acceptodds
Under review as a conference paper at ICLR 2027

SimpleTouch: Can Vision-Language-Action Models Master Contact-Rich Manipulation Without Tactile Policy Pretraining?

Abstract

Tactile sensing provides essential contact information for robotic manipulation, yet incorporating it into pretrained vision-language-action (VLA) models remains challenging. A common concern is that simply introducing touch during task-specific fine-tuning may fail to bridge the cross-modal gap, yielding limited gains or even reduced success. Consequently, existing methods often rely on large-scale tactile policy pretraining or separate visuotactile alignment, adding data requirements and training stages. We introduce SimpleTouch, a simple VLA extension that augments π0.5 with a tactile expert, to test whether these additional stages are necessary. Leveraging all tokens from a frozen pretrained tactile encoder, the expert learns from action supervision and multi-horizon prediction of future tactile latents. This single-stage training uses only task demonstrations, without additional tactile policy pretraining or separate alignment. With 50 demonstrations per task, SimpleTouch achieves the highest success rate among evaluated methods on all six UniVTAC tasks. Its average success rate reaches 77.5%, compared with 45.2% for FTP-π0.5 and 66.7% for FTP-1, corresponding to gains of 32.3 and 10.8 percentage points, respectively. Across four real-world tasks, it averages 71.3%, exceeding FTP-1 by 8.8 percentage points. These results demonstrate that, given pretrained VLA and tactile representations, additional tactile policy pretraining is not a prerequisite for strong performance on these tasks, offering a simpler route to contact-rich manipulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
tactile, vision-language-action models, robot manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5M7dOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SimpleTouch, a tactile extension of π0.5\pi_{0.5} for contact-rich manipulation without additional tactile policy pretraining or a separate visuotactile alignment stage. The method preserves all current tactile tokens from a frozen T3 encoder and uses a tactile expert to provide context for action generation. Multi-horizon future tactile prediction supplies auxiliary supervision during training, while inference uses only current observations. Experiments on six UniVTAC tasks and four real-robot tasks show improvements over the evaluated baselines with 50 demonstrations per task.

Critical strengths
  1. Learning to use touch from task demonstrations while reusing pretrained VLA and tactile representations is a useful practical direction. SimpleTouch provides a relatively simple recipe and outperforms FTP-1 in average success on both the simulation and real-robot evaluations.
  2. The separation between predictive supervision and action conditioning is clearly described. The simulation ablations examine spatial tactile detail, encoder freezing, and future prediction, while the Joint/IDM/Fast comparison supports using prediction during training without requiring future tactile generation at deployment.
Critical weaknesses
  1. The data-efficiency evaluation does not establish how the proposed recipe compares with tactile policy pretraining when task demonstrations are scarce. Section 4.4 compares only with FTP-π0.5\pi_{0.5}, which has not undergone tactile policy pretraining. Although the main experiments include FTP-1 at 50 demonstrations, its performance at 10 or 20 demonstrations is not reported. Since reducing the need for additional tactile policy pretraining is central to the paper, this comparison is important for understanding the conditions under which the proposed alternative remains effective.
  2. The component ablations in Section 4.3 are limited to simulation, leaving the benefit of future tactile prediction on real hardware unverified. Representative real-robot ablations would help isolate the contribution of the proposed tactile learning recipe.
  3. Appendix B.2 cites related predictive tactile methods, including OmniVTA, Dream-Tac, and UniTacVLA, but provides no experimental comparison with them. The internal Joint/IDM/Fast variants do not reproduce these complete methods and therefore do not establish the relative benefits of SimpleTouch over existing predictive tactile approaches. Direct comparisons with these related methods would strengthen the evidence.
Questions for authors
  1. Can the authors provide FTP-1 results at the lower demonstration budgets in Figure 6, particularly 10 and 20 demonstrations? This would clarify whether the advantage over a tactile-policy-pretrained model persists when task data are more limited.
  2. Does removing future tactile prediction reduce real-robot performance when the current-touch pathway and training data are held fixed? A comparison on representative tasks such as Wipe Board or Pick Up Chips would help establish whether the predictive objective contributes beyond the tactile input itself.
  3. Please supplement the evaluation with direct comparisons against related predictive tactile methods such as OmniVTA, Dream-Tac, and UniTacVLA.
  4. The authors are encouraged to cite VTAM [1] and TacWAM [2] and briefly clarify their relationship to SimpleTouch. VTAM incorporates touch into a pretrained video model without independent tactile pretraining, while TacWAM uses future tactile supervision without allowing the action branch to directly read future prediction tokens.

[1] Haoran Yuan et al. VTAM: Video-Tactile-Action Models for Complex Physical Interaction Beyond VLAs. arXiv:2603.23481, 2026.

[2] Lei Jin et al. TacWAM: Anchor-Guided World Action Model with Mechanics-Aware Tactile Prediction. arXiv:2607.28391, 2026.

7 comments
AuthorsOpenReview ↗

1. Low-Data Evaluation and the Central Question (W1 & Q1)

Thank you for recognizing the practical value of our work. SimpleTouch’s defining feature is simplicity and effectiveness in both architecture and training: it reuses a strong pretrained VLA and an off-the-shelf tactile encoder, adding touch through a single stage of task post-training. Tables 1–2 establish that 50 demonstrations per task suffice to exceed FTP-1 without additional tactile policy pretraining, while substantially outperforming strong baselines that also omit that stage. We provide the main results and controlled ablations below.

New FTP-1 results at 10/20 demonstrations complete the comparison with the original 50-demonstration results. At each budget, methods use identical demonstration subsets and update counts: 4k/8k/20k steps for 10/20/50 demonstrations, respectively, always reporting the final checkpoint and 100 evaluation trials per setting. Both FTP-1 and SimpleTouch improve monotonically from 10 to 50 demonstrations.

Table A shows that SimpleTouch exceeds FTP-1 in all six task–budget settings. At 20 demonstrations, for example, Put Bottle/Insert Hole achieve 38%/57% versus FTP-1's 30%/48%. The performance advantage without additional tactile policy pretraining therefore extends to the 10- and 20-demonstration settings, while the original advantage over FTP-π0.5 is retained.

Table A. Low-data comparison with added FTP-1 results (Figure 6 + new experiments; %; 100 trials per setting).

TaskMethod10 demos20 demos50 demos
Put BottleFTP-π0.552419
Put BottleFTP-1223047
Put BottleSimpleTouch233850
Insert HoleFTP-π0.5142047
Insert HoleFTP-1304864
Insert HoleSimpleTouch435785
AuthorsOpenReview ↗

2. Prediction Supervision and Hardware Evidence (W2 & Q2)

Unfortunately, organizational and resource changes beyond our control have made the original robot platform unavailable, preventing further hardware ablations. We deeply regret this situation. However, we had already conducted the requested comparison on Pick Up Chips while the platform was available: removing only the future tactile prediction loss, with the current-touch pathway, training data, and all other training and execution settings unchanged, reduced success from 70% (14/20) to 50% (10/20), a 20-pp drop. This agrees with the simulation ablations (Table B, from manuscript Table 3).

Table B. Prediction ablation with current touch fixed (manuscript Table 3; simulation, %).

Method / metricPull-out KeyPut BottleInsert HoleInsert TubeMean
No prediction6037839568.8
Full62508510074.3
Gain (pp)+2+13+2+5+5.5

We chose Pick Up Chips because it requires a careful balance between maintaining a stable grasp and avoiding damage to fragile chips. Crushing was the main failure mode after removing prediction supervision. This controlled comparison directly supports a benefit of the predictive objective beyond current tactile input. The failure pattern is also consistent with our motivation: training-time future-touch supervision encourages the shared expert to learn how actions affect subsequent contact, supporting more appropriate grasping actions. In simulation, removing only the prediction loss likewise lowers success on all four tasks, from 74.3% to 68.8% on average.

AuthorsOpenReview ↗

3. Direct Comparison with Predictive Tactile Methods (W3 & Q3)

We strongly value direct comparisons with these predictive tactile methods and pursued their reproduction before preparing the manuscript. However, OmniVTA has not released its model implementation, and the UniTacVLA repository still states “coming soon.” We also found no public UniVTAC implementations or evaluation results for either method. These missing resources prevented reliable reproduction and fair direct comparisons. For Dream-Tac, a closely related method with a public implementation, we reproduced it using Dream-Tac’s official implementation and training hyperparameters. On UniVTAC, both methods use the same 50 demonstrations per task and are evaluated over 100 trials per task (Table C).

Table C. Dream-Tac vs. SimpleTouch on UniVTAC (success rate, %).

MethodPull-out KeyPut BottleInsert HoleInsert TubeMean
Dream-Tac2644419150.5
SimpleTouch62508510074.3

SimpleTouch outperforms Dream-Tac on all four tactile-sensitive tasks, increasing mean success from 50.5% to 74.3%, a 23.8-pp gain (Table C), further validating the effectiveness of our simple post-training approach. We attribute this advantage to the combination of a strong backbone and effective tactile adaptation: π0.5 provides a strong foundation for vision-language understanding and action generation, while SimpleTouch equips it to act on tactile feedback through encoder freezing, full tactile tokens, and multi-horizon predictive supervision. The comparison with FTP-π0.5 using the same backbone and the controlled ablations further support the roles of these designs (Table 1, Figure 4, and Table 3).

Reviewer_E2wqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SimpleTouch, a tactile-enhanced extension of the pretrained π0.5 vision-language-action (VLA) model. It introduces a separate tactile expert that processes all tokens from a frozen pretrained T3 tactile encoder and integrates tactile features into action generation through layerwise attention. Multi-horizon future tactile prediction provides auxiliary training supervision without requiring additional tactile policy pretraining or explicit visuotactile alignment. With 50 demonstrations per task, SimpleTouch achieves an average success rate of 77.5% across six simulation tasks and 71.3% across four real-robot tasks. The results are promising, although the novelty, fairness of comparisons, and robustness of the empirical conclusions require further examination.

Critical strengths

1.The method offers a relatively simple way to incorporate tactile feedback into pretrained VLA models without additional tactile policy pretraining.

2.The tactile expert preserves spatial tactile information and provides a dedicated pathway for tactile-conditioned action generation.

3.Experiments support the benefits of retaining all tactile tokens, freezing the pretrained encoder, and using multi-horizon future tactile prediction.

4.The paper evaluates six simulation tasks and four real-robot tasks, reporting improvements over several relevant baselines.

Critical weaknesses

1.The approach primarily combines a separate tactile expert, layerwise attention, and auxiliary future prediction. Its fundamental advance over existing tactile-VLA and predictive-policy methods is not sufficiently established.

2.SimpleTouch uses shorter execution horizons for insertion tasks, while baselines use their respective default settings. Since feedback frequency substantially affects success, the main comparisons do not fully isolate the contribution of the proposed architecture.

3.Real-robot experiments use only 20 trials per task, and simulation results use 100 episodes per task. The paper does not sufficiently characterize uncertainty across training seeds or demonstrate the statistical significance of the reported gains.

4.SimpleTouch differs from FTP-π0.5 in several aspects, including encoder freezing, tactile token usage, and future-touch supervision. Although individual ablations are provided, the relative contributions and interactions of these components remain unclear.

5.Real-world experiments cover four tasks using one robot platform and GelSight tactile sensors. Generalization to unseen tasks, objects, sensor configurations, and robot embodiments remains untested.

Questions for authors

1.Can the authors provide comparisons under matched execution horizons and feedback frequencies to isolate the benefit of SimpleTouch's architecture?

2.How does SimpleTouch compare with recent tactile-VLA methods under identical backbones, demonstrations, training budgets, and inference settings?

3.Can the authors report results across multiple training seeds and provide confidence intervals or statistical tests for the main improvements?

4.What are the individual and combined contributions of encoder freezing, full tactile-token preservation, and multi-horizon prediction?

5.Does SimpleTouch generalize to unseen objects, tasks, tactile sensors, or robot embodiments without additional task-specific training?

8 comments
AuthorsOpenReview ↗

1. Contribution and Novelty (W1)

Thank you for recognizing the simplicity of our approach and the existing ablations. SimpleTouch’s central innovation is strong tactile manipulation through a simple architecture and simple training: single-stage post-training on only 50 demonstrations per task equips a strong existing VLA with touch and exceeds pretrained FTP-1 in mean success. The main results and individual and combined ablations below establish this advance and the contributions of the key designs.

To our knowledge, SimpleTouch is the first VTLA with an FTP-1-style tactile-expert architecture to exceed FTP-1 in mean success in both simulation and real-world evaluation without additional tactile policy pretraining (Tables 1–2). The innovation centers on a simple architecture, simple training, and strong performance: a shared tactile expert and off-the-shelf representations equip the strong π0.5 backbone with touch, achieving this result through single-stage post-training on only 50 demonstrations per task.

The shared high-level architecture makes the design choices particularly consequential: both methods use T3 and a tactile expert (Table 4), yet SimpleTouch raises simulation success from 45.2% for FTP-π0.5 to 77.5% (+32.3 pp; Table 1). Individual ablations establish the roles of full tokens, encoder freezing, and predictive supervision (Figure 4; Table 3). Our new combined ablation further shows that restoring the three designs within the same implementation raises four-task mean success from 50.0% to 74.3% (+24.3 pp; Table C). We introduce a novel framework that unifies tactile dynamics modeling and action learning within a VLA architecture. A shared tactile expert and asymmetric attention allow predictive supervision to improve action learning while deployment uses only current observations, without generating future touch. This extends Fast-WAM's training–inference decoupling principle to single-stage tactile learning within a VLA (Section 3.4; Figure 2).

AuthorsOpenReview ↗

2. Matched Execution Settings and Fair Comparisons (W2, Q1 & Q2)

To address comparison fairness, we have fully matched SimpleTouch to FTP-1 and FTP-π0.5: training H32, execution H20, and identical temporal ensembling (Table A). The original training H50 follows the official π0.5 default. FTP-π0.5 and SimpleTouch share the π0.5 backbone, T3 encoder, and the same 50 demonstrations per task; all methods train for 20k steps and report the final checkpoint.

Table A. Matched H32/H20 and temporal ensembling (four contact-sensitive tasks; %; 100 trials per setting).

MethodPull-out KeyPut BottleInsert HoleInsert TubeMean
FTP-π0.53019477242.0
FTP-14847647959.5
SimpleTouch66459410076.3

On the four contact-sensitive tasks where tactile feedback is particularly important—Pull-out Key, Put Bottle, Insert Hole, and Insert Tube—SimpleTouch achieves 76.3% mean success, versus 59.5% for FTP-1 and 42.0% for FTP-π0.5, gains of 16.8 and 34.3 pp. Together with the controlled ablations in Figure 4 and Table 3, this comparison shows that SimpleTouch's advantage comes from preserving full pretrained tactile representations and learning contact dynamics through multi-horizon predictive supervision, rather than differences in horizon or ensemble settings.

AuthorsOpenReview ↗

3. Training-Seed Stability and Evaluation Protocol (W3 & Q3)

Our main evaluation follows FTP-1: 100 simulation rollouts and 20 hardware trials per method/task, with main results reported as repeated-rollout success rates [1]. To directly examine sensitivity to training seeds, we nevertheless added two training seeds on each of Pull-out Key and Insert Hole, yielding three seeds per task including the original result (Table B).

Table B. SimpleTouch stability across training seeds (success rate, %; 3 seeds per task).

TaskOriginal seedAdded seed 1Added seed 2Mean ± sample SDRange (pp)
Pull-out Key62616462.3 ± 1.53
Insert Hole85858484.7 ± 0.61

Pull-out Key achieves 62%, 61%, and 64% across seeds, with a mean of 62.3% and sample standard deviation of 1.5 pp; Insert Hole achieves 85%, 85%, and 84%, with a mean of 84.7% and sample standard deviation of 0.6 pp. The ranges are only 3 and 1 pp, respectively, indicating limited seed sensitivity on these two tasks and supporting the stability of our post-training recipe. The main tables also include 4,800 simulation and 400 hardware evaluations; hardware uses 20 task-specific policies with 20 trials each, consistent with the repeated-rollout scale of VTAM, DT-TacVLA, and TacWAM [2–4].

References

[1] FTP-1: A Generalist Foundation Tactile Policy Across Tactile Sensors for Contact-Rich Manipulation.

[2] VTAM: Video-Tactile-Action Models for Complex Physical Interaction Beyond VLAs.

[3] DT-TacVLA: A Dual-Timescale Tactile-Enhanced Vision-Language-Action Model for Contact-Rich Manipulation.

[4] TacWAM: Anchor-Guided World Action Model with Mechanics-Aware Tactile Prediction.

Reviewer_3yGDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a simple and effective way to combine tactile signals with existing VLA models: a readily available tactile expert and encoder interact with the VL information to improve the action, thereby improving performance on fine manipulation tasks. The authors test the method on multiple fine manipulation tasks involving tactile sensing. The simulation results show that this approach achieves better performance and higher data efficiency than other methods, and the real robot experiments show good results.

Critical strengths
  1. The proposed method is simple and highly reproducible. It improves performance on fine manipulation tasks through a simple idea, and has high reproducibility and significance.
  2. The paper offers a meaningful discussion of tactile representation and efficiency, and is well presented. The authors provide some discussion of the tactile representation and of training/inference efficiency, which is very meaningful. The figures and tables are well presented, and the paper is clearly written and easy to understand.
Critical weaknesses
  1. The real robot evidence is too narrow, and the margins too small, to support the paper's central claim. The real robot experiments select only two baselines related to tactile sensing, and do not include FTP-π0.5 (the strongest tactile baseline without pretraining in simulation), nor VITaL, Tactile-VLA, and others. Moreover, for the baseline with the best performance, FTP-1, the improvements are rather limited on the remaining tasks apart from Wipe Board, so the paper's conclusions are not strongly supported.
  2. Some statements are not supported by sufficient evidence or justification. These include the reason for using the T3 Large encoder, whether next token prediction can truly learn latent dynamics, and the relationship between the effectiveness of future tactile prediction and task success rate, among others.
Questions for authors
  1. Is the next token prediction paradigm reasonable, and can it truly learn latent dynamics? Can some components of JEPA and world models play a role here? Would they theoretically be better than the current approach?
  2. How does the T3 Large encoder compare with other tactile encoders? Although the paper keeps the encoder fixed as a given condition and discusses the effectiveness of different representations in Section 4.3, for the tactile encoder I would still like the authors to compare with tactile encoders such as Sparsh-X and AnyTouch 2, because the authors do not give sufficient reasons for using the T3 Large encoder, nor reasons for not using these encoders.
  3. Have the authors tested speed optimization through asynchronous inference and temporal aggregation? During system execution, tactile signals are still closed loop in the form of chunks, which reduces the real time performance of the system. Although the paper optimizes inference efficiency, the timing of the tactile closed loop still has some limitations in real time responsiveness. Have the authors tested speed optimization through asynchronous inference and temporal aggregation?
  4. What is the relationship between the effectiveness of future tactile prediction and task success rate? The paper demonstrates the effectiveness of future tactile prediction through qualitative analysis, but does not show, experimentally or theoretically, the relationship between the effectiveness of this future prediction and task success rate. It is difficult to infer from better tactile prediction that it plays a critical role. For example, how does better tactile prediction affect the action?
  5. Is the real robot evidence sufficient to support the central claim? The real robot experiments select only two baselines related to tactile sensing, and do not include FTP-π0.5 (the strongest tactile baseline without pretraining in simulation), nor VITaL, Tactile-VLA, and others. Moreover, for the baseline with the best performance, FTP-1, the improvements are rather limited on the remaining tasks apart from Wipe Board, so the paper's conclusions are not strongly supported.
8 comments
AuthorsOpenReview ↗

1. Main Results and Hardware Evidence (W1 & Q5)

Thank you for your careful review and for recognizing the simplicity, reproducibility, and practical value of SimpleTouch. Our central contribution is a simple architecture and single-stage post-training recipe that equips a strong pretrained VLA with touch, using only 50 demonstrations per task and no additional tactile policy pretraining. Below, we address your concerns with evidence from the manuscript and additional experiments.

Main results and baseline selection. FTP-1 is the strongest baseline in manuscript Table 1, making it our key hardware comparator. Without additional tactile policy pretraining, SimpleTouch exceeds FTP-1 by 10.8 percentage points and FTP-π0.5 by 32.3 points in simulation (Table A).

Table A. Simulation results (manuscript Table 1; six-task mean). Gain = SimpleTouch minus baseline.

MetricACTVITaLTactile-VLAUniVTAC-ACTFTP-π0.5π0.5FTP-1SimpleTouch
SR (%)30.336.341.842.845.246.366.777.5
Gain (pp)+47.2+41.2+35.7+34.7+32.3+31.2+10.8—

Hardware gains. SimpleTouch improves mean success over FTP-1 by 8.8 percentage points, with three wins and one tie. Beyond Wipe Board, it gains 10 points on Pick Up Chips and 5 on Play Mahjong (Table B).

Table B. Hardware success rates (manuscript Table 2; %).

Method / metricPlay MahjongWipe BoardPick Up ChipsInsert USBMean
FTP-19565603062.5
SimpleTouch10085703071.3
Δ (pp)+5+20+100+8.8

Achieving higher success through simple, single-stage post-training while eliminating additional tactile policy pretraining is itself an important contribution. For context, FTP-1 reports that tactile policy pretraining improves over FTP-π0.5 by 21.5 points in simulation and 20.9/31.6 points on hardware with seen/unseen sensors (its Tables 1–3) [1]. SimpleTouch surpasses this pretrained reference while omitting that stage.

Evaluation scope. Five methods across four hardware tasks require 20 separately trained/fine-tuned task-specific checkpoints and 400 evaluations (20 per method/task). Including demonstration collection and physical resets, this is a substantial evaluation comparable in scale to representative studies [2,3] (Table C).

Table C. Main hardware evaluation scope, excluding additional ablations.

StudyMethodsEvaluation settingsTrials / pairTotal trials
VTAM [2]4420320
DT-TacVLA [3]5320300
SimpleTouch5420400

References

[1] FTP-1: A Generalist Foundation Tactile Policy Across Tactile Sensors for Contact-Rich Manipulation.

[2] VTAM: Video-Tactile-Action Models for Complex Physical Interaction Beyond VLAs.

[3] DT-TacVLA: A Dual-Timescale Tactile-Enhanced Vision-Language-Action Model for Contact-Rich Manipulation.

AuthorsOpenReview ↗

2. Tactile Encoder Choice (W2 & Q2)

SimpleTouch supports off-the-shelf encoders. Even a purely visual VAE achieves 67.8% mean success on the four tactile-sensitive tasks, exceeding FTP-1’s 59.5% (manuscript Figure 4, right, and Table 1).

We chose T3-Large for consistent input conditions: it uses only the current tactile frame and follows FTP-1’s encoder choice. We additionally evaluated AnyTouch 2 [1] and Sparsh (MAE) [2]. Sparsh-X [3] targets DIGIT 360, so we used the GelSight Mini-compatible Sparsh (MAE) instead (Table D).

Table D. SimpleTouch encoder substitutions; FTP-1 is a reference (%; 100 trials per setting).

Method / encoderPull-out KeyPut BottleInsert HoleInsert TubeMean
FTP-14847647959.5
visual VAE6351847367.8
T3-Large62508510074.3
Sparsh (MAE)68438910075.0
AnyTouch 269508710076.5

In our configurations, Sparsh (MAE) and AnyTouch 2 use tactile history and reach 75.0%/76.5%, versus 74.3% for single-frame T3. They use 300/398 tokens per sensor, versus T3’s 197. T3 therefore achieves comparable performance with fewer tokens while preserving the baselines’ single-frame input conditions. These results support both our simple default and compatibility with alternative encoders.

References

[1] AnyTouch 2: General Optical Tactile Representation Learning for Dynamic Tactile Perception.

[2] Sparsh: Self-supervised Touch Representations for Vision-based Tactile Sensing.

[3] Tactile Beyond Pixels: Multisensory Touch Representations for Robot Manipulation.

AuthorsOpenReview ↗

3. Predictive Learning and Policy Benefit (W2, Q1 & Q4)

The objective is conditional flow matching, not autoregressive next-token prediction. We supervise continuous tactile latents at ten future offsets, t+5 through t+50 (Section 3.4; Appendix A.2).

Prediction supervision improves control. With the current-touch pathway unchanged, removing only prediction supervision lowers four-task mean success from 74.3% to 68.8%, including a 13-point drop on Put Bottle (Table E).

Table E. Controlled prediction ablation (manuscript Table 3; %).

Method / metricPull-out KeyPut BottleInsert HoleInsert TubeMean
No prediction6037839568.8
Single horizon5344799868.5
Multi-horizon62508510074.3

We additionally report a controlled hardware ablation on Pick Up Chips: removing only the prediction loss, with the current-touch pathway, training data, and other settings unchanged, lowers success from 70% (14/20) to 50% (10/20). Crushing the chips is the main failure mode.

How prediction affects actions. The future-touch loss updates expert parameters shared with the current-touch pathway, improving the tactile context available to the action branch (Eq. 7). Actions never read future prediction tokens; future-touch generation is unnecessary at deployment. Figures 5 and 11 visualize the learned tactile evolution, while the controlled ablations establish its policy benefit.

Relation to JEPA/world models. JEPA shares the goal of learning representations through latent prediction; this does not imply superior performance under our low-data setting. General-purpose models such as V-JEPA 2 use large-scale pretraining—over one million hours of video [1]. SimpleTouch reuses existing VLA and tactile representations and learns from only 50 task demonstrations in one stage, avoiding additional tactile-policy or world-model pretraining.

References

[1] V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their feedback. Below, we summarize our contribution and the evidence addressing their concerns.

Core Contributions of Our Work

  1. Simple architecture. An off-the-shelf frozen tactile encoder and a shared tactile expert equip π0.5 with touch.

  2. Simple training. Only 50 demonstrations per task and single-stage post-training are needed, without additional tactile policy pretraining or separate modality alignment.

  3. Strong performance. Mean success exceeds FTP-1 by 10.8/8.8 percentage points in simulation/hardware, and FTP-π0.5 by 32.3 points in simulation (Tables 1–2).

Experimental Evidence Highlighted in Our Responses

  • Hardware evidence. FTP-1, the strongest baseline in the main table, is our key hardware comparator. Five methods across four tasks require 20 task-specific checkpoints and 400 evaluations, yielding three wins and one tie against FTP-1.

  • Design contributions and compatibility. Individual and joint ablations validate the key designs: restoring all three raises four-task mean success from 50.0% to 74.3%; removing prediction supervision lowers hardware Pick Up Chips success from 70% to 50%. Visual VAE, Sparsh (MAE), and AnyTouch 2 substitutions demonstrate encoder compatibility.

  • Comparison fairness and stability. With matched training/execution horizons and temporal ensembling, four-task mean success is 76.3% versus FTP-1’s 59.5%; the Dream-Tac comparison yields 74.3% versus 50.5%. SimpleTouch exceeds FTP-1 in all six low-data settings. Across three training seeds, success ranges are 61%–64% on Pull-out Key and 84%–85% on Insert Hole.

Clarifications and Additional Discussion

  • Scope. This work studies task-level tactile post-training; all evaluated methods adapt using the same task demonstrations.

  • Prediction mechanism. Future-touch conditional flow matching improves action learning through the shared expert; future-touch generation is unnecessary at deployment.

  • Related work. We will cite VTAM and TacWAM in the revision and clarify differences in backbones, tactile representations, and training procedures.

We will incorporate the relevant results and clarifications into the revised paper PDF. We welcome further feedback and respectfully invite reassessment in light of this evidence.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewers,

Thank you again for your time and thoughtful feedback. We have provided point-by-point responses and additional experimental results addressing your concerns.

As the discussion period approaches its end, we would greatly appreciate the opportunity to discuss our responses and findings with you. Please let us know if any concerns remain or if further clarification would be helpful. We would be happy to respond promptly and continue the discussion.

Thank you for your consideration.

AuthorsOpenReview ↗

What Makes Research Matter

We sincerely thank the reviewers and the Area Chair for their time and care throughout the review process, and for the thoughtful discussion during the rebuttal. We remain confident in SimpleTouch's value and have been encouraged by conversations with researchers and positive feedback from people who have tried it. We hope that more researchers will use SimpleTouch and find it useful in their own work. To us, the value of research lies both in its ability to help others and in the shared understanding that emerges from thoughtful discussion. The opportunity to explore concrete questions, clarify differing views, and deepen our understanding has been a particularly valuable part of this process. We hope that all of us will continue to pursue work of lasting value. Thank you again.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.