{"id":"b99424e5-9127-4c94-97c0-44beb52445fa","arxiv_id":"2506.10850","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An InEKF framework is introduced that integrates partial orientation measurements (roll/pitch from horizon, yaw from dual-antenna GPS) via a planar-frame projection and infinite covariance, and is validated in marine simulation.","lead":"This paper shows how to feed only partial orientation information, like roll and pitch from a boat's horizon view, into an invariant EKF for autonomous surface vessels. It combines this with GPS heading and shows in simulation that the filter stays accurate and converges quickly even from poor initial guesses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RollPitchProjection is not a group homomorphism; H=I is unsupported, and under large initial yaw error the ignored yaw residual couples into roll/pitch via non-commuting rotations, biasing updates exactly in the convergence experiment.","rationale":"The reader's weakest assumption correctly identifies the semi-planar linearization H=I as the load-bearing point. My analysis sharpens the mechanism: the problem is not only the magnitude of roll and pitch themselves, but the way a large yaw error—explicitly present in the convergence experiment's 60° initial orientation uncertainty—gets conjugated through the roll/pitch factor and contaminates the first two innovation components. The paper's infinite-covariance construction only zeros the yaw channel of the innovation; it cannot remove cross-coupling into roll/pitch that arises from the non-commutativity of rotations. Thus the central claim that partial-orientation integration 'preserves the InEKF's superior convergence properties' is not supported by the derivation as written, and the provided simulation evidence is insufficient because no code, data, or Monte Carlo seeds are released to verify that the effect is benign. This is a serious but not fatal gap: a corrected linearization or a numerical demonstration that the cross-coupling is negligible in the tested envelope would restore the claim. Since the reader already returned CONDITIONAL, my read does not change the verdict; it adds a more precise condition and a concrete falsification check.","tokens_in":11877,"tokens_out":10912,"duration_ms":131897,"concrete_test":"Run a numerical Jacobian audit of the exact innovation for the regime in Table II. For a grid of true attitudes with roll/pitch in {0°, ±5°, ±15°, ±30°} and yaw errors δψ in {0°, 10°, 60°}, set R_t = Rz(ψ)Rp and R_hat = R_t exp(ξ^) for small roll/pitch perturbations ξ1, ξ2 and a yaw perturbation ξ3 = δψ. Compute η = log(P(R_t)^{-1} P(R_hat))^vee directly and compare ∂η1,2/∂ξ3 with zero. If for roll/pitch ≤ 5° and δψ = 60° the cross-coupling exceeds roughly 0.05 rad, then H=I fails in the paper's own convergence scenario, and the infinite-covariance update cannot fully discard the yaw error. Optionally, re-run the Fig. 4 experiment with an EKF that linearizes the exact P at each step and release the Monte Carlo data; if that filter materially differs, the claimed convergence property is an artifact of the asserted H=I.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §III.C.1, the paper defines P(R)=Rz(ψ(R))^T R and then asserts that the innovation for a roll/pitch measurement has the same linearization H=[I3 0] as a full orientation measurement. This is the load-bearing step. Let R_hat = R_t exp(ξ^) be the left-invariant rotation error and let Rp = Rz(ψ_t)^T R_t be the true roll/pitch factor. A direct computation gives z^{-1} ẑ = P(R_t)^{-1} P(R_hat) = exp(-ξ^) · Rp^{-1} Rz(δψ) Rp, where δψ is the yaw error. The middle factor is not the identity unless Rp = I or δψ = 0; because Rp contains roll/pitch, the conjugated world-z rotation has roll/pitch components of order ||(φ,θ)|| · |δψ|. Assigning infinite covariance to the yaw component of the innovation removes only the third component; it does not remove this cross-coupled contamination from the first two components. In the convergence experiment (Fig. 4, Table II), initial orientation error is 60° while roll/pitch are around ±5°, giving a bias of roughly 5° in the roll/pitch updates—larger than the 2° measurement noise. The manuscript's own caveat that 'more thorough analysis is needed' concedes the point, but the abstract and conclusion still claim that partial-orientation integration preserves the InEKF's superior convergence properties.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Left-Invariant Extended Kalman Filter (InEKF) on SE2(3) for autonomous surface vessels that fuses partial orientation measurements: roll and pitch from a monocular horizon camera and yaw from dual-antenna GPS. The authors define a planar-frame projection P(R) that strips yaw from a full rotation, treat the resulting roll/pitch measurement as a full SO(3) update whose yaw component is given infinite covariance, and derive an analytical infinite-covariance update via the Woodbury identity. They compare the filter in the VRX simulation against an InEKF using lower-rate full orientation updates and against a multiplicative EKF, reporting comparable steady-state accuracy and faster convergence of the InEKF under large initial errors.","tokens_in":12228,"tokens_out":10411,"duration_ms":93475,"significance":"If the theoretical derivation were sound, the paper would fill a genuine gap: integrating partial orientation measurements into the invariant filtering framework is practically relevant for open-ocean ASVs, and the proposed framework is clearly motivated. The paper also provides a modular pipeline (horizon line detection, roll/pitch extraction) and a simulation-based comparison with two baselines, which is a good empirical starting point. However, the two load-bearing theoretical steps—the Jacobian of the nonlinear projection P(R) and the infinite-covariance matrix identity—are not established; one is asserted and the other is mathematically incorrect as stated. The empirical results cannot compensate for these gaps because the convergence experiment is exactly the regime where the unmodeled cross-coupling is largest. With corrected derivations, the approach may well be salvaged, but as it stands the paper's central claims are not supported by the presented analysis.","major_comments":[{"comment":"The claim that the RollPitchProjection P(R) has the same measurement Jacobian H=[I3 0] as a full orientation measurement is not derived and is not correct in general. Since P is not a group homomorphism, the left-invariant error structure of Section III.A does not carry over. For R_hat = R_t exp(ξ^) and Rp = Rz(ψ_t)^T R_t, the innovation satisfies z^{-1}ẑ = exp(-ξ^) · Rp^{-1} Rz(δψ) Rp (with δψ the yaw error), so the conjugated yaw error contaminates the roll/pitch components at order ||(φ,θ)||·|δψ|. Under the conditions of Fig. 4 and Table II (60° initial orientation error, ±5° true roll/pitch, 2° measurement noise), this contamination exceeds the measurement noise and introduces a bias in the updates that are claimed to provide fast convergence. The manuscript's own caveat that 'a more thorough analysis is needed' highlights the gap, yet the abstract and conclusion claim unqualified preservation of the InEKF's convergence properties. Please provide the true Jacobian of P(R) (including the yaw-error coupling) and quantify the resulting bias, or restrict the claims to the validated small-roll/pitch regime and report bias and convergence results there.","section":"Section III.C.1"},{"comment":"The Woodbury identity is misstated. For non-commuting matrices A and B, (A+B)^{-1} is not A^{-1}-A^{-1}(AB^{-1}+I)^{-1}; the standard identity is (A+UBV)^{-1}=A^{-1}-A^{-1}U(B^{-1}+VA^{-1}U)^{-1}VA^{-1}. As a result, the limiting expression for S^{-1} in Eq. (14) does not follow. For example, with R = Rx(π/2) and Σ̃ = I, Eq. (14) gives S^{-1}=diag(1/2,0,1/2), whereas the correct limit is diag(1/2,1,0). Because this formula is the mechanism that discards the yaw information in the roll/pitch update, the error is load-bearing; please re-derive S^{-1} using the standard identity (e.g., by writing S = (Σ̃+R^T M_φθ R) + L (R^T e3)(R^T e3)^T and taking the limit) or state that the implementation uses a finite, large covariance instead of the analytical formula.","section":"Section III.C.3, Eq. (14)"},{"comment":"The convergence claim attributed to the partial-orientation framework is not clearly supported by the experiments. In the with-horizon condition (6 Hz roll/pitch), the text states that the InEKF is only 'slightly faster' than the MEKF; the dramatic MEKF divergence that motivates the conclusion occurs in the second condition, where the proposed roll/pitch horizon measurements are removed and the filters use only heading and GPS updates. Thus the conclusion that 'our integration of partial orientation measurements, such as roll and pitch, or yaw alone, preserves the InEKF's superior convergence properties' conflates the baseline InEKF behavior with the effect of the proposed roll/pitch measurement model. Please report separate convergence statistics for the with-horizon condition (e.g., time to reach a threshold error, divergence counts, or a convergence-rate curve) and either substantiate or qualify the claim about the specific benefit of the proposed projection-based updates.","section":"Section V.C and Conclusion"}],"minor_comments":[{"comment":"The text in Section V.B says 'A series of 50 Monte Carlo simulations' while the caption of Fig. 3 says 'across 100 Monte Carlo simulations'; please reconcile the number.","section":"Section V.B / Fig. 3"},{"comment":"The camera declination angle is denoted by the same symbol ψ used throughout Section III for yaw, which is confusing; please use a different symbol (e.g., β) for the camera declination.","section":"Section IV.4"},{"comment":"There are two typos in the related-work paragraph on the underwater InEKF: 'Woodsbury identity' should be 'Woodbury identity' and 'meanigful' should be 'meaningful'.","section":"Section II.C"},{"comment":"The sentence 'By left multiplying RW p by Rp R' appears incomplete or mislabeled; the intended rotation composition (e.g., R_b^p = Ry(θ)Rx(φ) applied to Rz(ψ)) should be written explicitly to avoid ambiguity.","section":"Section III.C.1"},{"comment":"The figure caption says 'The boxplots illustrate the distribution of errors across 100 Monte Carlo simulations,' but the vertical axis and the text in Section V.B describe 'Trajectory Mean Abs. Error,' which mixes per-trajectory and per-timestep quantities; please clarify what each boxplot entry represents.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a workshop-style submission (ICRA Workshop on Field Robotics) and the theoretical gaps are substantial for a journal-level claim. The authors' earlier work [20] is used as the basis for the infinite-covariance technique, but the derivation here is not merely terse—it uses an incorrect matrix identity. I would encourage the editor to send the revised manuscript back to the same reviewers, since the core idea is promising and the empirical setup is reproducible, but the theoretical foundation must be corrected before the claims about convergence preservation can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does something genuinely useful: it shows how to feed roll/pitch-only and yaw-only measurements into a left-invariant EKF on SE2(3) using a planar-frame projection plus an analytical infinite-covariance update. The writing is clear, the figures help, and the authors are honest about the semi-planar assumption. That alone makes it worth a look for anyone working on ASV or marine state estimation.\n\nThe main soft spot is real and load-bearing. In Section III.C.1, they define P(R) and then assert that the innovation linearizes as H = I, same as a full orientation measurement. That is not derived, and it is not a group homomorphism. The stress-test note works out the actual expression: the yaw residual δψ gets conjugated by the roll/pitch factor, so the first two components of the innovation contain contamination of order (roll,pitch)·δψ. With ±5° roll/pitch and a 60° initial yaw error, that is about 5° of bias—larger than the 2° measurement noise. The authors themselves concede \"more thorough analysis is needed\" in that section, but the abstract and conclusion still claim the InEKF's superior convergence is preserved. The convergence experiment in Fig. 4 is exactly where this matters, so the claim is stronger than the evidence.\n\nThe comparison against the full-orientation InEKF is also weaker than it looks. They downsample the roll/pitch to 1 Hz to match the heading update, discarding the other 29 measurements per second, then show their filter beats that deliberately crippled baseline. The MEKF comparison is fairer, and there the improvement is modest but real.\n\nMinor issues: the Monte Carlo count is 50 in the text and 100 in the figure caption, and no code or data are released. These are fixable.\n\nBottom line: this deserves a serious referee. The core idea is new enough and the application is practical. But the paper needs major revision before acceptance: derive the Jacobian of P(R) or validate the H=I approximation carefully, rerun the convergence experiment with a fair full-orientation baseline, and either temper the abstract or back the claim with a correct derivation. I would want to see that revision, but I would not desk-reject this.","headline":"A useful, clearly written extension of InEKF to partial orientation measurements, but the paper's central convergence claim rests on an unproven linearization step and a favorable baseline comparison.","tokens_in":12679,"tokens_out":2807,"would_cite":false,"duration_ms":28224,"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":"An invariant EKF on SE2(3) can absorb roll, pitch, and yaw measurements as full-orientation updates by projecting the belief into a planar frame and giving unobserved components infinite covariance.","keywords":["invariant extended Kalman filter","partial orientation measurement","autonomous surface vessel","horizon-based attitude estimation","SE2(3) Lie group","semi-planar assumption","infinite covariance","roll-pitch-yaw"],"falsifier":"Run the filter on simulated trajectories whose roll or pitch excursions exceed the validated envelope—for example, wave-driven angles of ±10 or ±30 degrees—and compare the estimation error against ground truth; if the innovation becomes biased or the filter diverges where a full-orientation InEKF does not, the $H = I$ linearization of the RollPitchProjection is the cause. A direct calculation of the neglected second-order rotation terms as a function of roll and pitch magnitude would settle the claim without additional experiments.","tokens_in":11701,"feed_emoji":"⛵","tokens_out":8681,"duration_ms":85510,"temperature":0.7,"pith_summary":"This paper aims to establish that partial orientation measurements—roll and pitch from a horizon-looking monocular camera, plus yaw from dual-antenna GPS—can be fused into a left-invariant extended Kalman filter on the SE2(3) group without destroying the filter's geometric consistency. The proposed method projects the estimated rotation into a planar frame that contains only the measured components, then applies the standard full-orientation update while assigning an analytically infinite covariance to the unmeasured yaw component. The authors argue this matters for open-ocean autonomous surface vessels, where the horizon is the persistent visual reference and full orientation measurements are either unavailable or arrive at too low a rate. In Monte Carlo simulations, the partial-orientation InEKF matches a multiplicative EKF in steady-state accuracy, outperforms an InEKF that reconstructs full orientation at a lower rate, and converges faster and more reliably under large initial state errors.","feed_headline":"Horizon roll/pitch beat low-rate full orientation in an invariant EKF","feed_subtitle":"A planar-frame projection with infinite covariance lets an SE2(3) filter use horizon and dual-antenna GPS data directly.","key_machinery":"The load-bearing object is the planar-frame projection $P(R)$, which maps a full rotation to its X-then-Y roll–pitch component by extracting yaw as $\\psi = \\operatorname{atan2}(R_{1,0}, R_{0,0})$, constructing $R_z(\\psi)$, and left-multiplying its transpose into the belief. Together with the group homomorphism $h: SE2(3) \\to SO(3)$ and the analytic infinite-covariance limit for the unobserved rotation component, this lets a partial measurement reuse the full-orientation innovation $V^l = \\log(z^{-1}\\hat{z})^\\vee$ with $H = I$. The validity of this machinery rests on the semi-planar assumption that roll and pitch stay moderate, so that the non-commutativity of small rotations does not spoil the linearization.","core_discovery":"The central discovery, stated on the paper's own terms, is a measurement-integration pattern for left-invariant filters. A function $P(R)$ called the RollPitchProjection extracts the yaw of the current rotation belief, builds the planar frame $R_z(\\psi)$, and rotates the belief into that frame, leaving a residual rotation that can be handled as a full SO(3) measurement with $H = [\\,I_{3\\times3}\\;\\;0_{3\\times6}\\,]$. Because the yaw component of a roll/pitch measurement would otherwise incorrectly reinforce the filter's current yaw belief, that component is assigned infinite variance and the innovation covariance is evaluated analytically using a matrix-inversion identity, keeping the update well conditioned. The same construction is applied to yaw-only heading measurements by building the planar frame from the estimated heading. The paper reports that the resulting partial-orientation InEKF outperforms an InEKF using lower-frequency full orientation measurements, and that it retains superior convergence speed and reliability compared with a multiplicative EKF under high initial uncertainty.","pith_inferences":["The planar-frame projection pattern likely generalizes to other partial measurements whose unobserved directions form a subgroup of the state group, such as single-axis position updates on SE(3), by zeroing the unobserved directions and taking the same infinite-covariance limit.","A natural stress test is to drive roll or pitch excursions beyond the validated ±5 degree envelope; the paper's suggestion that the method may work up to ±30 degrees is explicitly not fully validated.","Because the camera-height estimate shifts the horizon pixel location by under a pixel over a 1–3 m height range, the pitch measurement may be sensitive to calibration error; a real-world dataset would be needed to confirm the simulation findings.","If the same pattern extends to other Lie groups with a planar subgroup, it could become a general recipe for 'partial invariant measurements' beyond orientation, such as partial velocity or partial position observations."],"forward_implications":["High-frequency roll and pitch from a horizon camera can be used at their native rate rather than being downsampled and stitched into a synthetic full orientation measurement.","A pure yaw or heading measurement can be fused without a magnetometer and without constraining roll and pitch, because the unmeasured components receive infinite covariance.","The filter keeps the InEKF's convergence behavior under large initial state errors, which the paper identifies as its practical advantage over a multiplicative EKF.","The framework applies to any semi-planar vehicle—surface vessels, aircraft in level flight, or wheeled robots on mild terrain—not only to the specific boat simulated here.","The reported simulation results support using the horizon as the primary orientation reference in open-ocean autonomy, where fixed landmarks are absent."],"supporting_citations":[{"why":"Introduces the left-invariant extended Kalman filter whose error dynamics and update structure this paper extends.","marker":"[14]"},{"why":"Establishes the stability and convergence properties of invariant filters that the paper claims to preserve.","marker":"[15]"},{"why":"Supplies the infinite-covariance approach used to ignore unmeasured state components while keeping the update well conditioned.","marker":"[20]"},{"why":"Provides the matrix-inversion identity used to derive the innovation covariance in the infinite-variance limit.","marker":"[21]"},{"why":"Gives the InEKF equations and operators that the paper builds on for its measurement update.","marker":"[22]"},{"why":"Provides the line-segment detector used to extract the horizon roll and pitch measurements.","marker":"[26]"},{"why":"Defines the multiplicative EKF baseline that the paper compares against for convergence and accuracy.","marker":"[28]"}],"fun_headline_variants":["Partial orientation InEKF beats low-rate full orientation for ASVs","Horizon roll/pitch fed into InEKF outperforms full orientation","Yaw-ambiguous roll/pitch integrated into invariant EKF for ASVs","InEKF with horizon-derived roll/pitch beats full orientation","Roll/pitch from horizon improves invariant EKF estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the semi-planar assumption: roll and pitch stay within moderate limits (validated at about ±5 degrees, claimed up to about ±30 degrees without full validation), so that projecting the rotation into a planar frame and linearizing with $H = I$ remains accurate despite the non-commutativity of rotations.","fun_headline_variants_meta":{"raw":{"variants":["Partial orientation InEKF beats low-rate full orientation for ASVs","Horizon roll/pitch fed into InEKF outperforms full orientation","Yaw-ambiguous roll/pitch integrated into invariant EKF for ASVs","InEKF with horizon-derived roll/pitch beats full orientation","Roll/pitch from horizon improves invariant EKF estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000457,"raw_usage":{"total_tokens":2340,"prompt_tokens":1037,"completion_tokens":1303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1213}},"tokens_in":653,"tokens_out":1303,"duration_ms":11589,"temperature":1.0,"reasoning_tokens":1213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:16:25.676431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the filter on simulated trajectories whose roll or pitch excursions exceed the validated envelope—for example, wave-driven angles of ±10 or ±30 degrees—and compare the estimation error against ground truth; if the innovation becomes biased or the filter diverges where a full-orientation InEKF does not, the $H = I$ linearization of the RollPitchProjection is the cause. A direct calculation of the neglected second-order rotation terms as a function of roll and pitch magnitude would settle the claim without additional experiments.","supporting_citations":[{"cited_title":"Left-invariant extended kalman filter and attitude esti- mation,","cited_arxiv_id":null,"evidence_quote":"Introduces the left-invariant extended Kalman filter whose error dynamics and update structure this paper extends."},{"cited_title":"The invariant extended kalman filter as a stable observer,","cited_arxiv_id":null,"evidence_quote":"Establishes the stability and convergence properties of invariant filters that the paper claims to preserve."},{"cited_title":"Invariant extended kalman filtering for underwater navigation,","cited_arxiv_id":null,"evidence_quote":"Supplies the infinite-covariance approach used to ignore unmeasured state components while keeping the update well conditioned."},{"cited_title":"Woodbury, Inverting Modified Matrices, ser","cited_arxiv_id":null,"evidence_quote":"Provides the matrix-inversion identity used to derive the innovation covariance in the infinite-variance limit."},{"cited_title":"An introduction to the in- variant extended kalman filter [lecture notes],","cited_arxiv_id":null,"evidence_quote":"Gives the InEKF equations and operators that the paper builds on for its measurement update."},{"cited_title":"LSD: a line segment detector,","cited_arxiv_id":null,"evidence_quote":"Provides the line-segment detector used to extract the horizon roll and pitch measurements."},{"cited_title":"Generalized multiplicative extended kalman filter for aided attitude and heading reference system,","cited_arxiv_id":null,"evidence_quote":"Defines the multiplicative EKF baseline that the paper compares against for convergence and accuracy."}],"review_version":1}