{"id":"c75c4850-fa93-4023-9192-02c7e5afe59f","arxiv_id":"2505.06451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A closed-loop force-torque feedback extension of few-shot imitation learning lets a robot adapt its wiping force to unseen surface heights and sponge properties, reaching 96% of reference force in real-world tests.","lead":"This paper adds real-time force-torque feedback to a few-shot imitation learning pipeline for robot wiping, letting the robot adjust its pressing force when the table height or sponge changes. In physical tests the robot delivered about 96% of the reference wiping force across 40 scenarios, versus 4% for the open-loop baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 96% headline ratio is measured against Table II reference forces for 9 unseen sponges, yet the paper documents demonstrations only for the normal sponge; the reference-force provenance must be established before the central claim is supportable.","rationale":"The reader's weakest assumption identified the provenance of the Table II reference forces as load-bearing, and my reading agrees. The paper reports 8 demonstrations with a normal sponge in IV-C2 and never describes collecting demonstrations with the 9 custom sponges, yet Table II lists reference forces for all of them. Since the main result is expressed as a percentage of these reference forces, the validity of the denominator determines whether '96% accuracy' is a comparison to human wiping performance or a comparison to an internal target. This is a reporting gap that can be closed by the authors, which is why the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. I do not see a more fundamental flaw: the method is concrete, the hardware experiments cover the claimed 40 scenarios, the ablations in the appendix address model depth, window size, and demonstration count, and the wall-wiping result provides a partial transfer check. The remaining concerns about single-trial evaluation and lack of an ablation isolating the pretrained encoder are secondary; they affect the precision of the claim and the attribution of the effect, but neither is as decisive as the reference-force question. The proposed check targets exactly the missing information: if per-sponge demonstration forces are available or can be collected, recomputing the ratios will confirm or refute the 96% headline. Until then, the central quantitative claim should be treated as conditional on reference-force provenance.","tokens_in":11908,"tokens_out":4869,"duration_ms":50305,"concrete_test":"Request or reproduce the reference-force collection protocol for all 10 sponges. Specifically, check whether Table II was produced by kinesthetically recording human demonstrations for each custom sponge (s1f1 through s3f3) under the same instructions as IV-C2, with the same table geometry. If the protocol exists, recompute the percentage ratios in Table I and Table III as (robot mean vertical force)/(per-sponge human demonstration mean vertical force)*100 using those directly measured references; report the average across the 40 scenarios and its spread. If the protocol does not exist, the reference forces should be measured with a force sensor during human wiping of each sponge, and the evaluation rerun; if the resulting average departs from 96% by more than about 5 percentage points, the headline claim is not supported by the current data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that the method applies, on average, 96% of the reference force across 40 scenarios (abstract; V-A; V-B). The reference forces used for this ratio are listed in Table II for every sponge, e.g., Normal -12.6 N, s1f1 -22.8 N, ..., s3f3 -36.7 N. However, the only demonstration collection protocol described in IV-C2 states: 'We collected 8 demonstrations using a normal sponge...' No protocol is given for obtaining reference forces for the 9 custom-made sponges. If those reference values were not measured from human demonstrations carried out with each sponge under the same instructions, then the denominator of the 96% ratio is not an independent external benchmark; it is an internally defined target, and the headline comparison to human wiping force is not established. This is load-bearing because the paper's abstract and both result sections use exactly this ratio to claim adaptation to unseen sponge properties. The concern is likely an omission in reporting rather than a deliberate defect, but it must be resolved before the central claim can be accepted on its quantitative evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a few-shot imitation-learning framework for contact-rich wiping that combines a pre-trained haptic representation of object properties with a real-time force-torque feedback loop. The method first pre-trains a VAE encoder on simulated exploratory FT trajectories to obtain a sponge-property latent code, then trains a motion-trajectory decoder and an FT feedback decoder on eight kinesthetic demonstrations with a single normal sponge. At deployment, the robot collects an exploratory FT trajectory for the current sponge, encodes it to obtain z_sponge, generates an offline planar wiping trajectory, and uses the FT feedback loop to predict online vertical displacements. The experiments cover 40 scenarios (10 sponges x 4 surface types, including a wall) and compare against a no-feedback imitation baseline and an admittance controller. The headline result is that the method applies, on average, 96% of a per-sponge reference force, whereas the baseline achieves 4% and admittance control 42%.","tokens_in":12142,"tokens_out":7609,"duration_ms":78083,"significance":"If the quantitative claims hold, the paper makes a useful empirical contribution: it shows that a small number of demonstrations, combined with a pre-trained haptic representation and online FT feedback, can adapt a wiping policy to unseen surface heights and unseen deformable-object properties on real hardware. The 40-scenario deployment study is substantial, and the ablation of the FT-loop depth, TCN window size, and number of demonstrations in the appendix is a positive feature. The comparison to a no-feedback imitation baseline is relevant and the wall-wiping test is an interesting stress test. However, the strength of the central 96% claim depends on the provenance of the per-sponge reference forces and on statistical support that the current single-trial presentation does not provide.","major_comments":[{"comment":"The 96% headline figure is computed relative to the reference forces in Table II, yet the paper documents human demonstrations only for the normal sponge ('We collected 8 demonstrations using a normal sponge', Section IV-C2). No protocol is given for obtaining the reference forces listed for the nine custom sponges (e.g., s1f1 -22.8 N, s3f3 -36.7 N). If those values were not obtained from genuine human demonstrations performed with each sponge under the same 'apply as much force as possible' instruction, then the denominator of the 96% ratio is not an independent external benchmark, and the abstract's claim of applying 'reference forces' is not established. Please specify the measurement protocol, including the number of demonstrations per sponge and whether the same demonstrator and instruction were used.","section":"IV-C2, Table II"},{"comment":"The FT feedback decoder θ_height is trained only on the normal sponge: the eight demonstrations (IV-C2) and the single real unlabeled trajectory (IV-C1) all use the normal sponge, so during training z_sponge is effectively a single constant vector. At deployment the same decoder receives z_sponge codes of nine unseen sponges. The paper provides no analysis or ablation showing that θ_height output varies appropriately with z_sponge; the appendix ablates FT-loop depth, TCN window size, and number of demonstrations, but not the pretrained encoder, and instead refers to [16] for pre-training ablations. Without evidence that the decoder responds meaningfully to z_sponge, the claim that the method adapts to unseen sponge properties through the pretrained representation is unsupported.","section":"III-B2, III-C, Table IV"},{"comment":"Each reported condition appears to be a single trial: the standard deviations in Table I are within-trajectory across time steps, not across independent runs. No confidence intervals or statistical tests are provided. Given the large within-trajectory variability in the proposed method (e.g., s1f1 low: -18.0 ± 11.9 N; s1f2 low: -28.7 ± 18.0 N), the claims that the method 'significantly outperforms' baselines and that the applied force does not 'significantly vary' with height are not statistically supported.","section":"Table I, V-A, V-B"},{"comment":"The wall-wiping evaluation compares force measurements in a gravity-neutral configuration against the Table II reference forces, which were measured during table wiping where the vertical force includes the sponge's weight (as acknowledged by the baseline's positive no-contact force in Section V-A). If no gravity compensation or offset was applied in the wall configuration, then the ratios in Table III (e.g., Normal 115%, s1f1 104%) are not directly comparable to the reference, and the claim of '104% of the reference force' for wall wiping is not established.","section":"V-C, Table III"},{"comment":"The admittance control baseline is given a target force defined as 'the force applied when the sponge is pressed by 1cm' (Section V). This target is arbitrary and not derived from the task or from human demonstrations, so the AC comparison does not demonstrate that admittance control cannot adapt to sponge properties; it only shows that this particular target choice yields lower forces than the reference. The comparison to the no-feedback baseline (Aoyama et al.) is more informative.","section":"V, Eq. (4)"}],"minor_comments":[{"comment":"The phrase '96% accuracy in applying reference forces' is misleading: the quantity is an average ratio of applied force to reference force, not a classification accuracy. Please rephrase to 'on average applied 96% of the reference force' or similar.","section":"Abstract"},{"comment":"The statement that 'the FT trajectories of the unlabeled data collected both in simulation and in the real world were similar' is unquantified; please provide a similarity metric or illustrative comparison.","section":"IV-C1"},{"comment":"The phrase '1 demonstration unlabeled data' is unclear; it should be 'one unlabeled trajectory' or 'one real-world unlabeled trajectory'.","section":"IV-C2"},{"comment":"The 'Average' row of Table I is difficult to read because it contains multiple numbers without clear column alignment; please reformat so that each column (Contact, Average [N], Std) is explicit for the average row as well.","section":"Table I"},{"comment":"The sentence 'with the standard deviation being only about 5% larger than that of human demonstrations' is ambiguous: it is not clear whether the comparison is across trials, across heights, or within trajectories. Please clarify and report the corresponding numbers.","section":"V-A"}],"recommendation":"major_revision","confidential_remarks":"The main gate for this paper is the provenance of the Table II reference forces. If those values cannot be traced to independent demonstrations with each sponge, the central 96% claim should be revised. I also recommend requiring the authors to state the number of trials per condition and to provide confidence intervals or raw data, since the current table reports single-run point estimates with large within-trajectory variance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a legitimate, practical extension of Aoyama et al.'s haptic-representation imitation framework, adding a closed-loop force-torque feedback loop that adjusts vertical position online. The real-robot experiments show a real difference: the baseline loses contact and essentially fails, while the proposed method maintains contact and applies forces in the right ballpark across height changes and unseen sponges. The wall-wiping test is a nice gravity-neutral check. That said, the headline \"96% of reference force\" is weaker than it looks, because the paper never documents where the reference forces for the 9 custom sponges come from.\n\nWhat's genuinely new is the integration of a pre-trained object-property encoder with a TCN-based FT feedback loop that predicts the next vertical displacement from recent FT history. That is a small but sensible step beyond [16]'s open-loop policy. The appendix ablations give some evidence that depth, window size, and number of demonstrations matter, and the authors honestly acknowledge the higher variance relative to admittance control.\n\nThe main soft spot is the provenance of Table II. Section IV-C2 describes collecting 8 demonstrations with a normal sponge, but Table II lists average forces for all 10 sponges. If those values came from human demonstrations with each sponge, that is fine but must be reported; if they came from simulation or a scaling rule, then the 96% is measured against an internally defined target, not an independent human benchmark. This is load-bearing for the central quantitative claim, so the authors must clarify it. Also, each condition appears to be a single trial, and the standard deviations in Table I are large, often 10-20 N against references of 20-35 N. The 96% average needs repeated trials or confidence intervals to be more than a point estimate. The missing ablation for the pre-trained encoder is less concerning since they defer to [16], but it would still strengthen the paper.\n\nWho should read this: people working on force-based imitation learning or contact-rich manipulation with deformable objects will find the experiment design useful. This is not a field-redefining breakthrough, but it is a solid engineering contribution with clearly stated limitations.\n\nI would send it to peer review. The main issue is a reporting gap, not a fatal flaw. The authors should be asked to document the reference-force collection protocol and to provide repeated trials or error bars. With those changes, the paper would be acceptable.","headline":"Useful closed-loop extension of a haptic few-shot imitation framework, but the 96% headline rests on reference forces for unseen sponges whose provenance is never documented.","tokens_in":12647,"tokens_out":2146,"would_cite":false,"duration_ms":21595,"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":"With eight demonstrations and real-time force-torque feedback, a robot wipes unseen surfaces and sponges at 96% of human reference force.","keywords":["imitation learning","contact-rich manipulation","force-torque feedback","deformable objects","representation learning","few-shot learning","force control","wiping task"],"falsifier":"Have independent human demonstrators wipe with each of the nine custom sponges on the same surfaces, measure their applied forces, and re-run the 40-scenario evaluation against those measured references; if the robot's average ratio falls well below 96% or varies strongly across sponges, the claim of adaptation to unseen sponge properties is not established by the current metric. A second check would be to ablate the sponge latent code during deployment: if force tracking is unchanged when $Z_{\\mathrm{sponge}}$ is randomized or fixed, the pre-trained representation is not the mechanism driving the adaptation.","tokens_in":11724,"feed_emoji":"🧽","tokens_out":12552,"duration_ms":103548,"temperature":0.7,"pith_summary":"The paper sets out to show that a few-shot imitation policy can adapt to changes that break open-loop wiping: different table heights and different sponge stiffness/friction. It closes the loop with a force-torque feedback module that takes the last five force-torque samples plus a pre-trained latent code for the sponge's physical properties and predicts the next vertical displacement of the end-effector. Across 40 scenarios on real hardware, the method kept contact 100% of the time and applied on average 96% of the reference force recorded from human demonstrations, while the open-loop baseline applied 4% and an admittance controller applied 42%. The same model, trained only on table-wiping demonstrations, also wiped a vertical wall while applying 104% of the reference force. If the result holds, contact-rich deformable-object manipulation can be made robust to environment shifts with very few demonstrations and no explicit target force or position.","feed_headline":"Force feedback lifts robot wiping force to 96% of human reference","feed_subtitle":"With only eight demonstrations, the robot adapts to unseen surface heights and sponges in every tested scenario.","key_machinery":"The load-bearing mechanism is the FT feedback loop $\\phi_{\\mathrm{ft}}-\\theta_{\\mathrm{height}}$: a two-layer temporal convolutional network ($\\phi_{\\mathrm{ft}}$) with 25 hidden channels encodes the five most recent force-torque samples into a six-dimensional latent state $Z_{\\mathrm{ft}}$, which is concatenated with the five-dimensional pre-trained sponge-property code $Z_{\\mathrm{sponge}}$ from the VAE encoder $\\phi_{\\mathrm{sponge}}$; a two-layer decoder $\\theta_{\\mathrm{height}}$ then outputs the next vertical displacement $\\Delta\\hat{h}_{t+1}$. The loop is trained on the few demonstrations by minimizing the MSE between predicted and demonstrated vertical displacements, and at deployment it runs online while the planar wiping trajectory is replayed open-loop. The frozen pre-trained encoder supplies the only object-property information the loop has for unseen sponges, so $Z_{\\mathrm{sponge}}$ carries the adaptation to stiffness and friction. This split between an offline planar path and an online one-dimensional vertical correction is what lets a small demonstration set and real-time force sensing coexist.","core_discovery":"The central claim is that coupling a pre-trained latent representation of deformable-object properties with an online force-torque feedback loop converts a few-shot imitation policy into an adaptive force controller. Given eight kinesthetic demonstrations of table wiping with one sponge, the robot replays the planar wiping path offline from the demonstrated motion while predicting the end-effector's next vertical displacement online from the past five force-torque readings and the latent code of the sponge currently in hand. This closed-loop vertical prediction is what allows the robot to keep a demonstration-comparable force when the table is low, high, sloped, or replaced by a vertical wall, and across sponges whose stiffness and friction were never seen in training. The paper reports an average of 96% of the reference force over all 40 scenarios, versus 4% for the open-loop baseline and 42% for admittance control, and 104% for the wall-wiping transfer.","pith_inferences":["The same latent-code-plus-feedback recipe could transfer to scrubbing, polishing, or in-hand soft-object manipulation, wherever a demonstrated path exists and the uncertain axis is the one in contact.","Ablating $Z_{\\mathrm{sponge}}$ during deployment, by randomizing or fixing it to a constant, would test whether the pre-trained representation actually carries the sponge adaptation; the paper does not report this experiment.","The reported averages across height settings may smooth over systematic errors on extreme sponges; checking the per-sponge ordering of applied force against each sponge's stiffness/friction would reveal whether the 96% average hides a compensating pattern."],"forward_implications":["A wiping policy trained on eight demonstrations of a single sponge can maintain full contact and near-demonstration force across surface heights it never saw, including a sloped table and a vertical wall.","For deformable, elastic tools, the method removes the need to specify a target force or target position in advance, which is the practical obstacle that makes admittance and impedance control awkward in this setting.","The offline/online split reduces the environment-adaptation problem to one axis, so the same architecture is a candidate template for other contact-rich manipulations with one uncertain contact axis.","The measured variability is about three times that of admittance control, meaning the near-target average force comes with a consistency cost that future versions need to reduce."],"supporting_citations":[{"why":"It supplies the pre-trained haptic representation and few-shot motion decoder that this work extends, and it is the open-loop baseline the experiments compare against.","marker":"[16]"},{"why":"It defines the impedance-control family used as a comparison point for closed-loop force control.","marker":"[17]"},{"why":"It provides the adaptive admittance-control method used as the non-learning baseline.","marker":"[18]"},{"why":"It grounds the variational autoencoder objective used to pre-train the sponge-property latent space.","marker":"[19]"},{"why":"It provides the temporal convolutional network used to encode force-torque history in the feedback loop.","marker":"[20]"},{"why":"It supplies the admittance-control update equation used to instantiate the AC baseline in the experiments.","marker":"[25]"}],"fun_headline_variants":["96% force accuracy: few-shot wiping adapts with FT feedback","Force-torque loop lifts wiping to 96% of reference from 4%","Adaptive wiping: 8 demos, unseen surfaces, 96% force accuracy","Few-shot wiping with FT feedback hits 96% force, beats baseline","Robot wiping adapts via FT feedback: 96% force vs 4% open-loop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 96% figure assumes the per-sponge reference forces in Table II are genuine human-demonstration benchmarks for each of the ten sponges, whereas the dataset section only documents collecting demonstrations with the normal sponge.","fun_headline_variants_meta":{"raw":{"variants":["96% force accuracy: few-shot wiping adapts with FT feedback","Force-torque loop lifts wiping to 96% of reference from 4%","Adaptive wiping: 8 demos, unseen surfaces, 96% force accuracy","Few-shot wiping with FT feedback hits 96% force, beats baseline","Robot wiping adapts via FT feedback: 96% force vs 4% open-loop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1335,"prompt_tokens":953,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":569,"tokens_out":382,"duration_ms":3882,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:42:09.934179+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have independent human demonstrators wipe with each of the nine custom sponges on the same surfaces, measure their applied forces, and re-run the 40-scenario evaluation against those measured references; if the robot's average ratio falls well below 96% or varies strongly across sponges, the claim of adaptation to unseen sponge properties is not established by the current metric. A second check would be to ablate the sponge latent code during deployment: if force tracking is unchanged when $Z_{\\mathrm{sponge}}$ is randomized or fixed, the pre-trained representation is not the mechanism driving the adaptation.","supporting_citations":[{"cited_title":"Few-shot learning of force-based motions from demonstration through pre-training of haptic representation,","cited_arxiv_id":null,"evidence_quote":"It supplies the pre-trained haptic representation and few-shot motion decoder that this work extends, and it is the open-loop baseline the experiments compare against."},{"cited_title":"Adaptive admittance control: An approach to explicit force control in compliant motion,","cited_arxiv_id":null,"evidence_quote":"It provides the adaptive admittance-control method used as the non-learning baseline."},{"cited_title":"A tutorial survey and comparison of impedance control on robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"It supplies the admittance-control update equation used to instantiate the AC baseline in the experiments."}],"review_version":1}