Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Tactile prediction improves contact-rich robot policies most when it supervises the middle of the action pathway, where future contact states are most linearly accessible, rather than the vision-language features or the final motor-output f

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 01:34 UTC pith:WEUD5KKM

load-bearing objection Useful design principle—attach future-tactile prediction to an intermediate action-expert layer—but the evidence needs error bars and a less in-sample interface selection before I'd trust the ordering. the 4 major comments →

arxiv 2607.14609 v1 pith:WEUD5KKM submitted 2026-07-16 cs.RO

Representation-Aligned Tactile Grounding for Contact-Rich Robotic Manipulation

Ruilin Chen , Jingkai Jia , Tong Yang , Xinyu Zhou , Qiao Sun , Jiangwei Zhong , Shizeng Zhang , Nuo Chen
show 3 more authors
Bailin He Wei Li Wenqiang Zhang
This is my paper
classification cs.RO
keywords tactile groundingvision-language-action modelfuture tactile predictionrepresentation alignmentlatent tactile predictorcontact-rich manipulationlinear probeauxiliary supervision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper studies where in a vision-language-action (VLA) policy future touch should be used as training supervision. Using frozen linear probes, it finds that future tactile states are most predictable from intermediate action-expert features, less so from the vision-language backbone or the final action state. Based on this, it introduces a Latent Tactile Predictor that predicts future tactile embeddings in a compact latent space and uses that as a training-only grounding loss. On five real-world contact-rich manipulation tasks with two different VLA backbones, grounding at this middle interface reaches 74% average success, versus 58% and 62% for less aligned interfaces. The paper's central claim is that where the tactile supervision is attached matters more than adding more supervision.

Core claim

The paper establishes that future tactile prediction is most effective when it supervises intermediate action-expert representations rather than earlier vision-language features or later motor-decoding features. Frozen linear probes trained on a frozen policy show a non-monotonic predictability curve: validation MSE for predicting future raw tactile sequences falls from the VLM side to a minimum at an intermediate action-expert layer, then rises near the final action state. The same pattern holds with and without tactile input and across model scales. A lightweight query-based predictor, the Latent Tactile Predictor, uses this intermediate feature to predict future tactile latents, avoiding

What carries the argument

The central object is the intermediate action-expert representation inside the VLA policy's action pathway. Two components carry the argument: (1) frozen linear probes that read out future tactile sequences from candidate representations and rank interfaces by validation MSE, and (2) a Latent Tactile Predictor, a lightweight query-based head that takes learnable queries and the diagnosed intermediate feature to predict future tactile latents. The probe provides the selection rule; the predictor provides the grounding loss that shapes the chosen representation during training without adding inference-time cost.

Load-bearing premise

The load-bearing assumption is that the representation from which future tactile states are most linearly predictable on a frozen policy is also the best place to attach a training-time grounding loss for the learned policy; this transfer is validated on a single task before being applied to the full suite.

What would settle it

A concrete falsifier: run the full interface-selection procedure per task instead of using one shared interface. If the per-task best interface differs from the intermediate action-expert layer, or if grounding at the layer with lowest probe MSE does not yield the highest success rate on a held-out task, the transfer assumption fails. Alternatively, a task where final-action states are far more linearly predictive than intermediate features would contradict the claimed universality.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the representation-alignment principle holds, future tactile prediction should be designed as an interface-selection problem, not just a loss-design problem.
  • The same probe-based selection procedure could be reused for other auxiliary prediction objectives, such as force, slip, or proximity, to locate the most contact-grounded representation.
  • The latent tactile objective is training-only, so the inference pathway stays unchanged, meaning no added latency or sensor overhead at deployment.
  • Broad multi-layer supervision is consistently worse than a single well-aligned interface, suggesting that mismatched supervision can actively harm learning.
  • The consistency across two different VLA backbones indicates the 'middle-of-action-pathway' result may be a general architectural property, not a quirk of one model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The linear-probe ranking is assumed to transfer to the best training interface, but this transfer is validated on only one task before being applied across the full suite; a stronger test would run the full interface selection per task and check that the lowest-MSE layer still yields the highest success.
  • The probe target is raw future tactile data while the training loss uses latent embeddings, so the method's success depends on the quality of the tactile encoder; a different sensor with different noise characteristics could shift the ranking.
  • The 'middle is best' pattern may reflect a general property of action-conditioned representations: contact dynamics become linearly decodable after motor intent is formed but before being compressed into executable commands. This could be tested by repeating the probe analysis with other future sensory modalities.
  • A direct replication on a task where final-action states are highly predictive of future touch would be the clearest way to probe the universality of the claimed ordering.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies where to apply future-tactile prediction in vision-language-action (VLA) policies. Using frozen linear probes on raw future tactile sequences, the authors find that intermediate action-expert features are most linearly predictive of future contact states, compared with VLM-side features and final action outputs. They then introduce a Latent Tactile Predictor (LTP) that predicts compact latent tactile embeddings from this intermediate representation, using a training-time auxiliary loss that does not affect inference. Experiments on five real-world contact-rich manipulation tasks with two VLA backbones (SmolVLA and π0) report that the proposed representation-aligned grounding achieves higher average success than less-aligned interfaces, raw tactile prediction, and multi-interface supervision.

Significance. If the claimed ordering is robust, the paper makes a useful contribution: it provides a principled answer to where auxiliary tactile prediction should be attached in VLA policies, and the LTP is lightweight and inference-free. The study design is thoughtful: the probe analysis is a simple diagnostic, the ablations isolate the interface location and target representation, and the evaluation includes two architectures and real hardware. The main limitations are statistical (no variance or significance on 20-trial evaluations), the limited validation of the raw-probe-to-latent-training transfer, and the fact that the interface is selected on the same task suite used for the headline results. These concerns do not invalidate the approach but currently prevent the results from being conclusive.

major comments (4)
  1. [§3.2 (Eq. 1) and §3.3 (Eq. 4)] The interface is selected by linear predictability of raw future tactile sequences (Eq. 1), but LTP is trained to predict latent tactile embeddings z_{t+1:t+H} (Eq. 4). The paper states that raw targets are used 'to keep the interface diagnosis independent of the latent tactile encoder,' yet the transfer from raw-probe ranking to latent-grounding effectiveness is validated only in Fig. 2D, on a single task (USB Drive Insertion) and without statistical significance. Since Eq. 4 is the objective actually used for policy training, the probe criterion and the training objective are mismatched. Please show that the same interface ranking is obtained when probing with latent tactile targets, or provide per-task validation that the intermediate interface is best.
  2. [§4.1, Evaluation protocol; Figs. 5-6] The headline results are based on 20 trials per task and are reported as point estimates without confidence intervals or significance tests. For 20 trials the standard error of a proportion is roughly 10 percentage points at 50% success, so several of the reported differences are within sampling noise (e.g., π0 +Tactile 54% vs VLM Pred 58% in Fig. 6). Please provide confidence intervals, bootstrap estimates, or pairwise significance tests (e.g., McNemar's test) to support the ordering intermediate > final > VLM and the claimed gains over baselines.
  3. [§3.2, Fig. 2D] The probe diagnosis and the downstream validation are both performed on the same task suite: the frozen policy is trained on the same demonstrations, and Fig. 2D validates the transfer on USB Drive Insertion, one of the five evaluation tasks. The intermediate interface is then applied to all tasks without a separate validation set. This creates a risk that the reported ordering is a selection artifact of the specific tasks rather than a general representation-alignment principle. Please clarify whether the probing policy uses the same data as the final experiments, and either hold out tasks for interface selection or report the sensitivity of the results to the choice of interface.
  4. [§3.3, Eq. (3); Appendix A] The method section does not specify which exact layer is used as the intermediate action-expert representation. The probe curves in Fig. 2A-C show a minimum in an intermediate region, but no layer index (e.g., layer 8 of 16) is reported for SmolVLA or π0. Since the entire method hinges on this interface, the exact selection rule and layer identity must be provided for reproducibility.
minor comments (5)
  1. [Section 1, before Fig. 1] There is a leftover non-English instruction '1.增加论文的motivation图' (add a motivation figure) in the body, immediately before Figure 1. This should be removed.
  2. [Fig. 2A-C] The caption says the curves are evaluated at normalized timesteps T, but the legend is not legible in the current rendering. Clarify which color corresponds to which T and what the marker 'VLM' denotes (the VLM-side representation error at a fixed width or layer?).
  3. [Figs. 5-6] The numeric labels on the bars are difficult to read and appear partially truncated in several places. Please ensure all percentages are legible and consistent with the text.
  4. [Appendix A] The LTP architecture is described only as a 'lightweight query-based prediction head.' Provide the number of layers, hidden dimension, and parameter count, or state that the implementation will be released.
  5. [References [25-26]] Some entries have incomplete author names ('Liu et al.', 'Xu et al.'); please format to include at least the first author's full name and initials.

Circularity Check

0 steps flagged

No significant circularity: probe selection and latent grounding are distinct steps, and the central claim rests on real-robot comparisons rather than an input-equivalent reformulation.

full rationale

The paper's derivation chain is: (i) frozen linear probes measure raw future-tactile predictability from candidate representations (Eq. 1); (ii) this diagnosis selects the intermediate action-expert layer; (iii) a separate Latent Tactile Predictor is trained at that layer with L_tac on encoded future tactile latents (Eqs. 2-4); (iv) the central claim is carried by real-robot success comparisons among interfaces. None of these steps reduces an output to an input. The probe target (raw r_{t+1:t+H}) and the LTP target (latent z_{t+1:t+H}=E_tac(r_{t+1:t+H})) are deliberately different; the paper states it uses raw probes "to keep the interface diagnosis independent of the latent tactile encoder used later for policy training." The concern that raw linear predictability may not transfer to latent-grounding effectiveness is a validity/generalization limitation (validated only on USB Drive Insertion in Fig. 2D before being applied to all five tasks), not a circular reduction. The headline results (74% vs 58/62) come from actual policy training and evaluation, not from the probe fit. The only author-overlapping reference appears to be ForceVLA [21], used in related work as one example of MoE-based tactile fusion; it is not load-bearing. Therefore no circular step is present; the low score reflects only the minor non-load-bearing self-citation and the in-sample interface-selection concern.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The method relies on the tactile encoder and the intermediate-layer selection. No free physical constants or invented entities are introduced; the main parameters are the chosen grounding layer and the supervision weight, both set by hand or by in-sample probe selection.

free parameters (2)
  • Grounding interface layer index = 'intermediate action-expert layer' (exact index not reported)
    Selected by linear-probe validation MSE on the same task suite; a discrete design choice fit to the evaluation data.
  • Tactile supervision weight lambda_tac = 1.0
    Set by hand across all experiments (Tables 1-2); no sensitivity analysis reported.
axioms (4)
  • domain assumption Linear-probe MSE ordering on a frozen policy predicts which internal representation is the best interface for auxiliary grounding during policy training.
    Section 3.2 uses probe curves to identify the intermediate interface; transfer is tested on one task (Fig. 2D) before being applied to all five tasks.
  • domain assumption Future tactile observations r_{t+1:t+H} provide supervision for action-induced contact dynamics that improves the policy when used as an auxiliary loss.
    Motivated by prior future-tactile prediction work [25-28], accepted as a premise rather than proven here.
  • domain assumption The tactile encoder E_tac used to compute latent targets preserves control-relevant contact structure and is an appropriate target space.
    Eq. (2) defines targets as E_tac(r_{t+1:t+H}); if the encoder discards control-relevant information, the loss cannot ground it.
  • domain assumption Intermediate action-expert features are causally prior to final motor decoding, i.e., 'before compression' claim.
    Architecture assumption about SmolVLA/π0 flow-based action experts; not formally shown.

pith-pipeline@v1.3.0-alltime-deepseek · 9781 in / 10576 out tokens · 113786 ms · 2026-08-02T01:34:54.655588+00:00 · methodology

0 comments
read the original abstract

Tactile-enhanced vision-language-action (VLA) policies have been introduced for contact-rich manipulation, where critical interaction states are often hidden from vision. Future tactile prediction is a promising way to use touch because it turns tactile outcomes into supervision for action-induced contact dynamics. Yet VLA policies contain representations with different roles, from perceptual encoding to motor prediction, making it unclear where this supervision should be applied. We study this as a representation-alignment problem. Through a linear probe analysis, we find that future tactile states are most predictable from intermediate action-expert features, rather than from vision-language features or final action states. Motivated by this observation, we introduce a lightweight Latent Tactile Predictor (LTP), which predicts compact future tactile embeddings from the identified intermediate representation. By avoiding direct prediction of noisy raw tactile signals, LTP provides an action-outcome grounding signal that aligns intermediate action representations with future contact consequences. Experiments on real-world contact-rich manipulation tasks show that representation-aligned tactile grounding outperforms less aligned or multi-interface tactile prediction, highlighting the importance of where tactile supervision is applied.

Figures

Figures reproduced from arXiv: 2607.14609 by Bailin He, Jiangwei Zhong, Jingkai Jia, Nuo Chen, Qiao Sun, Ruilin Chen, Shizeng Zhang, Tong Yang, Wei Li, Wenqiang Zhang, Xinyu Zhou.

Figure 1
Figure 1. Figure 1: Motivation of representation-aligned tactile grounding. Future tactile prediction is most effective when applied to a representation aligned with action-conditioned contact dynamics. VLM-side features are too perception-oriented, while final action features are already specialized for motor decoding. We therefore ground the intermediate action-expert representation, where future tactile information is most… view at source ↗
Figure 2
Figure 2. Figure 2: Linear-probe diagnosis and downstream validation. (A) Future tactile prediction MSE without tactile input. (B) The same probe analysis with current tactile as input. (C) The same analysis on the π0 backbone. (D) Downstream success when grounding VLM-side, intermediate action-expert, and final action representations. Curves in (A–C) are evaluated at normalized action-generation timesteps T of the flow-based… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of future latent tactile grounding. The VLA policy takes current observations (It, l, rt) - image, language, and tactile input—and predicts an action chunk aˆt+1:t+H. During training, the aligned action-expert feature, which is most grounded in future contact states, conditions a latent tactile predictor, which predicts future tactile latents zˆt+1:t+H from learnable queries q1:H. The targets zt+1… view at source ↗
Figure 4
Figure 4. Figure 4: Experimental task suite. The five blocks correspond to USB Drive Insertion, Whiteboard Wiping, Plug Insertion, Light Bulb Unscrewing and Deformable Object Grasping.Green check marks indicate successful executions. For deformable-object grasping, both a successful execution and a failure case, marked with a red cross, are shown. tactile queries Q = {q1, . . . , qH} to predict the future tactile latent seque… view at source ↗
Figure 5
Figure 5. Figure 5: Success rates on five real-world contact-rich manipulation tasks using the SmolVLA backbone. We report task success rates for input-level tactile conditioning, external predictive tactile representation learning, less-aligned tactile prediction interfaces, and our representation-aligned tactile grounding under identical settings. 4.1 Experimental Setup Tasks and data collection. All experiments are conduct… view at source ↗
Figure 6
Figure 6. Figure 6: Success rates with the π0 backbone on five real-world contact-rich manipulation tasks. We compare the standard π0 policy, tactile-conditioned π0, two less-aligned future tactile prediction interfaces (VLM Pred. and AE Final Pred.), and our representation-aligned tactile grounding under the same task setting. Evaluation protocol. We evaluate each method with 20 real-world trials per task and report task suc… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation results on multi-layer grounding applied to SmolVLA across all tasks. None adds tactile input without future tactile prediction. Multi-A predicts from all action-expert features, Multi-B predicts from every two action-expert layers (0/2/4/...), and Multi-C predicts from the middle layers 4–11 of the 16-layer SmolVLA action expert. representation aligned with contact dynamics, rather than simply ad… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative examples of successful real-world rollouts. Uniformly sampled frames from successful rollouts of Plug Insertion, USB Drive Insertion, Whiteboard Wiping, Deformable Object Grasping, and Light Bulb Unscrewing are arranged chronologically to illustrate the execution process. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 14 linked inside Pith

  1. [1]

    M. J. Kim et al. OpenVLA: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024. URLhttps://arxiv.org/abs/2406.09246

    Pith/arXiv arXiv 2024

  2. [2]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, J. Jakubczak, T. Jones, L. Ke, S. Levine, Z. Li-Bell, S. Mothukuri, S. Nair, K. Pertsch, F. Xia, U. Zhilinsky, et al. π0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. URLhttps://arxiv.org/abs/2410.24164

    Pith/arXiv arXiv 2024

  3. [3]

    Shukor et al

    M. Shukor et al. SmolVLA: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. URLhttps://arxiv.org/abs/2506.01844

    Pith/arXiv arXiv 2025

  4. [4]

    Higuera et al

    C. Higuera et al. Tactile beyond pixels: Multisensory touch representations for robot manipula- tion.arXiv preprint arXiv:2506.14754, 2025. URL https://arxiv.org/abs/2506.14754

    Pith/arXiv arXiv 2025

  5. [5]

    Pattabiraman et al

    V . Pattabiraman et al. Learning precise, contact-rich manipulation through uncalibrated tactile skins.arXiv preprint arXiv:2410.17246, 2024. URL https://arxiv.org/abs/2410.17246

    Pith/arXiv arXiv 2024

  6. [6]

    Hao et al

    P. Hao et al. TLA: Tactile-language-action model for contact-rich manipulation.arXiv preprint arXiv:2503.08548, 2025. URLhttps://arxiv.org/abs/2503.08548

    Pith/arXiv arXiv 2025

  7. [7]

    Huang et al

    J. Huang et al. Tactile-VLA: Unlocking vision-language-action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160, 2025. URL https://arxiv.org/ abs/2507.09160

    Pith/arXiv arXiv 2025

  8. [8]

    Zhang et al

    K. Zhang et al. TacVLA: Contact-aware tactile fusion for robust vision-language-action manipulation.arXiv preprint arXiv:2603.12665, 2026. URL https://arxiv.org/abs/ 2603.12665

    arXiv 2026

  9. [9]

    Li et al

    Y . Li et al. ForceVLA2: Unleashing hybrid force-position control with force awareness for contact-rich manipulation.arXiv preprint arXiv:2603.15169, 2026. URL https://arxiv. org/abs/2603.15169

    arXiv 2026

  10. [10]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  11. [11]

    Karamcheti, S

    S. Karamcheti, S. Nair, A. Balakrishna, T. Kollar, P. Liang, and D. Sadigh. Prismatic VLMs: Investigating the design space of visually-conditioned language models. InInternational Conference on Machine Learning (ICML), 2024

    2024

  12. [12]

    Marafioti, O

    A. Marafioti, O. Zohar, M. Farré, M. Noyan, E. Bakouch, P. Cuenca, C. Zakka, L. Ben Allal, A. Lozhkov, N. Tazi, V . Srivastav, J. Lochner, H. Larcher, M. Morlon, L. Tunstall, L. von Werra, and T. Wolf. SmolVLM: Redefining small and efficient multimodal models.arXiv preprint arXiv:2504.05299, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, J. Ibarz, B. Ichter, A. Irpan, T. Jackson, S. Jesmonth, N. J. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. Leal, K.-H. Lee, S. Levine, Y . Lu, U. Malla, D. Manju- nath, I. Mordatch, O. Nachum, C. Parada, J. Peralta, E. Perez, K. Pertsc...

    2023

  14. [14]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. Gonzalez Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. Leal, L. Lee, T.-W. E. Lee, S. Levine, Y . Lu, H. Michalewski, I. Mordatch,...

    2023

  15. [15]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y . L. Tan, L. Y . Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine. Octo: An open-source generalist robot policy. InRobotics: Science and Systems (RSS), 2024. URLhttps://arxiv.org/abs/2405.12213

    Pith/arXiv arXiv 2024

  16. [17]

    Zhu et al

    X. Zhu et al. Touch in the wild: Learning fine-grained manipulation with a portable visuo- tactile gripper.arXiv preprint arXiv:2507.15062, 2025. URL https://arxiv.org/abs/ 2507.15062

    arXiv 2025

  17. [18]

    Zhang et al

    C. Zhang et al. VTLA: Vision-tactile-language-action model with preference learning for insertion manipulation.arXiv preprint arXiv:2505.09577, 2025. URL https://arxiv.org/ abs/2505.09577

    Pith/arXiv arXiv 2025

  18. [19]

    Cheng et al

    Z. Cheng et al. OmniVTLA: Vision-tactile-language-action model with semantic-aligned tactile sensing.arXiv preprint arXiv:2508.08706, 2025. URL https://arxiv.org/abs/2508. 08706

    arXiv 2025

  19. [20]

    Bi et al

    J. Bi et al. VLA-Touch: Enhancing vision-language-action models with dual-level tactile feedback.arXiv preprint arXiv:2507.17294, 2025. URL https://arxiv.org/abs/2507. 17294

    Pith/arXiv arXiv 2025

  20. [21]

    J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai, C. Lu, and W. Zhang. ForceVLA: Enhancing vla models with a force-aware MoE for contact-rich manipulation.arXiv preprint arXiv:2505.22159, 2025. URL https://arxiv.org/abs/2505. 22159

    arXiv 2025

  21. [22]

    Morissette et al

    C. Morissette et al. TacFiLM: Tactile modality fusion for vision-language-action models.arXiv preprint arXiv:2603.14604, 2026. URLhttps://arxiv.org/abs/2603.14604

    Pith/arXiv arXiv 2026

  22. [23]

    Gubernatorov et al

    K. Gubernatorov et al. HapticVLA: Contact-rich manipulation via vision-language-action model without inference-time tactile sensing.arXiv preprint arXiv:2603.15257, 2026. URL https://arxiv.org/abs/2603.15257

    arXiv 2026

  23. [24]

    Huang et al

    Y . Huang et al. TaF-VLA: Tactile-force alignment in vision-language-action models for force- aware manipulation.arXiv preprint arXiv:2601.20321, 2026. URL https://arxiv.org/ abs/2601.20321

    arXiv 2026

  24. [25]

    Mla: A multisensory language-action model for robotic manipulation.arXiv preprint arXiv:2509.26642, 2025

    Liu et al. Mla: A multisensory language-action model for robotic manipulation.arXiv preprint arXiv:2509.26642, 2025. URLhttps://arxiv.org/abs/2509.26642

    arXiv 2025

  25. [26]

    Action-aware pretraining for tactile representations: Tpp / exumi

    Xu et al. Action-aware pretraining for tactile representations: Tpp / exumi. InConference on Robot Learning (CoRL), 2025. URLhttps://arxiv.org/abs/2509.14688

    arXiv 2025

  26. [27]

    Higuera et al

    C. Higuera et al. Visuo-tactile world models.arXiv preprint arXiv:2602.06001, 2026. URL https://arxiv.org/abs/2602.06001

    arXiv 2026

  27. [28]

    Ye et al

    G. Ye et al. Learning to feel the future: DreamTacVLA for contact-rich manipulation.arXiv preprint arXiv:2512.23864, 2025. URLhttps://arxiv.org/abs/2512.23864. 11 Appendix A Implementation Details For clarity, we summarize the training protocols used in our experiments in Tables 1 and 2. We train tactile-supervised policies on two architectures, SmolVLA a...

    Pith/arXiv arXiv 2025

  28. [2025]

    URLhttps://arxiv.org/abs/2504.16054

    Pith/arXiv arXiv