{"id":"b4530545-ce70-4883-8018-b979cc63c394","arxiv_id":"2608.00730","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Push-Wiper uses a diffusion policy to generate segmented sponge-pushing trajectories that aggregate viscous stains into a compact region, then removes the residue with post-processing primitives.","lead":"This paper presents Push-Wiper, a robot framework that cleans sticky spills by pushing them together with a sponge rather than wiping them across the surface. It reports cleaning scores up to 130% higher than baselines and zero-shot transfer to new stains and curved surfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Cleaning Score advantage depends on an unvalidated, morphology-based stain detector that may miss the thin residual films produced by baseline smearing, so the central quantitative claim is not yet established.","rationale":"The reader's weakest_assumption identifies exactly the most load-bearing concern: the Cleaning Score, which is the paper's headline quantitative result, depends entirely on a stain detector that is never validated. The paper explicitly describes the detector as 'denoised' and based on morphology and connected-component filtering, which are precisely the operations that would remove thin, low-contrast films left by smearing baselines. Since the baselines' failure mode is smearing and Push-Wiper's success mode is aggregation into a compact blob, the detector could systematically favor the proposed method. This is not a disagreement with the paradigm or an ad hominem concern; it is a measurement-instrument problem. The proposed test—manual or gravimetric ground truth on a subset of trials—would settle whether the reported CS gap is real or an artifact. Because the concern is serious but addressable, and the reader already assigned CONDITIONAL, no change to the reader's verdict is needed.","tokens_in":10167,"tokens_out":2728,"duration_ms":26466,"concrete_test":"Select 10 ketchup and 10 peanut butter trials per method, prioritizing runs where baselines leave visually thin smears; manually label residual stain pixels from the wrist-camera images (or weigh extracted residue) and recompute CS from these ground-truth masks. If the detector-based CS gap between Push-Wiper and baselines shrinks by more than about 20% or no longer supports the reported 130% improvement, Eq. (6) is biased and the central comparison must be re-run with a validated metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—CS up to 130% higher than baselines—is measured exclusively through the custom stain detector in Section IV-A. Eq. (6) defines CS as 1 - N_after/N_before, where N_before and N_after are pixel counts from a 'denoised stain mask' produced by fusing HSV/Lab/grayscale cues with morphology and connected-component filtering. The paper provides no calibration of this detector against ground-truth stain area. This matters because the baselines' characteristic failure mode is smearing: Fig. 4 and Table I show FC and PO leave thin, spread films and even negative CS, while Push-Wiper's aggregation produces a compact, high-contrast blob. A denoising/thresholding pipeline can systematically erase faint, low-contrast smears while preserving the aggregated blob, inflating the measured gap (89.88 vs 32.64 and 44.98). A secondary confound is that PO is trained on synthetic Pygame sweep trajectories rather than expert demonstrations, mixing strategy and data-quality effects, but the unvalidated detector is the more fundamental threat to the headline number.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Push-Wiper proposes a framework for robotic cleaning of viscous stains that reformulates the task as an aggregation problem. The system uses a Diffusion Policy over binary stain maps to generate segmented pushing trajectories, an Arbitrary Surface Pose Interpolator (ASPI) to map 2D actions onto 3D surfaces, and a hybrid force–position admittance controller for execution. The gathering phase is followed by a post-processing phase with predefined primitives. Experiments on ketchup and peanut butter compare Push-Wiper against two baselines (Full-Cover and PushAll-Onetime) using a Cleaning Score (CS), defined as the percentage reduction in detected stain pixels. The paper reports an overall average CS of 89.88 for Push-Wiper versus 32.64 and 44.98 for the baselines, and demonstrates zero-shot generalization to curved surfaces, solids, liquids, and unseen viscous stains.","tokens_in":10476,"tokens_out":4182,"duration_ms":39133,"significance":"If the reported results hold, the aggregate-then-finish paradigm is a valuable and potentially transformative contribution to robotic surface cleaning, where viscous stains have resisted simple wiping or scrubbing strategies. The decoupled design, in which the policy reasons only about 2D binary topology while geometric execution is handled by ASPI and force control, is elegant and leads to a genuinely impressive zero-shot transfer to curved surfaces and unobserved stain types without retraining. The experimental setup is fair in important ways: all methods share the same perception and execution stack, the same force controller, and the same wall-clock time budget per trial, with baselines evaluated at their best CS within that budget. The paper also makes concrete falsifiable claims and provides a project website, which supports reproducibility. However, the central quantitative claim rests on a custom stain detector that is not validated, and the headline improvement is not supported by statistical tests, so the significance of the numerical advantage is not yet firmly established.","major_comments":[{"comment":"The Cleaning Score is computed from an unvalidated stain detector. The detector fuses HSV/Lab/grayscale cues with morphology and connected-component filtering, but the paper provides no calibration of this detector against ground-truth stain area, no manual validation subset, and no sensitivity analysis. This is load-bearing because the baselines' characteristic failure mode is smearing into thin, low-contrast films (Fig. 4 shows severe smearing for Full-Cover and residual fragments for PushAll-Onetime), while Push-Wiper produces a compact, high-contrast aggregated blob. A thresholding/morphology pipeline can systematically erase faint smears while preserving the blob, inflating the measured CS gap (89.88 versus 32.64 and 44.98). The authors should report a validation of the detector on a held-out set with ground-truth stain area (e.g., known printed stain sizes or manual pixel annotation) and show that the reported qualitative conclusions are robust to detector parameters.","section":"Section IV-A, Eq. (6)"},{"comment":"No statistical significance tests are reported despite large baseline variance. For example, Full-Cover on peanut butter complex stains has mean 6.11 with standard deviation 38.45, so the mean is not even one standard error above zero, and the paper's claim that Push-Wiper 'significantly outperforms' the baselines is unsupported by the reported statistics. The paper should provide paired per-trial comparisons (since the same stain distributions are used across methods), or at least bootstrap confidence intervals for the mean differences, for each stain type and complexity level. Without such tests, the 'up to 130% higher' claim is a point estimate with unknown reliability.","section":"Table I"},{"comment":"The PushAll-Onetime baseline is trained on synthetic Pygame-generated global sweep trajectories, while Push-Wiper is trained on teleoperated expert demonstrations. This confounds the comparison: the performance gap could be due to the aggregation strategy itself, or due to the distribution shift and lower quality of synthetic supervision for the baseline. The paper states this design enables 'a controlled comparison' of global sweep versus aggregation-first, but the training data differ in a way that is not controlled. To support the attribution of the improvement to the segmented pushing strategy, the authors should either train PO on the same expert demonstrations (e.g., with experts performing global sweeps) or provide an additional ablation that controls for the supervision source.","section":"Section IV-B"}],"minor_comments":[{"comment":"The phrase 'up to 130% higher' is ambiguous. Table I shows much larger relative improvements (e.g., peanut-butter complex: 89.13 vs 6.11 is a roughly 1300% increase), so '130%' likely refers to the improvement over the average of the two baselines rather than over either individual baseline. Please state the reference used for the percentage improvement.","section":"Abstract"},{"comment":"The protocol of reporting the 'best CS each baseline attains at any stopping point within T' is a post-hoc selection that can inflate baseline performance and is not a standard evaluation protocol. Please specify how the stopping point is chosen (e.g., whether it is selected by an oracle after observing the full run) and discuss the effect on the comparison.","section":"Section IV-C"},{"comment":"The post-processing evaluation reports only five trials per stain type with no standard deviation. Please include variance or individual trial results, especially since the reported improvement from post-processing (e.g., peanut butter 85.12 to 98.51) is based on small samples.","section":"Table IV"},{"comment":"The aggregation objective J_t = D(S_t) + λK(S_t) is introduced as the formal goal, but it is not directly measured during experiments. Please clarify how λ and the stopping threshold ϵ are chosen and provide a sensitivity analysis, or explicitly state that this objective is only a design motivation.","section":"Equation (2)"},{"comment":"The curved-surface experiments use 10 trials per geometry (5 ketchup, 5 peanut butter), which is a small sample. Given the strong claims about zero-shot generalization, it would be helpful to report individual trial results or confidence intervals, and to include more than two geometries.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong fit for a robotics audience, and the aggregation-first paradigm is genuinely novel. The main risk is the unvalidated cleaning-score metric: if the detector systematically ignores thin smears, the headline advantage is an artifact. I would ask the authors to validate the detector against ground truth and to add significance testing before acceptance. The PO baseline confound is also worth addressing, but it is secondary to the metric issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Push-Wiper. The core idea—reformulate viscous-stain cleaning as aggregation via segmented pushes, then finish with wipes/dabs—is genuinely new and clearly motivated. The paper's design choice to decouple 2D topological planning from 3D execution through a binary stain map and a surface interpolator is clean, and it makes the zero-shot curved-surface results believable. The controlled comparison (same perception and execution stack, matched wall-clock budget, scraping after each trajectory) is fairer than most cleaning papers. I put the central quantitative claim on hold, though. The Cleaning Score depends entirely on a custom stain detector that is never validated against ground truth. The baselines' failure mode is smearing—thin, faint films—and the detector's morphology and connected-component filtering can systematically erase those films while keeping the compact aggregated blob that Push-Wiper produces. That would inflate the 89.88 vs 32.64/44.98 gap. This isn't a hypothetical nitpick; it's the load-bearing measurement. The PO baseline also mixes strategy with data source: it's supervised by synthetic Pygame sweeps, while Push-Wiper uses teleoperated expert demos. That alone could explain part of the gap. There are no significance tests, and the generalization experiments are 10 trials per cell. Minor: the post-processing section has five trials per stain and no error bars, fine for a sanity check. What holds up: the aggregation-first framing, the decoupling, and the zero-shot transfer to convex/concave surfaces and novel substances are real and worth pursuing. The paper is not a desk reject. I'd send it to review with a clear request: validate the stain detector against hand-labeled or physical ground truth (e.g., weighed residue), run the baseline with expert demonstrations, and report per-trial data or confidence intervals. If the detector holds, the contribution is solid. Who's it for: robotic cleaning and contact-rich manipulation researchers, especially people working with deformable or fluid-like materials. I would bring it to a reading group, and I'd cite it if I worked on cleaning. My recommendation: accept for peer review with major revisions.","headline":"A plausible new cleaning paradigm whose headline number is measured by an unvalidated stain detector; worth refereeing after the metric is grounded.","tokens_in":10914,"tokens_out":1725,"would_cite":true,"duration_ms":15787,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Viscous stains come off fastest when a robot first pushes them into one compact blob, the paper argues.","keywords":["robotic cleaning","viscous stains","aggregation","diffusion policy","hybrid force-position control","zero-shot generalization","stain segmentation","surface cleaning"],"falsifier":"Run matched ketchup and peanut butter trials for Push-Wiper and the two baselines, then measure residue by an independent method, for example weighing the surface before and after cleaning under controlled conditions or photographing under fixed illumination and hand-labeling the wetted area, and compare rankings against the pixel-based Cleaning Score. If full-coverage wiping leaves less physically removable material than the detector suggests, the aggregation advantage shrinks or vanishes.","tokens_in":9951,"feed_emoji":"🧽","tokens_out":6417,"duration_ms":57482,"temperature":0.7,"pith_summary":"The paper claims that the reliable way to clean viscous stains is not to wipe or scrub them away directly, but to first push the material inward until it forms one compact blob, then wipe that blob away. Push-Wiper implements this aggregate-then-finish idea with segmented pushing trajectories planned from a binary stain map and executed under hybrid force-position control. On ketchup and peanut butter trials, the reported Cleaning Score averages 89.88 out of 100, compared with 32.64 for full-coverage wiping and 44.98 for one-shot global pushing, an improvement the paper puts at up to 130%. The same trained system is reported to transfer without retraining to curved surfaces, solid debris, liquid spills, and previously unseen viscous stains.","feed_headline":"Blob-first cleaning beats wiping on sticky stains","feed_subtitle":"Aggregate-then-finish strategy reports up to 130% higher cleaning scores on ketchup and peanut butter.","key_machinery":"The load-bearing mechanism is the aggregate-then-finish loop with segmented pushing. At each macro-step, the policy infers a single complete pushing stroke that acts as a local optimizer for $D(S_t)+\\lambda K(S_t)$, the stain's maximal spatial diameter plus fragmentation; ASPI then converts the predicted 2D translation plus yaw into a 6D pose trajectory aligned with the local surface normal, and the admittance controller keeps the sponge in stable contact at a constant normal force. This decoupling separates 2D topological planning from 3D geometric execution, so the policy can be trained entirely on planar data and still generalize to curved surfaces.","core_discovery":"The paper's central claim is that cleaning performance for high-viscosity contaminants is determined less by contact modeling than by the topological strategy of aggregation: repeatedly pushing the stain's boundary inward shrinks its spatial diameter and merges disconnected fragments, converting a smeary removal problem into a localized removal problem. A Diffusion Policy trained on segmented expert strokes outputs low-dimensional actions $a=(x_b,y_b,\\Delta\\theta)$ from a texture-free binary stain map, deliberately ignoring visual appearance and 3D geometry. The Arbitrary Surface Pose Interpolator (ASPI) maps those 2D actions onto the measured surface normal to produce smooth 6D trajectories, and an admittance controller holds the normal force at a setpoint while tracking the remaining axes. The paper reports that this decoupling yields near-complete cleaning on planar, convex, and concave surfaces and argues that aggregation, not wiping or scrubbing, is the right abstraction for general-purpose cleaning.","pith_inferences":["Inference: The binary-map abstraction suggests the learned skill is geometric herding rather than material-specific contact behavior, so the same trajectory policy might aggregate granular debris or deformable objects whenever the tool can push them; the paper only demonstrates this for rigid disks.","Inference: The reported advantage over wiping may be partly detector-dependent: if the stain detector undercounts thin residual films, the method that leaves a compact blob will look cleaner than one that smears a thin layer. An independent mass-based or hand-labeled ground-truth measurement would separate true removal from perceptual convenience.","Inference: The finish phase could itself be learned or coverage-planned rather than fixed primitives, which may extend the framework to stains too large or too sticky to gather into one small region."],"forward_implications":["If the central claim is correct, coverage-style wiping, the default for many cleaning robots, is the wrong primitive for semi-solid stains; aggregation strokes should come first.","Because the policy plans only in 2D plus yaw, expert demonstrations collected on flat tabletops can be reused for convex and concave surfaces without additional curved-surface training data.","Keeping force feedback in the execution layer rather than in the policy means the same learned pushing behavior can be ported to different manipulators by adjusting the force setpoint.","The post-processing primitives, dab, scrape, rinse, squeeze, and final wipe, turn a near-complete gathering result into an almost complete clean, raising average Cleaning Score from 89.44 to 99.25 in the reported trials."],"supporting_citations":[{"why":"Supplies the Diffusion Policy used to generate segmented pushing actions from binary stain maps.","marker":"[12]"},{"why":"Supplies the unified admittance control law behind the hybrid force-position controller that maintains the 20 N normal contact.","marker":"[28]"},{"why":"Provides the apparent viscosity range of ketchup used to define the easier test stain.","marker":"[29]"},{"why":"Provides the higher viscosity of peanut butter used to justify the harder test stain.","marker":"[30]"},{"why":"Frames viscous stains as a fluid-mechanics challenge whose rheology makes wiping spread contamination, motivating aggregation.","marker":"[6]"}],"fun_headline_variants":["Aggregate, don't wipe: robot boosts sticky-stain cleaning by 130%","Push-Wiper: pushing stains together beats wiping them apart","Robot cleaner uses aggregation to remove viscous stains with 130% gain","Zero-shot transfer: Push-Wiper cleans unseen sticky messes by aggregation","Forget scrubbing: robot aggregates viscous stains into one easy-to-remove blob"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Section IV-A, is that the Cleaning Score, computed from a custom HSV/Lab/grayscale stain detector with morphology filtering, faithfully measures true stain-area reduction; if that detector misses the thin smear films the baselines leave while seeing Push-Wiper's compact blob, the reported advantage would be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Aggregate, don't wipe: robot boosts sticky-stain cleaning by 130%","Push-Wiper: pushing stains together beats wiping them apart","Robot cleaner uses aggregation to remove viscous stains with 130% gain","Zero-shot transfer: Push-Wiper cleans unseen sticky messes by aggregation","Forget scrubbing: robot aggregates viscous stains into one easy-to-remove blob"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2880,"prompt_tokens":962,"completion_tokens":1918,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1835}},"tokens_in":578,"tokens_out":1918,"duration_ms":15526,"temperature":1.0,"reasoning_tokens":1835,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:17:13.601532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run matched ketchup and peanut butter trials for Push-Wiper and the two baselines, then measure residue by an independent method, for example weighing the surface before and after cleaning under controlled conditions or photographing under fixed illumination and hand-labeling the wetted area, and compare rankings against the pixel-based Cleaning Score. If full-coverage wiping leaves less physically removable material than the detector suggests, the aggregation advantage shrinks or vanishes.","supporting_citations":[{"cited_title":"Rheological proper- ties of tomato ketchup","cited_arxiv_id":null,"evidence_quote":"Provides the apparent viscosity range of ketchup used to define the easier test stain."},{"cited_title":"Rheological properties of peanut butter,","cited_arxiv_id":null,"evidence_quote":"Provides the higher viscosity of peanut butter used to justify the harder test stain."},{"cited_title":"The fluid mechanics of cleaning and decontamination of surfaces,","cited_arxiv_id":null,"evidence_quote":"Frames viscous stains as a fluid-mechanics challenge whose rheology makes wiping spread contamination, motivating aggregation."}],"review_version":2}