{"id":"00edd70e-a832-4ab7-b4ef-146b3e72b1c9","arxiv_id":"2607.18210","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A haptic-calibration and visual-feedback pipeline lifts grasping success to 96.7% inside the calibrated region and 72.7% across the full workspace on the NICO humanoid.","lead":"Extends a low-cost haptic-calibration method from reaching to tabletop grasping on the NICO humanoid, adding YOLO detection, stereo vision from built-in fisheye cameras, and a visual hand-alignment loop. If the reported numbers hold, cheap robot platforms can improve sim-to-real grasping without RGB-D cameras or motion capture.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M4's full-workspace success may reflect pivot/z-yaw corrections tuned to the test grid rather than the visual-feedback loop.","rationale":"The reader's conditional verdict is appropriate. My focus is slightly narrower: the z/yaw corrections affect all models and therefore do not fully explain M3's inside-area improvement, but the pivot correction and the post-hoc 'SV OK' filtering specifically threaten M4's headline full-workspace result. The paper reports enough detail to raise the concern but not enough to rule it out. A held-out refitting test would settle whether the 72.7% is a genuine property of the visual-feedback loop or an artifact of corrections tuned to the evaluation grid. Until that test is done, conditional acceptance is the correct outcome.","tokens_in":9038,"tokens_out":5342,"duration_ms":50752,"concrete_test":"Re-run M4 and M3 after refitting all task-specific corrections on a disjoint subset: split the 22 evaluation positions into fit and holdout sets, estimate the pivot correction and RBF z/yaw maps using only the fit set, then evaluate on the holdout set. If M4's holdout success rate falls below M3's or below the reported 57.6% overall rate, the full-workspace claim depends on test-grid-tuned corrections. Additionally, report binomial confidence intervals for the 66-attempt comparisons; with n=66, 72.7% vs 57.6% is within sampling noise, so statistical significance should be checked before drawing the headline comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's cleanest generalization claim — that M4 gives the best full-workspace success (72.7%, and 94.1% when stereo hand localization is OK) — assumes that the stereo pivot correction (Sec. 3.3) and the RBF-interpolated z/yaw corrections (Sec. 3.5) are not fitted to the 22 evaluation positions. The text states that the pivot correction 'moves predicted positions toward a fixed point' and reports accuracy on a 9×4 grid, but it does not specify how the pivot parameters were chosen or whether the manual z/yaw reference points overlap the test grid. If these corrections were selected using the same grid positions, M4's advantage over M3 (which also uses the z/yaw corrections but not the pivot) is partly an artifact of memorized corrections, not of the visual-feedback loop. The 94.1% 'SV OK' rate also excludes 15 of 66 attempts post hoc, so the robust full-workspace claim rests on 72.7%, which itself depends on the pivot correction. The M3 inside-NN-area result (29/30) is less vulnerable because M1 shares the same z/yaw corrections and that area was defined independently, but the headline comparison across the full workspace is the load-bearing piece.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends a previous haptic-calibration method for 2D reaching to tabletop grasping on the NICO humanoid robot, using YOLO-based object/hand detection, stereo vision, and task-specific corrections. It compares three calibration models (M1, M2, M3) with a visual-feedback hand-alignment model (M4) over 66 attempts per model on a 22-point grid. The main reported results are that M3 achieves 96.7% success inside the neural-network correction area (29/30) and M4 achieves 72.7% success over the full workspace (48/66), with 94.1% success (48/51) when stereo hand localization is successful.","tokens_in":9384,"tokens_out":5764,"duration_ms":51141,"significance":"If the claims hold, the paper demonstrates that a calibration trained for reaching can transfer to grasping and that a simple stereo-vision hand-alignment loop can outperform calibration outside the trained region—all using only the robot's built-in cameras, without external sensors. The paper is honest about failure modes, explicitly acknowledging M3's poor extrapolation and stereo hand-localization failures, and it reports per-position grid results rather than only aggregate numbers. However, the evidence is statistically thin: 66 attempts per model, no confidence intervals, a post-hoc conditional success rate, and several manually fitted corrections whose selection process is not disclosed. These issues weaken the strength of the central comparative claims.","major_comments":[{"comment":"The headline comparison (M4 72.7% vs M3 57.6% vs M2 48.5%) is reported without any measure of uncertainty. With 66 attempts per model, a two-proportion test for the M4–M3 difference is not statistically significant (approximate p≈0.07), so the claim that M4 is 'best across the full workspace' is not supported by conventional significance. Provide Wilson confidence intervals, and use a paired test (e.g., McNemar or Wilcoxon signed-rank on per-position success counts) since the same 22 grid positions are used. Adjust the qualitative claims accordingly.","section":"Sec. 4.2, Table 2"},{"comment":"The pivot correction (Sec. 3.3) and the RBF-interpolated z and yaw corrections (Sec. 3.5) are manually fitted, but the paper does not specify how the reference points were chosen or whether they overlap the 22 evaluation positions, nor the criterion for selecting the pivot point. If these corrections were tuned using the same test grid, the reported success rates partly reflect hand-tuning rather than the calibration or feedback models. Please report the reference-point coordinates, the pivot-selection procedure, and a sensitivity analysis (e.g., results without the pivot correction, or with a validation split of the reference points).","section":"Sec. 3.3 and Sec. 3.5"},{"comment":"The 'Grasp Rate (SV OK)' of 94.1% (48/51) excludes 15 of 66 attempts (5 positions) where stereo hand localization failed. This is a post-hoc exclusion, and the unconditional full-workspace rate is 72.7%. The paper should clearly label the 94.1% as conditional and should map the failed positions to examine whether they cluster in a particular workspace region or with particular background depths. As it stands, the robust full-workspace claim rests on 72.7%, which itself is not shown to be statistically distinct from M3's 57.6% (see first comment).","section":"Sec. 4.2, Fig. 8"}],"minor_comments":[{"comment":"The text states mean displacement of 2.14 cm inside and 5.45 cm outside the NN area, while the figure caption reports 2.13 cm and 5.35 cm. Please reconcile the numbers.","section":"Sec. 4.1, Fig. 7"},{"comment":"M2 is only reported in Table 2; there is no per-grid visualization or discussion of why it underperforms M3. A sentence or two would help the reader verify the consistency with reaching results.","section":"Sec. 4.1"},{"comment":"The pipeline description does not explicitly specify whether M4 also uses the z and yaw corrections. If it does, that should be stated; if not, the description of the common lower/close phase is ambiguous.","section":"Sec. 3.5"},{"comment":"The figure lists 'Hand V OK' and 'Grasp Rate ( V OK)' without defining what 'V OK' stands for; the text uses 'SV OK'. Please standardize the terminology.","section":"Sec. 4.2, Fig. 8"},{"comment":"No confidence intervals are reported anywhere; even for the key 29/30 and 48/51 results, Wilson intervals would be a useful addition.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and addresses a practical sim-to-real transfer problem with a low-cost setup. The central idea is sound and the authors are transparent about limitations. However, the lack of statistical analysis and the undisclosed manual fitting of corrections are load-bearing issues that need to be resolved before the claims can be accepted. I recommend major revision rather than rejection, as these issues are addressable with additional analysis and reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You can safely read this as a follow-up to the authors' earlier ICANN paper, not as a new mechanism. The new content is a full grasping pipeline—YOLO object/hand detection, stereo localization, the old M1–M3 correction models used as modules, a simple stereo visual-feedback loop (M4), and an actual grasping evaluation on a physical NICO. The work is honest in structure: they compare four models on 66 attempts each, report failures visually, and show where M3 extrapolates badly. That is worth credit.\n\nThe strongest evidence is M3 inside the trained NN area: 29/30. That is consistent with the earlier reaching results and holds up as a calibration-transfer result, despite the small n. The M4 result is more interesting but also more fragile: 72.7% overall, rising to 94.1% only when stereo hand localization 'worked' (48/51). That conditioned number is not a full-workspace claim; the paper is upfront about that, but the headline in the abstract and conclusion leans on it.\n\nWhere I'd push back on the skeptic's worry: M1 shares the z/yaw corrections with M2/M3, and M1 still fails at 16.7%, so those corrections are not doing the heavy lifting for M4. The bigger issue is the stereo pivot correction (Sec. 3.3), which moves predicted positions toward a fixed point to compensate for camera drift. It is a fitted correction, and the paper never says whether the pivot parameters or the z/yaw reference points were chosen independently of the 22-position test grid. That is a real gap. If M4's advantage over M3 is partly a byproduct of corrections tuned to that workspace, the 'visual feedback is better outside the calibrated region' story gets weaker. I would not call it fatal, but it is a hole.\n\nThe smaller soft spots: 66 attempts per model gives no confidence intervals; differences like 72.7 vs 57.6 are probably real but not quantified. One object, one fixed hand pose, no retraining of M1–M3. The paper doesn't compare with a no-correction/no-feedback baseline except M1, which is a weak baseline.\n\nWho is this for: labs working with NICO or similar low-cost humanoids, and people designing sim-to-real calibration experiments. It deserves a serious referee; it's a reproducible, low-cost experiment with honest reporting. My recommendation is to send it out, with the request that the authors clarify exactly how the z/yaw and pivot corrections were selected, and ideally add a permutation or bootstrap confidence interval for the main comparison.\n\nWould I bring to reading group? Maybe, mostly as a case study in how post-hoc conditioning changes the interpretation of a system result.","headline":"A solid, plainly reported extension of haptic calibration from reaching to grasping on NICO; the results are believable but the paper under-specifies where the task-specific corrections come from, and the headline M4 numbers have an asterisk.","tokens_in":9837,"tokens_out":3331,"would_cite":true,"duration_ms":32824,"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":"This paper extends a touchscreen-based haptic calibration from 2D reaching to tabletop grasping on the NICO humanoid, and shows that a stereo-vision hand-alignment loop outperforms calibration across the full workspace.","keywords":["sim-to-real transfer","robotic grasping","haptic calibration","humanoid robot","stereo vision","visual feedback","YOLO object detection","tabletop grasping"],"falsifier":"Repeat the M3 and M4 evaluations using a leave-one-out protocol: derive the z/yaw and pivot corrections from reference points that exclude each evaluated grid position, then compare the success rates. If the inside-area success of M3 drops substantially, the reported rates depend on corrections selected with knowledge of the test grid.","tokens_in":8925,"feed_emoji":"🤖","tokens_out":4194,"duration_ms":41703,"temperature":0.7,"pith_summary":"The paper tries to establish that a calibration model trained on 2D touchscreen reaching can also correct grasp positions for a humanoid robot, and that a simple stereo-vision feedback loop can outperform calibration outside the trained region. On a 22-point tabletop grid, the fully nonlinear calibration model M3 reached 96.7% success inside the neural-network calibration area, while the visual feedback model M4 reached 72.7% across the whole workspace, and 94.1% when stereo hand localization succeeded. The pipeline uses only the robot's built-in low-resolution fisheye cameras, avoiding depth cameras, motion capture, and external tracking. If correct, this means a cheap, explicit calibration plus a simple feedback loop can close much of the sim-to-real gap for grasping.","feed_headline":"Touchscreen calibration lifts robot grasping to 96.7 percent","feed_subtitle":"A stereo-vision hand-alignment loop reaches 72.7 percent across the full tabletop, with no depth cameras or motion capture.","key_machinery":"Three correction models map real-world table coordinates to simulated inverse-kinematics targets: M1 is piecewise-linear, M2 predicts horizontal coordinates with a neural network while interpolating height, and M3 predicts all three coordinates with a fully nonlinear neural network. The visual feedback model M4 replaces calibration with an iterative hand-alignment loop that shifts the IK target by p_k = p_{k-1} + (p_object - p_hand)/2^{k-1} until the horizontal distance falls below 1 cm. Task-specific z-height and palm-yaw corrections, manually measured at irregular reference points and interpolated with radial-basis functions, are applied for grasp execution, and a stereo-vision pivot corre","core_discovery":"The central claim is that the authors' earlier haptic-calibration models, built for 2D reaching, transfer to tabletop grasping when combined with object detection, stereo-vision localization, and task-specific grasp corrections. The fully nonlinear model M3, which predicts a complete corrected 3D target, succeeds in 29 of 30 attempts inside the calibration area, while the visual feedback model M4, which uses no calibration, reaches 48 of 66 attempts across the full workspace, and 48 of 51 when stereo hand localization is correct. The paper interprets this as evidence that calibration-based correction is strongest where training data exist, while simple closed-loop visual alignment is more ro","pith_inferences":["A direct head-to-head comparison with domain-randomization methods on the same robot and grid would test whether explicit calibration plus simple feedback is competitive without learning an end-to-end policy; the reported numbers are suggestive but the evaluation setups differ.","The observed drift after ten minutes of camera operation suggests the fisheye calibration is time-varying; an online recalibration or a learned drift model could make the manual pivot correction unnecessary.","Because the z-height and yaw corrections are task-specific and manually measured, applying the pipeline to other object shapes or sizes would require re-measuring them; an automated measurement procedure could preserve the low-cost advantage.","A stereo-depth method with stronger background rejection, such as temporal filtering, would likely convert the 94.1% conditional success of the visual feedback model into a full-workspace rate."],"forward_implications":["Grasping can be improved without depth cameras, motion capture, or external tracking; the robot's built-in fisheye stereo is sufficient.","Inside the region where neural-network training data exist, a fully nonlinear correction reaches near-reliable grasping performance.","Outside that region, visual feedback beats calibration, showing the two approaches complement each other.","Stereo-vision failures in hand localization are the main bottleneck for visual feedback; fixing them would lift overall success from 72.7% toward the conditional 94.1%.","The gap between M2 and M3 shows that predicting the full 3D target, including height, matters for grasp success."],"fun_headline_variants":["Calibration hits 96.7% in zone, vision loop 72.7% overall","NICO grasps: 96.7% local with calibration, 72.7% full with stereo","Sim-to-real: calibration peaks at 96.7%, visual feedback wins overall","Robot grasping: calibrated spot beats vision, but vision wins workspace","No depth cameras: NICO's stereo loop reaches 72.7% across tabletop"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the manually measured z-height and palm-yaw corrections (Section 3.5) and the stereo pivot correction (Section 3.3) were not tuned on the same 22 test positions used to report success; if they were, the reported success rates partly reflect hand-tuning rather than the calibration or feedback models themselves.","fun_headline_variants_meta":{"raw":{"variants":["Calibration hits 96.7% in zone, vision loop 72.7% overall","NICO grasps: 96.7% local with calibration, 72.7% full with stereo","Sim-to-real: calibration peaks at 96.7%, visual feedback wins overall","Robot grasping: calibrated spot beats vision, but vision wins workspace","No depth cameras: NICO's stereo loop reaches 72.7% across tabletop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1476,"prompt_tokens":714,"completion_tokens":762,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":649}},"tokens_in":458,"tokens_out":762,"duration_ms":8448,"temperature":1.0,"reasoning_tokens":649,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:39:27.607807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the M3 and M4 evaluations using a leave-one-out protocol: derive the z/yaw and pivot corrections from reference points that exclude each evaluated grid position, then compare the success rates. If the inside-area success of M3 drops substantially, the reported rates depend on corrections selected with knowledge of the test grid.","supporting_citations":[],"review_version":1}