{"id":"3ad06d79-a84c-4acb-acdb-385cb5c5c4a7","arxiv_id":"2411.13310","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A decoupled moving horizon estimator for SLAM is shown to have robustly stable ego-state estimates and bounded landmark errors, with landmark updates restricted to informative horizons.","lead":"This paper analyzes a decoupled moving horizon estimator for SLAM, where the robot pose and landmark positions are estimated separately and each landmark is updated only when its measurements are informative. It proves a robust error bound for the ego-state and for landmark positions under detectability assumptions, with simulations on a planar robot showing much lower computation time than a coupled baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Prop. 1 relies on true trajectory feasibility for landmark NLP (7), but (7) predicts every horizon output with the single final ego-state estimate hat(x)^s_k rather than the per-time poses hat(x)^s_j, so for a moving robot the feasibility argument fails.","rationale":"The reader's weakest assumption concerned Assumption 1 (ego-state detectability) and the fact that condition (5) is not checked in the simulations. Both are real issues, but the more load-bearing problem is internal to the proof of Proposition 1: the landmark NLP (7) as written predicts all historical outputs using the terminal ego-state estimate hat(x)^s_k, whereas the true outputs depend on the per-time poses x^s_j. This makes the true trajectory infeasible for (7) in any moving-robot scenario, so the optimality argument that converts the informativity condition into the recursive bound (11) and then into (8) does not go through. This is not merely a verification problem; it is a mismatch between the gating condition (5), which uses a sequence of past ego-state estimates, and the optimization problem it gates, which uses only the current estimate. The correct fix is to write (7) with hat(x)^s_j and redo the proof, then rerun the simulations under that corrected model. Because the intended architecture is otherwise coherent and the fix appears local, a conditional accept with mandatory correction is the appropriate outcome rather than a rejection of the overall research direction.","tokens_in":15017,"tokens_out":10312,"duration_ms":115271,"concrete_test":"In the planar setup of Sec. VI, pick a time k where a landmark is visible over the horizon and evaluate the constraint residual of the true trajectory in (7) as written: r_j = y^{e,l}_j - h^e(hat(x)^s_k, x^{e,l}, xi^{e,l}_j) for j in [k-M, k-1]. If any r_j is nonzero for a moving robot, the 'true trajectory is feasible' step after Eq. (10) is invalid. Then re-run the simulations with (7) corrected to use hat(x)^s_j as the pose at each j; if the landmark error bound (8) is reproduced, the issue is a fixable typo, and if not, the proposed bound is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing step in Proposition 1 is the claim, after Eq. (10), that 'the true system trajectory is feasible in the MHE optimization problem', which justifies J(optimal) <= J(true) and the recursive landmark bound (11)/(8). This feasibility claim is false for NLP (7) as written: its constraint is yhat^{e,l}_{j|k} = a^{e,l}_k h^e(hat(x)^s_k, hat(x)^{e,l}_{k|k}, hat(xi)^{e,l}_{j|k}) for all j in I[k-N^{e,l}_k, k-1], i.e. every past measurement in the horizon is predicted using the single ego-state estimate at the horizon end, hat(x)^s_k. The actual data are generated as y^{e,l}_j = a^{e,l}_j h^e(x^s_j, x^{e,l}, xi^{e,l}_j), with the robot pose at time j. Unless the robot is stationary over the landmark horizon, the true trajectory does not satisfy the constraints of (7), so the feasibility inequality used in the proof does not hold and the derivation of (11)/(8) is not valid. This also explains a mismatch with the informativity condition (5), which is stated over the sequence {hat(x)^s_{k-M},...,hat(x)^s_{k-1}}, while the update (7) uses only hat(x)^s_k. Fixing Assumption 1 or verifying condition (5) does not repair this gap: even under perfect ego-state estimates, the landmark MHE must use the pose at each measurement time, not the final pose, for the true trajectory to be feasible. The manuscript should either correct (7) to use hat(x)^s_j inside the horizon and re-derive the proof, or present a different argument that does not rely on feasibility of the true trajectory.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a decoupled moving horizon estimation (MHE) scheme for SLAM. The robot pose is estimated from ego-sensors alone by solving the MHE problem (6), while each landmark position is updated by a separate, smaller MHE problem (7) only when a horizon-level informativity condition (5) is satisfied. Proposition 1 claims that, under an i-IOSS assumption on the ego-subsystem and a Lipschitz condition on the landmark measurement model, the pose estimator is robustly globally exponentially stable and each landmark error satisfies the bound (8), which decays with the number of informative landmark horizons. The paper also discusses how the assumptions relate to bearing-only and range sensors, proposes a recursive least squares simplification for range measurements, and reports simulations on two planar SLAM scenarios.","tokens_in":15416,"tokens_out":12419,"duration_ms":135552,"significance":"If the main result were correct, this would be a useful contribution: it would give robust MHE-style error bounds for SLAM under intermittent landmark visibility, with a naturally parallelizable landmark update structure and no fitted constants. The paper is also transparent about the restrictiveness of Assumption 1, explicitly noting in Section V-B that standard bearing-only or range-only SLAM without anchors is not observable and hence does not satisfy the assumption. However, the central proof currently breaks at a load-bearing feasibility step for the landmark MHE, so the claimed guarantees are not established as written. The simulation section does not verify the key informativity condition used by the algorithm, and the reported horizon and decay parameters are not reconciled with the sufficient condition of the cited stability theorem.","major_comments":[{"comment":"The proof step 'Because of the feasibility of the true system trajectory in the MHE optimization problem' after Eq. (10) is not valid for the NLP in (7). The constraints in (7) require \\hat y^{e,l}_{j|k} = a^{e,l}_k h^e(\\hat x^s_k, \\hat x^{e,l}_{k|k}, \\hat \\xi^{e,l}_{j|k}) for every j in the horizon, whereas the measurements are generated as y^{e,l}_j = a^{e,l}_j h^e(x^s_j, x^{e,l}, \\xi^{e,l}_j). Unless the robot is stationary over the horizon, the true trajectory (x^{e,l}, \\xi^{e,l}_j, y^{e,l}_j) does not satisfy the constraints, so the inequality J(optimal) ≤ J(true) does not follow. This inequality is exactly what produces the recursive bound (11) and hence the advertised bound (8), so Proposition 1(2) is not established as written. The landmark MHE should use per-time ego-state estimates \\hat x^s_{j|k} inside the horizon, with the detectability condition (5) aligned to the same per-time poses, or else the proof must be replaced by an argument that does not rely on feasibility of the true trajectory.","section":"IV, Eq. (7) and proof of Proposition 1"},{"comment":"The numerical validation does not verify the theoretical conditions under which Proposition 1 is stated. Algorithm 1 updates a landmark only when condition (5) holds, but the simulation section describes updates based on visibility and never reports that (5) was checked, so Figure 3 cannot be read as validating the theorem. In addition, Assumption 1 is only argued through observability in Section V-B; the paper does not exhibit an i-IOSS Lyapunov function or the matrices from Definition 1 for the simulated model. The reported parameters η = 0.99 and horizon length 20 are also difficult to reconcile with the sufficient condition 4 η^{N_s} λmax(\\bar U, U) < 1 from Theorem 1, since Definition 1 implies \\bar U ⪰ U and hence λmax(\\bar U, U) ≥ 1, which would require a horizon much longer than 20 for this decay rate. Either the simulations should check and report all conditions of Proposition 1, or the claim that they validate the theory should be substantially softened.","section":"VI and Algorithm 1"}],"minor_comments":[{"comment":"The visibility matrix inside the sum should be a^{e,l}_j, not a^{e,l}_k; as written, only the visibility at the current time multiplies the output terms from all past times in the horizon, which cannot be the intended condition.","section":"Definition 3, Eq. (5)"},{"comment":"The horizon parameter and the prior landmark estimate are inconsistent: Algorithm 1 line 11 passes N^{e,l} and \\hat x^{e,l}_{k-N^s_k}, while the cost in (7) uses M and the proof uses \\hat x^{e,l}_{k-M}. These quantities should be defined consistently.","section":"Algorithm 1 and Eq. (7)"},{"comment":"The displayed measurement model is ambiguous for the two sensor types: for the bearing-only case, the first block R(-θ_k)(x^{e,l}-p_k) has magnitude equal to the distance rather than a unit vector, contrary to Figure 1; for the range case, α = 1 makes the second block a constant rather than a range measurement. Please reconcile Eq. (14) with the text and Figure 1.","section":"Section V-A, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The feasibility gap in the proof of Proposition 1 is decisive: the landmark NLP (7) as written does not admit the true trajectory when the robot moves, so the main error bound is currently unproved. I still recommend major revision rather than rejection because a corrected per-pose landmark formulation plus a reworked bounding argument may be within the scope of the paper. The simulations should also be rerun or reported with all assumptions and the informativity condition actually verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the architecture, not for the proof. The paper proposes a decoupled MHE scheme for SLAM where the ego-state is estimated from ego-sensors alone and each landmark is updated only when a detectability condition holds. That is a useful structure: it allows parallel landmark updates and gives a principled reason to skip uninformative measurements. The informativity-index dependent bound (8) is a nice conceptual target, and the simulation section does show a large computational advantage over the coupled MHE baseline. The discussion of assumptions in Section V is honest: they admit Assumption 1 fails for bearing-only or range-only SLAM without anchoring, and they point to literature for verifying it.\n\nThe problem is the proof of Proposition 1. After Eq. (10) they claim 'the true system trajectory is feasible in the MHE optimization problem' and use that to get J(opt) <= J(true). For the landmark NLP (7), the constraint predicts every measurement in the horizon using the single final ego-state estimate \\hat{x}^s_k (and the current data association A_k). The true measurements are generated with the robot pose at each time j. Unless the robot is stationary over the horizon, the true trajectory does not satisfy (7), so the inequality and the recursive bound (11)/(8) do not go through. This is not cured by verifying Assumption 1 or checking condition (5); the landmark MHE itself would need to use per-time poses inside the horizon, or the proof needs a different argument that does not rely on feasibility.\n\nThe reader's concerns are also on point but secondary: Assumption 1 is load-bearing and not verified for the simulated system, condition (5) is not actually checked in the simulations, and the derivation of (8) defers to [13, Cor. 1]. None of those would be fatal if the feasibility claim were repaired, but as it stands the central stability result is unproven.\n\nThe honest assessment: this is a legitimate extension of MHE stability theory to a decoupled SLAM back-end, and the computational results are genuine, but the paper currently asks the reader to accept a theorem whose key step is false. The fix is identifiable—correct (7) to use \\hat{x}^s_j in the horizon and re-derive, or prove a looser bound without feasibility. That is refereeable; I would not desk-reject it. I would send it to review and ask for a corrected proof or a clear statement of what the theorem actually requires.","headline":"The decoupled MHE-SLAM architecture is sensible and the simulations show a real speedup, but the main proof currently rests on a feasibility claim that is false as written.","tokens_in":15994,"tokens_out":2314,"would_cite":false,"duration_ms":23562,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B53","93C55","93E10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Decoupled MHE for SLAM yields provable landmark error bounds under intermittent visibility.","keywords":["moving horizon estimation","SLAM","robust estimation","incremental input/output-to-state stability","landmark detectability","intermittent visibility","recursive least squares"],"falsifier":"Construct a simulation with a landmark that is visibly informative according to condition (5) but whose MHE update (7) produces an error that grows rather than decays across successive informative horizons, or run the scheme on a bearing-only SLAM setup without any anchoring measurement and observe that the ego-state error does not converge, contradicting Proposition 1.","tokens_in":14795,"feed_emoji":"🤖","tokens_out":1427,"duration_ms":16173,"temperature":0.7,"pith_summary":"This paper proposes a moving horizon estimation (MHE) scheme for simultaneous localization and mapping (SLAM) that splits the estimation problem into an ego-state update and separate landmark updates, and it aims to prove robust stability and bounded estimation errors even when landmarks are only intermittently visible. The central claim is that if the ego-state subsystem is detectable and the landmark measurement model is Lipschitz continuous, then the ego-state estimate is robustly globally exponentially stable and each landmark error bound decreases exponentially with the number of informative measurement horizons. The authors argue that this addresses the persistent-excitation requirement that plagues coupled SLAM estimators, and they support the claim with simulations on circular and corridor scenarios.","feed_headline":"Decoupled MHE gives SLAM provable error bounds","feed_subtitle":"When landmarks wink in and out of view, per-landmark updates stay bounded and ego-state estimates stay stable.","key_machinery":"The central machinery is the decoupling of MHE into an ego-state estimator (NLP (6)) and per-landmark estimators (NLP (7)), gated by the landmark detectability condition (5) derived from the exponential i-IOSS decrease condition applied to a static landmark model. The informativity index $m_l^k$ counts how many disjoint informative horizons for landmark $l$ have occurred up to time $k$, and it is the quantity that controls the exponential decay in the landmark error bound (8).","core_discovery":"The paper establishes Proposition 1: under Assumptions 1 and 2, the decoupled MHE scheme in Algorithm 1 with NLPs (6) and (7) yields (1) an ego-state estimate that is robustly globally exponentially stable (RGES), and (2) for each landmark, an estimation error bound of the form (8) that decays exponentially with the informativity index $m_l^k$, i.e., with the number of informative horizons for that landmark. The proof works by applying the i-IOSS-based MHE stability theorem to the ego-state subsystem and then, for each landmark, chaining the landmark detectability condition (5) over successive informative horizons, with the landmark estimate held fixed when no informative horizon occurs. The authors also show that when landmarks are periodically detectable the informativity index grows linearly in time, and they provide a streamlined recursive least squares formulation for the range measurement model.","pith_inferences":["A direct testable extension would be to report, in simulations, whether condition (5) was explicitly checked before each landmark update; the paper's simulations do not state that this check was performed, so the empirical validation of the gating mechanism remains indirect.","The bound (8) suggests a natural active-SLAM objective: plan trajectories that maximize the minimum informativity index across landmarks, turning the theoretical error bound into a planning cost and potentially improving worst-case mapping accuracy.","The decoupling could be applied to other joint state-and-parameter estimation problems beyond SLAM, wherever the parameters appear only in the measurement equation and the state subsystem is detectable on its own.","The authors' reliance on observability-based sufficient conditions for Assumption 1 hints that the scheme's practical reach depends on external anchoring measurements; a systematic way to verify i-IOSS for common SLAM models would broaden the class of systems covered."],"forward_implications":["If the decoupled MHE scheme is correct, SLAM can be performed with provable robust error bounds even when landmarks enter and leave the field of view, without requiring persistent excitation of every landmark.","Landmark updates become parallelizable, since each landmark's NLP (7) is independent once the ego-state estimate is fixed, reducing per-step computation compared to a coupled augmented-state MHE.","The ego-state estimation error, being RGES, provides a bounded uncertainty that is propagated into the landmark error bound through the Lipschitz assumption, so the landmark guarantees hold as long as the ego-state detector is sufficiently informative.","For range-based sensors, the landmark update reduces to recursive least squares, giving the same stability properties with significantly lower computational cost than a full MHE update.","When each landmark is periodically detectable, the informativity index grows linearly with time, making the landmark estimate RGES as well rather than merely bounded."],"supporting_citations":[{"why":"Supplies the MHE stability theorem (Corollary 1) that the paper uses to prove RGES of the ego-state estimate from Assumption 1.","marker":"[13]"},{"why":"Provides the decoupled block-coordinate MHE-SLAM structure that the paper adapts and analyzes.","marker":"[11]"},{"why":"Establishes that detectability of the augmented state requires parameter persistent excitation, motivating the decoupled scheme.","marker":"[18]"},{"why":"Shows practical stability of MHE without persistent excitation, a foundation the paper builds on for the landmark error analysis.","marker":"[20]"},{"why":"Defines i-IOSS detectability and the Lyapunov-function condition that Assumption 1 relies on.","marker":"[22]"},{"why":"Shows that without anchoring or an external position sensor even linear SLAM is unobservable, which the paper cites to justify Assumption 1.","marker":"[26]"},{"why":"Shows inherent unobservability of 2D range-based SLAM and how anchoring ensures observability, supporting verification of Assumption 1.","marker":"[29]"},{"why":"Provides the generalized recursive least squares stability result used for the streamlined range-measurement scheme.","marker":"[30]"}],"fun_headline_variants":["SLAM with decoupled MHE: provable bounds under intermittent landmarks","Decoupled MHE bounds SLAM errors even when landmarks vanish","Provable error bounds for SLAM via decoupled MHE","Robust SLAM: decoupled MHE ensures bounded landmark errors","MHE with per-landmark updates gives SLAM stability bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ego-state must be detectable from ego-sensors alone (Assumption 1); without an external position measurement or anchoring landmarks, even linear SLAM is unobservable, so the guarantee fails for many bearing-only or range-only setups.","fun_headline_variants_meta":{"raw":{"variants":["SLAM with decoupled MHE: provable bounds under intermittent landmarks","Decoupled MHE bounds SLAM errors even when landmarks vanish","Provable error bounds for SLAM via decoupled MHE","Robust SLAM: decoupled MHE ensures bounded landmark errors","MHE with per-landmark updates gives SLAM stability bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3045,"prompt_tokens":872,"completion_tokens":2173,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":2084}},"tokens_in":488,"tokens_out":2173,"duration_ms":15290,"temperature":1.0,"reasoning_tokens":2084,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:34:45.484188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a simulation with a landmark that is visibly informative according to condition (5) but whose MHE update (7) produces an error that grows rather than decays across successive informative horizons, or run the scheme on a bearing-only SLAM setup without any anchoring measurement and observe that the ego-state error does not converge, contradicting Proposition 1.","supporting_citations":[{"cited_title":"A lyapunov function for robust stability of moving horizon estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the MHE stability theorem (Corollary 1) that the paper uses to prove RGES of the ego-state estimate from Assumption 1."},{"cited_title":"Distributed block coordinate moving horizon estimation for 2d visual-inertial-odometry SLAM,","cited_arxiv_id":null,"evidence_quote":"Provides the decoupled block-coordinate MHE-SLAM structure that the paper adapts and analyzes."},{"cited_title":"A moving horizon state and parameter estimation scheme with guaranteed robust convergence,","cited_arxiv_id":null,"evidence_quote":"Establishes that detectability of the augmented state requires parameter persistent excitation, motivating the decoupled scheme."},{"cited_title":"Nonlinear detectability and incremental input/output-to-state stability,","cited_arxiv_id":null,"evidence_quote":"Defines i-IOSS detectability and the Lyapunov-function condition that Assumption 1 relies on."},{"cited_title":"The effects of partial observability in SLAM,","cited_arxiv_id":null,"evidence_quote":"Shows that without anchoring or an external position sensor even linear SLAM is unobservable, which the paper cites to justify Assumption 1."},{"cited_title":"On the observability and observability analysis of SLAM,","cited_arxiv_id":null,"evidence_quote":"Shows inherent unobservability of 2D range-based SLAM and how anchoring ensures observability, supporting verification of Assumption 1."},{"cited_title":"Generalized recursive least squares: Stability, robustness, and excitation,","cited_arxiv_id":null,"evidence_quote":"Provides the generalized recursive least squares stability result used for the streamlined range-measurement scheme."}],"review_version":1}