{"id":"0bc02f21-2055-4544-8709-b5436fcd291e","arxiv_id":"1908.02002","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Inference after each control step can be updated from the planner's already computed factorization, avoiding duplicate computation and cutting inference update time by at least half.","lead":"This paper shows that a robot can update its map and position estimates by reusing mathematical calculations already done during the planning phase, instead of redoing them after each new measurement. The result is at least two times faster inference updates with nearly identical accuracy, tested on synthetic data and the KITTI driving dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (35) is proven only for the first Gauss-Newton iteration; the paper's unconditional 'same estimation accuracy' claim lacks a bound when relinearization is required, and the experiments avoid that regime.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the equality R_{k+1|k+1}=R_{k+1|k} presupposes identical linearization points and a single Gauss-Newton update, while standard iSAM2 relinearizes and the paper gives no error bound when it does. My review of the manuscript confirms this is the central correctness risk. The algebra in Sections 3.4.1-3.4.4 is internally consistent under the stated assumptions, and the complexity analysis and KITTI timing results are plausible, with the caveat that only the inference block is timed while planning dominates overall cost. The experiments are carefully scoped to the single-iteration/no-relinearization regime: Section 4.1.1 states this explicitly, and Figure 1 times only the first inference update iteration. Thus the paper's stronger claims in the abstract and conclusions ('same estimation accuracy' as a general property) are not supported beyond that regime. Because the reader's conditional verdict already reflects this gap, no verdict change is needed; the paper should be accepted only as a conditional contribution pending an explicit error analysis or a demonstration that relinearization is unnecessary in the intended operating conditions.","tokens_in":39372,"tokens_out":5217,"duration_ms":52532,"concrete_test":"Run the KITTI active-SLAM setup with injected measurement noise (e.g., 10x the original sigma) so that the posterior after actual measurements requires relinearization. Compare UD-OTM-OO, which reuses the planning R for a single update, against iSAM2 run to full convergence with fluid relinearization. Log per-step differences in the MAP translation/rotation means and covariances, and record the number of iSAM2 relinearization events per step. If any step with relinearization shows a mean translation difference above numerical noise (say > 1e-6 m), Eq. (35) is not valid in the general regime and the 'same estimation accuracy' claim requires a correction or bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algebraic core is sound under the stated assumptions: with consistent DA and an identical first-iteration linearization point, R_{k+1|k+1}=R_{k+1|k} (Eq. 35) and only the RHS vector needs updating. However, Section 3.4 explicitly qualifies this equality as holding 'for the first iteration in the inference stage at time k+1.' Standard iSAM2 in normal operation performs fluid relinearization (Kaess et al., 2012); when the actual measurements shift the optimum away from the planning linearization point, a converged Gaussian-Newton estimate requires re-evaluating the Jacobian and re-factorizing R. RUB reuses the planning R unchanged and performs a single update, so outside the single-iteration regime the result is an approximate, not exact, MAP estimate. The paper provides no error bound for this approximation. The experiments specifically sidestep the issue: Section 4.1.1 states 'no re-linearization is necessary; hence, iSAM comparison is valid,' and the timing in Figure 1 covers only the first inference update iteration. The KITTI accuracy comparison is thus evidence only that no problematic relinearization occurred in that particular run, not that the method is generally exact. This is the load-bearing gap between the algebraic claim and the advertised paradigm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers an MPC plan-act-infer loop and observes that the future belief propagated during belief space planning, b[X_{k+1}|k], already contains most of the factorized calculations needed for the subsequent Bayesian update. Under data association that is consistent between planning and inference, the authors prove that the factorized Jacobians coincide at the first Gauss-Newton iteration (Eq. (35)), so only the RHS vector d needs to be updated. They present four RHS-update methods (OTM, OTM-OO, DU, DU-OO), analyze their computational complexity, and combine this with an iSAM2-style Bayes-tree modification for inconsistent data association (Algorithm 1). The approach is evaluated in synthetic full-SLAM simulations and on 1400 steps of KITTI monocular data, reporting inference-update speedups of up to two orders of magnitude and negligible accuracy differences relative to iSAM2.","tokens_in":39583,"tokens_out":7030,"duration_ms":75125,"significance":"If the exactness claim is accepted for the intended operating regime, this is a useful contribution: it turns a conceptual similarity between BSP and inference into a concrete computational shortcut, and the four RHS-update methods are derived cleanly from standard linear algebra. The empirical evaluation is substantial, includes real-world data, and compares a MATLAB implementation against a C++ iSAM2 baseline, which is a conservative comparison in the authors' favor. I see no circularity: the reuse equations follow from QR factorization rather than from the authors' earlier JIP framework. The principal weakness is that the exactness result is demonstrated only for the first Gauss-Newton iteration at an identical linearization point, while the paper advertises unconditional same-accuracy equivalence; this gap needs to be closed or the claim qualified.","major_comments":[{"comment":"The equality R_{k+1|k+1} = R_{k+1|k} is valid only for the first Gauss-Newton iteration at a common linearization point; the text says 'for the first iteration in the inference stage at time k+1' immediately before Eq. (35). Standard iSAM2 performs fluid relinearization as measurements arrive, and when the optimum shifts, the planning Jacobian is no longer the Jacobian at the current linearization point, so the reused R is stale and a single RHS update does not compute the exact MAP estimate. The paper's unconditional claims of 'same estimation accuracy' (Abstract and end of Section 3.4) and 'algebraically equivalent to estimation via iSAM' (Section 4.2.4) are therefore too strong. The authors should either restrict the claim to the single-iteration/no-relinearization regime, provide a bound on the approximation error when relinearization is needed, or extend the method to update R when the linearization point changes.","section":"Section 3.4, Eq. (35)"},{"comment":"The experiments deliberately avoid the regime in which the central equality could fail. Section 4.1.1 states 'no re-linearization is necessary; hence, iSAM comparison is valid,' and the timing includes only the first inference update iteration; the KITTI experiment likewise uses a single horizon step and reports no substantial loop closures or relinearization events. Consequently, the results support the claim that R reuse is exact when a single Gauss-Newton update suffices, but they do not provide evidence for the general claim that RUB inference matches standard Bayesian inference in systems that relinearize. An experiment with active relinearization, or a theoretical error bound, is needed to close this gap.","section":"Section 4.1.1 and Section 4.2.4"}],"minor_comments":[{"comment":"In Table 1, 'natch' should be 'match' in the description of daug_{t|k}.","section":"Table 1"},{"comment":"The legends of Figures 5 and 6 use 'UD' for the methods that the text calls DU and DU-OO; please make the notation consistent with the body and with Table 1.","section":"Figures 5 and 6"},{"comment":"Several typographical errors should be corrected: 'Given's' should be 'Givens' in Section 3.4.5, 'affectively' should be 'effectively' in Section 2.2, 'ssytem' should be 'system' in Section 5, and 'Beysian' should be 'Bayesian' in Figure 10.","section":"Throughout"},{"comment":"The phrase 'algebraically equivalent to estimation via iSAM' should be rephrased to state explicitly that the equivalence holds under the identical-linearization-point, first-iteration conditions identified in Section 3.4.","section":"Section 4.2.4"},{"comment":"The paper switches between 'iSAM' and 'iSAM2' when referring to the baseline; please use one name consistently or state explicitly that 'iSAM' denotes the iSAM2-based baseline throughout.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The four RHS-update methods are correct, and the KITTI experiment shows a credible speedup in the inference block—roughly a factor of two or better—with negligible accuracy loss. But the 'same estimation accuracy' claim is narrower than the abstract lets on: it is proven only for the first Gauss-Newton iteration at an identical linearization point, and once relinearization is needed the reused factorized Jacobian is stale and the method becomes approximate, with no error bound in the paper.\n\nThe core RUB idea appeared in the authors' 2017 ICRA paper, so the novelty is in the elaboration: four exact RHS update methods (OTM, OTM-OO, DU, DU-OO), a data-association update algorithm built on iSAM2, complexity analysis, and both synthetic and KITTI validation. The algebra is sound under the stated assumptions, and the DA update is the most interesting piece—it handles inconsistent associations between planning and inference by detaching and re-eliminating only the affected sub-Bayes-tree. The KITTI accuracy comparison is a good sanity check: translation error around 1e-14 and covariance differences that are practically zero.\n\nThe load-bearing gap is Eq. (35). The paper states in Section 3.4 that R_{k+1|k+1}=R_{k+1|k} holds for the first iteration in the inference stage, but the abstract and conclusions drop that qualifier. Standard iSAM2 performs fluid relinearization when the linearization point goes stale; in that regime the reused R is not the one standard inference would compute, and the update is approximate. The experiments sidestep this: the synthetic setup explicitly does a single inference update with no relinearization, and the KITTI comparison just happened not to require relinearization in that particular run. The paper needs either a carefully scoped exactness statement or an error analysis for the relinearization case. As a lesser issue, the 'paradigm shift' framing is overblown; this is an incremental but useful contribution to joint inference and planning. Also, the timing covers only the inference block, not the whole plan-act-infer loop, though the authors acknowledge the scope.\n\nThe paper is for researchers in active SLAM, belief space planning, and incremental factor-graph inference. The DA update algorithm has independent value. It deserves a serious referee: send it to review, and ask the referee to push for a precise exactness statement, an error bound or a relinearization-aware experiment, and toned-down language. That is a major revision, not a desk reject.","headline":"Reusing planning factorizations for inference updates is a sound, useful trick with a real KITTI speedup, but the exactness claim only holds for one Gauss-Newton iteration at the same linearization point and the paper owes an error bound for the rest.","tokens_in":40098,"tokens_out":4073,"would_cite":true,"duration_ms":39653,"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":"Planning-stage factorizations can be reused to perform Bayesian inference updates exactly, at substantially lower cost.","keywords":["Bayesian inference","belief space planning","incremental smoothing and mapping","square-root information matrix","data association","inference update","calculation reuse","active SLAM"],"falsifier":"Compare the $R$ matrix obtained by reusing the planning factorization against the $R$ matrix from a standard inference update on a sequence where the actual measurements are deliberately set to shift the linearization point across a re-linearization threshold; statistically significant divergence in the resulting MAP estimates would show that the equality $R_{k+1|k+1}=R_{k+1|k}$ does not survive re-linearization.","tokens_in":39160,"feed_emoji":"🤖","tokens_out":8689,"duration_ms":103264,"temperature":0.7,"pith_summary":"The paper proposes that the belief update at the heart of Bayesian inference need not be computed from scratch: when a robot has already planned an action, the planning step has already solved a nearly identical inference problem over the next-timestep belief. Under consistent data association and identical linearization points, the factorized Jacobian (square-root information matrix) obtained during planning equals the one standard inference would compute, so only the right-hand-side vector has to be updated. The authors give four exact methods for that update, plus a data-association correction routine for the realistic case where what the robot planned to see differs from what it actually sees. If correct, the approach yields the same estimation accuracy as standard Bayesian inference while reducing inference-update time by at least a factor of two, with larger gains on loop closures and high-dimensional states.","feed_headline":"Planning calculations recycled to cut inference update time","feed_subtitle":"The factorized Jacobian from planning matches the inference one, so only the right-hand side is recomputed.","key_machinery":"The load-bearing object is the QR-factorized Jacobian of the Gaussian nonlinear least-squares problem: the square-root information matrix $R$ and its RHS vector $d$, often stored as a Bayes tree. Planning at time $k$ produces the augmented linear system formed from the previous factorized Jacobian and the new factors, and the identity $R_{k+1|k+1}=R_{k+1|k}$ says the expensive factorization carries over to inference. Around that identity the paper constructs four RHS-update methods (OTM, OTM-OO, DU, DU-OO) and a data-association update that detaches the involved sub-Bayes-tree, repairs the corresponding sub-factor-graph, and re-eliminates it.","core_discovery":"On the paper's own terms, the central discovery is the identity $R_{k+1|k+1}=R_{k+1|k}$: for the first inference iteration at time $k+1$, the factorized Jacobian (square-root information matrix) of the posterior updated from actual measurements equals the factorized Jacobian of the belief propagated during the precursory planning stage, provided the same action is executed, data association is consistent, and both problems linearize at the same point. The only new computation needed is therefore the right-hand-side vector $d_{k+1|k+1}$, and the paper gives four algebraically equivalent methods to produce it: two that rotate the new residual with the rotation matrix already available from planning, and two that use information downdate/update formulas. When planning and inference disagree on data association, an incremental factor-graph/Bayes-tree procedure first removes planning-only factors and adds inference-only factors, after which the RHS update applies. The complete pipeline is compared with standard incremental Bayesian inference on synthetic active-SLAM scenarios and on a real-world driving dataset, reporting the same estimation accuracy and at least a two-fold reduction in inference-update time.","pith_inferences":["The exactness of the reused $R$ is tied to the first inference iteration and identical linearization points; an open question the paper leaves is how much accuracy is lost when real measurements trigger re-linearization, since no error bound is given.","The same RHS-update machinery could be applied to a curated library of precomputed beliefs rather than only the single planning-stage belief; the unsolved design problem is how to choose the stored belief closest to the needed posterior without spending the saved time on search.","If planning could predict newly appearing landmarks, the fraction of factors reused would rise above the level seen in the real-data experiment, which suggests the reported speedup is a lower bound that improves with prediction quality.","The paper's Gaussian/Gauss-Newton carrier object suggests a natural test for non-Gaussian beliefs: the equality would need a different summary statistic (e.g. the Fisher information) and the update-in-place step would likely become approximate."],"forward_implications":["The only required change to a belief-space planner is that its output includes the propagated future belief alongside the chosen action; the planner itself does not need to be replaced.","Because the factorized Jacobian is reused, the inference update avoids paying the same factorization cost twice, which is why loop closures and growing state dimension cost less when the relevant factors were already incorporated during planning.","The rotation-based method (OTM-OO) is the fastest in the reported experiments, while the downdate-based methods trade some speed for easier availability of the required matrices.","With the data-association repair step, the paradigm still outperforms the standard baseline in simulation even when only half of the predicted associations are correct, and on real-world data the full pipeline runs in less than half the inference-update time."],"supporting_citations":[{"why":"supplies the incremental smoothing and mapping baseline and the Bayes-tree machinery used for the data-association update.","marker":"(Kaess et al., 2012)"},{"why":"introduces the square-root information matrix / QR formulation of smoothing that the reuse identity exploits.","marker":"(Dellaert and Kaess, 2006)"},{"why":"provides the information downdate-update formulas underlying the DU and DU-OO methods.","marker":"(Cunningham et al., 2013)"},{"why":"defines the belief-space planning and predicted-measurement model that generates the planning-stage belief.","marker":"(Indelman et al., 2015)"},{"why":"is the preliminary version of this work that set out the joint inference-and-planning reuse concept.","marker":"(Farhi and Indelman, 2017)"},{"why":"supplies the real-world driving dataset for the final benchmark.","marker":"(Geiger et al., 2013)"}],"fun_headline_variants":["Reusing planning Jacobian halves inference update time","Inference update reuses planning factorized Jacobian","Planning's Jacobian speeds inference updates","Incremental inference via planning's leftover Jacobian","Bayesian update from planning calculations, no recompute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the planning and inference stages linearize at exactly the same point and one Gauss-Newton update is enough, so if the real measurements move the MAP estimate enough to make the system re-linearize, the reused factorized Jacobian $R$ is stale and the update is only approximate; the paper states the identity for the first inference iteration and gives no bound on the resulting error.","fun_headline_variants_meta":{"raw":{"variants":["Reusing planning Jacobian halves inference update time","Inference update reuses planning factorized Jacobian","Planning's Jacobian speeds inference updates","Incremental inference via planning's leftover Jacobian","Bayesian update from planning calculations, no recompute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1371,"prompt_tokens":1006,"completion_tokens":365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":622,"tokens_out":365,"duration_ms":4643,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:57:02.630479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the $R$ matrix obtained by reusing the planning factorization against the $R$ matrix from a standard inference update on a sequence where the actual measurements are deliberately set to shift the linearization point across a re-linearization threshold; statistically significant divergence in the resulting MAP estimates would show that the equality $R_{k+1|k+1}=R_{k+1|k}$ does not survive re-linearization.","supporting_citations":[],"review_version":1}