{"id":"8c85b2b1-33aa-414a-9a99-04ebde0e8209","arxiv_id":"2511.03078","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"3D Cal repurposes a 3D printer as an automated calibration rig and trains a lightweight CNN, TouchNet, to reconstruct depth maps for DIGIT and GelSight Mini.","lead":"This paper introduces 3D Cal, an open-source library that turns a low-cost 3D printer into an automated probe for collecting labeled training data from vision-based tactile sensors. It reports depth-reconstruction performance on DIGIT and GelSight Mini and gives data-volume guidelines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported unseen-object reconstruction errors are computed after fitting ground-truth depth scale and xy alignment to the predictions; the 156/205 μm figures are Type-2 averages from this favorable protocol, so the headline accuracy claim is not an unbiased estimate.","rationale":"The reader's explicit weakest assumption—that spherical-probe training generalizes to arbitrary unseen geometries—is a legitimate limitation, but it is not the most load-bearing concern. The central quantitative claim is the reported error of 156/205 μm, and that number is produced by an evaluation protocol that fits ground-truth depth scale and xy alignment to the predictions. This inflates apparent accuracy and makes the headline claim unverifiable from the manuscript alone. The abstract's numbers, while traceable to Type-2 averages, are presented as 'average reconstruction errors' without this caveat, which compounds the issue. A single controlled re-evaluation with known indentation depth and no scale fitting would settle whether the claimed accuracy is real. Since this concern is substantive but addressable, and the library/dataset contribution remains plausible, the reader's CONDITIONAL verdict is appropriate; no change is needed.","tokens_in":11864,"tokens_out":4480,"duration_ms":42762,"concrete_test":"Re-evaluate the three test objects by using the 3D printer to indent each object to a single known z-depth (no manual indentation), then compute Type 1 and Type 2 errors against the fixed CAD ground truth with only rigid xy registration and no depth-scale optimization. If the average Type 2 error remains below roughly 200 μm for both sensors, the claim survives; if it rises substantially, the reported 156/205 μm figures are artifacts of the fitting procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—'average reconstruction errors of 156 μm and 205 μm on unseen objects'—depends on the depth-error numbers in Section IV/Table I being meaningful. They are not, as reported. Section IV states that ground-truth CAD depth maps were aligned via 2D cross-correlation and that 'the indentation depths of the ground-truth CAD models were adjusted to minimize the mean squared error between the predicted and ground-truth depth maps.' This effectively fits a global depth-scale parameter to each prediction. Any systematic under- or over-estimation of indentation depth is removed before errors are computed, so the reported errors are lower bounds on true depth-reconstruction error, not unbiased estimates. The abstract's 156 μm and 205 μm are not the Overall Errors in Table I but the across-object averages of Type 2 errors; Type 2 excludes all zero-depth regions and is computed after this per-object depth fit. The claim 'comparable to state-of-the-art methods' is also unsupported: no baseline method is evaluated anywhere in the paper. The generalization limitation (spherical-probe training, only three test objects, acknowledged shadowing failures on the pawn) is real but secondary; the metric problem undermines even the three-object results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 3D Cal, an open-source software library that repurposes a low-cost FDM 3D printer as an automated probing device for collecting labeled tactile-sensor data, and TouchNet, a fully convolutional network that maps RGB sensor images plus coordinate embeddings to surface gradients, which are integrated into depth maps. The authors calibrate DIGIT and GelSight Mini sensors, perform a spatial-sampling ablation to recommend a minimum number of probe locations, and report depth-reconstruction errors on three unseen 3D-printed objects. The headline results are average Type 2 errors of 156 μm for DIGIT and 205 μm for GelSight Mini, which the paper states are comparable to state-of-the-art methods.","tokens_in":12125,"tokens_out":3097,"duration_ms":97921,"significance":"If the reported accuracy is reliable, the paper would make a useful contribution to tactile-sensor calibration: it lowers the hardware barrier by using an inexpensive 3D printer, releases code, pre-trained weights, and a large dataset of over 70,000 probe images, and provides practical guidance on the amount of calibration data needed. The spatial ablation and the open-source infrastructure are valuable regardless of the exact error numbers. However, the central accuracy claim is currently not supported by the evaluation protocol, and no baseline method is benchmarked, so the paper's main quantitative contribution needs substantial revision.","major_comments":[{"comment":"The reported depth errors are computed after fitting the ground-truth indentation depth to minimize the MSE against the predicted depth maps: 'the indentation depths of the ground-truth CAD models were adjusted to minimize the mean squared error between the predicted and ground-truth depth maps.' This is a per-object global depth-scale fit. Any systematic under- or over-estimation of indentation depth is therefore removed before errors are reported. The abstract's 156 μm and 205 μm figures are the averages of the Type 2 errors in Table I, not the Overall Errors, and thus inherit this favorable protocol. These numbers are not unbiased estimates of absolute depth-reconstruction accuracy; they are at best lower bounds. Please report errors without the per-object depth fit, or fix/measure the indentation depth independently, and report the fitted scale factors so readers can assess the bias.","section":"Section IV, Table I"},{"comment":"The claim of performance 'comparable to state-of-the-art methods' is unsupported because no state-of-the-art depth-reconstruction method is implemented, run, or compared anywhere in the paper. There is no baseline such as GelSight's classical photometric-stereo pipeline, other supervised tactile depth estimators, or a re-implementation of a published DIGIT/GelSight Mini model. Please include at least one baseline on the same test objects under the same (or clearly stated) evaluation protocol, or remove the comparative claim.","section":"Section IV and Abstract"},{"comment":"The generalization claim is based on training with only spherical probe indentations and evaluating on only three 3D-printed objects (hemispheres, pill, pawn), all with relatively smooth geometry. The authors themselves note that both sensors struggle to reconstruct the pawn's neck due to shadowing, showing the spherical-probe training already fails for a non-trivial concavity. The Discussion's statement that the model generates 'high-quality depth reconstructions on unseen and complex objects' overstates the evidence. A broader test set with varied curvature, sharp features, and concavities, or a more qualified claim, is needed.","section":"Section III–IV"},{"comment":"The protocol also performs a 2D cross-correlation to align the predicted and ground-truth depth maps in the xy-plane. This is reasonable for manual indentation, but combined with the depth-scale fit it removes both translational and depth-scale errors before accuracy is reported. This means the reported Overall, Type 1, and Type 2 errors all characterize residual shape error after a three-degree-of-freedom per-object alignment, not the raw output of the system. The paper should clearly state this and, ideally, provide the errors before alignment and scaling for at least one set of results.","section":"Section IV, Table I"}],"minor_comments":[{"comment":"The definitions of Type 1 and Type 2 errors are given in the text but would benefit from explicit equations. It would also help to state that the abstract's 156/205 μm figures are the across-object averages of the Type 2 errors, not the Overall Errors in Table I.","section":"Section IV"},{"comment":"The spatial ablation evaluates models on held-out probe coordinates from the same spherical-probe grid, so it measures interpolation within the calibration manifold, not generalization to unseen object geometries. This should be stated explicitly when presenting the '20% of coordinates' recommendation.","section":"Section III"},{"comment":"Some figures lack axis units or have clipped colorbars; for example, Figures 2 and 3 would benefit from clearer labels on the MSE values. The manuscript also repeatedly says 'high-resolution' depth maps without quantifying resolution; reporting the pixel pitch or effective spatial resolution of the output maps would be useful.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially useful open-source systems contribution, and the dataset/software release is a genuine strength. The main obstacle is the evaluation protocol: the headline accuracy numbers are computed after per-object depth scaling and xy alignment, so they are not honest estimates of reconstruction accuracy. This is fixable within the paper's scope by re-running the evaluation without the depth-scale fit and by adding baselines. I would not reject the paper, but the quantitative claims need to be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the tooling: 3D Cal turns a consumer 3D printer into a two-axis probing gantry, ships an open-source library, releases over 70,000 probe images plus pretrained models, and gives practical data-volume guidelines for two widely used sensors. That is genuinely useful and lowers a real barrier in tactile sensing. The ablation study is also solid — the spatial sampling analysis with statistical tests and the recommendation around 240 coordinates along a 0.5 mm grid is actionable and well presented. TouchNet itself is modest (coordinate embeddings, gradient output, Poisson integration), but it works well enough and runs at 30 fps on laptop hardware.\n\nThe soft spots are mostly in Section IV and the abstract. The evaluation protocol explicitly fits the ground-truth CAD depth maps to the predictions: xy alignment via cross-correlation and then per-object depth scaling that minimizes MSE between predicted and ground-truth depths. That removes global depth errors and translation errors before any accuracy number is computed. The 156 µm and 205 µm quoted in the abstract are the averages of the Type 2 errors in Table I, not an unbiased estimate of reconstruction accuracy; they are lower bounds. The paper also claims performance 'comparable to state-of-the-art methods' but evaluates no baseline method anywhere. The generalization test is narrow — three smooth 3D-printed objects, all indented manually, with the authors themselves noting shadowing failures on the pawn's neck. Spherical-probe training transferring to arbitrary geometry is a reasonable assumption to test, but three objects is a thin test.\n\nNone of this invalidates the library or the dataset. The engineering description is clear and reproducible, and the release itself is a contribution. But the headline accuracy claims need to be reworked: report errors without the per-object depth fit, include baselines, or at minimum state the protocol honestly in the abstract. The authors are honest about the shadowing limitation and the manual indentation, which makes me think this is fixable rather than a fundamental problem.\n\nThis paper deserves a serious referee. The evaluation issues are addressable, and the library will likely be used by the tactile sensing community regardless. I would want to see a revised version with unbiased metrics and at least one existing method for comparison, but I would not desk-reject it.","headline":"Useful open-source library and dataset for calibrating DIGIT/GelSight Mini with a cheap 3D printer, but the headline depth-error numbers are lower bounds computed after fitting ground-truth scale and alignment to predictions, and no baselines support the 'comparable to state-of-the-art' claim.","tokens_in":12609,"tokens_out":1847,"would_cite":true,"duration_ms":19099,"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 low-cost 3D printer can be repurposed into an automated rig that calibrates vision-based tactile sensors to micron-level depth accuracy.","keywords":["tactile sensing","depth reconstruction","sensor calibration","vision-based tactile sensors","TouchNet","3D printer probing","DIGIT","GelSight Mini"],"falsifier":"Press a concave or undercut object of known geometry (for example a threaded screw or a ring) into DIGIT and GelSight Mini, compute the TouchNet depth map, and compare to ground truth. If contact-region errors rise far above the reported 156–205 µm range, the claim that spherical probes alone suffice for arbitrary geometry would be falsified.","tokens_in":11743,"feed_emoji":"🖐️","tokens_out":7837,"duration_ms":77002,"temperature":0.7,"pith_summary":"The paper introduces 3D Cal, an open-source library that turns a low-cost 3D printer into an automated probing device for calibrating vision-based tactile sensors. The printer fabricates a socket that holds the sensor in a known location, then a spherical probe tip indents the gel in a regular 0.5 mm grid while images are captured and labeled automatically. The authors claim a lightweight convolutional network, TouchNet, trained only on these spherical-probe images, reconstructs depth maps of unseen objects with whole-surface average errors of roughly 16–52 µm and contact-region errors averaging about 156 µm (DIGIT) and 205 µm (GelSight Mini)—comparable to existing methods. They also report that reconstruction loss stabilizes when data from roughly 240 probe locations (20% of the grid) is used, providing a practical calibration recipe. If the approach holds, labs already owning a hobbyist 3D printer can skip expensive robot-arm or CNC calibration setups.","feed_headline":"A 3D printer plus neural net maps touch to micron-depth accuracy","feed_subtitle":"Open-source 3D Cal automates calibration for DIGIT and GelSight Mini, cutting contact-region depth error to about 156–205 µm.","key_machinery":"The central mechanism is the full 3D Cal pipeline: a G-code-driven FDM printer converted into a two-axis automatic probe; a 2 mm spherical probe tip that creates precisely known indentation geometry; and a dataset of images labeled with spatial coordinates. The learning model carrying the argument is TouchNet, a 9-layer fully convolutional network that takes an RGB image concatenated with a two-channel (x, y) coordinate embedding and outputs a surface gradient map (Gx, Gy); a fast Poisson integration reconstructs absolute depth. The coordinate embedding is what lets the model associate pixel responses with physical locations, a necessary step because illumination and response vary across the","core_discovery":"3D Cal's central claim is that the labor-intensive calibration of vision-based tactile sensors can be fully automated with hardware most robotics labs already own: an FDM 3D printer. The printer first prints a rigid base that fixes the sensor in its coordinate frame; a 2 mm spherical probe mounted on the printhead indents the gel in a 0.5 mm × 0.5 mm grid, capturing 30 images at each of roughly 1,200 locations per sensor. These coordinate-labeled images train TouchNet, a fully convolutional network that predicts a surface gradient map which a fast Poisson integrator converts to a depth map. Trained solely on spherical indentations, the model reconstructs three unseen 3D-printed objects—hemis","pith_inferences":["Extension: the same probe dataset could be reused to learn force, shear, or stress labels if a force-sensing probe replaces the rigid sphere, since the images already encode dense deformation information.","Extension: the strongest test of the spherical-probe assumption would be a concave or undercut object such as a threaded screw or ring; the pawn's shadowed neck already flags a failure region, so such tests could bound the generalization claim.","Extension: the 20% (≈240-point) guideline is calibrated to these two sensors; other gel sizes or optical setups may need grid density scaled by sensing area rather than copying the same fraction.","Extension: if the pre-trained weights transfer across sensors after fine-tuning, the pipeline effectively becomes a foundation-model approach for tactile depth estimation—a direction the paper names as future work but does not demonstrate."],"forward_implications":["Researchers with a hobbyist 3D printer can generate thousands of coordinate-labeled tactile images in about two hours per sensor, removing the need for industrial robot arms, CNC machines, or motion-capture rigs.","The ablation study indicates that roughly 240 probe locations along a 0.5 mm grid are sufficient for stable reconstruction, so a new sensor can be calibrated in a single automated session.","Because TouchNet inference runs in under 30 ms, the pipeline can produce depth maps at 30 fps on modest laptop hardware, making closed-loop tactile manipulation practical.","Releasing the datasets and pre-trained DIGIT/GelSight Mini weights lets other labs fine-tune their own models instead of starting from scratch, accelerating adoption of depth-sensing touch.","If the generalization beyond the three test objects holds, vision-based tactile sensors can deliver physical geometry estimates without per-object retraining, supporting grasp verification, surface inspection, and shape reconstruction tasks."],"fun_headline_variants":["Open-source 3D Cal turns a 3D printer into a tactile sensor calibrator","Automated tactile calibration: 3D printer + CNN hits micron depth accuracy","3D Cal: turn any 3D printer into an automated touch-sensor calibrator","Print your own tactile sensor calibration rig: 3D Cal automates it","Micron-accurate robot touch: 3D Cal works with a cheap 3D printer"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a network trained only on spherical probe indentations on a 0.5 mm grid generalizes to arbitrary unseen object geometries; the paper tests this on just three 3D-printed objects and already notes that shadowing breaks reconstruction at the pawn's neck.","fun_headline_variants_meta":{"raw":{"variants":["Open-source 3D Cal turns a 3D printer into a tactile sensor calibrator","Automated tactile calibration: 3D printer + CNN hits micron depth accuracy","3D Cal: turn any 3D printer into an automated touch-sensor calibrator","Print your own tactile sensor calibration rig: 3D Cal automates it","Micron-accurate robot touch: 3D Cal works with a cheap 3D printer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000797,"raw_usage":{"total_tokens":3364,"prompt_tokens":782,"completion_tokens":2582,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2470}},"tokens_in":526,"tokens_out":2582,"duration_ms":18187,"temperature":1.0,"reasoning_tokens":2470,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:59:04.664439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Press a concave or undercut object of known geometry (for example a threaded screw or a ring) into DIGIT and GelSight Mini, compute the TouchNet depth map, and compare to ground truth. If contact-region errors rise far above the reported 156–205 µm range, the claim that spherical probes alone suffice for arbitrary geometry would be falsified.","supporting_citations":[],"review_version":1}