{"id":"02156cce-ad7e-4353-b843-f013628e4621","arxiv_id":"2505.00842","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multi-robot localization is done with an EKF on matrix Lie groups using new stochastic composition, averaging, and fusion operations plus Mahalanobis rejection of faulty ArUco pseudo-pose measurements.","lead":"Robot teams can keep track of where each member is by sharing camera and motion measurements, but bad or noisy readings can ruin the shared estimate. This paper builds a Lie group Kalman filter that fuses inertial, wheel odometry, and ArUco marker data, rejects faulty marker readings, and reports experiments with up to three wheeled robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The filter is specified on a 15-D state manifold and a 9-D measurement manifold, but the prediction equations, noise covariance, adjoint, and measurement noise are 18-D and 12-D; as written, the algorithm is not well-defined on the claimed groups.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the dimensions in the filter machinery do not match the dimensions of the stated Lie groups. This is load-bearing because every EKF equation depends on m: if m = 15, the 18D vectors and matrices in Eqs. (51)-(53) are undefined for GX; if m = 18, the algorithm runs on G = SE2(3) × SE2(3), not on GX, and the claimed manifold is wrong. The 12D measurement covariance Rl similarly does not correspond to the 9D GZ. The paper provides no code, data, or explicit statement identifying which dimensions are correct, so the central experimental claim cannot be traced to a well-defined algorithm. A secondary issue is the averaging identity in Section III-D: the step X_i^{α_i} = \\bar{X}_i^{α_i} exp(α_i[ζ_i]^) is generally false for non-commuting factors, but this affects mainly the fusion demonstration, whereas the dimension mismatch affects the main prediction and update filter used in every experiment. The experiments are also qualitative, without error bars or an external baseline, which further supports the conditional verdict rather than full acceptance. Since my concern substantiates the reader's concern without identifying a different outcome, the verdict should remain CONDITIONAL.","tokens_in":26922,"tokens_out":10384,"duration_ms":108933,"concrete_test":"Re-derive the EKF on the 15D group GX by writing the Lie-algebra coordinate as ζ = [φ_R; ρ_p; ρ_v; b_g; b_a] and computing F, Q, H, K, and the state correction from Eqs. (14)-(21) on this 15D algebra. If the resulting 15D recursion is not exactly the restriction of the printed 18D equations (51)-(53) to the zero bottom-rotation subspace, then the stated manifold and the filter matrices describe different algorithms. Apply the same reduction to GZ: express Rl on the 9D Lie algebra and check whether the εI3 block vanishes from the reduced noise model or is required to keep HPH^T + R invertible; if it is required, the measurement model includes three unmodeled directions that are not in GZ.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states GX := SE2(3) × R3 × R3, which is 15-dimensional, yet Eq. (51) defines f ∈ R18, Eq. (52) defines w ∼ N(0_18, Q) with Q ∈ R18×18, and Eq. (53) uses this 18D noise in the discrete-time propagation. Appendix A correspondingly derives Ad_X and J_r as 18×18 for the block group G, and only says GX is the special case where the second block's rotation is identity. The explicit X in Eq. (50) indeed fixes that rotation to I3, so the bottom rotation coordinate is constrained; but the paper never states that the extra three covariance directions are zero. The experiments set P(0|0) = 10^-2 I18, which assigns nonzero uncertainty in a direction that leaves GX. The measurement side has the same problem: GZ := SE(3) × R3 is 9-dimensional, but Section IV-C1 defines Rl = diag([Rpl, εI3, Rvl]) ∈ R12×12, and the εI3 block corresponds to a rotation coordinate of the velocity block that is fixed to I3 in Zl. Since the general EKF of Section II-C is stated for an m-dimensional Lie group, m is inconsistent between the manifold and the matrices. If this is a typo, the paper does not say which of the 15D/18D or 9D/12D objects are correct, so the reported experimental results cannot be checked against the stated theory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops stochastic operations on matrix Lie groups—composition, inversion, difference, averaging, fusion, and constrained fusion—for correlated and uncorrelated estimates, and embeds them in a multi-robot EKF localization framework. The prediction step runs on G_X = SE_2(3) × R^3 × R^3 using IMU measurements; updates fuse wheel-odometry velocity and pseudo-pose measurements constructed from ArUco markers on G_Z = SE(3) × R^3. A Mahalanobis-distance fault detection module gates pseudo-pose updates for leader, follower type 1, and follower type 2 robots. The paper reports experiments with three AgileX LIMO robots and a PhaseSpace motion tracker, claiming real-time, scalable, fault-tolerant localization and showing that IMU/WO fusion failures are recovered by ArUco/FD updates.","tokens_in":27267,"tokens_out":5522,"duration_ms":55910,"significance":"If the proposed operations and filter are correct, the paper offers a unified Lie-group treatment of multi-robot sensor fusion: it generalizes prior pose-only operations to correlated multi-estimate composition and fusion, and it demonstrates the framework on physical hardware with three robot types rather than in simulation. The fault detection module addresses a practical issue in marker-based updates, and the experiments use a motion-capture ground truth. However, part of the theoretical contribution is located in the authors' prior paper [44], and the dimension inconsistencies described below must be resolved before the reported experiments can be reliably tied to the stated mathematics.","major_comments":[{"comment":"The prediction filter is not well-defined on the stated 15-dimensional state group. Section IV-A defines G_X := SE_2(3) × R^3 × R^3, which has dimension 15 (9 for pose/velocity plus 6 for biases), but the prediction uses f ∈ R^18 in Eq. (51), w ∼ N(0_18, Q) with Q ∈ R^{18×18} in Eq. (52), and Eq. (53) propagates this 18-dimensional noise. Appendix A derives Ad_X and J_r for the 18-dimensional group G = SE_2(3) × SE_2(3) and states that G_X is the special case where the second block's rotation equals I_3; the explicit X in Eq. (50) indeed fixes that rotation. The paper never states that the three covariance directions corresponding to that fixed rotation are zero. The experiments set P(0|0) = 10^{-2} I_18 in Section V-D1, which assigns nonzero uncertainty in directions that leave G_X. As written, Eqs. (13)–(18) are therefore not defined on the claimed manifold, and the trajectories cannot be reproduced from the stated theory. The authors should either define the filter on the full 18-dimensional group G, or explicitly state the 15-dimensional parameterization with zero-noise/zero-covariance constraints on the fixed rotation and correct Eq. (53) and P(0|0) accordingly.","section":"§IV-A, Appendix A, §V-D1 (Eqs. 50–53, 62)"},{"comment":"The measurement update has the same dimension inconsistency. The text states G_Z := SE(3) × R^3, which is 9-dimensional, but defines R_l = diag([R_pl, ε I_3, R_vl]) ∈ R^{12×12}, with analogous R_f1 and R_f2. The ε I_3 block corresponds to the rotation coordinate of the second block, which is fixed to I_3 in the measurement members Z_l, Z_f1, and Z_f2. The general EKF in Section II-C is written for a q-dimensional group, so q is inconsistent between the stated manifold and the matrices. If the measurement model is intended to live in SE(3) × SE(3) (12-dimensional), the text should say so and remove the placeholder ε block; if it is intended to be SE(3) × R^3, the innovation and noise must be 9-dimensional with the fixed rotation component removed. The authors need to state which objects are correct and adjust the equations and experiments accordingly.","section":"§IV-C1–C3 (measurement model and R_l, R_f1, R_f2)"},{"comment":"The averaging and fusion operations are not self-contained, despite being central to the paper's claimed contribution (i) and to the fusion experiments in Section V-D4. The averaging operation is introduced via X_i^{α_i} = \\bar{X}_i^{α_i} exp(α_i [ζ_i]^∧) without proof; fractional powers of a noisy Lie-group element are not canonical, so this equality needs a precise definition (e.g., through the exponential of α_i times the logarithm) and a consistency argument before the covariance formula (33) is valid. The fusion operation's closed-form solution, Eqs. (35)–(38), is taken directly from the authors' prior paper [44] with only the final result reproduced, and the cross-covariance recursion is also quoted from [44] without derivation. Since the overall filter consistency and the fusion experiments depend on these steps, the manuscript should provide self-contained proofs or clearly state which parts of the claimed novelty are delegated to [44].","section":"§III-D, §III-E, §V-D4"}],"minor_comments":[{"comment":"The contribution list jumps from (ii) to (iv); the missing item (iii) should be restored or the list renumbered.","section":"§I-B"},{"comment":"Equation (52) first writes w with −w_g and −w_a and then states an equivalence to +w_g and +w_a; because these noises are zero-mean, the distributions coincide, but the double definition is confusing and should be clarified.","section":"§IV-A, Eq. (52)"},{"comment":"The appendix says the proof for G_Z := SE(3) × R^3 follows similarly and is left to the reader; given that the dimension of this group is central to the issues above, the explicit group law, logarithm, and Jacobians for G_Z should be included.","section":"Appendix A"},{"comment":"Figure 9 is described qualitatively as showing better or worse localization for different P(0|0), but no error metric or axis definition is given; please state what quantity is plotted so the claim that the proposed method is insensitive to P_0 can be checked.","section":"§V-D1, Fig. 9"},{"comment":"The group SE_2(3) is used both as the 9-dimensional pose-velocity group and as a factor of the block-diagonal group G_X; the embedding used in Eq. (50) should be defined once and used consistently throughout.","section":"General notation"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core leans heavily on the authors' own prior work [44] for the fusion operation and cross-covariance recursion, so the novelty and self-citation pattern should be checked carefully during revision; the missing contribution (iii) suggests a last-minute edit that should be verified. The experimental platform and fault-detection evaluation are solid, and the direction is publishable after the dimension inconsistencies and the missing derivations are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The batch stochastic operations—composition of n correlated estimates, averaging, and constrained fusion—go beyond Mangelson et al.'s pairwise machinery, and the three-role leader/follower localization with fault detection is a practical combination that hasn't been put together this way. But as written, the central filter is not well-defined. The prediction state is declared to live on SE2(3)×R^3×R^3, which is 15-dimensional, yet the process model f, the noise w, the covariance Q, and the adjoint matrices in Appendix A are all 18-dimensional. The measurement group G_Z = SE(3)×R^3 is 9-dimensional, but the measurement noise covariances R_l, R_f1, R_f2 are 12×12. The paper doesn't say that the extra directions have zero variance—in fact P(0|0) is set to 10^{-2} I_{18}, which puts nonzero covariance in directions that leave G_X. If this is a typo, the manuscript doesn't indicate which set of equations is correct, so the experimental results can't be checked against the stated theory.\n\nThe averaging operation has a more basic problem. The paper asserts that X_i^{α_i} = \\bar{X}_i^{α_i} exp(α_i [ζ_i]^∧). That identity is false for general elements of SE(3); it holds only if the mean and the noise perturbation commute. The experiments use averaging to form the reference member for fusion, so this is load-bearing, not a harmless aside.\n\nThe fusion operation and the cross-covariance recursion are taken from the authors' preceding RA-L paper [44] without derivation. That's defensible as a follow-up, but it moves some of the claimed novelty outside this manuscript and makes self-contained verification difficult.\n\nWhat's genuinely good: the experimental design is serious—motion-capture ground truth, camera and IMU calibration, and a fault-detection demonstration that separates good and bad ArUco updates. The qualitative plots do show the proposed filter tracking ground truth where IMU/WO fusion drifts. But there are no error bars, no comparison against a standard EKF on SE(3) or an existing multi-robot method, and no released code or data, so the headline claims about efficiency, reliability, and scalability are unsupported.\n\nMy take: the underlying idea is plausible and worth building on, but the manuscript has load-bearing correctness issues that must be fixed. I'd send it to review, but with the expectation of major revisions and a request for a corrected, self-contained statement of the stochastic operations.","headline":"A plausible multi-robot Lie-group localization framework, but the manuscript has a load-bearing dimension mismatch and an invalid averaging identity; it needs major revision before the results can be trusted.","tokens_in":27761,"tokens_out":5914,"would_cite":false,"duration_ms":54942,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E11","68T40"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Lie-group EKF with newly defined stochastic operations fuses IMU, wheel odometry, and fiducial-marker pseudo-poses, keeping multi-robot teams localized where plain IMU/odometry fusion diverges.","keywords":["multi-robot localization","matrix Lie groups","extended Kalman filter","stochastic operations on Lie groups","fault detection","fiducial markers","multi-modal sensor fusion","collaborative localization"],"falsifier":"Compute the intrinsic dimension of $G_X=\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$ and compare it with the dimension of $Q$ in Eq. (52): if the filter uses the 18-dimensional matrix in the 15-dimensional covariance update, the predicted covariance is mathematically undefined on the stated state manifold, so the method cannot be implemented exactly as written. A separate behavioral falsifier: drive a follower through a sharp turn while its marker is fully occluded for a known interval; if the gated filter's error is not significantly smaller than the ungated filter's error, the fault-detection claim fails.","tokens_in":26667,"feed_emoji":"🤖","tokens_out":12722,"duration_ms":119472,"temperature":0.7,"pith_summary":"This paper proposes a fault-tolerant, multi-modal localization framework for teams of wheeled robots, built on new stochastic operations for matrix Lie groups. Its central claim is that by representing each robot's pose, velocity, and IMU biases as a single element of the extended-pose group $\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$, and by constructing pseudo-pose measurements from fiducial markers through Lie-group composition, one can fuse IMU, wheel odometry, and relative-pose data in a single EKF that scales across a leader-follower network. A squared-Mahalanobis-distance fault-detection gate admits only trustworthy marker updates. The paper reports real-time experiments in which this IMU/WO/ArUco/FD filter recovers ground-truth trajectories where IMU-only or IMU-plus-odometry fusion diverges, especially during turns, sharp maneuvers, and uneven terrain.","feed_headline":"Lie-group filter keeps robot teams on track when odometry fails","feed_subtitle":"New stochastic operations fuse IMU, wheel data, and fiducial-marker pseudo-poses across a robot network.","key_machinery":"The central object is the matrix Lie group $G_X=\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$ for prediction (an extended pose group whose elements carry rotation, position, velocity, and IMU bias terms) together with the measurement group $G_Z=\\mathbb{SE}(3)\\times\\mathbb{R}^3$. The argument is carried by newly defined stochastic operations: composition, inverse, difference, averaging, and fusion, in which Lie-algebra uncertainties $\\zeta\\sim\\mathcal{N}(0,P)$ are mapped through the Adjoint $\\mathrm{Ad}_X$ and the right Jacobian $J_r$, so covariance propagation is done entirely on the Lie algebra. A squared Mahalanobis distance $D^2(k)$ on the innovation gates whether a pseudo-pose update is admitted.","core_discovery":"On its own terms, the discovery is that every operation a filter needs—composition, differencing, inversion, averaging, and fusion—can be defined for correlated and uncorrelated stochastic members of a matrix Lie group by pushing Gaussian uncertainties through the Adjoint map and the right Jacobian, and that these operations make it possible to synthesize pseudo-pose measurements that propagate through a multi-robot network. The prediction step runs on $G_X=\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$, where each robot carries rotation, position, velocity, and IMU biases; the update consumes body velocity plus pseudo-pose on $G_Z=\\mathbb{SE}(3)\\times\\mathbb{R}^3$. The paper claims that this combination runs in real time, remains accurate as leader, follower type 1, and follower type 2 estimates are cascaded, and recovers ground-truth trajectories in cases where the IMU/wheel-odometry filter alone fails.","pith_inferences":["The same stochastic operations should apply to cooperative localization with UWB, LiDAR, or camera-relative measurements, since the update only needs a relative-pose or relative-velocity model on a Lie group; the paper demonstrates only fiducial markers.","A controlled fault-injection study—occluding a follower's marker for known intervals during a sharp turn—would let practitioners measure the gate's detection latency and false-alarm rate, quantities the paper does not report.","The Lie-group fusion rule may also serve object-level multi-robot SLAM, where landmarks and robots both carry SE(3) estimates with cross-correlations, because the fusion operation is designed for correlated pose estimates."],"forward_implications":["A team of robots can maintain globally consistent localization with only one stationary fiducial marker for the leader while followers observe mobile markers on their neighbors, with no GPS required.","A robot whose proprioceptive IMU/odometry filter has begun to diverge can be pulled back to the ground-truth trajectory whenever a chain of trusted relative-pose measurements reaches it.","Because the stochastic operations are defined on an arbitrary matrix Lie group, the same EKF machinery transfers to other product-manifold state spaces and other relative-pose sensors.","Fusing two pseudo-pose measurements at a follower yields an estimate that the paper reports as better than either individual estimate and better than the IMU/odometry filter."],"supporting_citations":[{"why":"Supplies the closed-form fusion of correlated pose estimates and the cross-covariance recursion that Section III-E uses to combine a follower's multiple pseudo-pose measurements.","marker":"[44]"},{"why":"Provides the discrete-time EKF prediction and update equations on Lie groups that the paper adapts to its product manifolds.","marker":"[70]"},{"why":"Establishes the exponential-coordinate uncertainty representation for poses on SE(n) that underlies all of the paper's stochastic operations.","marker":"[26]"},{"why":"Characterizes the uncertainty of jointly distributed correlated pose pairs in the Lie algebra, the basis for the correlated composition and difference operations.","marker":"[29]"},{"why":"Provides the V-map, adjoint, right Jacobian, and BCH identities used throughout the derivations.","marker":"[23]"},{"why":"Introduces the kappa-direct isometries group SE_kappa(3), the family from which G_X and G_Z are built.","marker":"[66]"},{"why":"Supplies the continuous-discrete EKF-on-Lie-groups formulation whose process and measurement models the paper adopts.","marker":"[69]"},{"why":"Defines the Mahalanobis distance that the fault-detection module uses to gate updates.","marker":"[72]"},{"why":"Demonstrates Mahalanobis-distance-based anomaly diagnosis in mobile robots, motivating the fault-detection design.","marker":"[73]"},{"why":"Establishes fiducial markers as a simple, reliable, cost-effective source of relative pose estimates for the pseudo-pose updates.","marker":"[56]"}],"fun_headline_variants":["Lie-group filter keeps robot swarms on track when sensors fail","Fault-tolerant multi-robot localization on matrix Lie groups","Matrix Lie groups enable robust multi-robot position fusion","Robot teams beat odometry failure with Lie-group localization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the filter equations are dimensionally consistent on the stated Lie groups: the prediction group $\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$ has 15 degrees of freedom, while the noise covariance $Q$, the adjoint, and the right Jacobian used in prediction are written as $18\\times18$, and the measurement group $\\mathbb{SE}(3)\\times\\mathbb{R}^3$ has 9 degrees of freedom while the measurement noise matrices are $12\\times12$; without an unstated correction, the equations are not well-defined as printed.","fun_headline_variants_meta":{"raw":{"variants":["Lie-group filter keeps robot swarms on track when sensors fail","Fault-tolerant multi-robot localization on matrix Lie groups","Matrix Lie groups enable robust multi-robot position fusion","Robot teams beat odometry failure with Lie-group localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2910,"prompt_tokens":1024,"completion_tokens":1886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":1820}},"tokens_in":640,"tokens_out":1886,"duration_ms":12543,"temperature":1.0,"reasoning_tokens":1820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:35:02.905067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the intrinsic dimension of $G_X=\\mathbb{SE}_2(3)\\times\\mathbb{R}^3\\times\\mathbb{R}^3$ and compare it with the dimension of $Q$ in Eq. (52): if the filter uses the 18-dimensional matrix in the 15-dimensional covariance update, the predicted covariance is mathematically undefined on the stated state manifold, so the method cannot be implemented exactly as written. A separate behavioral falsifier: drive a follower through a sharp turn while its marker is fully occluded for a known interval; if the gated filter's error is not significantly smaller than the ungated filter's error, the fault-detection claim fails.","supporting_citations":[{"cited_title":"Consistent fusion of correlated pose estimates on matrix lie groups,","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form fusion of correlated pose estimates and the cross-covariance recursion that Section III-E uses to combine a follower's multiple pseudo-pose measurements."},{"cited_title":"Discrete extended kalman filter on lie groups,","cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time EKF prediction and update equations on Lie groups that the paper adapts to its product manifolds."},{"cited_title":"Associating uncertainty with three- dimensional poses for use in estimation problems,","cited_arxiv_id":null,"evidence_quote":"Establishes the exponential-coordinate uncertainty representation for poses on SE(n) that underlies all of the paper's stochastic operations."},{"cited_title":"Characterizing the uncertainty of jointly distributed poses in the lie algebra,","cited_arxiv_id":null,"evidence_quote":"Characterizes the uncertainty of jointly distributed correlated pose pairs in the Lie algebra, the basis for the correlated composition and difference operations."},{"cited_title":"Continuous- discrete extended kalman filter on matrix lie groups using concentrated gaussian distributions,","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous-discrete EKF-on-Lie-groups formulation whose process and measurement models the paper adopts."},{"cited_title":"On tests and measures of group divergence,","cited_arxiv_id":null,"evidence_quote":"Defines the Mahalanobis distance that the fault-detection module uses to gate updates."},{"cited_title":"Detecting anomalies in un- manned vehicles using the mahalanobis distance,","cited_arxiv_id":null,"evidence_quote":"Demonstrates Mahalanobis-distance-based anomaly diagnosis in mobile robots, motivating the fault-detection design."},{"cited_title":"Fiducial markers for pose estimation: Overview, applica- tions and experimental comparison of the artag, apriltag, aruco and stag markers,","cited_arxiv_id":null,"evidence_quote":"Establishes fiducial markers as a simple, reliable, cost-effective source of relative pose estimates for the pseudo-pose updates."}],"review_version":1}