{"id":"c70219c2-91c3-407e-b835-5cacdab4722d","arxiv_id":"2603.08560","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In contact-rich robotic disassembly, compact force-field tactile representations (TacFF) outperform vision-only and high-resolution tactile-image policies, especially in tight-tolerance and deformable tasks; naive fusion of two tactile representations degrades performance.","lead":"This paper tests whether adding touch sensors helps robots take things apart by pulling, sliding, and releasing clips, and finds that a compact force-field representation of touch beats both camera-only control and high-resolution touch images. The work is a systematic comparison across five simulated and five real disassembly tasks, and may help guide sensor and representation choices for recycling, maintenance, and remanufacturing robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing matched-resolution control: TacFF's advantage over TacRGB may be due to 10×14 compactness, not force-field content; extraction algorithm also unspecified.","rationale":"The reader's CONDITIONAL verdict already captures the main risk. My read strengthens the specificity: the unspecified TacFF extraction is one component, but the deeper issue is the missing matched-resolution control. TacRGB and TacFF differ simultaneously in semantic content and in input dimensionality, so the comparison in Tables I–II cannot identify which factor drives the result. In the small-data regime, this is a genuine threat to the central 'rather than high-resolution tactile imagery' claim. That said, the paper has independent support: the same qualitative ordering appears in simulation with large N, the task suite is thoughtfully varied, the fusion condition (Table V) is a useful negative result, and the dim-lighting robustness result is internally consistent. These do not outweigh the confound, but they make rejection unnecessary; the paper can be made acceptable by adding the control and specifying the TacFF extraction/calibration. Hence the reader's conditional verdict remains appropriate.","tokens_in":10729,"tokens_out":8599,"duration_ms":88302,"concrete_test":"Add a condition Vision + TacRGB-lowres in which the GelSight RGB frame is downsampled/cropped to 10×14 (same as TacFF) and fed through the identical ResNet-18 encoder, trained on the same 50 demonstrations and evaluated on the same real-world trials for R2–R5 (ideally n≥50 per cell). If TacFF does not statistically beat TacRGB-lowres, the reported advantage is attributable to compactness/input size rather than force-field content; if it does, the force-semantics claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that TacFF's success-rate advantage is caused by structured force semantics, but the experiments do not control for observation dimensionality or preprocessing. In §II.C/§III.A, TacRGB is a 240×160 RGB image while TacFF is a 10×14 grid of shear/normal channels, and both are encoded by the same ResNet-18. With only 50 demonstrations and 400 epochs, a high-dimensional raw image can underperform because of overfitting or input noise regardless of modality. The paper reports no TacRGB condition downsampled to 10×14 to isolate compactness from force semantics. This is not merely a missing baseline: it is the specific alternative explanation for Tables I–II and the stronger failure of TacRGB in R2/R3/R4. Additionally, §II.B says TacFF is 'computed from optical flow and depth reconstruction' but gives no algorithm or calibration, and sim (TacSL) vs. real (GelSight R1.5) consistency is unverified. Without a matched-size raw-image control and a reproducible extraction description, the conclusion 'structured force fields, not high-resolution imagery' is confounded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates tactile sensing for robotic disassembly using a unified Diffusion Policy framework. It introduces five simulated rigid-body tasks (S1–S5) and five real-world tasks (R1–R5, including deformable components), and compares three sensing configurations: Vision Only, Vision + TacRGB (high-resolution tactile image), and Vision + TacFF (compact 10×14 force-field grid). Across simulation and real-world experiments, TacFF-based policies achieve the highest success rates, with the largest margins in contact-dependent and deformable tasks. The paper also reports that naive concatenation of TacRGB and TacFF degrades performance, arguing that structured force representations are more effective than high-dimensional tactile imagery.","tokens_in":11014,"tokens_out":3410,"duration_ms":35567,"significance":"If the central claim holds, the paper makes a valuable, actionable contribution to tactile representation design for contact-rich manipulation: it suggests that compact force-field encodings can outperform high-resolution tactile images, and it provides a disassembly benchmark spanning increasing contact complexity. The strengths are the unified policy framework (all modalities compared under identical training protocols), a relatively large simulation evaluation (1500 rollouts per configuration per task), and real-world tasks that include deformable contact interactions, which are under-represented in prior tactile manipulation benchmarks. The consistent pattern across simulation and real-world for the shared tasks (S1/R1, S2/R2, S3/R3) is a useful indicator. However, the load-bearing conclusion about representation semantics is currently underdetermined by the experiments: the comparison does not control for input dimensionality, and the TacFF representation is not specified at a level that permits reproduction or assessment of sim-to-real fidelity.","major_comments":[{"comment":"The central claim is that TacFF's advantage over TacRGB is caused by structured force semantics, not by lower input dimensionality. But the comparison is confounded: TacRGB is a 240×160 image, while TacFF is a 10×14 grid, and both are encoded by the same ResNet-18. With only 50 demonstrations and 400 training epochs, a higher-dimensional input may underperform due to overfitting or representation sparsity rather than because force content is worse. I ask for a matched-resolution control: downsample TacRGB to 10×14 (or upsample/project TacFF to a comparable visual form) and retrain. Without such a control, the abstract and conclusions' claim that 'structured force fields, not high-resolution imagery' is the active ingredient is not supported.","section":"§III.A and Tables I–II"},{"comment":"The TacFF representation is described only qualitatively: 'computed from optical flow and depth reconstruction' with shear components and normal force over a grid. No algorithm, equations, calibration procedure, or mapping between GelSight R1.5 measurements and the 10×14 grid is given. It is also unclear whether the simulation (TacSL) and the real GelSight sensor use the same extraction pipeline and normalization; Fig. 3 suggests visual similarity but does not quantify agreement. Because Table II's central comparison depends on the quality and consistency of TacFF, the absence of this detail is a reproducibility and validity gap. Please provide the extraction method (or cite a public implementation), the calibration steps, and at least a qualitative/quantitative validation that the sim and real force fields correspond.","section":"§II.B, TacFF definition"},{"comment":"The real-world results rest on 20 trials per configuration per task, with no confidence intervals or statistical tests. Several cell-to-cell swings are large (e.g., R2: TacRGB 30% vs TacFF 70%; R4: 5% vs 75%), but with n=20 the standard error is around 10 percentage points at p≈0.5, so some differences that drive the 'consistently highest' claim may not be statistically reliable. The dim-lighting and fusion results (Tables IV and V) use the same small n. Table I reports 1500 rollouts per configuration, but no standard errors or seed-level variability are shown, and some margins (e.g., S3: 53.5 vs 50.0; S5: 61.9 vs 59.5) are small. Please include confidence intervals, per-seed breakdowns, and/or significance tests for the key comparisons, and if possible increase the real-world rollout count or clearly state the precision limitation.","section":"§III.C, Tables II, IV, V"}],"minor_comments":[{"comment":"The sentence 'Robotic disassembly is a critical capability for recycling, maintenance, and remanufacturing.' is duplicated verbatim.","section":"§I, first paragraph"},{"comment":"The bar chart's horizontal axis is labeled 'Modalities Design', which is unclear; the individual bars are not labeled with the task or modality names. Please make the figure self-contained.","section":"Fig. 1"},{"comment":"The action is described as 'restricted to 3D translation and yaw rotation' but is then defined as a 5D command including a gripper command; clarify the dimension count and what the gripper command encodes.","section":"§III.A, action space"},{"comment":"The paper states that all observations are 'normalized to [−1, 1]', but it does not say how TacRGB and TacFF are normalized (per-channel min–max, per-sensor calibration, etc.). This matters for reproducibility of the modality comparison.","section":"§II.B / §III.A"},{"comment":"The combined TacRGB+TacFF configuration is only tested on two real tasks with 20 rollouts each. The conclusion that 'naive concatenation dilutes information' would be stronger with more tasks and more trials, or with a companion simulation experiment where rollouts are plentiful.","section":"§III.D.2, Table V"}],"recommendation":"major_revision","confidential_remarks":"The study is well-motivated and the benchmark design is thoughtful, but the key claim about representation semantics needs an additional matched-resolution control and a reproducible TacFF definition. I do not see an irreparable error; the requested experiments and specification are within the scope of a revision. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a genuinely useful study comparing tactile representations for robotic disassembly under a unified diffusion policy. The main result — that a compact force-field encoding (TacFF) beats both vision-only and high-res tactile RGB (TacRGB), especially in contact-heavy and deformable tasks — is plausible and consistent across simulation and real hardware. The fusion-failure finding (adding TacRGB to TacFF degrades performance) is the most interesting piece; it's a real negative result that many people won't expect.\n\nWhat's good: the task suite is designed with progressively tighter tolerances and deformable elements; simulation uses 1500 rollouts per cell with 3 seeds, which is solid. Real-world trends line up with simulation, which adds credibility. The dim-lighting ablation is a nice touch. The paper is honest about the execution-time tradeoff, even if the interpretation is a bit one-sided.\n\nSoft spots, in proportion: First, the missing matched-resolution control. TacRGB is 240x160; TacFF is a 10x14 grid. The central claim is that structured force semantics beat high-resolution imagery, but you can't rule out that any compact representation would work as well with only 50 demos. A downsampled TacRGB at the same resolution would isolate that. This is not a minor quibble — it's the specific alternative explanation for the tables. Second, TacFF extraction is described qualitatively (\"optical flow and depth reconstruction\") with no equations or calibration. In sim it comes from TacSL, but in real from GelSight R1.5; the mapping is unspecified. That's a reproducibility problem. Third, real-world cells are 20 trials each, no error bars. Some headline gaps (R2, R4) are large enough to be meaningful, but several are 1–2 trial swings. Fourth, no code or models released.\n\nI don't think the paper is broken. The pattern is consistent, and the fusion failure suggests TacRGB is actively adding noise — that's a genuine effect. But the representation comparison is underdetermined as written. I'd want the resolution-matched baseline and a more precise TacFF definition before treating the main claim as established.\n\nFor you: it's a good reading-group paper — the confound debate is instructive. If you work on tactile policies, cite it for the fusion-negative result, not for the strongest claim. Send it to peer review; a good reviewer will request the control, and the paper deserves the revision cycle.","headline":"A well-structured empirical study on tactile representation for disassembly, but the main claim is confounded by missing resolution-matched controls and an underspecified force-field extraction.","tokens_in":11543,"tokens_out":3022,"would_cite":true,"duration_ms":29492,"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":"Compact tactile force fields beat high-res touch images for robotic disassembly","keywords":["tactile sensing","robotic disassembly","force-field representation","diffusion policy","multimodal manipulation","contact-rich tasks","deformable objects","imitation learning"],"falsifier":"A reader could settle the modality claim by training the policy with (1) the raw optical-flow field before it is reduced to the 10x14 grid, and (2) a spatially shuffled or permuted TacFF grid, on the same tasks. If the raw or shuffled versions match TacFF's success rates, the compact structured encoding is not the cause. Separately, if a TacFF grid computed from simulated depth reconstruction fails to produce comparable success when the real sensor is replaced with a different shear/normal calibration, the sim-to-real consistency assumption is falsified.","tokens_in":10610,"feed_emoji":"🤖","tokens_out":4831,"duration_ms":41603,"temperature":0.7,"pith_summary":"This paper tries to establish that, for contact-rich robotic disassembly, the representation of tactile information matters more than its presence. It compares vision alone, vision plus high-resolution tactile images (TacRGB), and vision plus a compact tactile force field (TacFF) under a single imitation-learning policy. Across five simulation and five real-world tasks, TacFF consistently yields the highest success rates, with the biggest gains in tight-tolerance, asymmetric-resistance, and deformable scenarios. The paper also shows that naively combining TacRGB and TacFF degrades performance, indicating that unstructured and structured tactile signals dilute each other. A sympathetic reader would care because disassembly is a contact-dominated, force-sensitive process where visual cues are often ambiguous, and the paper offers a concrete representation choice that improves robustness.","feed_headline":"Compact touch force fields beat high-res tactile images in disassembly","feed_subtitle":"A 10x14 shear-and-normal grid lifts success on tight and deformable tasks; naive fusion backfires.","key_machinery":"TacFF (tactile force field) is the central object: a 10x14 grid in which each cell encodes shear-x, shear-y, and normal force components, computed from optical flow and depth reconstruction of the contact surface. It is the load-bearing modality in the paper's comparison, because the success-rate gains appear only when this structured force summary is included, not when high-resolution tactile images (TacRGB) are added. The policy is a diffusion-model-based visuomotor controller that encodes front and wrist RGB, optional tactile inputs, and end-effector pose with ResNet-18 backbones, concatenates two-step history features, and denoises action chunks (horizon 16, execution of the first 8 step","core_discovery":"The central claim is that a compact 10x14 force-field representation of touch — encoding shear direction and magnitude plus normal force at each cell — is the effective tactile modality for contact-rich disassembly, and that adding it to vision outperforms both vision alone and vision plus high-resolution tactile imagery. The evidence includes simulation results where TacFF raises success in a tight socket from 36.9% to 44.3% (S2) and real-world results where it reaches 70% in the matching R2 task, and 55% on a deformable clip (R5) where vision-only collapses to 0% under dim lighting. The paper's task-by-task pattern shows tactile sensing matters most when geometric constraints tighten, resi","pith_inferences":["A direct test of the representation claim would be to train the same policy with a spatially shuffled TacFF grid; if shuffled grids retain the gains, the spatial structure of the force field is not what carries the information.","The paper leaves the TacFF computation unspecified; if the optical-flow-to-force-field algorithm is published and shown to transfer across different GelSight-like sensors, the result would likely extend to other force-sensitive manipulation skills such as insertion, prying, and cable mating.","The fusion failure suggests a likely research direction: learning a gating or attention mechanism over tactile streams could recover the lost gains, a hypothesis the paper itself flags."],"forward_implications":["For contact-rich disassembly, policies should be built on compact structured force-field tactile input rather than high-resolution tactile imagery.","Tactile sensing acts as a stabilizing signal: in dim lighting, TacFF maintained success on a deformable clip task while vision-only dropped to zero.","Naive concatenation of heterogeneous tactile representations can reduce success rates, so multimodal fusion for force-sensitive tasks needs structured or attention-based integration.","The task-dependent pattern generalizes: gains concentrate in tight-tolerance, asymmetric-resistance, and deformable scenarios, not in loose geometry-dominant ones."],"fun_headline_variants":["Force-field touch beats high-res tactile images in disassembly","Compact force fields outperform rich tactile images for robots","Naive fusion of tactile signals backfires in disassembly","Tactile force fields key for tight and deformable disassembly","Tactile force field rescues deformable disassembly when vision fails"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire comparison rests on the assumption that TacFF is computed faithfully from optical flow and depth reconstruction — an algorithm the paper never specifies — such that the 10x14 grid in simulation and on the real sensor carries the same physical meaning; if the computation is fragile or inconsistent, the modality comparison is confounded by representation quality.","fun_headline_variants_meta":{"raw":{"variants":["Force-field touch beats high-res tactile images in disassembly","Compact force fields outperform rich tactile images for robots","Naive fusion of tactile signals backfires in disassembly","Tactile force fields key for tight and deformable disassembly","Tactile force field rescues deformable disassembly when vision fails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":2871,"prompt_tokens":758,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":2043}},"tokens_in":502,"tokens_out":2113,"duration_ms":12577,"temperature":1.0,"reasoning_tokens":2043,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:29:33.368051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the modality claim by training the policy with (1) the raw optical-flow field before it is reduced to the 10x14 grid, and (2) a spatially shuffled or permuted TacFF grid, on the same tasks. If the raw or shuffled versions match TacFF's success rates, the compact structured encoding is not the cause. Separately, if a TacFF grid computed from simulated depth reconstruction fails to produce comparable success when the real sensor is replaced with a different shear/normal calibration, the sim-to-real consistency assumption is falsified.","supporting_citations":[],"review_version":1}