{"id":"f74db8f8-a9ef-4cf3-b597-686a0133954e","arxiv_id":"2411.17521","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"BESTAnP decouples the acoustic-n-point problem into range-based translation and azimuth-based rotation estimation, yielding the first closed-form six-degree sonar pose solution with asymptotic CRLB optimality.","lead":"BESTAnP is a new algorithm that estimates a sonar's position and orientation from distance and azimuth measurements of known points. It breaks the problem into two closed-form steps and, after one refinement step, is claimed to reach the statistical accuracy limit while running over ten times faster than prior AnP solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's unqualified claim that BESTAnP gives a closed-form six-DOF AnP solution and approaches the CRLB breaks down for coplanar/near-coplanar point sets, because the range-only translation step (6) becomes singular/ill-conditioned; no degeneracy analysis is provided.","rationale":"I read the paper in good faith: the decoupling of range-based translation and azimuth-based rotation is a genuine algorithmic contribution, the experiments are extensive, and the empirical CRLB tracking in Fig. 3 supports the single-GN claim for the tested non-degenerate configurations. The most load-bearing soft spot, in my view, is not the delegation of Theorem 3 to [27] (a checkable companion result) but the unqualified scope of the closed-form and optimality claims with respect to point geometry. Eq. (6) requires non-coplanar points to invert A^T A; the paper states this in passing but does not analyze the coplanar/near-coplanar regime, and planar scenes are common in the sonar setting. The reader's formulation is partially right but imprecise: for coplanar points with the sonar off the plane, the azimuth nullspace of Q* is not generally enlarged; the rank failure is in the translation design matrix A. I would keep the CONDITIONAL verdict and ask for a degeneracy analysis or an explicit restriction in the abstract/problem statement, plus a near-coplanar experiment.","tokens_in":76,"tokens_out":28278,"duration_ms":329188,"concrete_test":"Run T=1000 Monte Carlo trials with the same noise levels as Fig. 5 (σ_d=1e-3 m, σ_θ=1e-3, n=1000) for three point configurations: (i) exactly coplanar points on a plane not containing the sonar; (ii) near-coplanar points with elevation spread 0.1 deg; (iii) the paper's non-coplanar baseline. Record the rank/condition number of A^T A, the fraction of trials in which \\hat t_BE is finite, and the RMSE of \\hat t_BE and \\hat R_GN against the CRLB. If configuration (i) has no unique \\hat t_BE and (ii) is orders of magnitude above the CRLB at n=1000, the abstract needs an explicit non-coplanarity/generic-position caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"BESTAnP's core decoupling rests on the range-only linear least squares (6) in Section IV-A, which inverts A^T A. If the 3D points are coplanar, A has a null vector coming from the plane equation, so A^T A is singular and \\hat t_BE is not uniquely defined; for near-coplanar points the estimator is ill-conditioned, with errors that can be large for any realistic n. The problem statement (Section III) and the abstract do not state this restriction, and Section IV gives no degeneracy analysis, fallback, or conditioning criterion. Since planar structures (walls, floors) are common in forward-looking-sonar scenes, the claimed closed-form full-pose solution and the single-GN asymptotic-optimality result (Theorem 3) are not established for a practically important class of AnP instances. I would add one correction to the reader's weakest_assumption: the failure mechanism is the rank deficiency of A in the translation step, not generally an enlarged nullspace of Q*; for a planar wall with the sonar off the plane, the azimuth matrix Q* can still have the required rank 5.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes BESTAnP, a two-step estimator for the acoustic-n-point (AnP) problem. The first step estimates the sonar translation from distance-only measurements via a bias-eliminated linear least-squares formulation, and the second step estimates the rotation by eigendecomposition of a bias-corrected Gram matrix built from azimuth observations. A single Gauss-Newton iteration is then applied to refine the pose. The paper claims that the two-step estimator is √n-consistent, that one GN iteration yields an estimator with the same asymptotic distribution as the maximum-likelihood estimator, that the covariance approaches the Cramér-Rao lower bound, and that the method is the first closed-form six-degree-of-freedom AnP solver. The claims are supported by Monte Carlo simulations, a Raspberry Pi timing study, real pool experiments, and an embedded sonar-based odometry demonstration.","tokens_in":15111,"tokens_out":8595,"duration_ms":83194,"significance":"If the technical claims hold, BESTAnP would be a practically valuable AnP solver: it is non-iterative up to a single GN refinement, has an explicit consistency analysis, and is demonstrated to be over ten times faster than state-of-the-art methods while matching their accuracy. The manuscript is commendable for shipping open-source code and for validating the algorithm in both synthetic and real underwater scenarios, including a sonar-only odometry pipeline. The main caveats are that the headline statistical-optimality result is not proved in this paper, the CRLB is used without derivation, and the non-coplanarity/generic-position assumption behind the closed-form solution is stated only locally and is not analyzed as a degeneracy limitation.","major_comments":[{"comment":"The closed-form translation step inverts A^T A in Eq. (6), and the rotation step requires rank(B*)=5 in Eq. (9). The manuscript states, immediately before Eq. (6), that the 3D points must be spatially non-coplanar, and it assumes 'generic positions' after Eq. (9), but the abstract and Section III present the six-degree closed-form solution and the asymptotic-optimality claim without these preconditions. For coplanar point sets, such as a sonar observing a wall or floor, A is rank-deficient and \\hat t_BE in Eq. (7) is not uniquely defined; for near-coplanar sets the estimator is ill-conditioned. This is a load-bearing restriction on the central claim. Please state the assumption in the problem formulation and abstract, and add a degeneracy/conditioning analysis or a fallback procedure for planar scenes.","section":"Section IV-C, Theorem 3"},{"comment":"Theorem 3 is the basis for the 'statistically optimal' and 'covariance approaching the CRLB' claims, but its proof is not contained in this manuscript. The text says 'For the proof of Theorem 3, please refer to the proof of Theorem 7 in [27]', where [27] is the authors' own arXiv preprint on camera-robot pose estimation from points and lines. Because the observation model, parameterization, and estimation problem differ from the AnP likelihood (5), the reduction is not immediate. Please either provide a self-contained proof of Theorem 3 or state precisely the mapping from the AnP problem to Theorem 7 of [27], including all regularity conditions that must be verified.","section":"Section V-A, Figs. 2–6"},{"comment":"The CRLB curves in Figs. 2–6 are displayed without a definition or derivation. The paper does not give the CRLB expression for the observation model (4), nor does it state the conditions under which the single-GN estimator attains it. Since 'statistically optimal' is a headline contribution, please include the CRLB formula (or a precise citation with the exact expression) and explain how it is computed for the plotted experiments.","section":"Section IV-A, Eq. (7)"}],"minor_comments":[{"comment":"The first sentence reads 'We has introduced the noise-free sonar measurement model'; it should be 'We have introduced'.","section":"Algorithm 1"},{"comment":"The input list includes only 'azimuth angle measurements tan θi', but step 4 determines the sign by checking cos θ_i and the reprojected cos θ̂_i. Please specify whether the signed azimuth angle (or its cosine) is available as input, since tan alone does not determine the sign of cos.","section":"Appendix A"},{"comment":"The proof of Lemma 1 refers to 'Lemma 3' before Lemma 3 has been stated (it appears in Appendix B); please reorder or renumber the lemmas for readability.","section":"Fig. 8 caption"},{"comment":"The caption contains the typo 'Dual-palne'; it should be 'Dual-plane'.","section":"Section V-A.3"},{"comment":"The sentence 'the App method performs even better than the CRLB' should be clarified as an effect of bias in the App estimator, since the CRLB is a bound for unbiased estimators; otherwise the statement may confuse readers.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' prior work: Theorem 1 is from [26] and the key asymptotic-optimality proof is deferred to [27]. The editor may wish to verify that [27] is available and that the reduction is valid; if [27] is not yet published, the proof of Theorem 3 should be included in this paper. The 'first closed-form six-DOF AnP solver' claim should also be checked against the related-work literature, and the authors should be asked to make the non-coplanarity assumption explicit in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does deliver a genuinely new algorithmic result: it decouples the AnP problem into a range-only translation estimate and an azimuth-only rotation estimate, producing a closed-form six-degree pose. That is the first such solver in the AnP literature, and the experimental section backs it up — Monte Carlo scaling, pool tests, and an odometry integration, all with open-source code. The claimed speedup (over 10x, about 1 ms on a Pi 5) is plausible from the algorithm's structure.\n\nSecond, the headline \"statistically optimal\" claim is not actually proved here. Theorem 3, the single-Gauss-Newton optimality result, is deferred to the authors' own unpublished preprint [27], and the CRLB curves appear without derivation. The paper is honest about this, but it means you cannot take the optimality label on faith. The consistency part (Theorem 2) is proved in the appendices, and the translation step is directly from [26], which is a published, peer-reviewed paper.\n\nThe real soft spot is degeneracy. The translation step inverts A^T A, which requires the 3D points to be non-coplanar. The paper states this in Section IV-A, but the abstract and problem statement do not. For planar or near-planar scenes — walls, floors, which are common in sonar — the translation estimate is singular or ill-conditioned, and the whole pipeline breaks. The stress-test note is right that the failure mechanism is rank deficiency in A, not an enlarged nullspace of Q*. This is a genuine limitation and should be analyzed or at least discussed. The experiments carefully use non-coplanar structures, so the reported results are not misleading, but the scope is narrower than the abstract implies.\n\nMinor issues: the noise model adds noise to tan(theta) rather than theta; the paper justifies this by simulation, which is fine. The single-GN refinement means \"closed-form\" is slightly generous, but the initial estimator is closed-form, so that's more a wording quibble.\n\nOverall, this is a useful, well-executed paper. It deserves peer review. I would send it to a robotics or estimation venue, with a request that the authors either prove Theorem 3 in the paper or clearly mark it as a dependency, and add a degeneracy discussion.","headline":"Real new AnP solver with strong experiments, but the optimality proof is outsourced to an unreviewed preprint and the coplanar degeneracy is unaddressed.","tokens_in":15732,"tokens_out":2547,"would_cite":true,"duration_ms":23619,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The acoustic-n-point problem is solved in closed form by separating ranges and azimuths, with one Gauss-Newton step reaching the Cramer-Rao bound.","keywords":["acoustic-n-point","2D forward-looking sonar","pose estimation","range-based localization","bias elimination","consistent estimator","Cramer-Rao lower bound","sonar odometry"],"falsifier":"Run the method on synthetic scenes where all 3D points lie in a single plane while the sonar looks at them with realistic pose and noise levels. If BESTAnP's translation RMSE no longer decreases at the predicted rate with more points, or the rotation estimate stays biased while the plane geometry is held fixed, the non-coplanarity premise is doing the work; if the RMSE still tracks the Cramer-Rao bound on coplanar configurations, the premise can be relaxed.","tokens_in":14664,"feed_emoji":"📡","tokens_out":5111,"duration_ms":49799,"temperature":0.7,"pith_summary":"The paper claims that estimating a sonar's full six-degree pose from distance and azimuth measurements of known 3D points does not require iterative search. It argues that distance measurements alone determine the sonar's position through range-based localization, and that azimuth measurements then determine its rotation through an eigendecomposition, after correcting for measurement noise bias. A single Gauss-Newton step from this closed-form starting point is claimed to produce an estimator whose covariance approaches the Cramer-Rao lower bound as the number of points grows. If correct, this makes real-time acoustic pose estimation practical on low-power onboard computers and enables sonar-only odometry in unknown environments.","feed_headline":"First closed-form six-degree pose from sonar point measurements","feed_subtitle":"Distances solve for position, azimuths for orientation; one Gauss-Newton step reaches the Cramer-Rao bound.","key_machinery":"The carrying object is the $6 \\times 6$ matrix $\\mathbf{Q}_{BE} = (1/n)\\mathbf{B}^\\top\\mathbf{B} - \\hat{\\sigma}_\\theta^2 \\mathbf{S}$, where the rows of $\\mathbf{B}$ are $[\\tan\\theta_i(\\mathbf{W}p_i - \\hat{\\mathbf{t}}_{BE})^\\top,\\ (\\hat{\\mathbf{t}}_{BE} - \\mathbf{W}p_i)^\\top]$ and $\\mathbf{S}$ contains the sample scatter of the translated 3D points. Its noise-free counterpart $\\mathbf{Q}^*$ has a nullspace containing $\\mathbf{r}^* = [\\mathbf{r}_1^{*\\top},\\ \\mathbf{r}_2^{*\\top}]^\\top$, so the unit eigenvector of the smallest eigenvalue of $\\mathbf{Q}_{BE}$ is the consistent estimate of the rotation rows; subtracting $\\hat{\\sigma}_\\theta^2 \\mathbf{S}$ removes the bias caused by angle noise. Translation comes separately from the bias-eliminated linear least-squares problem $\\min_{\\mathbf{x}\\in\\mathbb{R}^4}\\|\\mathbf{A}\\mathbf{x}-\\mathbf{b}\\|^2$ with $\\mathbf{A} = [-2\\mathbf{W}p_i^\\top,\\ 1]$ and $b_i = d_i^2 - \\|\\mathbf{W}p_i\\|^2$.","core_discovery":"The paper's central claim is that the measured partial spherical coordinates split the acoustic-n-point problem cleanly: the true distance $d_i^* = \\|\\mathbf{W}p_i - \\mathbf{t}^*\\|$ makes translation a range-only localization problem, while the azimuth relation $\\tan\\theta_i^* = \\mathbf{r}_2^*(\\mathbf{W}p_i - \\mathbf{t}^*) / \\mathbf{r}_1^*(\\mathbf{W}p_i - \\mathbf{t}^*)$ turns rotation into a linear null-space problem once translation is fixed. The paper constructs a bias-eliminated matrix $\\mathbf{Q}_{BE} = \\mathbf{Q} - \\hat{\\sigma}_\\theta^2 \\mathbf{S}$ whose smallest eigenvector yields the first two rows of the rotation matrix consistently, then projects the result onto $SO(3)$. It proves $\\sqrt{n}$-consistency of both the translation and rotation estimators and shows that a single Gauss-Newton iteration from this start is asymptotically equivalent to the maximum-likelihood estimate, so the covariance reaches the Cramer-Rao lower bound as $n$ grows.","pith_inferences":["A natural extension the paper leaves implicit is a minimal-point version of the AnP solver: the nullspace structure of $\\mathbf{Q}_{BE}$ suggests that a small set of correspondences could seed RANSAC-style robust estimation, analogous to minimal solvers in the PnP literature.","The non-coplanarity premise means planar underwater structures such as walls, floors, and flat ship hulls are a degeneracy risk; the paper does not analyze this regime, so a practical system would need to detect and fall back to constrained models for such scenes.","The decoupling logic may transfer to any range-and-single-angle sensor, such as radar or ultrawideband localization, where distance and one angle are measured but elevation is unavailable; the same two-step bias-eliminated estimator could apply after adapting the measurement model.","The paper's comparison treats angle noise as added to $\\tan\\theta$; for sonars with wide azimuth apertures or high noise levels, the equivalence to direct angle noise weakens, so a useful extension would be a bias correction that works directly on $\\theta$ residuals."],"forward_implications":["Full six-degree sonar pose estimation becomes practical on embedded platforms, since the closed-form core uses only linear least squares and an eigendecomposition, with a single Gauss-Newton refinement step.","The accuracy of acoustic pose estimation approaches the Cramer-Rao lower bound in the large-point regime, giving users a statistically grounded baseline rather than an ad hoc iterative solution.","Sonar-only odometry becomes feasible: the paper demonstrates trajectories estimated by alternately running AnP pose tracking and 3D triangulation without external sensors.","Existing AnP methods that iterate on part of the translation or rely on planar or weak-perspective approximations are no longer needed for full-pose accuracy in generic scenes."],"supporting_citations":[{"why":"Supplies the bias-eliminated range-based localization estimator used for the translation step, including the $\\sqrt{n}$-consistency theorem and the concentration lemma used in later proofs.","marker":"[26]"},{"why":"Provides the eigendecomposition-based rotation estimation, the sign-correction trick, and the theorem that a single Gauss-Newton iteration from a $\\sqrt{n}$-consistent start is asymptotically equivalent to the maximum-likelihood estimate.","marker":"[27]"},{"why":"Supplies the bias-elimination construction for pose estimation with noisy angle measurements, the SO(3) projection step, and the tangent-space Gauss-Newton parameterization.","marker":"[13]"},{"why":"Defines the AnP problem and supplies the App, Non-app, and Combined+CIO baseline algorithms that the experiments compare against.","marker":"[4]"},{"why":"Provides the ANRS triangulation routine that the paper uses to build its novel sonar-only odometry demonstration.","marker":"[17]"}],"fun_headline_variants":["Bi-step sonar pose: distance for translation, azimuth for rotation","First closed-form six-degree pose from sonar point matches","Sonar pose estimation ten times faster with statistical optimality","Optimal sonar pose via one Gauss-Newton step"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the 3D points being spatially non-coplanar and in generic positions, because the translation step needs a unique sphere intersection and the rotation step needs the azimuth matrix's nullspace to be exactly one-dimensional; near-coplanar point sets, which occur naturally with walls and floors underwater, are not analyzed.","fun_headline_variants_meta":{"raw":{"variants":["Bi-step sonar pose: distance for translation, azimuth for rotation","First closed-form six-degree pose from sonar point matches","Sonar pose estimation ten times faster with statistical optimality","Optimal sonar pose via one Gauss-Newton step"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3919,"prompt_tokens":1004,"completion_tokens":2915,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2846}},"tokens_in":620,"tokens_out":2915,"duration_ms":21048,"temperature":1.0,"reasoning_tokens":2846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:02:11.674813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on synthetic scenes where all 3D points lie in a single plane while the sonar looks at them with realistic pose and noise levels. If BESTAnP's translation RMSE no longer decreases at the predicted rate with more points, or the rotation estimate stays biased while the plane geometry is held fixed, the non-coplanarity premise is doing the work; if the RMSE still tracks the Cramer-Rao bound on coplanar configurations, the premise can be relaxed.","supporting_citations":[{"cited_title":"Optimal camera-robot pose estimation in linear time from points and lines","cited_arxiv_id":"2407.16151","evidence_quote":"Provides the eigendecomposition-based rotation estimation, the sign-correction trick, and the theorem that a single Gauss-Newton iteration from a $\\sqrt{n}$-consistent start is asymptotically equivalent to the maximum-likelihood estimate."},{"cited_title":"Cpnp: Consistent pose estimator for perspective-n-point problem with bias elimination,","cited_arxiv_id":null,"evidence_quote":"Supplies the bias-elimination construction for pose estimation with noisy angle measurements, the SO(3) projection step, and the tangent-space Gauss-Newton parameterization."},{"cited_title":"Acoustic-n-point for solving 2d forward looking sonar pose estima- tion,","cited_arxiv_id":null,"evidence_quote":"Defines the AnP problem and supplies the App, Non-app, and Combined+CIO baseline algorithms that the experiments compare against."},{"cited_title":"Application of forward-scan sonar stereo for 3-d scene reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the ANRS triangulation routine that the paper uses to build its novel sonar-only odometry demonstration."}],"review_version":1}