{"id":"348e2b91-ecba-436e-8443-3b3715509a34","arxiv_id":"2509.03889","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-arm robot folds and hangs crumpled shirts in mid-air using confidence-aware visual correspondences and touch-supervised grasp affordance.","lead":"This paper presents a two-armed robot that folds and hangs crumpled shirts while they hang in the air, using pixel-level matching to a flat reference shirt and touch sensors to choose safe grasps. It matters because it removes the usual need to flatten garments first, a step that limits most prior cloth manipulation systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Confidence calibration is the load-bearing assumption; the paper's own limitations admit overconfidence and threshold transfer variability, so the deferral mechanism may be brittle.","rationale":"The reader identified confidence calibration as the weakest assumption, and the manuscript's own limitation statements support that. The system's claimed novelty is confidence-aware deferral; the simulation comparisons (Fig 4) show distributional losses improve pixel matching but do not establish calibration on real garments. The small trial counts and missing real-world baselines are real weaknesses, but they weaken the evidence rather than the mechanism itself. The calibration issue is load-bearing because if confidence estimates are miscalibrated, the central 'defer action in low-confidence states' claim fails regardless of other components. The paper deserves credit for a reproducible simulated dataset, real robot deployment, and explicit failure analysis; this is not a rejection. A calibration-focused re-analysis would settle whether the threshold is robust. The reader's verdict of CONDITIONAL remains appropriate, so no change is recommended.","tokens_in":15380,"tokens_out":6100,"duration_ms":60754,"concrete_test":"Re-analyze the real-world correspondence data behind Table 1 and Appendix 7.4: for every query, record the predicted peak softmax probability and whether the region classification is correct. Build a reliability diagram (predicted confidence vs empirical accuracy) across confidence bins, separately for each shirt and for table vs suspended configurations, and compute expected calibration error. Then check whether the operating point 6e-6 corresponds to a consistent accuracy level across shirts; if calibration error is high or the optimal threshold varies substantially, the deferral mechanism is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the distributional correspondence model's confidence values reliably separating trustworthy from untrustworthy matches at a fixed threshold of 6e-6. Section 6 concedes that 'the network occasionally overestimates its certainty in challenging configurations,' and Appendix 7.4 reports that real-world transfer of thresholds 'introduced high variability'; the 6e-6 value is selected post hoc as a 'clear inflection point' on real images. Because these confidences are softmax probabilities over roughly 500k pixels trained with a hand-set Gaussian target, their absolute scale need not correspond to empirical accuracy. If the threshold does not transfer across shirts, lighting, and deformations, the deferral policy either grasps wrong points with high confidence or rejects useful grasps, and the claimed ability to handle occluded table-top and in-air states becomes a hand-tuned heuristic rather than a calibrated capability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a dual-arm robotic system for garment folding and hanging that operates directly on crumpled table-top and in-air garments. It couples a ResNet-based dense visual correspondence network trained with a distributional (KL) loss to output pixel-wise correspondence probabilities and confidence, a visuotactile grasp-affordance network fine-tuned via tactile labels, and a reactive state machine that defers grasping when confidence or affordance are below thresholds. Experiments report simulation correspondence accuracy, real-world region-level classification (e.g., 73.3% for suspended net, 68.9% safe decisions for combined), tactile classifier accuracy (99.7%), affordance precision@80 (76.3%), and end-to-end folding (6/10) and hanging (7/10) success. The paper also demonstrates extracting canonical grasp points from human videos.","tokens_in":15632,"tokens_out":6807,"duration_ms":63582,"significance":"If the reported results hold, the work would demonstrate a meaningful new capability: reactive in-air garment manipulation from crumpled states without flattening or full-state estimation, enabled by confidence-gated deferral and tactile feedback. The distributional loss for garment correspondences and the simulated dataset with realistic seams/hem details are useful contributions, and the tactile-grounded affordance fine-tuning is a sensible approach to mitigate sim2real gaps. However, the paper's central claim hinges on the calibration of the learned confidence estimates, which the authors themselves note is imperfect, and the empirical evaluation is thin (10 trials per task, no baselines). The contributions are potentially significant but require additional validation to be convincing.","major_comments":[{"comment":"The central mechanism is confidence-based deferral, but the paper does not establish that the confidence values from Eq. (1) are calibrated. Eq. (1) is a softmax over descriptor distances; training by KL divergence to a Gaussian target does not guarantee the resulting probabilities track empirical correctness. Section 6 concedes the network 'occasionally overestimates its certainty in challenging configurations,' and App. 7.4 reports that real-world transfer of thresholds 'introduced high variability,' with the deployed 6e-6 threshold chosen post hoc as a 'clear inflection point.' As a result, the reported safe-decision rates (68.9%, 70.8%, 80%) may reflect threshold tuning rather than calibrated confidence. Please provide calibration curves, a threshold sweep, and an ablation without deferral to substantiate the load-bearing claim.","section":"§6, App. 7.4, Eq. (1)"},{"comment":"The main task claims rest on 10 trials per task, with no confidence intervals or significance tests. Folding success 6/10 and hanging 7/10 change by 10% with one trial. The comparison 'without affordance fine-tuning' is also a single 10-trial run. There is no real-world comparison to prior correspondence-based folding ([1], [28]) or to a non-reactive baseline using the same network. For a journal-level claim of robust in-air manipulation, please report more trials, statistical treatment, and at least a no-deferral ablation.","section":"§4, Table 1, App. 7.2"},{"comment":"The real-world correspondence evaluation conflates correct rejection with calibration. A 'safe decision' is defined as either correct or below the 6e-6 threshold, but that threshold's calibration is exactly what is at issue. Region-level labels (Fig. 13) are coarser than pixel correspondences, and the inverse-direction evaluation only includes human-verifiable points, potentially biasing the sample. Please report precision-recall as a function of threshold, include all query points, and separate correct classifications from low-confidence rejections in each region.","section":"§4, App. 7.4"},{"comment":"The tactile classifier is both the supervision source for fine-tuning the affordance network and the in-loop validator for grasp success during execution. App. 7.5 acknowledges that this classifier cannot determine whether the grasped region corresponds to the intended visual target, so some positive labels are noisy (grasping cloth in front of the target). While the end-to-end folding/hanging metrics are independent, the affordance quality metric and the in-loop recovery both rely on the same classifier, creating potential bias. Please provide a human-verified evaluation of grasp success on a subset of trials, and report the tactile classifier's confusion matrix with respect to the intended-target criterion.","section":"§3.3, §4, App. 7.5"}],"minor_comments":[{"comment":"The notation for the predicted distribution switches between \\hat{p}_b and \\hat{p}_{b_i}; clarify the subscript convention.","section":"Sec. 3.2"},{"comment":"The grasp count is given as 8,500 in Sec. 3.3 but 8,000 in App. 7.5; reconcile the numbers.","section":"Sec. 3.3 vs App. 7.5"},{"comment":"The definition of 'correspondence success' for low-confidence outcomes is unclear; a trial with no grasp attempt is counted as correspondence success in Table 1. Explain how a non-grasp is scored as a correct correspondence.","section":"Table 1 caption"},{"comment":"The axis label says 'cumulative fraction of images' but the text says 'cumulative fraction of pixels'; make consistent.","section":"Fig. 4(a) caption"},{"comment":"The relationship between '6 out of 10 trials' and '30 total grasps attempted' is unclear; clarify whether the 6 trials are those in which at least one graspable keypoint was found.","section":"Sec. 4, folding results"},{"comment":"The human-video demonstration is described as 'outside the primary focus' yet is listed as a contribution in the abstract; consider moving it to future work or clearly labeling it as a preliminary illustration.","section":"App. 7.6"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a promising system, but the evidence base is thinner than the claims. The confidence-calibration concern is the main risk; the authors should be required to share calibration analyses and ideally code/data. Also, the lack of real-world comparisons to prior dense-correspondence folding methods (e.g., UniGarmentManip) is a gap for a journal submission. If the calibration and trial-count issues are addressed, the work could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid systems paper with a real new capability—folding and hanging shirts in air from crumpled states, without flattening—but the evidence base is thin and the confidence threshold that makes the whole reactive strategy work is hand-tuned and acknowledged by the authors as brittle. Worth refereeing, not worth treating as definitive.\n\nWhat's genuinely new: the distributional loss from MMGSD is extended to garments with realistic seams, and the confidence estimates are used inside a state machine that defers grasps, rotates the garment, and falls back to tactile validation. The visuotactile affordance, fine-tuned on 8,500 real grasps with a tactile classifier, is a nice piece of engineering. The paper also does real ablations: contrastive vs distributional, symmetric vs not, descriptor dimension, sigma, seams on/off. The authors are honest about limitations—they admit overconfidence, threshold transfer variability, and the human-demo pipeline being immature. That honesty earns credit.\n\nThe soft spots are real. The headline tasks run 10 trials each, no confidence intervals, no statistical comparison, and the real robot experiments don't compare against an existing folding or hanging system. The correspondence evaluation on real images uses region-level labels rather than pixel accuracy, which is reasonable but makes the numbers harder to interpret. And the 6e-6 threshold is chosen post hoc after noticing an inflection point; the paper itself says real-world threshold transfer was variable. That is the load-bearing piece. The stress-test note is on target: if the confidence values don't separate trustworthy from untrustworthy matches on a new shirt, the deferral behavior degrades into either wrong grasps or no-grasp loops. The system's rotation and tactile recovery soften this, but the central claim rests on a number that is essentially hand-set.\n\nThat said, I don't think this is a fatal flaw. The system does something prior work didn't: it operates on suspended, highly occluded garments and uses confidence to decide when to act. The limitations are clearly stated, and the simulation details are reproducible. I'd take it for peer review. A serious referee could ask for more trials, a real baseline comparison, and a robustness test of the threshold across shirts. The paper is fine for a conference with revision. I'd like to see it again after those additions.","headline":"Solid systems paper with a real in-air capability; thin stats and a hand-tuned confidence threshold make the headline claim provisional.","tokens_in":16058,"tokens_out":1982,"would_cite":true,"duration_ms":20816,"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":"A dual-arm robot folds and hangs crumpled shirts directly in midair by deferring actions when visual confidence is low.","keywords":["deformable object manipulation","clothing manipulation","dense visual correspondence","distributional loss","confidence-aware planning","visuotactile affordance","robotic folding","robotic hanging"],"falsifier":"Collect many real suspended-shirt queries, record the predicted peak confidence for each match, and compare whether accepted-by-threshold matches land in the correct region much more often than rejected ones. If the high-confidence set is not clearly more accurate than the low-confidence set, or no single threshold separates them reliably on real data, confidence-aware deferral is not doing the causal work. Alternatively, run the same folding task with deferral disabled by forcing a grasp at the first candidate regardless of confidence: if success does not drop, deferral is not load-bearing.","tokens_in":15341,"feed_emoji":"👕","tokens_out":7140,"duration_ms":66484,"temperature":0.7,"pith_summary":"Most robotic clothing systems flatten the garment first or assume key features are visible. This paper argues that a robot can fold and hang crumpled shirts directly—on a table and suspended in air—if it knows when its visual correspondences are trustworthy and can touch-check its grasps. The system pairs a dense correspondence network trained with a distributional loss, which outputs a probability heatmap rather than a single forced match, with a visuotactile grasp affordance network and a tactile classifier that validates contact. When correspondence confidence falls below a threshold, the robot rotates the shirt and re-evaluates instead of committing to a guess. The reported result is folding success in 6 of 10 trials and hanging success in 7 of 10, from highly occluded configurations without flattening or full-state estimation.","feed_headline":"Robot folds crumpled shirts midair, pausing when unsure","feed_subtitle":"Confidence-aware vision plus tactile feedback lets a dual-arm robot fold and hang garments without flattening them first.","key_machinery":"The distributional dense correspondence network and the confidence-based state machine it feeds. The network maps pixels to descriptors and, via a softmax over descriptor distances, outputs a probability heatmap over the canonical shirt; a KL loss against a multimodal Gaussian target lets symmetric regions count as multiple valid matches, producing confidence estimates. Those estimates gate every grasp: only above-threshold correspondence and affordance lead to an action, otherwise the shirt rotates and is re-evaluated. The tactile classifier plays a complementary role, supervising the affordance network, validating grasp success in real time, and signaling when the shirt is under tension.","core_discovery":"The claim is that confidence-aware dense correspondence plus visuotactile affordance lets a dual-arm robot manipulate garments in air directly from crumpled states. A ResNet-based network maps each pixel of a deformed shirt to a descriptor and predicts a probability distribution over matches on a canonical shirt, trained with a KL divergence to a multimodal Gaussian target that allows symmetric valid matches; the result is per-pixel confidence heatmaps. A state machine gates every grasp: both correspondence confidence and grasp affordance must clear thresholds, otherwise the shirt is rotated 30 degrees and re-evaluated, with a fallback grab to unfold it. High-resolution tactile sensing super","pith_inferences":["An implicit design choice is a single fixed confidence threshold; a runtime-adaptive threshold that reads the shape of the whole probability heatmap would likely improve the low-confidence cases the paper reports most, such as bottom regions and heavily folded sleeves.","The deferral principle is not garment-specific: any deformable object with symmetric or ambiguous correspondences, such as ropes, bags, or upholstery, could use the same confidence-gated state machine provided a canonical template exists.","Because the tactile classifier supervises affordance and validates grasps online, the system could collect large amounts of self-supervised grasp data during normal operation, plausibly improving both affordance and confidence calibration over time.","The human-video pipeline is limited by hand occlusion and skeleton errors, but it points to a scalable data loop: correspondence plus confidence could automatically filter unreliable human demonstrations."],"forward_implications":["Folding and hanging become feasible directly from crumpled and suspended states, removing flattening as a required preprocessing step.","Confidence thresholds make the system robust to self-occlusion: when features are hidden, the robot rotates instead of guessing, converting some hard perceptual errors into recoverable events.","The same task-agnostic grasp affordance module serves both folding and hanging, so a new task does not require retraining the grasp selector from scratch.","Tactile validation turns empty grasps, excess fabric, and slipped cloth into recoverable failures rather than terminal errors.","The dense descriptor map can act as a shared interface for other planners, such as translating human video demonstrations into robot grasp targets."],"supporting_citations":[{"why":"Supplies the distributional multimodal Gaussian shape descriptor formulation that the correspondence model adapts to shirts with seams and hems.","marker":"[29]"},{"why":"Provides the visuotactile affordance U-Net architecture and tactile classifier training procedure that the paper extends from towels to garments.","marker":"[19]"},{"why":"Introduces dense visual descriptors for task-agnostic manipulation, the representation baseline this work builds on and compares against.","marker":"[26]"},{"why":"Prior garment dense-correspondence manipulation framework used as a contrastive-loss baseline and canonical-template comparison.","marker":"[28]"},{"why":"Contrastive dense-descriptor fabric smoothing and folding method whose table-top results frame the real-world comparison for picking from the table.","marker":"[1]"},{"why":"The tactile sensor hardware that provides the high-resolution contact images used for grasp validation and tension detection.","marker":"[36]"},{"why":"Simulation software used to generate the large deformed-shirt dataset with varied geometry, lighting, and configuration.","marker":"[33]"},{"why":"Adds hems, stitches, and seams to the simulated shirts, giving the network the structural cues needed for real-world correspondence.","marker":"[34]"}],"fun_headline_variants":["Dual-arm robot folds crumpled shirts midair with tactile grip","Robot pauses and rotates when unsure folding crumpled shirts midair","Visuotactile AI handles hang-fold of crumpled garments without flattening","Confidence-aware robot folds crumpled shirts midair, defers uncertain grabs","Tactile-supervised robot folds hanging crumpled shirts, gated by confidence"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the confidence numbers produced by the correspondence network are trustworthy on real, unseen garments, so that a single fixed threshold separates dependable matches from unreliable ones; if the network systematically overestimates its certainty, the deferral strategy loses its reason to wait and the claimed in-air robustness collapses.","fun_headline_variants_meta":{"raw":{"variants":["Dual-arm robot folds crumpled shirts midair with tactile grip","Robot pauses and rotates when unsure folding crumpled shirts midair","Visuotactile AI handles hang-fold of crumpled garments without flattening","Confidence-aware robot folds crumpled shirts midair, defers uncertain grabs","Tactile-supervised robot folds hanging crumpled shirts, gated by confidence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1193,"prompt_tokens":735,"completion_tokens":458,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":360}},"tokens_in":479,"tokens_out":458,"duration_ms":5194,"temperature":1.0,"reasoning_tokens":360,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:33:11.247468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect many real suspended-shirt queries, record the predicted peak confidence for each match, and compare whether accepted-by-threshold matches land in the correct region much more often than rejected ones. If the high-confidence set is not clearly more accurate than the low-confidence set, or no single threshold separates them reliably on real data, confidence-aware deferral is not doing the causal work. Alternatively, run the same folding task with deferral disabled by forcing a grasp at the first candidate regardless of confidence: if success does not drop, deferral is not load-bearing.","supporting_citations":[{"cited_title":"Ganapathi, P","cited_arxiv_id":null,"evidence_quote":"Supplies the distributional multimodal Gaussian shape descriptor formulation that the correspondence model adapts to shirts with seams and hems."},{"cited_title":"Sunil, S","cited_arxiv_id":null,"evidence_quote":"Provides the visuotactile affordance U-Net architecture and tactile classifier training procedure that the paper extends from towels to garments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior garment dense-correspondence manipulation framework used as a contrastive-loss baseline and canonical-template comparison."},{"cited_title":"Ganapathi, P","cited_arxiv_id":null,"evidence_quote":"Contrastive dense-descriptor fabric smoothing and folding method whose table-top results frame the real-world comparison for picking from the table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The tactile sensor hardware that provides the high-resolution contact images used for grasp validation and tension detection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Simulation software used to generate the large deformed-shirt dataset with varied geometry, lighting, and configuration."},{"cited_title":"Albisser","cited_arxiv_id":null,"evidence_quote":"Adds hems, stitches, and seams to the simulated shirts, giving the network the structural cues needed for real-world correspondence."}],"review_version":1}