{"id":"acbef00e-bd2b-4319-a782-dd0d6f20403d","arxiv_id":"2608.01824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A vision-language-action robot policy that recursively refines its predicted future tactile latents from live tactile feedback improves measured success on seven real-robot dexterous tasks by 18 to 24 points over baselines.","lead":"ReTouch is a robot-control model that predicts future touch signals and updates those predictions during a task, helping a five-fingered hand adapt to slips and changing contact. On seven real-world manipulation tasks with 900 demonstrations, it reports success-rate gains of 18 to 24 percentage points over prior tactile robot policies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 'success rate' is a graded partial-credit task score (App. B.3), so the headline 18.4/23.8 pp gains may not be binary success gains; no error bars or significance tests are provided.","rationale":"In good faith, the paper has real support: closed-loop real-robot comparisons, consistent ablations, and an offline diagnostic isolating recursive refinement. The reader's CONDITIONAL verdict is sensible. I would not elevate to REJECT because the method may be sound and the metric issue is fixable by reporting binary rates and uncertainty. However, I do not fully agree that the privileged-distillation assumption is the single most load-bearing weakness. The FAE is jointly trained with a direct action-generation loss, and the ablations (Table 3) and D.3 comparison already isolate the refinement mechanism through action re-inference with latest tactile history; the risk is more about how much of the latent gain transfers, not whether the mechanism is circular. The more decisive issue is that the central quantitative claim is expressed as 'success rate' when the metric is a graded task score, and no statistical support is reported for 20-rollout averages. A strict binary recomputation is a concrete, feasible check that would settle whether the headline magnitudes are meaningful. Thus the conditional acceptance should remain, with binary metrics and uncertainty reporting as conditions.","tokens_in":17844,"tokens_out":13175,"duration_ms":123381,"concrete_test":"Recompute the headline comparison under a strict binary definition: for every rollout, count only a full task score of 1.0 as success, treating all partial-credit scores as failure. Recompute ReTouch versus the strongest baseline in the standard and challenging settings, and report the macro-average binary success difference with a bootstrap or permutation confidence interval over the 20 rollouts per task. If the binary advantage is materially smaller than 18.4 and 23.8 percentage points, or its confidence interval overlaps zero, the central claim as worded is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 4.2 frame the central claim as average success rate, but Appendix B.3 states that the primary metric is the mean normalized task score, with additive stage-based partial credit for six of seven tasks and binary completion only for Button Press. The numbers in Tables 1–3 therefore are not frequencies of completed tasks, and the magnitude of the headline gain depends on the credit schedule. For example, Sponge Wipe awards 0.5 for partial mark removal, so an 87.5% mean can occur without a single fully clean wipe; partial credit can also compress or expand differences between policies. Additionally, each cell is 20 rollouts with no reported variance, confidence intervals, or significance tests, and the offline diagnostics in Table 6 aggregate 100 trajectories without counts or intervals. Together these make the quantitative central claim less secure than the paper's wording suggests. The privileged-distillation assumption identified by the reader is a real secondary concern, but it is partially mitigated by the direct action loss on the FAE and by real-robot ablations that isolate recursive refinement; the metric and statistical issue is the more load-bearing threat to the stated claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReTouch, a vision-language-action model for contact-rich dexterous manipulation that combines structured tactile patch encoding with online refinement of predicted tactile latents and action chunks. A Tactile-Patch Encoder partitions each finger's 120 taxels into five functional patches and pools them into finger tokens, preserving finger identity and local contact structure. During training, a privileged Hindsight Action Expert (HAE) produces action-relevant future-tactile latent targets from ground-truth future tactile signals, and a Foresight Action Expert (FAE) is trained, via cosine alignment and action losses, to predict and recursively refine those latents using only deployable observations. At execution, the FAE runs at 36 Hz and, at each refinement offset, updates the future tactile latents and regenerates the remaining action chunk, conditioned on the latest tactile feedback. The authors collect XHT-Dataset, 900 demonstrations across seven contact-rich tasks on an XHand--UR7e platform, and evaluate ReTouch in closed-loop real-robot experiments, reporting average normalized task scores of 83.6% versus 65.2% for the strongest baseline under standard settings and 73.1% versus 49.3% under challenging settings.","tokens_in":18103,"tokens_out":6031,"duration_ms":56256,"significance":"If the reported results hold, ReTouch would be a meaningful advance for contact-rich dexterous manipulation: it demonstrates that future tactile latents can serve as an execution-time control state that is recursively refined from incoming tactile feedback, and it provides a real-robot evaluation platform and dataset that are valuable to the community. The paper's strengths include genuinely closed-loop real-robot experiments with a dense tactile stream, a carefully decomposed ablation set that isolates individual mechanisms, and sample-matched offline diagnostics. The main quantitative claim, however, rests on a graded partial-credit task score rather than binary completion, and the reported numbers come from 20 rollouts per cell without uncertainty quantification. The mechanism-level validation of tactile-latent fidelity is also self-referential, since it compares predicted latents to HAE latents produced by the same training pipeline. These issues are addressable and do not, in my assessment, invalidate the central idea, but they do need to be fixed before the paper's stated claims can be considered fully supported.","major_comments":[{"comment":"The primary metric is the mean normalized task score with additive stage-based partial credit for six of seven tasks; only Button Press uses a binary criterion. The paper repeatedly calls this metric 'success rate,' and the headline gains of 18.4 and 23.8 percentage points are differences in this graded score, not in binary completion frequencies. For example, Sponge Wipe awards 0.5 when part of the marks is removed but visible traces remain, so an 87.5% score is compatible with no fully cleaned wipe. Because the partial-credit schedule can compress or expand differences between policies, the central claim is not as secure as the wording suggests. Please report binary task-completion rates alongside the graded scores, at least for the main comparison and the key ablations, and discuss how the credit schedule affects the reported gaps.","section":"Section 4.1, Appendix B.3, Tables 1-3"},{"comment":"Each cell in Tables 1-3 is based on 20 real-robot rollouts, but no variance, confidence intervals, or significance tests are reported. Table 6 aggregates 100 trajectories without counts or intervals. The observed gaps, including the 23.6, 15.2, and 14.5 percentage point ablation drops, may be real, but with 20 trials per cell they are not statistically secure. Please add per-task and per-condition confidence intervals, or explicitly acknowledge the exploratory nature of the numbers and avoid unsupported precision in the abstract.","section":"Tables 1-3 and Section 4.2"},{"comment":"The mechanism-level claim that predicted tactile latents carry action-relevant contact information is validated only by cosine similarity to HAE latents produced by the same training pipeline. At deployment the HAE is removed, leaving the FAE latents as the sole carrier of future-contact information, so this validation is, as it stands, self-referential for the physical-fidelity claim. I note that the direct action loss on the FAE and the real-robot ablations partially mitigate this concern. Still, please add a direct check of predicted latents against ground-truth future tactile signals, for example by decoding predicted latents into patch-level forces and comparing them with measured future taxels, or explicitly state this limitation in the paper.","section":"Section 3.3, Eq. (5), Fig. 5, Table 6"},{"comment":"Training samples refinement offsets only from {4, 8, 12}, but deployment invokes the FAE at the 36 Hz control rate and applies the unexecuted suffix at every call, which in principle includes intermediate offsets not seen during training. The paper does not justify that random-offset training on three offsets transfers to the full range of deployed offsets. Please clarify whether deployment calls are in fact restricted to offsets {4, 8, 12} within each 16-step chunk, or provide evidence that the model generalizes to all intermediate offsets.","section":"Section 3.4, Appendix A.3"}],"minor_comments":[{"comment":"The term 'success rate' is used throughout for what Appendix B.3 defines as a normalized task score with partial credit. Please use 'normalized task score' or 'mean task score' in the abstract and main text wherever the graded metric is meant, and reserve 'success rate' for the binary Button Press task.","section":"Abstract and Sections 4.1-4.2"},{"comment":"The panel labels in Figure 5 do not match the caption: the caption describes future tactile prediction and remaining-action MSE, while the figure contains repeated 'Pipette Press' labels. Please correct the figure so that the labels correspond to the caption and the plotted quantities.","section":"Figure 5"},{"comment":"The mask M_t is not defined precisely. Please specify its length, how it selects the latent prefix corresponding to the elapsed portion of the action chunk, and how it is constructed for the initial call and for deployment calls at arbitrary offsets.","section":"Eq. (6)"},{"comment":"The notation for the number of eligible samples N_o is not fully specified. Please state how samples are filtered and how cases with no eligible future-tactile tokens are handled in the aggregate statistics.","section":"Appendix D.3, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The graded-metric concern raised in the review is the main threat to the central claim, and it is fixable: reporting binary completion counts and confidence intervals would greatly strengthen the paper. The privileged-distillation issue is secondary but should be addressed, since the mechanism-level evidence is currently circular. I did not find the paper internally inconsistent, and the real-robot evaluation with a new dataset is a solid empirical contribution that fits the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nReTouch is worth a look, but read the numbers with a grain of salt. The genuinely new idea is treating future tactile latents as a recursively maintained control state: the model re-estimates its tactile forecast at 36 Hz from incoming feedback and regenerates the action chunk with each update. I haven't seen that exact mechanism in the predictive-reactive baselines they cite, and the real-robot ablations back it up: removing intra-chunk refinement costs 23.6 points, and the fixed-prediction variant sits 15.2 points below full ReTouch. The paper is also more honest than many in this area: the tactile-patch encoder is a sensible engineering contribution, and the latency table shows they actually thought about the control loop.\n\nThe soft spots are real, though. The headline 'success rate' is a graded, partial-credit task score, not binary completion frequency. That's buried in Appendix B.3. For example, Sponge Wipe awards 0.5 for partial mark removal, so a mean of 87.5% is possible with no fully clean wipe. The 18.4 and 23.8 percentage-point gains are valid under their rubric, but they are not success-rate gains in the ordinary sense, and the abstract contributes to the overstatement. Compounding that, each cell is 20 rollouts with no variance, confidence intervals, or significance tests. The ablation deltas (23.6, 15.2, 14.5) are large enough that I doubt they'd vanish with error bars, but I'd like to see them.\n\nThe privileged-distillation assumption--aligning FAE latents to HAE latents from the same training pipeline--is a real secondary concern, but it's partially mitigated by the direct action loss on the FAE and by the real-robot ablation that isolates refinement. A physical-fidelity check (predicted contacts vs. actual contacts) would be a fair request.\n\nThe work is also not reproducible as-is: no public code, no public dataset. That's increasingly the norm in this corner of robotics, but it still caps the contribution until they ship the data.\n\nWho is this for? Anyone working on tactile policy learning, especially on how to reuse predictive latents in a closed loop. The metric and stats issues are fixable, and the core mechanism is genuinely interesting. If they release data/code and add confidence intervals, this would be a solid systems paper. I'd send it to serious peer review now--a good reviewer will push on the metric overstatement--but I would not treat the headline numbers as established yet.\n\nRegards.","headline":"Genuinely new control mechanism (recursive refinement of predicted tactile latents at 36 Hz) with strong real-robot ablations, but the headline gains are graded partial-credit scores, not binary success rates, and there are no error bars.","tokens_in":18640,"tokens_out":2389,"would_cite":true,"duration_ms":20477,"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":"ReTouch claims that treating future tactile predictions as a continuously refined control state—updated at 36 Hz from incoming contact feedback—raises average success by 18.4 percentage points over the strongest baseline in seven…","keywords":["dexterous manipulation","tactile sensing","vision-language-action model","tactile prediction","online refinement","contact-rich manipulation","real-robot learning","XHT-Dataset"],"falsifier":"Two checks would settle it: reproduce the Appendix D.3 sample-matched comparison on held-out trajectories and see whether recursive refinement still beats one-shot given the same fresh tactile history—if the latent cosine and action-MSE gains vanish, the refinement mechanism is not the carrier; and decode the refined latents into taxel-level contact maps, comparing them with the real future tactile frames, to verify that the alignment gain corresponds to physical contact fidelity rather than agreement with the privileged expert's code alone.","tokens_in":17612,"feed_emoji":"🤖","tokens_out":9045,"duration_ms":75039,"temperature":0.7,"pith_summary":"ReTouch is a vision-language-action model for five-finger manipulation that treats the predicted future touch state as something to be re-estimated during execution, not fixed once at the start of an action. As the hand executes, new tactile readings arrive at high rate and the model repeatedly refines its forecast of the next contact states and its remaining action chunk together, at 36 Hz. The paper argues this closed-loop refinement keeps the tactile forecast aligned with the physical interaction, so slips and contact shifts can be corrected before they cascade. On a new 900-demonstration dataset of seven contact-rich tasks, ReTouch reports an average success rate of 83.6%, beating the strongest baseline by 18.4 percentage points, and 73.1% under out-of-distribution and disturbance settings, beating the strongest baseline there by 23.8 points. If the mechanism is right, future touch is better used as a live control state than as a fixed prediction.","feed_headline":"Robot hand refines touch forecasts at 36 Hz for 18-point success gain","feed_subtitle":"Tactile forecasts re-estimated with incoming feedback let a five-finger hand recover from slips and contact shifts.","key_machinery":"The load-bearing object is the future tactile latent: a compact internal code, aligned by cosine distance to a privileged expert's layer-12 representations, that stands in for the next contact states and conditions action generation. Three parts carry the argument. The Tactile-Patch Encoder splits each finger's 120 taxels into five functional patches (tip, center, base, left, right), pools them with soft contact weights, and adds finger-identity and patch-position embeddings so each finger token keeps local contact structure. During training, the Hindsight Action Expert uses ground-truth future tactile patch features to create target latents, and the Foresight Action Expert is trained to predict those latents from the cached VLM context, chunk-start state, and observed tactile history, with a directional attention mask so action tokens cannot leak into the tactile latents. At execution, a masked carry-over operation re-inserts the already-elapsed part of the previous latent estimate into the Foresight queries, letting the model recursively refine the forecast and the remaining action chunk at 36 Hz against the VLM's slower 9 Hz semantic updates; this is the mechanism the ablations credit for most of the reported gain.","core_discovery":"The paper's central claim is that the limiting factor in tactile-aware dexterous manipulation is not how well the model can predict touch one more time, but whether it re-estimates that prediction as contact evolves. ReTouch therefore keeps the future tactile state inside the control loop: during training, a privileged Hindsight Action Expert sees the ground-truth future tactile frames and produces action-relevant target latents, while a Foresight Action Expert must predict the same latents from current observations; during deployment, only the Foresight expert runs, and at every 36 Hz call it updates the future tactile latents and re-predicts the remaining action chunk from the latest tactile history. The paper reports that this beats the strongest baseline by 18.4 points on average (83.6% vs 65.2%), by 23.8 points under challenging conditions, and that removing intra-chunk refinement drops average success by 23.6 points, while keeping the initial tactile prediction fixed but re-generating actions drops it by 15.2 points. The supporting offline diagnostic, on 100 held-out trajectories, shows the recursive refinement raises in-contact latent cosine similarity by 0.908% and lowers action-suffix error by 2.139% relative to a one-shot prediction that still re-infers actions from fresh tactile history.","pith_inferences":["The cosine-alignment objective supervises latents only against the privileged expert's own internal code, so the method actually certifies action-relevant agreement, not physical fidelity of the predicted touch; a natural extension would be to decode the refined latents back to taxel-level contact maps and check them against real future frames.","The same recursive-latent pattern could apply to any fast sensor stream—force/torque, proximity, strain—inside a slow semantic backbone, so the paper's contribution may generalize beyond tactile sensing to general asymmetric-rate control.","Because gains concentrate in tasks with sustained or shifting contact (Sponge Wipe, Liquid Transfer, pulling recovery) and ReTouch does not clearly win the visually dominated Cabinet Retrieval task, a reader should expect the method's advantage to depend on how much contact drift a task contains.","Treating the refined touch latent as a control state invites a comparison against an explicit model-based tactile predictor (e.g., a learned forward model that consumes the same tactile history); such a test would reveal how much of the gain comes from the learned latent representation versus the closed-loop re-estimation schedule."],"forward_implications":["A tactile forecast generated once at chunk start goes stale during contact-rich execution; ReTouch claims the refinement of the forecast itself—not just re-inference of actions from new touch—is what restores accuracy, with a 15.2-point drop when the prediction is frozen.","Robustness gains are largest exactly where contact changes: 30.0 points over the strongest baseline under a post-grasp pulling disturbance, 20.0 points under lighting change, and 25.0 points on sustained-contact Sponge Wipe.","Structured finger/patch encoding contributes 14.5 points; flattening raw taxels into an MLP input removes the finger identity and local contact structure the policy needs.","The timing of refinement matters: delaying the joint update by roughly one action step (non-blocking variant) costs 7.8 points, so the value lies in updating before the next command is sent.","The 9 Hz VLM / 36 Hz action-expert split is itself part of the claim: high-rate contact correction can happen without re-running the vision-language backbone."],"supporting_citations":[{"why":"Provides the pretrained π0 VLA checkpoint ReTouch is initialized from and defines the action flow-matching formulation it inherits.","marker":"Black et al. 2024"},{"why":"Provides the π0.5 checkpoint used to build the π0.5 and π0.5+tactile baselines that define the challenging-condition comparison.","marker":"Physical Intelligence et al. 2025"},{"why":"Tactile-VLA is the strongest average baseline at 65.2%, the direct comparison behind the +18.4 percentage-point claim.","marker":"Huang et al. 2025b"},{"why":"ViTacFormer supplies the ACT-style visuo-tactile baseline with future tactile prediction and is the best baseline on Cabinet Retrieval.","marker":"Heng et al. 2026"},{"why":"RDP is the slow-fast tactile-reactive baseline whose fixed high-rate action refinement ReTouch contrasts with prediction refinement.","marker":"Xue et al. 2025"},{"why":"TouchWorld represents the predictive-reactive design pattern ReTouch claims to depart from by re-estimating the tactile reference inside the chunk.","marker":"Zhou et al. 2026"}],"fun_headline_variants":["Tactile predictions refined in-loop boost robot hand success by 18%","Online-refined touch forecasts lift dexterous hand success by 18.4 points","ReTouch re-estimates touch in real time, boosting hand dexterity by 18%","Closed-loop tactile prediction refinement ups dexterous manipulation by 18.4%","At 36 Hz, ReTouch re-projects tactile future and lifts success by 18 points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cosine-aligning the predicted tactile latents to the internal representations of a training-only expert that sees the true future touch transfers the contact information the actions actually need, even though the alignment is never checked against physical fidelity of the predicted touch; if that transfer is lossy, the online refinement loop would guide actions with a distorted forecast.","fun_headline_variants_meta":{"raw":{"variants":["Tactile predictions refined in-loop boost robot hand success by 18%","Online-refined touch forecasts lift dexterous hand success by 18.4 points","ReTouch re-estimates touch in real time, boosting hand dexterity by 18%","Closed-loop tactile prediction refinement ups dexterous manipulation by 18.4%","At 36 Hz, ReTouch re-projects tactile future and lifts success by 18 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000981,"raw_usage":{"total_tokens":4223,"prompt_tokens":1060,"completion_tokens":3163,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":3050}},"tokens_in":676,"tokens_out":3163,"duration_ms":20019,"temperature":1.0,"reasoning_tokens":3050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:04:08.563275+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two checks would settle it: reproduce the Appendix D.3 sample-matched comparison on held-out trajectories and see whether recursive refinement still beats one-shot given the same fresh tactile history—if the latent cosine and action-MSE gains vanish, the refinement mechanism is not the carrier; and decode the refined latents into taxel-level contact maps, comparing them with the real future tactile frames, to verify that the alignment gain corresponds to physical contact fidelity rather than agreement with the privileged expert's code alone.","supporting_citations":[],"review_version":2}