{"id":"92d775cd-c797-45ff-b925-ee306ad4cf75","arxiv_id":"2507.19851","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hand-eye calibration method that solves the camera-to-arm transform by detecting the same flat plane in multiple depth-cloud views, using a closed-form estimate refined by Lie-group Gauss-Newton optimization.","lead":"PlaneHEC calibrates the pose between a robot's depth camera and its arm by observing any flat surface, like a wall or table, from multiple viewpoints. It avoids calibration targets and complex models, and claims calibration errors below the camera's own depth error in a few seconds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The closed-form rotation solution (Eq. 11) has an uncharacterized observability condition: without sufficient rotational diversity in robot poses, the stacked coefficient matrix is rank-deficient, SVD returns an arbitrary null-space vector, and the resulting X is meaningless; the paper neither…","rationale":"The reader's weakest assumption matches my assessment: the uncharacterized rank condition in Eq. 11 is the single most load-bearing technical gap. It directly determines whether the closed-form solution, and hence the entire algorithm, can recover X at all. I considered whether the Gauss-Newton Jacobian derivation (Eqs. 17-19) is a more fundamental flaw, but on inspection the column-splitting trick treats direction vectors as differences of points, so the point-perturbation formula applies by linearity; the Jacobian is defensible. The baseline comparison and the absence of the claimed simulation experiments are real editorial weaknesses, but they are about evidence quality rather than about whether the method can work. The degeneracy issue is more fundamental: it threatens the core mechanism in a realistic failure mode and is easily overlooked because SVD always returns 'a' solution. The proposed synthetic rank test is cheap and would settle the matter directly, which is why I keep the reader's CONDITIONAL verdict unchanged rather than escalating or downgrading it.","tokens_in":10845,"tokens_out":13786,"duration_ms":160522,"concrete_test":"Run a synthetic simulation with known ground-truth X and plane, generating clean M_i from chosen A_i. Test three motion plans: (1) all A_i have identical rotation and only translations vary; (2) rotations are about the plane-normal axis only; (3) random diverse rotations. For each, form the stacked Eq. 11 matrix, compute its singular-value spectrum and numerical rank, and compare the SVD-plus-orthogonalization R_X to ground truth. If configurations (1) and (2) give rank strictly below 8 and large rotation errors while (3) works, the degeneracy concern is confirmed. Also report the ratio of the 8th to 9th singular values as a practical conditioning diagnostic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central linear stage solves Eq. 9 in vectorized form: each consecutive pair contributes a 3×9 block (R_{A,j}^T ⊗ n_j^T − R_{A,j+1}^T ⊗ n_{j+1}^T) vec(R_X) = 0. The true R_X lies in the null space. For the SVD-based least-squares solution to recover a unique rotation (up to sign), the stacked matrix must have rank exactly 8; otherwise the minimal-singular-value vector is an arbitrary mix of null-space directions, and the subsequent orthogonalization does not repair the estimate. This rank condition requires the robot orientations to produce sufficiently diverse camera-frame plane normals. It fails in practical configurations: pure translations (constant R_A), rotations about a single axis, or rotations about the plane normal all leave the rows linearly dependent and the system underdetermined. The paper claims a 'generalized' and 'universal' calibration method but gives no observability analysis, no pose-selection guidance, and no simulation that sweeps pose diversity. Table I uses one presumably well-conditioned real dataset, so it cannot expose this failure mode. This is load-bearing because the closed-form solution is the initial value for the Gauss-Newton stage; if the initial R_X is wrong, the refinement cannot be trusted to recover the true X. Without a degeneracy characterization, a user following the method with a conservative motion plan can obtain a confident-looking but meaningless calibration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PlaneHEC is a hand-eye calibration method for eye-in-hand RGB-D cameras that uses observations of an arbitrary planar surface. For each view, a RANSAC plane fit produces a 1x4 row vector M_i, and the robot pose A_i is read from the controller; these are combined into the constraint Y = M_i X A_i, where X is the unknown hand-eye transform. The paper derives a closed-form rotation estimate from vectorized pairwise differences of these constraints, solves for the translation, and then refines both with a Gauss-Newton iteration on SE(3). Experiments on a FLEXIV-Rizon 4 robot with a RealSense D435 report rotation and translation errors of 0.13 deg and 3.35 mm, along with reconstruction consistency checks and a runtime of 0.28 s, and the paper claims that PlaneHEC is the fastest and most accurate among compared point-cloud-based methods.","tokens_in":11167,"tokens_out":9501,"duration_ms":113530,"significance":"The central idea is practically attractive: no calibration object is needed, plane parameters aggregate many depth points against sensor noise, and the closed-form stage avoids the initialization problem of ICP-based approaches. The use of 1x4 plane-row constraints with a Kronecker-product linear solution is a useful reformulation of the hand-eye problem for depth cameras. If the derivation and evaluation are corrected, the method could be a solid engineering contribution to camera-robot calibration. The paper does not provide code or data, but it does validate on held-out reconstruction frames, which is a reasonable consistency check rather than a circular fit to the calibrated transform. The main weaknesses are technical inconsistencies in the optimization derivation, an uncharacterized degeneracy condition in the closed-form rotation stage, and a mismatch between the claimed simulation evaluation and the actual experiments.","major_comments":[{"comment":"The residual blocks in Eq. (17) use inconsistent subscripts (i/i+1 in the first line versus j/j+1 in the others) and include spurious M X (0;1) terms that do not cancel because M_j and M_{j+1} differ. Eq. (18) mirrors the same problems in the Jacobian. As written, these equations do not equal the columns of g_j = M_j X A_j - M_{j+1} X A_{j+1}, so the Gauss-Newton stage is not reproducible and may be incorrect. Please rewrite the residual and Jacobian with consistent indices and verify the perturbation signs.","section":"III-D (Eqs. 17-18)"},{"comment":"The claimed normalization is not what right multiplication by A_average does: A_i times A_average has translation -R_{A,i} t_average + t_{A,i}, not t_{A,i} - t_average, so the resulting translation is not centered. More importantly, replacing A_i by a modified pose changes Y = M_i X A_i unless Y is redefined accordingly, so the plane constraint is not invariant under this operation. State precisely how the translations are normalized and prove that the plane equations remain satisfied; if left multiplication was intended, the effect on the constraint must still be addressed.","section":"III-A (Eq. 6)"},{"comment":"The rotation stage solves a homogeneous linear system whose unique least-squares solution requires the stacked coefficient matrix to have rank exactly 8. The paper gives no observability analysis, no pose-selection guidance, and no simulation that sweeps pose diversity. Degenerate motions such as pure translations, rotations about a single axis, or rotations about the plane normal make the coefficient matrix rank-deficient, in which case the SVD solution is an arbitrary vector in a null space and the resulting R_X is meaningless. This closed-form R_X is used as the initial value for Gauss-Newton, so the degeneracy directly undermines the overall method. Please add a rank/degeneracy condition and validate it with simulated or real pose-diversity experiments.","section":"III-C (Eq. 11)"},{"comment":"The abstract and conclusion state that PlaneHEC was evaluated in both simulated and real-world environments, but the manuscript contains no simulation experiment at all. All numerical accuracy results come from a single real platform, and the values in Table I are standard deviations over 50 random subsets rather than errors against a known ground-truth X. Thus the headline figures (0.13 deg, 3.35 mm) demonstrate repeatability or consistency, not absolute calibration accuracy. Add a simulation study with known X and controlled plane-detection noise, or explicitly qualify the accuracy claims as consistency measures.","section":"IV (Experiments)"},{"comment":"Table II compares numbers taken from different publications with different sensors, robots, and accuracy levels, using et/max(Ec,ER) as a normalization; this does not establish that PlaneHEC is more accurate than the baselines under the same protocol. The runtime claim is also unclear: Section IV-D says the average processing time for 500 frames was approximately 0.01 s per frame, while Table II lists a runtime of 0.28 s and says the majority of runtime is plane detection. Clarify what the 0.28 s includes and, if feasible, run at least one baseline on the same platform before claiming state-of-the-art accuracy and speed.","section":"Table II and Section IV-D"}],"minor_comments":[{"comment":"The claim of evaluation in 'simulated and real-world environments' is contradicted by the absence of any simulation section; this should be corrected in the abstract and conclusion.","section":"Abstract/Conclusion"},{"comment":"The caption and text refer to rotation and translation 'errors', but the reported quantities are standard deviations of repeated estimates computed from random subsets; the terminology should be changed to variability or consistency.","section":"Table I"},{"comment":"The procedure for resolving the sign of the RANSAC plane normal so that Y equals M_base rather than -M_base is not described; please specify how the direction of the normal vector is determined.","section":"III-A"},{"comment":"The runtime accounting needs consistency: 0.01 s per frame over 500 frames implies 5 s, not 0.28 s; state which stages are included in the reported runtime and which are excluded.","section":"IV-D"},{"comment":"The axes and legends of Figures 3 and 5 are not described in the text or captions, making the plots difficult to interpret; please provide units and legend definitions.","section":"Figures 3 and 5"}],"recommendation":"major_revision","confidential_remarks":"The headline comparison in Table II mixes numbers from prior systems with different sensors and robots, so the state-of-the-art claim is not supported as presented. The mismatch between the abstract's promised simulation experiments and the actual content should be resolved. I would also ask the editor to encourage the authors to release code and data, since the derivations in Eqs. (17)-(18) would otherwise be hard for readers to verify."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the Y = M X A row-vector plane constraint is a real extension of the classic AX=XB, and the paper shows it works on a real robot with a low-cost depth camera, with held-out validation. Second, the paper is sloppier than it should be: the abstract promises simulations that don't exist, the 'universal' claim ignores a real observability condition, and the baseline comparison in Table II is not apples-to-apples.\n\nThe new bit is the constraint form. Instead of tracking a feature point or a full pose, they track a plane as a 1x4 vector M, giving Y = M X A. The closed-form via Kronecker products and the Gauss-Newton refinement on SE(3) are standard building blocks, but combining them this way is clean, and the derivation is mostly correct. I checked Eq.11 carefully; the 3x9 blocks are right, and with enough pose diversity the SVD solution is well-posed. The real experiments are the strongest part: 300 point clouds, 50 random trials, a sensible held-out plane reconstruction test, and a qualitative robot-arm point cloud alignment. The reported 0.13 deg and 3.35 mm are plausible for a D435, and the runtime is genuinely fast.\n\nThe soft spots are in the claims and the missing analysis. The degeneracy issue is real: if the robot only translates, or rotates about a single axis, the stacked coefficient matrix for R_X is rank-deficient and you get a confident-looking but meaningless answer. The paper doesn't characterize this, doesn't offer pose-selection guidance, and then calls the method 'generalized' and 'universal.' That's overreach. It doesn't sink the method for typical use, but a user following the instructions with a poor motion plan can silently get garbage. The promised simulation experiments are absent; if they exist in a supplementary file, that should be stated. Table II compares different platforms and different sensors, then normalizes by camera accuracy in a way I don't find fully convincing; 'state-of-the-art' is stronger than the evidence. The subscripts in Eq.17/18 and the A_average normalization explanation are minor but point to a lack of polish.\n\nWho gets value: robotics engineers who want a fast, markerless, targetless calibration with an RGB-D camera. It's a good conference/workshop paper. The right call is to send it to peer review, with the expectation that the authors add an observability analysis (or at least a clear warning), reconcile the simulation claim, and soften the baseline comparison. I'd accept the core idea as a contribution.","headline":"A genuinely new plane-constraint calibration formulation with solid real-robot validation, but the paper overclaims universality and omits the observability analysis that would make the method safe to use.","tokens_in":11698,"tokens_out":4036,"would_cite":true,"duration_ms":42131,"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":"The paper shows that the hand-eye transform $X$ of a robot-mounted depth camera can be recovered from a single flat surface by solving $Y = M_i X A_i$, using a closed-form linear solve refined by Gauss–Newton on SE(3).","keywords":["hand-eye calibration","point cloud plane detection","RANSAC plane fitting","Kronecker product","Lie algebra optimization","depth camera","SE(3)","robot vision"],"falsifier":"Collect point clouds of one plane from several views while keeping the robot end-effector orientation essentially constant (rotational diversity near zero). The paper's derivation predicts the coefficient matrix in Eq. 11 becomes rank-deficient and the recovered $X$ should scatter across trials; if repeated runs still give a stable, accurate $X$, the claimed need for non-parallel consecutive views is contradicted, whereas unstable $X$ would confirm the degeneracy.","tokens_in":10665,"feed_emoji":"🤖","tokens_out":7365,"duration_ms":75698,"temperature":0.7,"pith_summary":"This paper tries to establish that hand-eye calibration for a robot-mounted depth camera can be done with any flat surface, such as a wall or tabletop, with no special calibration object or geometric model. It reformulates the task as a set of planar constraints $Y = M_i X A_i$, where $M_i$ is a $1\\times 4$ plane vector and $A_i$ a robot pose, and solves for the unknown hand-eye transform $X$ in two stages: a closed-form linear solution, then a Gauss–Newton refinement on SE(3). On a low-cost structured-light camera mounted on a robotic arm, the authors report 0.13 degrees rotation error and 3.35 mm translation error, with a runtime of about 0.28 seconds, and claim both higher accuracy and speed than the compared point-cloud-based calibration methods. If right, this makes fast, marker-free hand-eye calibration practical in ordinary environments.","feed_headline":"Flat wall calibration hits 0.13-degree hand-eye error","feed_subtitle":"Closed-form solve plus Gauss-Newton refinement on SE(3) beats point-cloud baselines for accuracy and speed.","key_machinery":"The central object is the plane-constraint equation $Y = M_i X A_i$, in which $M_i$ is a $1\\times 4$ row vector (unit normal and signed distance) of the fixed plane observed from view $i$, $A_i$ is the robot pose from forward kinematics, and $X$ is the unknown camera-to-tool transform. The equation encodes the invariance of the plane in the robot base frame. The solution machinery is a Kronecker-product linear system for $\\mathrm{vec}(R_X)$ whose least-squares solution is orthogonalized by SVD, a linear solve for $t_X$, and a Gauss–Newton update using the Lie-algebra perturbation model $(Xp)^\\odot$, iterating until $\\Delta X$ is small.","core_discovery":"The central claim is that the hand-eye transformation $X$ can be fully recovered by observing one static plane from multiple robot poses and solving $Y = M_i X A_i$, where $M_i = (v_i^T, d_i)$ is the plane row vector from RANSAC and $A_i$ is the robot pose. The rotation part is rewritten as a homogeneous Kronecker-product system $(R_{A,j}^T \\otimes v_j^T - R_{A,j+1}^T \\otimes v_{j+1}^T)\\mathrm{vec}(R_X)=0$, solved by SVD; the translation is then a linear least-squares problem. A Gauss–Newton iteration on SE(3) refines both jointly, using the closed-form result as initialization. The paper reports a mean rotation error of 0.13 degrees and a mean translation error of 3.35 mm on a RealSense-D435 camera on a FLEXIV-Rizon 4 arm using 30 views, and states that the translation error falls below the camera's depth accuracy because plane fitting averages many points.","pith_inferences":["The same plane-constraint equations could be extended to eye-to-hand setups or to multiple planes in a single view; the paper does not test these configurations, but the algebra does not depend on the mounting side.","The authors leave the degeneracy condition uncharacterized; a quantitative threshold on the minimum rotation angle between consecutive views would let users verify their data collection before solving.","Because the closed-form step is linear, a recursive update of the plane vector $M_i$ could turn PlaneHEC into an online calibration procedure; the paper only reports a batch solution."],"forward_implications":["Hand-eye calibration loses the need for calibration boards, markers, or precise geometric models; any sufficiently flat surface suffices.","Because plane fitting uses many points, the reported 3.35 mm translation error is below the camera's depth noise, so calibration accuracy is not limited by raw sensor error.","Rotation error below 0.7 degrees in all trials, with 0.13 degrees at 30 views, makes the method usable where orientation accuracy matters.","With about 8.6 iterations and 0.28 seconds total runtime, the method is fast enough for online or repeated recalibration."],"supporting_citations":[{"why":"Introduces the kinematic-loop formulation AX = XB that PlaneHEC generalizes to planar constraints.","marker":"[19]"},{"why":"Supplies the RANSAC plane-detection algorithm used to obtain the plane vectors M_i from depth point clouds.","marker":"[36]"},{"why":"Provides the Kronecker-product least-squares technique that the closed-form solution is built on.","marker":"[34]"},{"why":"Gives the Lie-group solution method for AX = XB that underlies the Gauss-Newton refinement on SE(3).","marker":"[35]"},{"why":"Demonstrates planar-constraint calibration for line laser scanners, which PlaneHEC extends to depth cameras.","marker":"[17]"},{"why":"Reconstruction-based point-cloud calibration baseline whose error metrics and evaluation strategy PlaneHEC adopts.","marker":"[14]"},{"why":"Learning-based 3D vision baseline LRBO that PlaneHEC compares against for accuracy and runtime.","marker":"[28]"}],"fun_headline_variants":["One flat wall, 30 robot views: 0.13° calibration accuracy","Wall-only hand-eye calibration beats point-cloud baselines","0.13° error: PlaneHEC calibrates from any planar surface","PlaneHEC: Calibrate robotic hand-eye with any wall or table","No model needed: PlaneHEC uses any plane for hand-eye calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the robot's reported poses $A_i$ are exact and that the observed views provide enough rotational diversity for the homogeneous linear system for $\\mathrm{vec}(R_X)$ to have a unique least-squares solution; if the arm's orientation barely changes between views, the coefficient matrix becomes rank-deficient and the solved rotation is arbitrary.","fun_headline_variants_meta":{"raw":{"variants":["One flat wall, 30 robot views: 0.13° calibration accuracy","Wall-only hand-eye calibration beats point-cloud baselines","0.13° error: PlaneHEC calibrates from any planar surface","PlaneHEC: Calibrate robotic hand-eye with any wall or table","No model needed: PlaneHEC uses any plane for hand-eye calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000959,"raw_usage":{"total_tokens":4100,"prompt_tokens":975,"completion_tokens":3125,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":3042}},"tokens_in":591,"tokens_out":3125,"duration_ms":23798,"temperature":1.0,"reasoning_tokens":3042,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:59:10.249079+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect point clouds of one plane from several views while keeping the robot end-effector orientation essentially constant (rotational diversity near zero). The paper's derivation predicts the coefficient matrix in Eq. 11 becomes rank-deficient and the recovered $X$ should scatter across trials; if repeated runs still give a stable, accurate $X$, the claimed need for non-parallel consecutive views is contradicted, whereas unstable $X$ would confirm the degeneracy.","supporting_citations":[{"cited_title":"Calibration of wrist-mounted robotic sensors by solving homogeneous transform equations of the form ax= xb,","cited_arxiv_id":null,"evidence_quote":"Introduces the kinematic-loop formulation AX = XB that PlaneHEC generalizes to planar constraints."},{"cited_title":"Ransac for robotic applications: A survey,","cited_arxiv_id":null,"evidence_quote":"Supplies the RANSAC plane-detection algorithm used to obtain the plane vectors M_i from depth point clouds."},{"cited_title":"Solving the robot-world/hand-eye calibration problem using the kronecker product,","cited_arxiv_id":null,"evidence_quote":"Provides the Kronecker-product least-squares technique that the closed-form solution is built on."},{"cited_title":"Robot sensor calibration: solving ax= xb on the euclidean group,","cited_arxiv_id":null,"evidence_quote":"Gives the Lie-group solution method for AX = XB that underlies the Gauss-Newton refinement on SE(3)."},{"cited_title":"Six dof eye-to-hand calibration from 2d measurements using planar constraints,","cited_arxiv_id":null,"evidence_quote":"Demonstrates planar-constraint calibration for line laser scanners, which PlaneHEC extends to depth cameras."},{"cited_title":"Reconstruction-based hand–eye cal- ibration using arbitrary objects,","cited_arxiv_id":null,"evidence_quote":"Reconstruction-based point-cloud calibration baseline whose error metrics and evaluation strategy PlaneHEC adopts."},{"cited_title":"Automatic robot hand-eye calibration enabled by learning-based 3d vision,","cited_arxiv_id":null,"evidence_quote":"Learning-based 3D vision baseline LRBO that PlaneHEC compares against for accuracy and runtime."}],"review_version":1}