrated on its own. The calibration time therefore matters for the manufacturing setting you mention. A full force sweep takes about 14 minutes per taxel, which is about 11 hours per array. Three references take 30 s to 2 min per taxel, which is 24 minutes to 1.6 hours per array. This could be shortened further by pressing faster.

Other sensor types. Our test setup applies normal force only, and the sensors available to us are of one type, so other sensor types are beyond what we could test. The structure is modular, however. The physical characteristics of another sensor type could in principle be added as a further expert, without changing the rest.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 1 of 4)

(Part 1/4)

We sincerely thank the reviewers for their careful reading. Three concerns were shared by the reviewers. First, the results were reported without confidence intervals over sensing locations and training seeds. Second, it was unclear how much fitting the expert weights at each location actually helps. Third, it was unclear how much force measurement calibration requires in practice. We address these three here and respond to the remaining points under each review. None of the numbers reported in the paper has changed. Points 1 to 4 below are new analyses of the submitted model. The experts, the three training seeds, and the calibration settings are exactly those of the submission, and nothing was retrained. Point 5 is a new ablation experiment. We retrained the experts w/ and w/o the proposed training objective, keeping everything else the same, to measure how much that objective contributes.

Point 1. Uncertainty (srNx W1, TPUB Results, dtin Q2)

The reviewers asked for uncertainty estimates on the reported errors. Every K=3 result now comes with a 95% confidence interval. We also show how each method varies across training seeds and how often ours wins across the sensing locations.

How the numbers are computed. Errors are averaged as in the paper. For each confidence interval, we redraw the 14 locations and the three seeds at random with replacement, recompute the average, and repeat this 10,000 times. The interval is the range that holds the middle 95% of these values (a percentile bootstrap). When two methods are compared, both use the same redrawn locations and seeds, so harder and easier locations do not distort the difference (a paired comparison). The p-values are from a two-sided Wilcoxon signed-rank test over the 14 locations.

MethodFull: RMSE [95% CI]Full: Δ vs Ours [95% CI]Short: RMSE [95% CI]Short: Δ vs Ours [95% CI]Ours wins at (full / short)Seed SD (full / short)
Classical PI0.820 [0.702, 0.943]+0.445 [+0.350, +0.535]0.631 [0.529, 0.742]+0.234 [+0.195, +0.272]14/14, 14/14— / —
Raw + GRU1.058 [0.696, 1.685]+0.683 [+0.358, +1.298]0.987 [0.749, 1.273]+0.590 [+0.342, +0.886]14/14, 14/140.421 / 0.149
Concat features + MLP0.769 [0.607, 0.927]+0.393 [+0.268, +0.511]0.915 [0.724, 1.119]+0.518 [+0.389, +0.653]14/14, 14/140.112 / 0.032
Concat features + GRU0.748 [0.583, 0.939]+0.372 [+0.225, +0.556]1.395 [0.768, 2.135]+0.998 [+0.358, +1.718]14/14, 14/140.128 / 0.664
TDNN + OT0.680 [0.614, 0.758]+0.304 [+0.221, +0.386]0.850 [0.802, 0.898]+0.453 [+0.359, +0.542]13/14, 13/140.001 / 0.009
D_Former0.963 [0.819, 1.114]+0.588 [+0.475, +0.704]0.963 [0.820, 1.117]+0.566 [+0.472, +0.662]14/14, 14/140.021 / 0.021
CAVIA0.652 [0.537, 0.770]+0.277 [+0.198, +0.358]0.607 [0.496, 0.729]+0.210 [+0.140, +0.283]14/14, 14/140.084 / 0.070
Ours0.375 [0.314, 0.440]—0.397 [0.314, 0.490]——0.012 / 0.012

RMSE in N, K=3. "Ours wins at" is the number of locations, out of 14, where our error is lower. "Seed SD" is how much the overall error changes when only the training seed changes. The classical PI model has no randomness, so it has no seed SD.

Under both windows, our method is more accurate than every external baseline, and the 95% confidence interval of each difference excludes zero. It is better at 13 or 14 of the 14 locations. With the full window, the RMSE of our method varies by only 0.012 N across training seeds. This is 7 times less than for CAVIA and 9 to 35 times less than for the GRU- and MLP-based baselines. We added the same information to the ablation tables. They are shown in Point 2 below and in Appendix E of the revised paper.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 2 of 4)

(Part 2/4)

Point 2. What local weight fitting contributes (srNx W1, TPUB Results, dtin W2/Q2)

We agree with the reviewers, and we thank them for raising this point. Most of the improvement over a single expert comes from combining the four experts with a fitted gain and offset, and fitting the weights at each location adds a small amount on top. The submission notes this in Sections 5.5 and 6, but it did not quantify it. Following the reviewers' comments, we looked at it more closely. The extra gain is small but real. Fitted weights win over equal weights at 11 of the 14 locations under both calibration windows, and they need no extra training. The table below compares each variant with our method, and the three bullets after it explain what the numbers mean.

VariantFull: Δ vs Ours [95% CI]Ours wins atpShort: Δ vs Ours [95% CI]Ours wins atp
Equal weights (gain + offset fitted)+0.007 [+0.004, +0.011]11/140.004+0.004 [+0.002, +0.007]11/140.017
Source-learned proportions-0.009 [-0.022, +0.002]9/140.583+0.011 [+0.003, +0.020]12/140.013
w/o static-0.003 [-0.013, +0.005]9/140.426+0.015 [+0.007, +0.023]12/140.005
w/o path+0.012 [-0.004, +0.028]9/140.119+0.033 [+0.021, +0.045]12/140.001
w/o temporal relaxation-0.008 [-0.023, +0.008]9/140.583+0.009 [-0.002, +0.021]12/140.042
w/o hysteresis+0.044 [+0.022, +0.067]11/140.003+0.064 [+0.042, +0.087]13/14<0.001
Path expert alone+0.111 [+0.044, +0.189]10/140.013+0.128 [+0.049, +0.213]11/140.005

Δ is the variant's RMSE minus ours, in N. A positive value means ours is lower.

  • Fitted weights versus equal weights. Fitted weights are better. The gain in RMSE is small (0.007 N with the full window, 0.004 N with the short window), but it shows up at 11 of 14 locations under both windows, and both confidence intervals exclude zero.
  • Fitted weights versus weights learned on the source arrays. Fitted weights are actually the safer choice. With the full window the two perform the same (the confidence interval includes zero). With the short window fitted weights are clearly better, at 12 of 14 locations. They are also easier to use when many sensors need calibration, because they come directly from the three references at each new location and need no training beforehand.
  • Removing an expert. All four experts are worth keeping. With the short window, removing any one of them makes the error larger at 12 or 13 of 14 locations. With the full window, removing the static or temporal relaxation expert makes no real difference (0.003 and 0.008 N, within the confidence intervals). Removing the hysteresis expert makes the error clearly larger under both windows, which makes it the most important of the four.

Why equal weights come close. The reviewers noticed that equal weights almost match fitted weights. They are right, and looking into the reason taught us something useful for calibrating sensors in large numbers. The four experts are mixed in almost the same proportions at every location. Across the 14 locations each expert's share stays within a narrow band (static 0.18 to 0.22, path 0.25 to 0.30, relaxation 0.20 to 0.25, hysteresis 0.27 to 0.33). What changes from one location to the next is the overall gain, which ranges from 1.3 to 3.5, and the offset, which ranges from 0.06 to 0.69 N. In other words, locations differ mainly in how sensitive they are, and this sensitivity is exactly what calibration is meant to measure at each sensing location. With only three references, the calibration keeps the proportions close to equal and uses the data mainly to set the gain and offset. A new location therefore mainly needs its gain and offset. This makes the calibration step simpler than we had described, and it makes the method easier to scale. The shared experts take care of the memory of the loading history, so it does not have to be learned again at each location. The response model is shared, and each new location only needs an output calibration. This is the main idea of the paper, and the reviewers' comments helped us state it more clearly than in our submission. We thank them for this.

We have added this finding to Section 5.5 and to the conclusion of the revised paper.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 3 of 4)

(Part 3/4)

Point 3. How much force measurement calibration needs (srNx W2, dtin W1/Q1)

The reviewers asked how much force measurement calibration needs in real use, given that our experiments recorded force continuously, and whether the method really saves effort. The short answer is YES. Calibration at a new location needs only three force readings, each with a rough time stamp. It does not need a continuous force recording.

In our experiments we did record force continuously, and this may have caused the confusion. The continuous recording was used for two other things. It is the ground truth against which we measure the error, and with force-based selection it is also used to detect contact and to choose the three references. Two results below show that the calibration itself needs neither a continuous recording nor precise timing.

Result 1. Time-based selection is already in Table 1 in the paper. The three references are taken at fixed times, with no force threshold and no contact detection, and 20 of the 57 references fall at or below 0.1 N. The RMSE is 0.415 N, compared with 0.597 N for CAVIA using the same references. Ours is better at 12 of 14 locations.

Result 2. To answer how precise these three readings must be, we ran a new test, perturbing the three references. We deliberately corrupt the three (time, force) pairs and leave everything else unchanged. A timing error means that the force value is read some seconds before or after the sensor frame it is paired with. We test a constant offset shared by all three references, and also a separate random offset for each reference. We also add noise to the force values and a scale error to the force gauge. Random cases are repeated 50 times for every recording and seed. A good result in this table means two things. The numbers stay close to the first row, which has no error added, and they stay far below the last row, which is the best baseline with exact references.

Perturbation of the three referencesForce-based selectionTime-based selection
None (as in the paper)0.3750.415
Constant timing error, ±0.5 s0.375–0.3760.409–0.424
Constant timing error, ±1 s0.372–0.3760.404–0.428
Constant timing error, ±2 s0.372–0.3780.397–0.431
Independent timing error per reference, uniform ±1 s0.3750.417
Independent timing error per reference, uniform ±2 s0.3750.417
Force noise, SD 0.05 N0.3760.416
Force noise, SD 0.2 N0.3810.422
Force scale error, ±5%0.366–0.3910.416–0.418
Force scale error, ±10%0.362–0.4110.421–0.425
Timing ±0.5 s and force noise 0.05 N together0.3760.416
For comparison, CAVIA (best baseline) with exact references0.6520.597

RMSE in N (ours, K=3, full window). A range gives the result for the two signs of the error.

A timing error of ±2 s increases the RMSE by at most 0.016 N, and force noise of 0.2 N by at most 0.007 N. A scale error in the force gauge carries over to the estimates, which is true for any calibration. In practice, a new location therefore needs three force readings with time stamps accurate to about a second. Section 5.1 and the new Appendix D.3 now make clear what the method needs and what our experimental setup happened to record.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 4 of 4)

(Part 4/4)

Point 4. Sensor and data description (srNx W2, TPUB Writing, dtin W3)

The reviewers asked for a description of the sensor and of the recordings. Both are now in the appendix of the revised paper.

Sensor. Each tactile array is a piezoresistive sensor made of a commonly used conductive fabric placed between two electrode layers, a construction that is well suited to production at scale. The sensing elements lie on a 7 by 8 grid with a pitch of 3.5 mm. The two upper corners, each a 2 by 2 block, are left out, which gives the 48 sensing elements. The array is about 4 cm by 3 cm and covers the whole finger, including the fingertip and its sides. The digital reading is the ADC readout. It follows the resistance of the sensing element and increases with force. Force is applied to one sensing element at a time with a plastic indenter that has a flat tip of 3 mm diameter. The full description, with photographs of the setup, is in the new Appendix D.1 (new Figure 5).

Sensing location. A sensing location is a single sensing element of an array, pressed by the indenter. The revised paper gives this definition where the term first appears, in Section 3.

Recordings. The peak force in the evaluated part ranges from 1.6 to 6.5 N across the 19 long recordings. The loading is slow to moderate, with the force changing by 0.11 N/s at the median and by at most 1.50 N/s. Our RMSE of 0.375 N is about 10% of the peak force, compared with 17.5% for CAVIA, the best baseline. The details are in Appendix D.2 and the new Figure 6, and in our response to Reviewer dtin.

Point 5. New ablation on calibration-aware training

A reviewer asked what the paper contributes to machine learning. One of our contributions is the way the experts are trained. Each expert is trained to be accurate after calibration, and not simply to predict force. The submission did not test how much this matters, so we ran a new ablation.

We trained the four experts twice from the same starting point. One version uses our training objective (Eq. 9 in the paper, called calibration-aware below). The other uses plain regression on force. Everything else is identical, including model, data, order of training samples, optimizer, number of updates (1,800), and seeds. We fixed all of these before looking at any result. The evaluation on the target arrays is the same as in the paper.

How the experts are trainedFull window RMSE [95% CI]Short window RMSE [95% CI]Seed SD (full / short)
Plain regression0.652 [0.546, 0.764]0.679 [0.576, 0.794]0.028 / 0.032
Our training objective (Eq. 9)0.376 [0.315, 0.442]0.404 [0.318, 0.500]0.007 / 0.012
Improvement from our objective0.275 N (42% lower), better at 14/140.275 N (41% lower), better at 14/14
Model reported in the paper, for comparison0.3750.3970.012 / 0.012

RMSE in N on the three target arrays. The last row is the model from the submitted paper, which the new training run reproduces.

Our training objective lowers the RMSE by 0.275 N under both windows (42% with the full window and 41% with the short window), and it is better at all 14 locations. Three more observations support this.

  • Plain regression fits the source arrays better but works worse on new arrays. The gap also grows with more training (0.071 N after 450 updates, 0.201 N after 900, and 0.275 N after 1,800). Plain regression learns the absolute scale of the source arrays, and a new location has a different scale.
  • The result does not depend on the regularization strength used at calibration. We tried four values (0.01, 0.1, 1, and 10), and plain regression was best at 1, the value used in the paper.
  • Training from scratch with our objective gives almost the same error as the submitted model (0.376 against 0.375 N, and 0.404 against 0.397 N). The numbers in the paper therefore do not depend on one particular set of trained weights.

This experiment is in Section 5.5 and Appendix F.4 of the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.