{"id":"94679680-5bd2-454c-b3e8-6f2be14e9fa0","arxiv_id":"2608.04043","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A pressure-array model matches a supervised CNN on 27-object tactile recognition using only frozen text embeddings and a small-data recipe.","lead":"Tactus lets a robot recognize objects from cheap pressure-sensor touch data alone, by comparing the pressure pattern with text descriptions. It matches a much heavier supervised model on a standard 27-object test, while staying open to new wording of queries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline comparison to STAG CNN rests on an unverified tuple-level protocol equivalence; recording-level anchor 0.722 is below the 0.76 baseline, so the 'matching/exceeding' claim may not survive an exact-harness rerun.","rationale":"The reader's weakest-assumption identification matches my own: Section V's protocol mismatch is the point where the headline result is least secure. The paper is unusually transparent—it reports the caveat and gives the recording-level anchor—but transparency does not make the comparison valid. The concrete test is feasible because the paper says the protocol 'mirrors' STAG's, so the original code or a faithful reimplementation can be run on Tactus's exact tiling. If the test shows equivalence, the central claim stands; if not, the result is still a useful open-vocabulary pressure-to-language model, but the abstract's baseline comparison should be weakened. I therefore keep the reader's CONDITIONAL verdict: the issue is specific and addressable, not a fundamental flaw.","tokens_in":8862,"tokens_out":5111,"duration_ms":55574,"concrete_test":"Run the STAG reference classifier under Tactus's exact fixed-seed cluster-tuple test harness, using the same tiling and tuple-level cosine ranking, and compare its top-1 to 0.76; equivalently, run Tactus's frozen embeddings through STAG's original evaluation script. If the STAG CNN does not yield approximately 0.76 on Tactus's tuple protocol, or if Tactus's tuple-level top-1 does not beat the baseline under a shared protocol, the abstract's 'matching/exceeding' claim should be downgraded to a protocol-relative observation and the recording-level 0.722 should be reported as the honest anchor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—0.771±0.062 top-1 matches or exceeds STAG's 0.76 supervised CNN—requires that Tactus's tuple-level evaluation be commensurable with the STAG baseline's evaluation. Section V discloses that the harness 'mirrors STAG's cluster-sampling test protocol but is not the authors' byte-identical harness.' This matters because Section IV itself warns that cluster sampling changes the test population, so tuple-level accuracies are comparable only within one protocol; and the paper's protocol-independent recording-level anchor is 0.722, below the baseline. If STAG's 0.76 was computed at frame/recording level or with a different cluster draw, Tactus's tuple-level 0.771 is not comparable, and the primary comparison becomes 'open-vocabulary model below the supervised baseline on the protocol-independent metric.' Also, 0.771±0.062 versus a point value of 0.76 is within one standard error; even under an exact protocol, 'exceeding' is not statistically supported, only 'matching.' The paper's phrasing is careful, but the load-bearing premise—protocol equivalence—is asserted, not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Tactus, an encoder that maps short windows of low-cost resistive pressure-array frames (from the STAG 548-taxel glove) into the frozen text-embedding space of a multimodal model, enabling open-vocabulary object recognition by cosine ranking against natural-language class phrases. On the STAG 27-object held-out test split, the paper reports a tuple-level top-1 of 0.771±0.062 over four runs (top-3 0.935), which it describes as matching, and at best exceeding, the dataset's supervised closed-set CNN baseline of 0.76. The paper also presents a measured recipe ladder (calibration affine, cluster-sampled grasp windows, same-sensor masked-autoencoder pretraining), a detailed error analysis showing that residual confusions are contact-structured rather than language-structured, quantified negative results (cross-sensor pooling, vision co-training, attention at small scale), and a forensic account of an input-normalization defect. Weights, code, and a session-memory integration layer are released.","tokens_in":9054,"tokens_out":5072,"duration_ms":56974,"significance":"If the core comparison is valid, Tactus is a notable demonstration that a cheap, low-dimensional pressure array can be connected to a frozen language space with accuracy comparable to a supervised closed-set CNN, without a trained classifier head. The paper is exemplary in transparency: it reports run-to-run variance, reproduces its own release evaluation to four decimal places, and documents negative results and a subtle data-path defect with enough detail to be actionable. The finding that the sensor's calibration affine contributed more accuracy than any architecture change is a practically valuable lesson for the tactile-sensing community. However, the headline comparison rests on a tuple-level evaluation protocol that the paper itself describes as 'mirroring' but not byte-identical to the STAG baseline harness, and the protocol-independent recording-level anchor (0.722 mean, 0.741 released) is below the 0.76 baseline. Whether the central 'matching/exceeding' claim survives an exact-harness rerun is therefore unresolved, which tempers the significance of the result as currently stated.","major_comments":[{"comment":"The central claim that Tactus matches or exceeds the STAG supervised CNN (0.771±0.062 vs. 0.76) is not supported by the evidence as presented because the two numbers come from different evaluation protocols. The paper states that the evaluation 'mirrors STAG's cluster-sampling test protocol but is not the authors' byte-identical harness,' and Section IV itself warns that cluster sampling changes the test population, making tuple-level accuracies comparable only within one protocol. The recording-level anchor, which the paper calls protocol-independent, is 0.722 (mean) and 0.741 (released checkpoint), both below 0.76. To make the comparison valid, please run the original STAG CNN under the exact same tuple-level protocol used for Tactus, or at minimum report the STAG baseline under a clearly identical protocol, and then report both tuple-level and recording-level numbers for both models. Without this, the headline statement is an unsupported comparison across protocols.","section":"Section V, Table I"},{"comment":"Even under an identical protocol, the statistical support for 'exceeding' is absent. The difference between 0.771 and 0.76 is 0.011, well within the reported standard deviation of 0.062 over four runs. The paper's wording 'matching, and at best exceeding' is careful, but the empirical basis only supports a claim of approximate parity, not superiority. Please provide a confidence interval or a paired comparison if possible, and phrase the contribution as 'comparable to' rather than 'exceeding' unless new evidence is supplied.","section":"Section V, Table I"},{"comment":"The recording-level score is described as the 'protocol-independent anchor,' but its ability to arbitrate the comparison with the STAG baseline depends on knowing what metric the STAG CNN's 0.76 was computed with. The paper does not state whether STAG's 0.76 is a frame-level, tuple-level, or recording-level number, nor does it report a recording-level number for the STAG baseline. Without this information, the fact that Tactus's recording-level 0.722 is below 0.76 cannot be interpreted as a valid baseline comparison either. Please clarify the metric used in the original STAG evaluation and, if possible, recompute that baseline under the recording-level protocol.","section":"Section VI, Table I"},{"comment":"The diagnostic comparing a learnable linear classifier head (0.33) against the frozen-text-target head (0.396) is performed on the defective input pipeline, because both numbers appear in the regime above the rule in Table II, where 'absolute values are depressed by defective input scaling.' This means the conclusion that the fixed text targets are not costing accuracy is confounded by the normalization defect. The Spearman-rho analysis in Section VI is conducted on the released, correctly normalized pipeline and is supportive, but the classifier-head comparison should be rerun after the calibration affine is restored before this negative result is reported as conclusive.","section":"Section VII, 'The text targets are not the bottleneck'"}],"minor_comments":[{"comment":"The title contains a typo: 'Open-V ocabulary' should be 'Open-Vocabulary'.","section":"Title page"},{"comment":"The claim that removing the class centroid spreads the targets to an off-diagonal mean of -0.038, described as 'the maximum for 27 unit vectors,' is mathematically the minimal possible mean (the regular-simplex arrangement gives -1/26), not a maximum in the usual sense. Please rephrase to avoid confusion.","section":"Section IV, Text targets"},{"comment":"The limitation that 'open-vocabulary here means text queries over the evaluation categories, not validated open-set generalization to novel object classes' directly qualifies the title's 'open-vocabulary' claim; consider reflecting this scope in the title or abstract to avoid overclaiming.","section":"Section X, Limitations"},{"comment":"The recording-level column for the STAG supervised CNN is left blank. Filling this cell (or explicitly stating that the baseline does not have a recording-level number) would help readers interpret the protocol-independent anchor.","section":"Table I"},{"comment":"The sentence 'The 13.5M-trunk model overfit catastrophically when trained past its optimum without augmentation' would benefit from a quantitative detail (e.g., the training/validation accuracy gap or the number of steps past optimum) to support the claim.","section":"Section VII, Capacity without regularization"},{"comment":"The figure caption states 'two diverse frames recover 89% of the eight-frame accuracy,' but the underlying numbers in the text (0.762 vs. 0.858) give 88.8%; please ensure the percentage is rounded consistently.","section":"Section VI, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and methodologically transparent, and the central engineering recipe (calibration-first, same-sensor MAE, cluster-sampled windows) is presented with a level of detail that is rare in this area. However, the primary quantitative claim—that Tactus matches or exceeds the STAG supervised baseline—rests on an unverified protocol equivalence. The protocol-independent recording-level numbers are below baseline, so the headline claim may not survive a faithful rerun of the STAG harness. This is fixable by running the baseline under identical protocols and softening the claim if needed. I would also note that the text side is the author's own fusion-embedding space (arXiv:2607.18666), which is not yet widely known; the evaluation's reproducibility therefore depends on that space being stable and publicly accessible, which is worth verifying during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTactus is worth your time even if you don't work in tactile sensing: it's the first open model that takes low-cost pressure-array frames and answers text queries in a frozen multimodal embedding space, and it ships code, weights, and a reproducible replay (recording-level 0.7407 matches the release note). On STAG's 27-object benchmark, the mean of four runs is 0.771±0.062 top-1—nominally above the dataset's supervised closed-set CNN at 0.76, but within one standard error. The authors carefully phrase the claim as 'matching, and at best exceeding.'\n\nThe paper's real strengths: (1) the recipe ladder in Table II, which shows the sensor's own calibration affine was worth more than every architecture change combined; (2) a quantified error analysis showing confusions are contact-structured, not linguistic (Spearman ρ≤0.05 over 702 pairs); (3) a forensic account of a normalization bug that silently discarded 97% of dynamic range—a cautionary tale for anyone working with resistive arrays; (4) honest negative results (cross-sensor pooling gives nothing, vision co-training hurts). The small-data recipe (187 recordings, MAE pretraining) is a genuinely useful data point.\n\nThe soft spot is the headline comparison. Section V states the evaluation 'mirrors STAG's cluster-sampling test protocol but is not the authors' byte-identical harness,' and the protocol-independent recording-level anchor is 0.722—below the 0.76 baseline. So the 'matching/exceeding' claim rests on an untested equivalence between their tuple-level protocol and STAG's. If the tuple draws differ in any systematic way, the comparison is not apples-to-apples. The paper is upfront about this, which I respect, but it remains the one thing I'd want fixed: rerun the baseline under their protocol, or report their model under STAG's exact harness. The missing error bars in the pre-correction rows of Table II are minor since those rows are trajectory, not final results.\n\nAlso note the explicit limitation: 'open-vocabulary' here means queries over the evaluation categories, not validated open-set generalization. That's an honest scope statement.\n\nVerdict: the paper is novel, reproducible, and unusually candid. It deserves a serious referee. I'd send it to review and push the authors to align the evaluation protocol with the baseline's. That single change would put the central claim on solid ground.","headline":"Tactus is a genuinely useful, honest open pressure-array-to-language model; its headline match with the supervised CNN is within noise and rests on an unverified protocol equivalence.","tokens_in":9619,"tokens_out":3489,"would_cite":true,"duration_ms":35308,"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":"Tactus maps low-cost pressure-array touch data into a frozen multimodal text space, reaching 0.771 top-1 on the STAG benchmark's 27 held-out objects—matching, and at best exceeding, the dataset's supervised closed-set CNN at 0.76—with no…","keywords":["tactile sensing","pressure arrays","open-vocabulary recognition","zero-shot classification","masked autoencoder","sensor calibration","multimodal embedding","small-data learning"],"falsifier":"Run the released checkpoint on STAG's held-out test recordings using the dataset's original byte-identical test-harness code and compute tuple-level top-1; a mean at or below the paper's protocol-independent recording-level 0.722 would indicate the claimed match to the 0.76 baseline was an artifact of the evaluation protocol.","tokens_in":8615,"feed_emoji":"✋","tokens_out":9692,"duration_ms":81804,"temperature":0.7,"pith_summary":"The paper claims that the cheapest tactile sensor class—resistive pressure arrays—can be made open-vocabulary: a small trained encoder projects pressure windows into a frozen multimodal embedding space, and recognition is just cosine ranking against natural-language grasp phrases. On the STAG benchmark's 27 held-out objects, four runs average $0.771$ top-1 (top-3 $0.935$), matching and at best exceeding the dataset's supervised closed-set CNN at $0.76$. The recipe relies on the sensor's own calibration affine, cluster-sampled grasp windows drawn from 187 training recordings, and masked-autoencoder pretraining on 144k unlabeled same-sensor frames; the calibration fix alone recovered more accuracy than every architecture change combined. The paper also documents failures with equal precision: cross-sensor pretraining pooling gave no gain, vision co-training degraded touch, and a mis-normalized input pipeline had silently discarded 97% of the sensor's dynamic range while still producing plausible results. The significance, if the result holds, is that touch becomes a first-class, language-searchable channel for low-cost robot memory without a supervised label set.","feed_headline":"Pressure-array touch model matches supervised CNN on 27 objects","feed_subtitle":"Open-vocabulary model scores 0.771 top-1 on held-out grasps with no trained classifier head.","key_machinery":"The central machinery is the frozen fusion-embedding space: a shared multimodal embedding space (hosting text, image, video, audio, and motion) that Tactus enters by training a 16.2M-parameter head—a ResNet-18-width trunk, a learned $1 \\times 1$ temporal fusion over eight frame-feature maps, and a $512 \\to 1024 \\to 2048$ projector—never updating the language side. Recognition becomes cosine ranking of the projected pressure window against the 27 class phrases, so touch plugs into the same memory layer as the other senses. Two supporting mechanisms carry the argument: the sensor's calibration affine $\\mathrm{clip}((\\mathrm{raw}-500)/150,0,1)$, which restored the dynamic range and outweighed every architecture change, and same-sensor masked-autoencoder pretraining with per-patch normalized targets, whose mean-predicting collapse is pinned at reconstruction loss 1.0 as a visible diagnostic.","core_discovery":"The core discovery is that open-vocabulary recognition from low-cost pressure arrays is not only possible but competitive with a supervised closed-set classifier on the same data: Tactus reaches $0.771 \\pm 0.062$ top-1 (top-3 $0.935$) across four runs against the STAG CNN's $0.76$, with no trained decision boundary for the label set. The model projects a window of $32 \\times 32$ pressure frames, fused by a learned $1 \\times 1$ convolution over $K=8$ cluster-diverse frames, into the frozen text space of a multimodal embedding model; prediction is cosine ranking against class phrases with centered targets. The dominant recipe elements are data-path corrections—calibrated normalization and cluster sampling—plus same-sensor masked-autoencoder pretraining, with architecture playing a minor role. Analyses of the released checkpoint show the residual errors concentrate in contact-ambiguous pairs (kiwano→chain, screw driver→mug) and are uncorrelated with text-target geometry (Spearman $\\rho \\le 0.05$ over 702 pairs); query phrasing, including bare names, moves accuracy by about one point.","pith_inferences":["The paper leaves implicit a general rule: for taxel arrays, pretraining should be treated as sensor-family-specific, and any pooled pretraining should be validated against a same-sensor-only control, a testable extension to any resistive-array dataset.","The silent data-defect story generalizes beyond this sensor: any resistive array with a resting pedestal is vulnerable to corpus-max normalization that flattens dynamic range, so adding a dynamic-range regression gate and a same-sensor control is a transferable diagnostic for other low-cost-sensor efforts.","A natural next experiment, not run in the paper, is whether the fusion-embedding route scales to a larger vocabulary or to truly novel object classes; the paper only validates text queries over the evaluation categories.","The closeness of the match to the supervised baseline ($0.771$ vs $0.76$) may be sensitive to evaluation protocol; an independent reimplementation of the original STAG harness would settle whether the parity is genuine or a protocol artifact."],"forward_implications":["If the central claim holds, low-cost pressure arrays—not just optical tactile sensors—can be integrated into language-searchable robot memory, letting robots answer 'what am I holding?' without a camera or a supervised classifier.","Two diverse frames already recover 89% of the eight-frame accuracy, so deployment latency can be very short: two contacts within a fraction of a second nearly saturate recognition.","Tuple voting over four independently sampled grasp windows adds about four top-1 points with no retraining, making test-time sampling a free accuracy lever for open-vocabulary touch.","Because errors are contact-structured and query-insensitive, further gains should come from better tactile encoders and data hygiene rather than from the text side.","Cross-sensor pretraining pooling gave no gain; the supported recipe is same-sensor pretraining, which is a direct constraint on future multi-sensor tactile foundation-model efforts."],"supporting_citations":[{"why":"Supplies the 27-object grasp dataset, the supervised CNN baseline at 0.76, and the cluster-sampling protocol that training and evaluation reuse.","marker":"[1]"},{"why":"Defines the frozen multimodal embedding space into which Tactus projects pressure windows and from which text queries are drawn.","marker":"[2]"},{"why":"Provides the masked-autoencoder pretraining method with per-patch normalized targets used on the 144k same-sensor frames.","marker":"[3]"},{"why":"Supplies the cosine-ranking zero-shot scoring paradigm that open-vocabulary recognition inherits.","marker":"[14]"}],"fun_headline_variants":["Open-vocab touch model matches supervised CNN on pressure arrays","Tactus: open-vocabulary recognition from low-cost pressure data","Zero-shot tactile recognition with 0.771 top-1 on 27 objects","Calibration and masking beat architecture in tactile recognition","Pressure arrays alone answer text queries: 0.771 top-1"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison rests on the assumption that the paper's tuple-level evaluation protocol is statistically equivalent to the original STAG harness that produced the 0.76 CNN baseline; the paper itself says it is not byte-identical, and the protocol-independent recording-level anchor is 0.722, below the baseline.","fun_headline_variants_meta":{"raw":{"variants":["Open-vocab touch model matches supervised CNN on pressure arrays","Tactus: open-vocabulary recognition from low-cost pressure data","Zero-shot tactile recognition with 0.771 top-1 on 27 objects","Calibration and masking beat architecture in tactile recognition","Pressure arrays alone answer text queries: 0.771 top-1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1690,"prompt_tokens":1046,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":554}},"tokens_in":662,"tokens_out":644,"duration_ms":6479,"temperature":1.0,"reasoning_tokens":554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:45:13.205242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released checkpoint on STAG's held-out test recordings using the dataset's original byte-identical test-harness code and compute tuple-level top-1; a mean at or below the paper's protocol-independent recording-level 0.722 would indicate the claimed match to the 0.76 baseline was an artifact of the evaluation protocol.","supporting_citations":[{"cited_title":"Learning the signatures of the human grasp using a scalable tactile glove,","cited_arxiv_id":null,"evidence_quote":"Supplies the 27-object grasp dataset, the supervised CNN baseline at 0.76, and the cluster-sampling protocol that training and evaluation reuse."},{"cited_title":"Masked autoencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"Provides the masked-autoencoder pretraining method with per-patch normalized targets used on the 144k same-sensor frames."}],"review_version":1}