{"id":"5e2ba3f6-ec64-4158-8ac3-c48477f45995","arxiv_id":"2504.21619","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Hand-eye calibration can be done in about six seconds without a calibration target by registering a single 3D scan of the robot base to a CAD model using a learned point cloud registration network.","lead":"This robot-vision paper calibrates a camera to a robot arm in seconds by letting the camera look at the robot base once and using a trained point cloud registration network to find the base pose. It replaces the usual chessboard and multiple arm movements with a single scan, and reports accuracy close to a commercial calibration product.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-world comparability claim depends on the unverified assumption that the reference CAD frame coincides with the physical robot base frame (Eq.","rationale":"The reader's verdict is CONDITIONAL, and the weakest assumption identified is exactly the unverified alignment between the reference model frame and the physical robot base frame. My stress-test analysis converges on the same load-bearing concern: the real-world comparison lacks an independent ground truth, and the simulation cannot expose CAD-to-real frame deviations because the reference and ground-truth data share the same CAD source. The paper does provide useful contributions: a generic dataset generation method, a broad simulation study across 14 cobot models, and publicly available code and data, which are genuine supporting evidence. However, the headline claim of 'performance comparable to existing commercial hand-eye calibration solutions' is only as strong as the validity of Eq. 5 and the associated pre-transformation. Since the paper does not describe how the reference frame is aligned to the physical robot base, nor does it verify this alignment experimentally, the claim remains conditional. I agree with the reader that this is a specific, addressable issue rather than a fundamental flaw: if the authors supply an independent ground-truth comparison or a verification procedure for the CAD-to-real alignment, the claim would be substantially strengthened. Therefore, I do not propose changing the verdict; CONDITIONAL remains appropriate until the concern is resolved.","tokens_in":12414,"tokens_out":2518,"duration_ms":27633,"concrete_test":"Obtain an independent ground-truth camera-to-base transform for the UR10e setup using a high-accuracy external reference, such as a laser tracker or a calibrated target firmly attached to the robot base, measured from the camera. Recompute the hand-eye transform with the proposed method and compare against this ground truth, in addition to the BM-HEC result. If the proposed method's error relative to ground truth is within the same 2-3 mm / 0.001-0.002 rad tolerance reported in Table III, the comparability claim holds; if the error exceeds the reported offset or is dominated by a systematic bias consistent with CAD-frame misalignment, the central claim is not supported. A complementary check is to perturb the reference model frame by a known rotation/translation in simulation and measure the induced error in t_cT, quantifying sensitivity to the assumed alignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed method achieves performance comparable to commercial hand-eye calibration in real-world conditions. The method's accuracy rests on Eq. 5, where the camera-to-base transform is derived from the registered camera-to-reference-model transform under the assumption that 'the frame of the reference data is aligned with the frame of the robot base in the real world' (Section III-D). This assumption is never verified. In practice, the reference model point cloud Q is generated from CAD/URDF/OBJ files; the actual robot base's coordinate frame and scale can differ from the CAD frame due to manufacturing tolerances, mounting variations, or model inaccuracies. The paper acknowledges a scale difference and introduces a pre-transformation ref'_refT (Eq. 4), but does not specify how this pre-transformation is obtained or validated. Any error in this pre-alignment, whether in rotation, translation, or scale, propagates directly into b_cT and hence into the final hand-eye transform t_cT = b_tT^{-1} c_bT^{-1}. The simulation study cannot detect this issue because in simulation the reference model and the ground-truth robot base are generated from the same CAD source, so CAD-to-real deviations are absent by construction. The physical experiment compares only against one commercial solution on one robot (UR10e), reporting offsets of 2-3 mm and 0.001-0.002 rad. Without an independent ground-truth measurement of the camera-to-base transform, agreement with a single commercial product does not establish absolute accuracy; if the commercial method itself has a similar bias, or if the CAD frame is misaligned, the claimed comparability could be spurious. This is the most load-bearing weakness because it directly threatens the validity of the headline real-world result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a 3D vision-based hand-eye calibration method that avoids external calibration objects by using the robot base itself as the target. The camera observes the base, a learned point-cloud registration network (PREDATOR) aligns the captured point cloud to a reference CAD/URDF model, and the resulting camera-to-model transform is combined with forward kinematics to obtain the hand-eye transform. The authors extend their prior work by proposing a generic dataset generation procedure (hemisphere viewpoints plus realistic joint configurations), evaluating the method in simulation on 14 collaborative robot arms from 9 brands, and performing a real-world experiment on a UR10e with a Zivid 2+ MR60 camera, comparing against a commercial eye-in-hand calibration solution (BM-HEC). They report simulation mean errors of 1.29 mm and 0.39 degrees, physical offsets of 2-3 mm and 0.001-0.002 rad relative to the commercial product, and a calibration time of about 6 seconds.","tokens_in":12639,"tokens_out":6364,"duration_ms":65777,"significance":"If the accuracy claims hold, the method offers a practical, fast, and calibration-object-free alternative to classical AX=XB hand-eye calibration. The paper has concrete strengths: the code and dataset are publicly released, the simulation study covers a wide range of cobot models, and a physical experiment against a commercial product is included. However, the evidence as presented is not fully convincing. The simulation uses the same CAD/URDF models for both training and test, so the reported accuracy mainly validates the registration pipeline under distribution-matched conditions. The real-world comparison has no independent ground truth and only shows agreement with one commercial solution. Most importantly, the derivation of the camera-to-base transform in Eq. (5) rests on an unverified assertion that the reference model frame coincides with the physical robot base frame, and the pre-transformation in Eq. (4) is not explained or validated. These gaps are load-bearing for the central claim of comparable real-world performance. The paper is likely of interest to the robotics community but requires substantial revision before its claims are fully supported.","major_comments":[{"comment":"The derivation of b_cT in Eq. (5) depends entirely on the statement in Section III-D that 'the frame of the reference data is aligned with the frame of the robot base in the real world.' No procedure is given for establishing or validating the pre-transformation ref'_refT introduced in Eq. (4); the manuscript does not state how this transform is computed, which degrees of freedom it contains (rotation, translation, and/or scale), or how its accuracy is verified. Any error in this pre-alignment propagates directly into t_cT through Eq. (7) and biases the final hand-eye result. In the simulation, the reference model and the ground-truth robot base are generated from the same CAD source, so this effect is absent by construction. Please specify the method for obtaining ref'_refT, provide a validation of the reference-frame alignment on the physical robot, or quantify the sensitivity of the calibration to a known misalignment of the reference frame.","section":"III-D, Eqs. (4)-(5)"},{"comment":"The simulation study trains the PREDATOR registration network on point clouds generated from the same 14 robot CAD/URDF models that are later used to generate the test point clouds (Table I lists per-robot dataset sizes and Table II lists the same 14 robots for evaluation). The reported mean errors of 1.29 mm and 0.39 degrees therefore largely characterize the registration pipeline under near-ideal, distribution-matched conditions rather than the generalizability of the method to unseen robot models. To support the claimed generality, please add a leave-one-out or held-out evaluation (e.g., train on 13 robots and test on the 14th) or introduce synthetic geometry perturbations (mounting offsets, model deviations) and show that accuracy degrades gracefully.","section":"IV-B, Table I"},{"comment":"The real-world comparison uses a single commercial solution (BM-HEC) as the reference, without any independent ground-truth measurement of the camera-to-flange transform. The reported offsets of 2-3 mm in position and 0.001-0.002 rad in rotation therefore cannot distinguish errors in the proposed method from errors in the commercial product, especially since the commercial solution is itself an estimate from an AX=XB formulation with its own noise characteristics. The paper's own text in Section IV-B acknowledges that 'discrepancies between the CAD model and the scanned robot base can further contribute to calibration errors,' but the physical experiment does not quantify this contribution. Please add an independent accuracy check (e.g., an external tracking system, a precisely machined test tool, or a multi-pose consistency metric) or clearly characterize the reported offsets as relative agreement rather than absolute accuracy.","section":"IV-C, Table III"}],"minor_comments":[{"comment":"The sentence 'the final calibration result is obtained by averaging the rotation matrices in quaternion space' should clarify the exact averaging procedure, including how quaternion sign ambiguity is handled and whether the average is computed over the 30 repeated scans or over the 30 poses; the current text is ambiguous.","section":"IV-B"},{"comment":"The dataset sizes in Table I are listed without a description of the generation parameters, such as the number of hemispheres, the number of camera viewpoints per radius, and the proportion of data coming from the hemisphere strategy versus the realistic-joint-configuration strategy. Please add these details to the dataset-generation description.","section":"IV-A, Table I"},{"comment":"Several figures contain garbled Unicode symbols in the manuscript PDF, which makes the plots difficult to read. Please regenerate the figures with proper font embedding and ensure the axis labels, legends, and captions are legible.","section":"Figures 4, 5, 9, 11"},{"comment":"The novelty relative to the authors' prior work [22] should be stated more explicitly. The current contribution list emphasizes the dataset generation, the 14-arm simulation study, and the real-world comparison, but does not clearly state whether the core registration pipeline or the use of the robot base as a target is new or unchanged from [22].","section":"Introduction, Contributions"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the authors' prior work [22]. The main new elements are the dataset generation strategy, the large-scale simulation across 14 cobots, and the real-world comparison. The underlying method (point-cloud registration of the robot base followed by kinematic composition) appears largely unchanged. The editor may wish to consider whether the incremental contribution is sufficient for a journal publication, especially given that the central real-world accuracy claim currently rests on an unverified frame-alignment assumption and a comparison against a single commercial baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible engineering extension, not a conceptual breakthrough. The new dataset generation procedure and the 14-arm simulation campaign are concrete additions over the authors' earlier paper, and the public code and data make the results checkable. But the real-world comparability claim leans on an assumption that is stated, not verified: the reference CAD model's frame is taken to coincide exactly with the physical robot base frame.\n\nWhat the paper actually does well: it turns the 'look at robot base' idea into something more generic. Generating training point clouds from hemisphere views plus realistic joint configurations is a sensible way to cover multiple robot geometries. Running 900 calibration evaluations across 14 arms from 9 brands is a solid breadth test. The physical experiment on a UR10e with a Zivid camera is a legitimate attempt to show the pipeline works outside simulation, and the comparison to a commercial AX=XB solver gives a useful anchor. The code and dataset are public, which is real evidence.\n\nThe soft spots are real but not fatal. The biggest one is the alignment assumption behind Eq. 5. The paper says the reference frame is aligned with the physical base frame, and introduces a scale pre-transformation in Eq. 4, but never says how that pre-transformation is obtained or validated. Any rotation, translation, or scale error between the CAD frame and the actual base leaks straight into the hand-eye result. Because the simulation generates both training and test clouds from the same CAD source, the reported 1.29 mm / 0.39° errors cannot detect this bias. That does not invalidate the method, but it means the simulation numbers are optimistic for absolute accuracy.\n\nThe real-world experiment compares against one commercial product on one robot. Offsets of 2–3 mm and 0.001–0.002 rad look good, but there is no independent ground truth, so the 'comparable' claim should be read as 'agrees with this particular commercial implementation under these conditions.' If the commercial solver shares the same kinematic bias, agreement is less meaningful. I would want at least one independent measurement, or a clear statement acknowledging this limitation.\n\nMinor point: Table III reports single values per method; I would prefer the scatter across the four runs rather than just the mean.\n\nOverall, the math is straightforward, the paper is honest about being an extension, and the reproducibility support is there. This deserves a serious referee. I would recommend conditional acceptance: keep the simulation and dataset contributions, and require either a stronger real-world validation or a more careful framing of what the experiment can show.","headline":"A credible engineering extension of the base-as-target idea, with a genuinely useful dataset-generation method and a broad simulation campaign, but the real-world 'comparable to commercial' claim rests on an unverified CAD-to-physical frame alignment assumption.","tokens_in":13273,"tokens_out":2235,"would_cite":false,"duration_ms":23125,"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":"Point-cloud registration of the robot base alone produces hand-eye calibration comparable to commercial solutions.","keywords":["hand-eye calibration","3D vision","point cloud registration","collaborative robot","robot base","eye-in-hand calibration","calibration-object-free","dataset generation"],"falsifier":"Use an independent high-precision measurement, such as a laser tracker or a precision fixture attached to the robot base, to obtain the true camera-to-base transform on a robot whose physical base geometry is deliberately made to differ from its CAD model; if the calibration error tracks that geometric mismatch rather than staying within the reported few millimeters, the frame-alignment assumption is the limiting factor and the central claim is refuted.","tokens_in":12170,"feed_emoji":"🤖","tokens_out":6881,"duration_ms":69511,"temperature":0.7,"pith_summary":"The paper claims that hand-eye calibration for a collaborative robot can be performed with a single 3D view of the robot base, without any external calibration object such as a chessboard. It reports that this approach is accurate enough to match a commercial calibration solution, while cutting the procedure from about two minutes and fifty seconds to under six seconds. This is useful because frequent recalibration in industrial settings is currently slow and inconvenient, and the method turns recalibration into a quick scan-and-go step. The paper supports the claim with simulation across 14 robot arms from 9 brands and with a physical experiment on a collaborative robot arm.","feed_headline":"Robot hand-eye calibration in seconds by looking at the base once","feed_subtitle":"A single 3D scan of the robot base matches commercial calibration accuracy without chessboards or repeated poses.","key_machinery":"The mechanism is a closed kinematic chain built around the robot base instead of a removable target. The registration module, trained on a generated dataset of robot-base point clouds, estimates the transform $^{c}_{ref'}T$ from the camera to the reference model; applying the predefined $^{ref'}_{ref}T$ gives the camera-to-base transform, and the chain $I = {}^{b}_{t}T {}^{t}_{c}T {}^{c}_{b}T$ converts this into the eye-in-hand calibration. The dataset is generated by sampling camera viewpoints on hemispheres of radius 0.5 m and 0.7 m and by adding robot poses sampled under realistic joint limits, so the method is not tuned to one robot geometry.","core_discovery":"The central discovery, stated on the paper's own terms, is that the robot base can serve as the calibration target. A camera mounted on the flange captures a point cloud of the base; a learned registration network aligns this scan to a reference model of the base whose frame is defined to coincide with the robot base frame. The resulting camera-to-base transform $^{b}_{c}T$, combined with forward kinematics through the closed chain $I = {}^{b}_{t}T {}^{t}_{c}T {}^{c}_{b}T$, yields the camera-to-flange transform. The paper reports a simulation mean error of $1.29$ mm in position and $0.39$ degrees in rotation over 14 robot arms, and a physical comparison in which the calibration differs from a commercial solution by roughly 3 mm, 2 mm, and 3 mm in position and 0.001 to 0.002 radians in rotation, with a single pose and a single frame of point cloud sufficient.","pith_inferences":["Beyond the paper: if CAD-to-real geometric deviations are the dominant error source, the method's accuracy ceiling will vary from robot unit to robot unit; a verification step that uses the robot's own joint motions to check consistency between multiple base views could expose that bias without an external target.","Beyond the paper: the same camera-to-base estimate is the full eye-to-hand calibration, so the method should extend to fixed cameras directly, which the paper notes is computationally simpler than the validated eye-in-hand case.","Beyond the paper: the reported simulation errors of about 1.29 mm sit close to the 2 mm voxel size of the point clouds, suggesting that denser scans or a higher-resolution reference model would lower the error floor further.","Beyond the paper: one could build a self-diagnostic calibration test by comparing calibrations from a few distinct base-looking poses; agreement across poses would indicate that registration noise, not base-frame misalignment, dominates."],"forward_implications":["Recalibrating a vision-guided collaborative robot becomes a seconds-long operation: point the camera at the base, capture one point cloud, and the hand-eye transform follows.","No external calibration object, board, or target-holding fixture is needed, removing a source of setup error and spare-part dependency.","The method's accuracy, as reported, is within a few millimeters and roughly a tenth of a degree of a commercial multi-pose calibration, placing it in the usable range for many industrial manipulation tasks.","Because the simulation spans 14 arms from 9 brands, the dataset-generation recipe may transfer to new robot models without per-robot manual tuning.","A single 3D frame is enough for a stable calibration, so the method is also suitable for cameras that cannot hold a static pose for long."],"supporting_citations":[{"why":"establishes the prior method this work extends, using the robot base as the calibration target.","marker":"[22]"},{"why":"supplies the learned low-overlap point cloud registration network used to align the captured scan to the reference model.","marker":"[35]"},{"why":"defines the classical AX=XB formulation that the commercial calibration baseline and the paper's comparison rely on.","marker":"[25]"},{"why":"ties point cloud registration to hand-eye calibration, the methodological lineage of this approach.","marker":"[24]"},{"why":"surveys hand-eye calibration techniques and their constraints, motivating the target-free fast-calibration goal.","marker":"[11]"},{"why":"defines the simultaneous robot/world and tool/flange formulation AX=YB that frames the eye-to-hand variant.","marker":"[26]"}],"fun_headline_variants":["Single base scan calibrates robot hand-eye in seconds","One glance at robot base yields hand-eye calibration in seconds","Drop chessboard: robot hand-eye calibration via single base scan","Calibrate robot hand-eye in seconds with one look at base","Robot hand-eye calibrated in seconds, one base scan, no chessboard"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reference model point cloud used for registration is exactly aligned in frame and scale with the physical robot base, so any mismatch between the CAD model and the real arm's geometry shows up directly as calibration error.","fun_headline_variants_meta":{"raw":{"variants":["Single base scan calibrates robot hand-eye in seconds","One glance at robot base yields hand-eye calibration in seconds","Drop chessboard: robot hand-eye calibration via single base scan","Calibrate robot hand-eye in seconds with one look at base","Robot hand-eye calibrated in seconds, one base scan, no chessboard"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001737,"raw_usage":{"total_tokens":6879,"prompt_tokens":974,"completion_tokens":5905,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":5819}},"tokens_in":590,"tokens_out":5905,"duration_ms":42174,"temperature":1.0,"reasoning_tokens":5819,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:57:25.586934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use an independent high-precision measurement, such as a laser tracker or a precision fixture attached to the robot base, to obtain the true camera-to-base transform on a robot whose physical base geometry is deliberately made to differ from its CAD model; if the calibration error tracks that geometric mismatch rather than staying within the reported few millimeters, the frame-alignment assumption is the limiting factor and the central claim is refuted.","supporting_citations":[{"cited_title":"Automatic robot hand-eye calibration enabled by learning-based 3d vision","cited_arxiv_id":null,"evidence_quote":"establishes the prior method this work extends, using the robot base as the calibration target."},{"cited_title":"Predator: Registration of 3d point clouds with low overlap","cited_arxiv_id":null,"evidence_quote":"supplies the learned low-overlap point cloud registration network used to align the captured scan to the reference model."},{"cited_title":"Calibration of wrist-mounted robotic sensors by solving homogeneous transform equations of the form ax= xb","cited_arxiv_id":null,"evidence_quote":"defines the classical AX=XB formulation that the commercial calibration baseline and the paper's comparison rely on."},{"cited_title":"Point cloud registration- enabled globally optimal hand–eye calibration","cited_arxiv_id":null,"evidence_quote":"ties point cloud registration to hand-eye calibration, the methodological lineage of this approach."},{"cited_title":"A comparative review of hand-eye calibration techniques for vision guided robots","cited_arxiv_id":null,"evidence_quote":"surveys hand-eye calibration techniques and their constraints, motivating the target-free fast-calibration goal."},{"cited_title":"Simultaneous robot/world and tool/flange calibration by solving homogeneous trans- formation equations of the form ax= yb","cited_arxiv_id":null,"evidence_quote":"defines the simultaneous robot/world and tool/flange formulation AX=YB that frames the eye-to-hand variant."}],"review_version":1}