{"id":"6299ece4-97da-4f17-8306-7e23df14a3fe","arxiv_id":"2608.01102","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Contact-aware attention scaling and dynamic tactile masking improve visuo-tactile imitation-learning success rates across ACT, Diffusion Policy, and pi0.","lead":"A new plug-in module, CAAT, tells robot policies to trust vision before contact and touch during contact, instead of learning when to switch. Across three policy backbones, it raised average task success by 18 points in simulation and by 21.1 points over the strongest baseline on a real tactile gripper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that contact-aware switching drives the gains is not isolated: the only learned contact-aware variant performs at baseline level, and no fixed-weight control tests whether hand-picked (0.3,0.7) scaling—rather than contact-state dependence—is what helps.","rationale":"The reader's conditional verdict remains appropriate, but the decisive missing control is more specific than the reader's stated weakest assumption. The reader emphasized reliability and privileged supervision of the contact classifier; my reading points instead to the absence of a fixed-weight control that isolates the contact-state dependence itself. Table 1's learnable CAAT variant underperforms a simple gate, and the numeric variant is chosen after benchmark evaluation, so the reported 18.0/10.0 pp gains are not yet attributable to the contact-state signal. The proposed fixed-weight check would settle this; if constant weights perform as well, the central 'contact-aware' claim is not supported. Because this is an experimental omission rather than a demonstrated failure, the conditional verdict should stand with this additional required condition.","tokens_in":13402,"tokens_out":9708,"duration_ms":84034,"concrete_test":"Run a fixed-weight version of Eq. 13 with z_t removed, using gamma=(0.3,0.7) at every timestep, and a second fixed-weight version with gamma=(0.7,0.3) at every timestep, on the five UniVTAC tasks with 150 demonstrations, same DTM, same training protocol, and at least five seeds; report means with standard errors. If either constant-weight version matches Ours (Numerical) at 69.6% within confidence intervals, the contact-state dependence is not the driver of the improvement; if both fall to the roughly 50% level, the contact-aware switching is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Table 1, Ours (Binary) and Ours (Learnable) score 51.8% and 50.6%, essentially at the level of ACT+DTM (51.6%) and below Gated ACT+DTM (59.6%). Only Ours (Numerical), with hand-specified weights gamma_v=0.7/gamma_tau=0.3 before contact and gamma_v=0.3/gamma_tau=0.7 during contact (Eq. 13; Appendix CAAT-Specific Settings), reaches 69.6%. The paper says this numerical strategy was adopted for real-world experiments after seeing the simulation results, so the headline gains are selected on the same benchmark used to evaluate them. More importantly, no reported baseline uses the same modality-specific CAAT readout with constant, contact-independent scaling; the ablation that 'removes contact-aware scaling' is described only qualitatively and its readout weights are not specified. The 18.0/10.0 pp simulation gains and 21.1 pp real-world margin could therefore come from the manually chosen numerical weights plus Dynamic Tactile Masking, not from the contact-state estimate z_t that is the central novelty. In addition, the ground-truth contact labels in Eq. 16 are never specified for real-world data, so the contact signal itself is not yet reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAAT, a modular framework for Transformer-based visuo-tactile policies. CAAT has two components: Dynamic Tactile Masking (DTM), which suppresses tactile tokens whose cosine similarity to a non-contact reference exceeds a threshold (Eqs. 3–6), and Contact-Aware Attention Scaling, which computes modality-specific attention readouts after joint self-attention and reweights the visual and tactile readouts according to an estimated binary contact state (Eqs. 8–13). The contact state is produced by a CNN–MLP classifier trained with ground-truth contact labels (Eqs. 15–17). The authors report that CAAT with fixed numerical scaling (0.7/0.3 before contact, 0.3/0.7 during contact) improves average success rate by 18.0 percentage points over direct concatenation and 10.0 points over gated fusion on the UniVTAC benchmark with ACT, and achieves a 60.0% average real-world success rate across ACT, Diffusion Policy, and pi0, outperforming the strongest baseline by 21.1 percentage points. The paper also presents data-efficiency curves, ablations of DTM, and attention visualizations. The central claim is that explicit contact priors improve data-efficient visuo-tactile policy learning.","tokens_in":13687,"tokens_out":4384,"duration_ms":40544,"significance":"If the result holds, CAAT is a useful inductive bias: it is architecture-agnostic, adds negligible compute, and its data-efficiency gains are consistent across demonstration budgets. The real-world UMI-style tactile gripper and the evaluation across three policy families are also valuable assets. However, the current evidence does not establish the central claim. The headline numbers come from the single variant selected after seeing the same benchmark's results, the contact-state signal uses ground-truth labels not available to baselines, and no error bars or statistical tests are reported. These are fixable with additional experiments and analysis, but they are load-bearing for the paper's conclusions.","major_comments":[{"comment":"The headline gain of 18.0 percentage points is reported only for the 'Ours (Numerical)' variant, whose weights gamma_v=0.7/gamma_tau=0.3 before contact and 0.3/0.7 during contact were adopted after comparing binary, learnable, and numerical scaling on the same UniVTAC benchmark. The two other CAAT variants perform essentially at the level of the concatenation baseline (51.8% and 50.6% vs. 51.6%), so the reported advantage is not robust to the choice of scaling parameterization. Please report results for all variants with variance estimates, and either specify the numerical weights in advance using a validation split or a separate held-out set, or correct for the multiple variants tested.","section":"§4.2, Table 1 and 'Comparison of scaling strategies'"},{"comment":"The contact classifier is trained with ground-truth contact labels y_t, but the paper never states how these labels are obtained in real-world experiments. In simulation, UniVTAC may provide them, but the real-world acquisition procedure is absent. Because the concatenation and gating baselines do not receive contact labels, CAAT effectively uses privileged supervision for its switching signal, which confounds the comparison. Please specify the real-world labeling protocol, or compare against a baseline that also receives the same contact-state information, and additionally report the performance of a rule-based contact detector (e.g., thresholding the tactile difference D_t) to show that the learned classifier is not the source of the gain.","section":"§3.3, Eq. (16) and Appendix 'CAAT-Specific Settings'"},{"comment":"The ablation that removes contact-aware scaling is described only qualitatively: 'Removing attention scaling substantially degrades performance in both settings,' but no success rates are given. Table 3 reports CAAT with and without Dynamic Tactile Masking, but it does not report the fixed-fusion baseline (DTM with fixed, contact-independent scaling). Please provide the actual numbers for this control in simulation and real-world, together with the exact readout weights used, so the contribution of the contact-state-dependent switching can be quantified separately from the contribution of the hand-picked numerical weights.","section":"§4.4 'Ablation Studies', paragraph on Contact-Aware Attention Scaling"},{"comment":"No error bars, confidence intervals, or statistical tests are reported anywhere. The simulation protocol uses 100 rollouts across 100 random seeds and the real-world protocol uses 20 trials per task, which is sufficient for a difference-of-proportions test or a bootstrap confidence interval. With such granularities, the reported 18.0 and 21.1 percentage point gaps need error estimates to rule out run-to-run variation. Please include per-task standard errors or confidence intervals and, where possible, a paired or per-task significance test.","section":"§Appendix 'Evaluation Protocol' and Tables 1–2"},{"comment":"The fixed numerical weights (0.7, 0.3) and (0.3, 0.7) are user-specified and not swept. Given that the binary and learnable variants fail to beat the gated baseline, the method's success appears to hinge on a narrow choice of the scaling coefficients. Please provide a sensitivity analysis of the numerical weights (e.g., varying each gamma by 0.1 or 0.2) to show that the performance gain is not a narrow peak, and justify the chosen values from first principles or from a validation set.","section":"§4.2, Table 1, and Appendix 'CAAT-Specific Settings'"}],"minor_comments":[{"comment":"The notation D_t = |T_t - T_ref| is ambiguous: the text says the classifier input is the difference between tactile images, but the equation could be read as a scalar norm; please clarify the exact input representation.","section":"Eq. (15)"},{"comment":"The learnable variant is said to use state-dependent scaling, but the definition gamma_L_tau(z_t) = 1 - gamma_L_v(z_t) forces the two weights to sum to one; this is a strong constraint that may explain why the learnable variant underperforms. Please state whether the numerical variant also imposes this constraint and discuss the modeling choice.","section":"Eqs. (20)–(22)"},{"comment":"The contact estimator is described as a 'frozen CNN–MLP binary classifier,' but Eq. (16) implies it is trained; please specify on which data it is trained, when it is frozen, and how the reference image T_ref is selected in episodes where the first timestep already involves contact.","section":"Appendix 'CAAT-Specific Settings'"},{"comment":"The data-efficiency curves are informative, but the 25- and 50-demonstration points for ACT+DTM in Table 4 are non-monotonic (21.6, 25.4, 22.2); please comment on this variance or add error bars to Figure 4.","section":"Figure 4 and Table 4"},{"comment":"The real-world table reports only point estimates, and the attention heatmap analysis is purely qualitative; please state the number of trials visualized and whether the attention-shift pattern was consistent across rollouts.","section":"Table 2 and Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is attractive and the modular design is a strength, but the evaluation as it stands does not isolate the proposed contact-aware mechanism from the hand-picked numerical scaling or from the privileged contact labels. I would not accept the paper without a fixed-weight control, error bars or statistical tests, a specification of the real-world contact-label acquisition, and a clear statement of how the numerical weights were chosen relative to the benchmark. These issues are addressable through additional experiments and analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2608.01102. The paper is worth a serious referee, but the main claim needs a control it doesn't currently have.\n\nWhat's genuinely new: readout-level contact-aware scaling and dynamic tactile masking, packaged as a plug-in for ACT/DP/pi0, plus a two-finger tactile UMI rig that lets the same gripper be used for demos and deployment. The ablation of dynamic tactile masking is done carefully—it is applied consistently across all fusion strategies and helps each one, which is a clean result. The data-efficiency curves are also a nice touch.\n\nThe soft spot is central. In Table 1, the binary and learnable variants of CAAT land at 51.8% and 50.6%, essentially the same as ACT+DTM (51.6%) and below Gated ACT+DTM (59.6%). Only the \"Numerical\" variant with hand-set weights (γv=0.7, γτ=0.3 before contact; 0.3/0.7 during) reaches 69.6%. That pattern is consistent with the gains coming from the particular fixed weights, not from the contact-state estimate. The paper says it adopted numerical scaling after seeing the simulation results, and the ablation that \"removes contact-aware scaling\" is only described qualitatively, without specifying the fixed readout weights. Without a control that uses the same readout and constant weights equal to the numerical values, we can't isolate contact-state dependence. That's a fixable but important gap.\n\nTwo smaller issues. The contact classifier is trained with ground-truth contact labels (Eq. 16); for the real-world data the paper never says how those labels were obtained. Also no error bars or significance tests are reported anywhere, which matters when the simulation protocol is 100 rollouts but real-world is only 20 trials per check.\n\nThe citation pattern looks fair, and the writing is clear. I don't see the circularity problem the reader flagged as severe—the scaling weights are openly user-specified, not fitted. The learnable variant failing is itself useful information, even if it undercuts the \"contact-aware\" interpretation.\n\nWho should read it: anyone working on visuo-tactile imitation learning or UMI-style hardware. It's not a home run, but the modular idea and the hardware are worth a look.\n\nRecommendation: send to peer review. Ask for the constant-weight control, error bars, and a description of real-world contact labels. With those, the paper could be solid.","headline":"A modular, well-tested fusion idea with a valuable hardware contribution, but the contact-state-dependence claim is undercut by the paper's own ablations.","tokens_in":14239,"tokens_out":2977,"would_cite":true,"duration_ms":24948,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CAAT hard-codes a contact phase switch into Transformer-based policies, emphasizing vision before contact and touch during contact, and reports gains of up to 21.1 percentage points over standard visuo-tactile fusion.","keywords":["contact-rich manipulation","visuo-tactile fusion","imitation learning","Transformer policies","attention scaling","tactile masking","data efficiency","tactile sensing"],"falsifier":"A decisive experiment is to train CAAT on the same demonstrations but replace the first-frame reference image with one taken mid-contact, or to start episodes already in contact, and measure success; if the average success rate does not collapse toward the gated-fusion baseline, the reference-image assumption is not the source of the gain, but the paper reports no such sensitivity check.","tokens_in":13226,"feed_emoji":"🤖","tokens_out":11875,"duration_ms":93222,"temperature":0.7,"pith_summary":"This paper argues that visuo-tactile policies fail at contact-rich manipulation under limited data because they leave the hard part—deciding when touch matters—to be learned from demonstrations, where contact occupies only a tiny fraction of the trajectory. CAAT instead hard-codes the phase structure: a binary contact estimate switches attention scaling so that vision dominates before contact and touch dominates during contact, while a masking step deletes tactile background regions that have not changed since the start of the episode. The paper reports that this explicit prior improves average success by 18.0 percentage points over direct visuo-tactile concatenation and 10.0 over learned gating in simulation, and by 21.1 over the strongest baseline in real-world trials, across ACT, Diffusion Policy, and $\\pi_0$. A sympathetic reader should care because the components are modular and add negligible overhead, which makes the claimed gains a plausible recipe for making tactile feedback usable in low-data regimes.","feed_headline":"Explicit contact priors beat learned fusion in tactile robot policies","feed_subtitle":"Switching attention from vision to touch at contact lifted success by 18-21 points across sim and real robot tasks.","key_machinery":"The load-bearing machinery is the contact-aware attention readout: after joint self-attention, the action query computes separate cross-attention readouts over the visual, tactile, and proprioceptive token subsets and combines them as $\\hat{r}_t = \\gamma_v(z_t) r^v_t + \\gamma_\\tau(z_t) r^\\tau_t + r^\\Upsilon_t$, where $z_t$ is a binary contact state from a CNN-MLP classifier on $|T_t - T_{\\rm ref}|$. The scaling factors are user-fixed functions satisfying $\\gamma_v(0) > \\gamma_\\tau(0)$ and $\\gamma_\\tau(1) > \\gamma_v(1)$, so vision is weighted before contact and touch after it. The mechanism's job is to replace a demonstration-hungry learned modality weighting with a structural prior, while retaining full cross-modal representation learning in the shared self-attention layers; Dynamic Tactile Masking then supplies the spatial counterpart by zeroing tactile tokens similar to the first-frame reference.","core_discovery":"The central discovery is that a task-level prior about contact—rather than more capacity or more data—is what unblocks visuo-tactile fusion. After joint self-attention, the action query reads out visual, tactile, and proprioceptive tokens separately, and the readouts are combined under a contact-conditioned weight $\\hat{r}_t = \\gamma_v(z_t) r^v_t + \\gamma_\\tau(z_t) r^\\tau_t + r^\\Upsilon_t$, with the weights flipped between $(0.7,0.3)$ and $(0.3,0.7)$ as the estimated contact state switches. This is paired with Dynamic Tactile Masking, which compares each current tactile patch token with the corresponding first-frame reference token and zeroes tokens whose cosine similarity exceeds $\\rho=0.8$, suppressing static background. The paper shows, in simulation and on a real two-finger tactile gripper, that this combination improves success rates over direct concatenation and learned gating, and that the fixed numerical scaling beats binary and learnable scaling variants. In the authors' own terms, the finding is that explicit contact-phase priors make Transformer-based visuo-tactile policies data-efficient without changing their action decoders.","pith_inferences":["Editorial inference: because the paper fixes Dynamic Tactile Masking across all baselines, the reported margins are joint effects; the individual contribution of contact-aware scaling alone is not isolated by a full ablation.","Editorial inference: the reference-image dependence suggests a natural robustness test—episodes beginning in contact or a corrupted reference frame—that the paper does not report; a reader should not assume the gains survive those conditions.","Editorial inference: the readout-scaling mechanism is not tied to binary contact; the same design could grade emphasis continuously from a force estimate or grasp-state signal, which the paper does not explore.","Editorial inference: the fixed numerical weights introduce a user choice; a small per-task search over $\\gamma$ values would test whether the exact constants matter, an experiment the paper leaves open."],"forward_implications":["Adding CAAT to an existing Transformer policy requires no change to its action decoder, so the same contact prior can be layered onto future policy backbones beyond the three tested.","The prior pays off most where demonstrations are scarce: with 25, 50, and 100 demonstrations in simulation, CAAT outperforms both baselines at every budget, so the gain is a data-efficiency gain, not just a ceiling-raising gain.","Dynamic Tactile Masking helps all fusion strategies, not only CAAT, which implies that suppressing static tactile background is an independently useful preprocessing step for visuo-tactile policy learning.","The fixed numerical scaling $(0.7,0.3)$ to $(0.3,0.7)$ beats both binary and learned scaling, suggesting that the exact weights are less important than having a smooth, phase-dependent preference rather than a hard switch.","In real-world trials the largest gains appear on open-box and powerbank-extraction tasks, where contact forces and local interaction matter most, so the mechanism's value scales with how contact-dominated the task is."],"supporting_citations":[{"why":"Supplies the simulation benchmark and datasets, and defines the direct visuo-tactile concatenation fusion baseline that CAAT is measured against.","marker":"Chen et al. 2026"},{"why":"Provides the learnable-gating visuo-tactile fusion baseline that CAAT outperforms by 10.0 points in simulation and in real-world trials.","marker":"Zang et al. 2026"},{"why":"Supplies the ACT Transformer policy backbone that carries the main simulation comparisons and one of the real-world backbones.","marker":"Zhao et al. 2023"},{"why":"Supplies the Diffusion Policy backbone used to test whether the contact prior generalizes across architectures.","marker":"Chi et al. 2025"},{"why":"Supplies the $\\pi_0$ backbone on which CAAT is fine-tuned for the third real-world policy family.","marker":"Black et al. 2024"},{"why":"Provides the UMI hand-held demonstration-collection paradigm that the paper's two-finger tactile gripper is designed from.","marker":"Chi et al. 2024"},{"why":"Supplies the PPTac tactile sensors mounted on the gripper fingertips that produce the tactile images used by the policy.","marker":"Lin et al. 2025"}],"fun_headline_variants":["Explicit contact priors beat learned fusion in tactile policies","Tactile masking and contact scaling lift visuo-tactile policy success","Contact-aware attention improves robot touch: 18-21% higher success","Switch to tactile at contact for data-efficient robot learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole switching mechanism rests on the assumption that a reliable binary contact state can be computed by comparing the current tactile image with a fixed non-contact reference image taken at the first timestep; if contact is already present at the start, the sensor drifts during an episode, or contact changes the tactile image only subtly, the scaling and masking will switch at the wrong time, and the paper reports no failure analysis for these cases.","fun_headline_variants_meta":{"raw":{"variants":["Explicit contact priors beat learned fusion in tactile policies","Tactile masking and contact scaling lift visuo-tactile policy success","Contact-aware attention improves robot touch: 18-21% higher success","Switch to tactile at contact for data-efficient robot learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001572,"raw_usage":{"total_tokens":6335,"prompt_tokens":1062,"completion_tokens":5273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":5200}},"tokens_in":678,"tokens_out":5273,"duration_ms":30557,"temperature":1.0,"reasoning_tokens":5200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:11:57.127104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment is to train CAAT on the same demonstrations but replace the first-frame reference image with one taken mid-contact, or to start episodes already in contact, and measure success; if the average success rate does not collapse toward the gated-fusion baseline, the reference-image assumption is not the source of the gain, but the paper reports no such sensitivity check.","supporting_citations":[],"review_version":2}