{"id":"8c6d516d-6d54-47a8-94d9-156b7fac3713","arxiv_id":"2607.16196","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 13.2k-parameter dilated 1D CNN classifies 18 affective touch gestures on a plush toy with 75% random-split and 85% leave-one-subject-out accuracy, with estimated 20 Hz operation on an ESP32.","lead":"This paper trains a compact 1D CNN to recognize affective touch gestures on a soft plush toy, publishing a 25-participant dataset and an open MATLAB pipeline. If the reported accuracy holds, such toys could interpret gentle versus negative touches on-device without cloud processing, which matters for privacy-preserving therapeutic companions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 20 Hz real-time claim rests on a possibly doubled-counted MCU benchmark; the paper's own 4-5x quantization speedup citation is for a different architecture family and does not transfer to ESP32.","rationale":"The reader's weakest_assumption focuses on exactly this: the real-time claim is derived from benchmarks, not measurements. I agree with that. The additional technical detail I contribute is that the quantization speedup reference is double-counted: the paper already applies a 137 MMAC/s Int16 rate from ESP-DSP, then cites CMSIS-NN's 4-5x quantization speedup as an extra factor, but CMSIS-NN numbers are not on top of a benchmark that already assumes integer operations. Moreover, the literature cited [65] is specifically for Arm Cortex-M, not the Xtensa LX6 in the ESP32. This makes the 23 ms estimate more fragile than the reader's general concern about memory/OS overhead. However, the central ML contribution—the dataset and the CNN with LOSO accuracy—is not damaged by this. The paper explicitly flags the estimate as theoretical and defers hardware measurement (Section 4.7), so the correct verdict remains CONDITIONAL rather than REJECT. The paper's own limitation statement is actually a strength: they do not overclaim measured performance, only theoretical compatibility. The verifiable core of the paper is the open dataset and reproducible pipeline, which is a solid contribution. The only serious change needed is to soften the '20 Hz real-time operation' language in the abstract and Section 5.1 until the direct measurement exists.","tokens_in":22933,"tokens_out":1816,"duration_ms":14354,"concrete_test":"Measure end-to-end inference time for the quantized 14d1_39d2_41d4 model on the actual ESP32-WROVER-E (or a faithful clone) using a representative C implementation with CMSIS-NN-like kernels adapted to Xtensa, reporting per-layer and total time. If the measured time exceeds 50 ms per window, the central 20 Hz feasibility claim fails; if it is below 50 ms, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central feasibility claim (Section 4.7) is that quantized inference on the ESP32-WROVER-E takes 23 ms, supporting 20 Hz operation. This number comes from applying published ESP-DSP benchmark MAC/s rates—58 MMAC/s FP and 137 MMAC/s Int16—to the 3.2 MMAC model. The critical issue: the selected architecture 14d1_39d2_41d4 is dilated (d1,d2,d4). Dilated convolution with stride 1 does not reduce arithmetic below the equivalent dense convolution; the receptive field grows but every filter tap is still computed at every output position, so the 3.2 MMAC count may already include volume dilation correctly. The real problem is the benchmark transfer: ESP-DSP MAC/s figures are peak library-kernel numbers for optimized FFT/filter operations, not end-to-end CNN inference on the same MCU. Applying them to a full network, with memory access, activation, batch-norm folding, and the softmax layer, overestimates throughput. Moreover, the paper's own source [65] for a '4-5x quantization speedup' is CMSIS-NN for Arm Cortex-M CPUs, not the Xtensa LX6 in the ESP32; there is no evidence this speedup applies. The 137 MMAC/s Int16 figure versus 58 MMAC/s FP already implies only a 2.4x gain; the additional 4-5x is not additive on top of that—it appears to be double counting. If the true achievable MAC/s on the ESP32 is closer to the 58 MMAC/s FP figure, the 23 ms estimate becomes over-optimistic by roughly 2.5x, and 20 Hz operation (50 ms budget) becomes marginal or infeasible once other tasks (sensor sampling, Bluetooth, control) are accounted for. The paper acknowledges the estimate is not a direct measurement and defers hardware integration (Section 4.7), so the feasibility conclusion is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a framework for affective touch classification in soft plush companions, built around a newly collected FAIR-compliant dataset of 1326 labelled gesture sequences from 25 participants and an open-source MATLAB pipeline. Through systematic exploration of 468 CNN variants, the authors select a compact dilated 1D CNN (14d1_39d2_41d4, 13.2k parameters) that achieves 75% test accuracy on a random window-level split and 85% mean accuracy under leave-one-subject-out cross-validation. The paper further estimates a quantized inference cost of 3.2 MMAC per 2.5 s window and, using published ESP32 DSP benchmarks, derives a 23 ms inference time, concluding that 20 Hz real-time operation on an ESP32-WROVER-E is feasible. A PC-based real-time simulation with the physical toy is used to qualitatively compare the CNN against a prior heuristic algorithm, leading to a proposed hybrid pipeline. The paper is positioned as a model-development and validation study, with physical hardware integration explicitly deferred to future work.","tokens_in":23366,"tokens_out":6816,"duration_ms":54838,"significance":"If the central results hold, this paper demonstrates that lightweight 1D CNNs can bring affective touch classification to microcontroller-class hardware, which is relevant for privacy-preserving and autonomous social assistive devices. The strongest concrete contributions are the public, FAIR-compliant dataset, the complete open-source MATLAB pipeline, and the systematic 468-model architecture search; these are reproducible and reusable assets. The LOSO-CV evaluation is a meaningful step beyond the random-split results, and the qualitative real-time comparison with the heuristic baseline provides useful engineering insight. However, the headline 20 Hz real-time claim rests on unmeasured benchmark transfer with a likely double-counting of quantization speedup, and the generalization claim is weakened by a macro-F1 score near 0.69 despite 85% mean accuracy. The paper is honestly qualified in places, but those qualifications need to be more consistently reflected in the abstract and conclusions.","major_comments":[{"comment":"The 23 ms inference estimate and the resulting 20 Hz real-time claim are not supported by the cited evidence. The 137 MMAC/s Int16 figure already corresponds to quantized int16 operations; applying an additional '4–5x quantization speedup' from [65] on top of it double-counts the benefit, and [65] is for Arm Cortex-M CMSIS-NN kernels, not the ESP32's Xtensa LX6. The ESP-DSP benchmarks are also DSP-library peak rates, not end-to-end CNN inference; memory access, activation, batch-norm folding, softmax, and firmware overhead are omitted. The paper acknowledges these are estimates, but the abstract and §5.1 present 20 Hz operation as a demonstrated capability. The authors should either measure inference on the target MCU or explicitly relabel the 20 Hz claim as an idealized projection pending direct measurement.","section":"§4.7"},{"comment":"The LOSO-CV protocol uses a different, improved training schedule (initial LR 0.003, piecewise decay, early stopping, up to 150 epochs) than the architecture-search protocol (Adam, LR 0.001, no early stopping). The 85% LOSO accuracy therefore does not evaluate the same model configuration that produced the 75% random-split test accuracy, and the architecture itself was selected on the random window-level split with acknowledged subject leakage. The paper should state explicitly that the 85% figure applies to a retrained variant, and should ideally evaluate the final architecture under both protocols with identical training schedules to permit a clean comparison. As written, the distinction is buried in §4.4 while the abstract highlights 85% without this caveat.","section":"§4.4 vs §3.6"},{"comment":"The headline LOSO metric is mean accuracy 0.850, but macro F1 is 0.687 and macro recall 0.717. For 18 imbalanced classes, accuracy overstates per-class performance; the claim in §5.1 that the model 'generalizes well to unseen users' is weakened by per-class F1 scores below 0.5 for Tail hit 2, Tail scratch 2, and Tail tap 2. The paper should report class-balanced metrics prominently, explain why accuracy is the primary headline, and temper the generalization conclusion to reflect the macro-F1 result.","section":"§4.4/§5.1"},{"comment":"The 3.2 MMAC calculation is asserted without the intermediate arithmetic. The number of input channels is ambiguous: §3.1 replaces the accelerometer X/Y/Z with a magnitude (11 channels), while §3.6.1 refers to '13-channel signals'. Include a reproducible table or formula specifying input channels, kernel sizes, dilations, padding, stride, and layer-wise MACs, so that the memory (14–53 KB) and MAC totals can be verified by readers.","section":"§4.7 / §3.6.2"}],"minor_comments":[{"comment":"The text says '13-channel signals' but §3.1 and Table 1 imply 11 channels after replacing accelerometer X/Y/Z with a magnitude. Please reconcile.","section":"§3.6.1"},{"comment":"The sentence 'LOSO-CV on the selected CNN architecture 14d1_39d2_41d4 with 8, 12, and 16 filters in the three convolutional layers' is ambiguous: clarify that the 13.2k-parameter configuration uses 8, 12, and 16 filters, and specify which configuration produced the reported LOSO results.","section":"§4.4"},{"comment":"Reference [64] should specify whether the 58 MMAC/s and 137 MMAC/s figures apply to the ESP32-WROVER-E specifically or to a generic ESP32, and under what clock/optimization settings.","section":"§4.7"},{"comment":"Reaction times were measured manually with ±0.3 s uncertainty. Consider automated timestamp logging from the simulator to make these latency comparisons more reliable and reproducible.","section":"§4.6"},{"comment":"The phrase '50ms steps, exactly as the models were tested in the real-time simulator' would be clearer if the simulator's prediction-rate setting in §3.5 is explicitly reported for the experiments in §4.6.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is refreshingly honest about its limitations, but the abstract and conclusions push the 20 Hz real-time claim further than the evidence warrants. Given the journal's applied scope, I would encourage the authors to either provide a direct inference-time measurement on the ESP32-WROVER-E or reframe the contribution as a model-development pipeline with hardware validation explicitly pending. The dataset and code release are genuinely valuable and likely sufficient for acceptance after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read for you. The paper's substance is the dataset and pipeline, not the CNN novelty. A FAIR 1326-sequence tactile gesture dataset from 25 participants across three age groups, paired with a complete MATLAB pipeline and a 468-model search, is a genuinely useful resource for the plush-robot/affective-touch community. The 13.2k-parameter dilated 1D CNN hits 75% on a random window split and 85% mean LOSO accuracy, which are respectable numbers given the class confusion they document (tail taps vs. scratches, same gesture held vs. on table). The authors also deserve credit for being upfront that the random-split results are upper bounds and that the embedded timings are estimates, and for doing a subject-independent evaluation at all.\n\nThe soft spot is the embedded feasibility arithmetic in Section 4.7. The 23 ms estimate comes from dividing 3.2 MMAC by a 137 MMAC/s Int16 benchmark from the ESP-DSP library, which measures optimized DSP kernels, not end-to-end CNN inference. Then they cite CMSIS-NN's 4-5x quantization speedup as additional evidence, but that is for Arm Cortex-M, not the Xtensa core in the ESP32, and the 137 vs. 58 MMAC/s gap already is the quantization gain — layering another 4-5x on top is double counting. If realistic throughput is closer to 58 MMAC/s, inference is around 55 ms, which already violates the 50 ms per-window budget before accounting for sensor reads, Bluetooth, and control. I wouldn't call the central model result broken, but the headline \"20 Hz real-time on ESP32\" is not supported by the evidence presented. It is a plausible target for future hardware work, not a validated result.\n\nAlso minor: architecture selection happened on the random split before the LOSO evaluation, and LOSO used a different training schedule (adaptive LR, early stopping), so the 85% is not a strictly out-of-sample confirmation of the 75% selected model. And the 75% number has no confidence interval. These are fixable and do not sink the paper.\n\nNet: this deserves a serious referee. It ships reproducible data and code, and the application is genuinely new. But the authors should either soften the embedded claim to \"theoretically plausible pending measurement\" or go measure it before publication. Recommend sending to peer review with a request to fix the Section 4.7 arithmetic and add error bars.","headline":"The public dataset and open pipeline are the real contributions; the 20 Hz embedded feasibility claim rests on a benchmark transfer that likely overstates real ESP32 performance.","tokens_in":23982,"tokens_out":3417,"would_cite":true,"duration_ms":29240,"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 13.2k-parameter 1D CNN can classify affective touch from a plush toy's sensors, reaching 85% accuracy on unseen users and a 20 Hz real-time budget on a low-power microcontroller.","keywords":["affective touch","1D CNN","embedded machine learning","gesture recognition","soft interactive companion","leave-one-subject-out cross-validation","dilated convolution","tactile sensing"],"falsifier":"Measure wall-clock inference time of the fully quantized 14d1_39d2_41d4 model with an 11×250 input window on the target microcontroller. If per-window latency exceeds 50 ms (that is, throughput falls below 20 Hz) under realistic memory and battery conditions, the paper's embedded-feasibility claim is refuted. A secondary falsifier: run a larger subject-independent evaluation with many more unseen participants; if mean LOSO accuracy drops well below 75%, the generalization claim weakens.","tokens_in":22848,"feed_emoji":"🧸","tokens_out":5081,"duration_ms":40501,"temperature":0.7,"pith_summary":"The paper aims to show that affective touch—gentle stroking, scratching, holding, hitting—can be classified from the multichannel capacitive and accelerometer sensors inside a soft plush companion by a compact 1D convolutional network. It reports that a three-layer dilated network with about 13,200 parameters reaches 75% accuracy on a held-out test split and a mean 85% accuracy on entirely unseen users via leave-one-subject-out cross-validation. It further argues, from published microcontroller benchmark data, that quantized inference costs about 3.2 million multiply-accumulates per 2.5-second window, which fits a 20 Hz real-time budget on the target low-power microcontroller. The practical claim is that emotionally meaningful touch interpretation can run entirely on-device, without cloud or external compute, and that a hybrid of a fast threshold filter for energetic negative gestures plus the CNN for subtle ones is the right embedded design.","feed_headline":"13k-parameter CNN reads plush-toy touch at 75–85% accuracy","feed_subtitle":"Subject-independent tests hit 85% accuracy, and quantized inference fits a real-time budget on a low-power microcontroller.","key_machinery":"The load-bearing object is the '14d1_39d2_41d4' architecture: three 1D convolutional layers with 14, 39, and 41 filters and dilation rates 1, 2, and 4. In the symbolic notation, a number is the filter count and 'd' a dilation coefficient. Dilated convolutions widen the temporal receptive field—covering the full 2.5-second, 250-sample window—without adding parameters, which is what lets a 13.2k-parameter model match much larger ones. The second piece of machinery is the sliding-window protocol: 250-sample windows with a 50-sample step, loop-padded rather than zero-padded for short sequences, turning roughly 1,300 raw records into about 5,600 training windows. The third is the hybrid inference","core_discovery":"On its own terms, the paper's central discovery is that a deliberately small dilated 1D CNN—layer filter counts 14, 39, 41 with dilation rates 1, 2, 4—matches or beats far larger networks on this tactile gesture task. Contrary to the usual scaling assumption, models with tens of thousands of parameters performed in the same accuracy band as models with hundreds of thousands, and dilated layers consistently led the rankings. Subject-independent evaluation (LOSO-CV, 25 folds) gave a mean accuracy of 0.850 ± 0.072 and macro-F1 of 0.687 ± 0.114, with confusions concentrated in gestures that have similar signal profiles rather than across emotional categories. The authors interpret this as eviden","pith_inferences":["The 75% random-split figure versus the 85% LOSO figure is puzzle for a strict 'random split is an upper bound' reading; the two protocols are not directly comparable, and window-level leakage can inflate training variance in unexpected ways.","The real-time estimate hinges on benchmark-derived throughput; until the quantized model is actually run on the target microcontroller, a two-to-threefold slowdown from memory access or scheduling would push inference past the 50 ms frame budget.","The 'merge similar classes' suggestion points toward a valence-only (positive/negative/neutral) classifier as a more robust product, which the confusion patterns already support.","A testable extension: train the same 14d1_39d2_41d4 architecture on other soft-sensor systems with different sensor layouts to see whether the dilated 1D-CNN result transfers or is specific to this 11-channel capacitive-plus-accelerometer setup."],"forward_implications":["On-device affective touch recognition becomes plausible for soft therapeutic toys: no cloud link, no laptop, and privacy-preserving local interpretation of a child's touch.","The open dataset and pipeline let other groups train and compare affective-touch models without re-collecting subject data.","The accuracy-complexity analysis suggests that for this sensor and data regime, very small dilated CNNs are sufficient, so deployment cost and power budget can be low.","The hybrid heuristic-plus-CNN pattern gives immediate reaction for high-force negative gestures (0.1–0.2 s) while the CNN resolves subtle gestures the old heuristic missed (e.g., activation in ~6% vs ~100% of subtle-gesture trials).","Confusion clusters are within same-emotion families, implying that merging similar gesture classes could push accuracy higher without collecting new data."],"fun_headline_variants":["13k-param CNN reads plush-touch at 85% subject-independent accuracy","Compact 1D CNN classifies affectionate touch on plush toys in real time","85% LOSO accuracy from 13k-param CNN for plush toy touch","Plush companion touch: tiny CNN nails subtle gestures at 85%","Real-time affective touch recognition on plush toys via compact CNN"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The real-time feasibility rests on published microcontroller benchmark speeds, not on a measured run of the actual quantized model; if the real device runs even two or three times slower than those benchmarks, the 20 Hz target is missed.","fun_headline_variants_meta":{"raw":{"variants":["13k-param CNN reads plush-touch at 85% subject-independent accuracy","Compact 1D CNN classifies affectionate touch on plush toys in real time","85% LOSO accuracy from 13k-param CNN for plush toy touch","Plush companion touch: tiny CNN nails subtle gestures at 85%","Real-time affective touch recognition on plush toys via compact CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3048,"prompt_tokens":844,"completion_tokens":2204,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2112}},"tokens_in":588,"tokens_out":2204,"duration_ms":15057,"temperature":1.0,"reasoning_tokens":2112,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T16:09:03.601716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure wall-clock inference time of the fully quantized 14d1_39d2_41d4 model with an 11×250 input window on the target microcontroller. If per-window latency exceeds 50 ms (that is, throughput falls below 20 Hz) under realistic memory and battery conditions, the paper's embedded-feasibility claim is refuted. A secondary falsifier: run a larger subject-independent evaluation with many more unseen participants; if mean LOSO accuracy drops well below 75%, the generalization claim weakens.","supporting_citations":[],"review_version":1}