{"id":"895c3117-30f2-4a51-86c0-4634e30d312f","arxiv_id":"2607.00145","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"IterIEKF applied to landmark-aided 3D inertial navigation outperforms SO(3)-EKF, IterSO(3)-EKF and IEKF by ~80% in position/velocity MAE and yields better NEES consistency in simulation.","lead":"The paper derives and tests an iterated invariant Kalman filter for drone navigation that fuses IMU data with known 3D landmarks. In Monte Carlo simulations it cuts position and velocity error by roughly 80% versus the standard invariant filter while improving consistency.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged continuous-visibility assumption.","rationale":"The paper’s strongest empirical claim is tightly scoped to a controlled simulation with continuous landmark visibility, perfect data association and known map positions. Under those conditions the multi-landmark Gauss-Newton update (Eqs. 28–37) and the SE2(3) right-invariant error dynamics produce the reported accuracy and consistency gains; the mathematics and the Monte-Carlo protocol are mutually consistent. The reader already identified the continuous-visibility assumption as the weakest external condition; that is precisely the point at which the manifold-intersection property (Theorems 1–2 of Goffin et al.) ceases to apply. No deeper internal flaw—incorrect Jacobian, inconsistent covariance update, or numerical instability—emerges from the derivations or the tabulated results. Consequently the CONDITIONAL verdict remains appropriate: the engineering contribution is solid and correctly scoped, yet remains simulation-only and dependent on idealized sensing. The concrete test above simply quantifies how far the reported gains degrade once the idealized sensing assumption is relaxed, without requiring any change to the present verdict.","tokens_in":15595,"tokens_out":524,"duration_ms":5093,"concrete_test":"Re-run the identical N=50 Monte-Carlo suite of Section 9 while randomly dropping each landmark independently with probability p=0.3 at every 1 Hz update (keeping perfect association when visible). If IterIEKF position MAE rises above ≈0.25 m or NEES systematically exceeds 20, the headline 80% gain is regime-specific and the claim should be qualified; otherwise the result is robust to intermittent visibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (≈80% MAE reduction for position/velocity and NEES near 15 under the Monte-Carlo protocol of Section 9) is internally consistent with the derived multi-landmark IterIEKF equations (Eqs. 28–37) and the reported tables/figures. The continuous full-visibility, perfect-association, known-map regime (K=3 always, 1 Hz) is the dominant external assumption that makes the manifold-intersection guarantee operative; the paper already states this setup explicitly and the reader correctly isolates it. No additional load-bearing mathematical inconsistency, hidden linearization error, or numerical artifact that would overturn the reported gains under the stated conditions is apparent. Accelerometer-bias observability limitations are candidly acknowledged (Section 9.3) and do not undercut the position/velocity claims.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper formulates and evaluates the Iterated Invariant Extended Kalman Filter (IterIEKF) for 3D landmark-aided inertial navigation of a MAV. The state is embedded in SE2(3)×R6 with a right-invariant error; IMU strapdown kinematics supply the prediction (Eqs. 11, 16–20), and relative 3D landmark positions supply the measurement model (Eqs. 12–15). The multi-landmark Gauss–Newton update (Eqs. 28–37) is derived so that, in the low-noise regime, the estimate is driven onto the intersection of the observed submanifolds while uncertainty remains in the tangent space. Monte-Carlo trials (N=50) on a EuRoC V2_01_easy trajectory with three always-visible known landmarks show that IterIEKF reduces position and body-velocity MAE by roughly 80% relative to the single-step IEKF (0.096 m vs 0.511 m; 0.095 m/s vs 0.48 m/s) while keeping average NEES near the state dimension 15, outperforming SO(3)-EKF, IterSO(3)-EKF and IEKF in accuracy and consistency. Accelerometer-bias observability limitations are acknowledged.","tokens_in":15826,"tokens_out":1127,"duration_ms":8720,"significance":"If the reported gains hold under the stated conditions, the work supplies a concrete, fully derived multi-landmark IterIEKF for a standard MAV localization setting and demonstrates that combining Lie-group invariance with iterated measurement updates yields large accuracy and consistency improvements over both classical multiplicative EKFs and the single-step IEKF. The complete algorithmic derivation (prediction and multi-landmark Gauss–Newton update), transparent Monte-Carlo protocol on a public trajectory, and joint reporting of MAE and NEES are clear strengths. The continuous full-visibility, perfect-association, known-map regime is a genuine limitation of scope, but it is stated explicitly and does not invalidate the internal comparison. The paper is a solid incremental contribution that makes the IterIEKF usable for landmark-aided inertial navigation.","major_comments":[{"comment":"Section 9 (and the measurement model of Eqs. 12–15) assumes K=3 landmarks remain continuously visible at 1 Hz with perfect data association and known world positions for the entire trajectory. The manifold-intersection guarantee that underpins the IterIEKF update (motivating example of Section 4 and Theorems cited from Goffin et al.) is operative only under this regime. The central claim of an ~80% MAE reduction is therefore demonstrated only for this idealized sensing schedule. The manuscript should either (i) add at least one ablation with intermittent visibility / partial landmark sets, or (ii) state more prominently in the abstract, contributions list and conclusion that the reported gains are conditioned on continuous full visibility of a known map.","section":null},{"comment":"Section 9.3 and Fig. 6 candidly show that accelerometer biases remain poorly observed (relative error persistently above 1) for every filter, including IterIEKF, because appending biases as an R6 factor breaks the strict group-affine property. While this does not overturn the position/velocity claims, the abstract and contribution bullets emphasize overall superiority without mentioning this structural limitation. A short clarifying sentence in the abstract or contributions list would keep the claim proportionate to the evidence.","section":null}],"minor_comments":[{"comment":"Eq. (37) updates the covariance with the gain and Jacobian from the first iteration (K0, H0) rather than the final iteration used for the state. This is a common practical choice but should be stated explicitly so that readers can reproduce the NEES numbers.","section":null},{"comment":"The stopping criterion δ=10−4 and the optional loss-increase check are given only in prose (Section 7.2). Adding them to a short algorithm box or pseudocode would improve reproducibility.","section":null},{"comment":"Fig. 1 and Fig. 2 are helpful geometric illustrations but the caption of Fig. 1 does not define the green/blue/gray objects; a one-line legend would help.","section":null},{"comment":"Table 1 reports relative errors to IEKF; adding absolute MAE for roll/pitch/yaw (or a short attitude MAE column) would make the attitude discussion in Section 9.2 fully quantitative.","section":null},{"comment":"A few typographical inconsistencies appear (e.g., “Coordenação” encoding, occasional missing spaces around citations). A light copy-edit pass would clean them.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a clean application paper that correctly cites the authors’ own prior IterIEKF quadruped work for the iteration scheme and supplies the missing multi-landmark MAV Jacobians. Novelty is incremental but legitimate for a robotics journal. The continuous-visibility assumption is the only material external limitation; once it is flagged more clearly the paper is ready for publication. No integrity or citation-pattern concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper that finally writes down the multi-landmark IterIEKF equations for SE2(3)\times R6 inertial navigation and shows they work. The novelty is domain-specific: first complete derivation and head-to-head Monte-Carlo comparison (N=50, EuRoC V2_01_easy trajectory) against SO(3)-EKF, IterSO(3)-EKF and plain IEKF. The math in Sections 5–8 is fully spelled out, matches the invariant literature, and the Gauss-Newton loop (Eqs. 28–37) is adapted correctly for stacked landmarks. Table 1 and the NEES plots are transparent: position MAE drops from 0.511 m (IEKF) to 0.096 m, velocity from 0.48 m/s to 0.095 m/s, NEES stays near 15. They also candidly flag that accelerometer biases remain poorly observed because the R6 bias embedding breaks group-affinity.\n\nWhat it does well is the controlled experiment design and the geometric motivation (Section 4). The continuous-visibility, perfect-association, known-map regime (K=3 always at 1 Hz) is the main external assumption that makes the manifold-intersection guarantee operative; the paper states it explicitly and the stress-test correctly isolates it. No hidden linearization error or numerical artifact overturns the reported gains under those conditions. Soft spots are real but proportionate: simulation-only, no code release, free noise/P0/δ parameters, and the result is incremental relative to Goffin et al. and the authors’ own quadruped IterIEKF work. Circularity is low—forward simulation against known ground truth.\n\nThis is for people already working with IEKF-style filters on MAVs or VIO who want the concrete multi-landmark update and a clear accuracy/consistency baseline. It is not a new technology class, but it is formally grounded enough and evidentially sharp enough that a serious editor should send it to referees rather than desk-reject. I would cite the equations and the MAE numbers if I am writing on invariant VIO this year, and I would bring the paper to reading group for the derivation and the bias discussion.","headline":"Solid first multi-landmark IterIEKF derivation for SE2(3) MAV navigation; clear ~80% MAE gains over IEKF in controlled sim, but simulation-only under continuous perfect visibility.","tokens_in":16431,"tokens_out":550,"would_cite":true,"duration_ms":5056,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"An iterated invariant Kalman filter cuts position and velocity error by about 80% for 3D landmark-aided inertial navigation.","keywords":["Iterated Invariant EKF","landmark-aided inertial navigation","visual-inertial odometry","false observability","Lie-group filtering","SE2(3)","Monte Carlo consistency","MAV localization"],"falsifier":"Re-run the same Monte Carlo suite but randomly occlude one or more landmarks for multi-second intervals (or inject occasional association swaps) and check whether the IterIEKF still retains an ~80% MAE advantage and NEES near 15 over the plain IEKF.","tokens_in":16520,"feed_emoji":"🛸","tokens_out":698,"duration_ms":5805,"temperature":0.7,"pith_summary":"This paper claims that the Iterated Invariant Extended Kalman Filter (IterIEKF), previously used on legged robots, works better than standard filters when a flying robot estimates its 3D pose and velocity from an IMU and known landmarks. Classical SO(3)-based EKFs become overconfident along directions that are actually unobservable (false observability), which hurts accuracy. The plain Invariant EKF fixes that by using Lie-group error dynamics, but its single linearization of the landmark measurement can still leave residual error. The IterIEKF repeatedly refines the update so that, in the low-noise regime, the estimate lands on the intersection of the observed manifolds while uncertainty stays in the tangent space. On Monte Carlo trials built from a EuRoC MAV trajectory with three always-visible landmarks, the authors report roughly 80% lower position and body-velocity mean absolute error than the plain IEKF, better consistency (NEES near the state dimension), and faster attitude convergence.","feed_headline":"Iterated invariant filter cuts drone pose error by ~80%","feed_subtitle":"On EuRoC-style trials with known landmarks, IterIEKF beats IEKF and classical SO(3) EKFs on accuracy and consistency.","key_machinery":"The IterIEKF Gauss-Newton loop on the right-invariant error: each iteration re-linearizes the multi-landmark innovation on SE2(3)\times R^6 so the estimate is driven onto the intersection of the observed submanifolds while the covariance remains in the tangent space.","core_discovery":"Formulated for the first time for landmark-aided 3D inertial navigation, the IterIEKF outperforms SO(3)-EKF, IterSO(3)-EKF and the ordinary IEKF on accuracy and consistency. In 50 Monte Carlo runs on a EuRoC-derived trajectory it reduces position MAE from 0.511 m to 0.096 m and velocity MAE from 0.48 m/s to 0.095 m/s relative to the IEKF, while keeping average NEES near 15.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["IterIEKF cuts landmark-aided 3D pose MAE ~80% versus IEKF","First IterIEKF for 3D landmark inertial nav beats SO(3)-EKF","IterIEKF trims position error from 0.51 m to 0.10 m on EuRoC runs","Landmark IterIEKF holds NEES near 15 while lifting accuracy","Iterated invariant filter tops IEKF and SO(3) EKFs in consistency"],"cache_read_input_tokens":128,"weakest_assumption_plain":"All three landmarks stay continuously visible at 1 Hz with perfect data association and known map positions for the whole trajectory; if any landmark drops out or is mis-associated, the manifold-intersection guarantee and the reported error cuts no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["IterIEKF cuts landmark-aided 3D pose MAE ~80% versus IEKF","First IterIEKF for 3D landmark inertial nav beats SO(3)-EKF","IterIEKF trims position error from 0.51 m to 0.10 m on EuRoC runs","Landmark IterIEKF holds NEES near 15 while lifting accuracy","Iterated invariant filter tops IEKF and SO(3) EKFs in consistency"]},"model":"grok-4.5","effort":"low","cost_usd":0.008458,"raw_usage":{"total_tokens":2011,"prompt_tokens":806,"num_sources_used":0,"completion_tokens":117,"cost_in_usd_ticks":84580000,"prompt_tokens_details":{"text_tokens":806,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1088,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":806,"tokens_out":117,"duration_ms":8051,"temperature":1.0,"reasoning_tokens":1088,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T07:09:13.865915+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same Monte Carlo suite but randomly occlude one or more landmarks for multi-second intervals (or inject occasional association swaps) and check whether the IterIEKF still retains an ~80% MAE advantage and NEES near 15 over the plain IEKF.","supporting_citations":[],"review_version":2}