{"id":"652b7344-5c00-4cd1-8eb3-f218be42c6be","arxiv_id":"2501.04170","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A split state-space Bayesian filter estimates full staircase geometry and segments stair surfaces reliably from noisy, occluded, cluttered point clouds.","lead":"This paper presents a Bayesian filter that models staircases as infinite lines plus finite endpoints to estimate their geometry from noisy, occluded LiDAR data. The method also segments clutter-free stair surfaces, and tests on a Spot robot show lower estimation errors than simple averaging baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (4) lacks the η sign on d_s/d_e, so the printed model cannot predict preceding stairs for straight staircases; this is load-bearing because occlusion handling and descending-staircase claims rely on bidirectional extrapolation.","rationale":"The reader's weakest_assumption (constant parameters) is a real scope limitation and is honestly acknowledged in the conclusion. But the more load-bearing issue is internal: Eq. (4), as printed, cannot predict a preceding stair even for a perfectly regular straight staircase. This is not about irregular geometry; it is about the sign of the displacement. The model is explicitly bidirectional in Sections III-B and IV-B, and the ability to extrapolate unseen stairs is what the paper claims enables operation under partial observations and occlusions. If the implementation matches the printed equation, then descending staircases and any scenario where new stairs appear below the current field of view would be estimated in the wrong direction, and the reported accuracy improvements would not be explainable. If the implementation differs, the paper still needs a corrected equation and an explicit test of bidirectional extrapolation. The real-robot results are a point in the paper's favor, but without code or data the reader cannot tell which version is real. I therefore keep the conditional verdict and add this geometric check as the decisive test.","tokens_in":12144,"tokens_out":11625,"duration_ms":113704,"concrete_test":"Analytic check: take a straight ascending staircase with known depth d and ψ_s=ψ_e=90°, initialize stair j=3, set i=2 and Δψ=0, then evaluate Eq. (4). It predicts y_s^(2,3)=y_s^(3)+d instead of y_s^(3)-d, i.e., stair 4 rather than stair 2. To settle whether this is only a typo, run the same straight staircase through the released code or obtain the implemented endpoint update from the authors and check the predicted endpoints for a preceding stair. Additionally, simulate a 5-step curved staircase with ground truth, mask the bottom two steps, and compare iterative Eq. (4) predictions with ground truth; if the fan angle accumulates an extra Δψ per step, the indexing is also off by one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B defines ψ_s and ψ_e as orientations perpendicular to the stair line in the ascending direction, and Eq. (4) is used to predict stair i from adjacent stair j. For a preceding stair, j=i+1 and η=-1, so the x/y increments are d_s cos(ψ_e+η(j+1)Δψ) and d_s sin(...). For a straight staircase Δψ=0, the increment is +d in the direction ψ_e, i.e., further up the staircase. Starting from stair j and moving 'up' cannot land on the lower stair i; it lands on stair j+1. The same issue appears for non-adjacent stairs via iterative application. The z-update in Eq. (3) correctly includes ηh, showing the sign pattern is intentional there and absent in Eq. (4). Unless an unstated convention makes ψ_e point downward, contradicting Section III-B, the printed extrapolation for preceding/new-below stairs is geometrically wrong. Since the method's occlusion and field-of-view claims depend on predicting unobserved stairs in both directions (Section IV-B: 'new preceding stairs' and 'new succeeding stairs'), this is a load-bearing internal inconsistency, not merely a scope limitation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a Bayesian estimation framework for staircases in cluttered environments. The authors propose a split state-space representation: an infinite-line state describing each stair edge with polar coordinates, and an endpoint state giving the physical endpoints of each stair. Six constant staircase parameters (height, depth, width, start/end yaw, curvature) are estimated jointly with the states using an EKF, and the refined staircase estimate is used to segment stair surfaces from a LiDAR point cloud. The method is evaluated on 17 real-world and 6 simulated staircases with manually labeled ground truth, and compared against two averaging/maximizing estimation baselines and a cloth-simulation segmentation baseline. The reported results show 67-89% reductions in parameter RMSE and about 30% reductions in location RMSE, with real-time execution on an NVIDIA Jetson AGX Orin.","tokens_in":12481,"tokens_out":6365,"duration_ms":62959,"significance":"If the model equations were corrected, this would be a practically useful contribution: the split representation is a sensible way to handle partial observations, the experimental protocol is solid (23 staircases, manual ground truth, two baselines), and the segmentation results are clearly reported. The paper also gives concrete timing results on an embedded platform, which supports the real-time claim. The central limitation, the constant-parameter assumption, is explicitly acknowledged in the Conclusion, so it does not by itself undermine the paper. However, the printed endpoint extrapolation contains a sign error that breaks backward prediction, and the process model contains an un-derived heuristic averaging step; both must be addressed before the empirical claims can be fully credited.","major_comments":[{"comment":"The endpoint prediction for a preceding stair omits the sign of η on the displacement terms. With η=-1 and a straight staircase (Δψ=0), the update x_s(i,j) = x_s(j) + d_s cos(ψ_e) moves in the ascending direction, so iterating (4) from stair j produces stair j+1 rather than stair j-1. Because Section IV-B and Eq. (12) use exactly this update to initialize 'new preceding stairs,' the printed model cannot extrapolate unobserved lower stairs or handle descending staircases as claimed. The z-update in Eq. (3) correctly uses ηh, so the omission is an internal inconsistency rather than a deliberate convention. The equation should use η d_s and η d_e (or an equivalent sign correction), and the experiments should be re-run with the corrected model.","section":"III-B, Eq. (4)"},{"comment":"The matched-stair process model is an un-derived equal-weight average of the current state and the predictions from both adjacent stairs. No justification or ablation is provided for the equal weights, although this heuristic is the core of the EKF prediction and therefore affects all reported accuracy numbers. Please provide a derivation (e.g., from a latent parameter model) or a sensitivity/ablation study demonstrating that the equal-weight choice is not responsible for the improvements over the baselines.","section":"IV-B, Eq. (11)"},{"comment":"The endpoint update selects the two endpoints that maximize XY Euclidean distance among predicted and measured candidates. When a measurement covers only part of a stair width because of occlusion or limited field of view, this rule can systematically overestimate the width and shift the endpoint state. Since endpoint accuracy feeds the reported location RMSE and the width parameter feeds the parameter RMSE, the authors should analyze this bias, for example by reporting results separately for occluded versus fully observed stairs.","section":"IV-C"}],"minor_comments":[{"comment":"The row labeled 'Stair Curvature RMSE (cm)' should use angular units (degrees or radians); as printed, the curvature error is assigned a length unit.","section":"Table I"},{"comment":"The sentence '67-89% reduction in estimated staircase parameter errors' should be tied to the specific parameters in Table I (height and width) to avoid overstating the result.","section":"Section I"},{"comment":"The Jacobians G_z and G_Ω are written without their evaluation points; stating that they are evaluated at the current measurement and pose would make the linearization precise.","section":"III-C, Eq. (9)"},{"comment":"The sentence 'if a stair has no match in the new measurement, EKF is not applied and its state remains unchanged' should also state whether the covariance remains unchanged, since this affects future gating and prediction.","section":"IV-C"},{"comment":"The constant-parameter assumption is mentioned in the Conclusion as a limitation, but it would help readers to see this scope warning next to the assumption statement in Section III-B.","section":"III-B and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be an accepted RA-L manuscript; the sign error in Eq. (4) may be a preprint artifact, but the authors should be asked to confirm that the implementation matches the corrected equation and to provide a corrected version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a competent engineering paper on staircase perception, and the strongest part is the real-robot evaluation across 23 staircases. The split state-space model (infinite-line state plus endpoints) fused through an EKF is a genuine new combination, and the reported gains over the authors' own averaging baseline are large. The segmentation pipeline is also sensible and achieves good numbers. If you work on legged navigation in cluttered indoor environments, this is worth a look.\n\nBut there is a load-bearing sign error in the printed math that the reader's report missed. Equation (4) is the extrapolation rule for predicting a new stair from a neighbor. When predicting a preceding stair, eta = -1 and psi_o = psi_e, but the x/y increments use +d_s, +d_e with no eta factor. For a straight staircase (Delta_psi = 0), that places the 'preceding' stair further up the staircase, not below. The z-update in Eq. (3) correctly uses eta*h, so the sign pattern is intentional there and absent in Eq. (4). Since the paper's occlusion and field-of-view claims explicitly rely on predicting 'new preceding stairs' in both directions, this is not cosmetic. It may be a typo or an unstated convention difference, but as printed the model cannot do what it claims. The authors need to correct this and ideally show the corrected equations reproduce the experimental results.\n\nOther soft spots are minor by comparison. No code or data is released, so the quantitative claims can't be independently checked. The baselines are only the authors' prior averaging method and a maximizing variant; a comparison with a recent model-based estimator such as Qing et al. would be more convincing. The constant-parameter assumption for stair geometry is explicitly acknowledged in the conclusion, and it is a real limitation for irregular stairs, but it's not a fatal flaw for the intended use case.\n\nFor a reader: this is aimed at robotics researchers doing real-time LiDAR-based stair perception, especially on legged platforms. It deserves a serious referee, but with a required major revision to fix the sign issue and ideally release code/data. My recommendation: send it to review, but request a corrected model and a reproducible evaluation.","headline":"Solid real-robot staircase perception paper with a novel state-space formulation, but Eq. (4) has a sign error that undermines the bidirectional prediction claim.","tokens_in":12948,"tokens_out":4485,"would_cite":false,"duration_ms":42737,"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":"A Bayesian split-state model reconstructs full staircase geometry from partial, cluttered LiDAR measurements, cutting parameter errors by 67-89% and location error by 30%.","keywords":["Bayesian inference","Extended Kalman Filter","staircase estimation","LiDAR point cloud","ground segmentation","cluttered environments","mobile robots","probabilistic robotics"],"falsifier":"Measure the estimator's predicted step positions on a staircase whose step depth changes partway, such as five 30 cm treads followed by five 25 cm treads, with only the lower half observed; if the predicted upper steps show a systematic, growing offset from ground truth, the constant-parameter assumption is the cause.","tokens_in":11969,"feed_emoji":"🪜","tokens_out":5706,"duration_ms":52168,"temperature":0.7,"pith_summary":"The paper argues that a staircase can be tracked reliably even when most of it is out of view or buried under clutter, by representing it twice: as an infinite line per step-edge plus a finite endpoint per step. These two representations are fused in an Extended Kalman Filter whose process model extrapolates unseen steps from six constant staircase parameters. On real and simulated staircases the estimator reduces parameter errors by 67-89% and location error by about 30% relative to earlier averaging and maximizing detectors, and the same estimate drives a ground-parallel RANSAC plane fit that segments stair treads from debris. The central tradeoff is that the six parameters are assumed constant along the whole flight, which is exactly where irregular staircases would break the model.","feed_headline":"Split-state Bayesian filter rebuilds occluded staircases in real time","feed_subtitle":"Combining infinite-line and endpoint states cuts parameter errors by 67-89% and location error by 30%.","key_machinery":"The load-bearing object is the split state vector: an 'infinite-line staircase state' ($\\mathcal{L}X_k$) holding, for each step, polar coordinates $(r^{(i)}, \\phi^{(i)}, z^{(i)}_s, z^{(i)}_e)$ of that step's tread-riser edge as an infinite line, together with an 'endpoint state' ($\\mathcal{P}X_k$) holding the Cartesian start and end points of each observed stair edge. Six constant parameters (height, depth, width, start yaw, end yaw, curvature angle) let the process model extrapolate neighbouring and unseen steps through equations (3)-(4). An Extended Kalman Filter fuses noisy LiDAR detections with this prediction, using Mahalanobis matching to decide whether a measurement is an existing, preceding, or succeeding stair, and a user-set covariance $R$ on the parameters controls how much the filter trusts the model versus the measurements. The segmentation step then turns the filtered state into crop boxes and a ground-parallel RANSAC plane.","core_discovery":"The paper claims that the full geometry of a staircase, including steps the sensor has never seen, can be estimated online by a split state-space model in which each stair edge is an infinite polar line (radius, normal angle, start and end z) and each stair also carries finite Cartesian endpoints in the world frame. The infinite-line state is updated by an EKF that averages, for each matched stair, its current estimate with predictions made from its two neighbours using the constant parameters; newly appearing stairs are extrapolated from the nearest observed stair. Endpoints are then refined by maximizing stair width and reprojected onto the filtered line, yielding a complete, consistent estimate that still runs under 30 ms per cycle. The same state estimate defines per-stair crop boxes, and a RANSAC plane fit constrained parallel to the ground separates tread surfaces from clutter, beating a cloth-simulation filter in accuracy on the reported data.","pith_inferences":["A natural next test is irregular staircases: because the six parameters are assumed constant, any flight with a mid-way change in step depth or height should produce a growing prediction bias, and adapting $R$ or running an EKF ensemble is the paper's own suggested remedy, not a demonstrated result.","The split representation may transfer to other quasi-regular structures such as ramps, bleachers, or conveyor lines, where an infinite-line prior with finite endpoints absorbs partial observations in the same way.","The segmented stair surfaces are natural foothold candidates; the paper lists legged motion planning as future work, so the connection to locomotion is an inference, not a demonstrated outcome."],"forward_implications":["Staircase estimation can continue through large occlusions, as in the reported example where a cone hides part of the stairs and the full staircase is still estimated.","The estimator outputs complete geometry including unobserved steps rather than only matching detected steps, which is what enables the reported parameter-error reductions.","Segmentation of stair surfaces from flat debris improves because the crop boxes and ground-parallel constraint are grounded in the filtered model, not just local geometry.","Real-time operation at 20 Hz is feasible on an embedded GPU: total estimation and segmentation stays under 30 ms even for staircases with 20 steps."],"supporting_citations":[{"why":"Supplies the staircase detection algorithm that produces the measurements used by the estimator, and provides the averaging and maximizing baselines it is compared against.","marker":"[8]"},{"why":"Provides the Cloth Simulation Filter baseline used for stair surface segmentation comparison.","marker":"[15]"},{"why":"Weighted line fitting algorithm used inside the detection module, which motivates the design choice of not maintaining a covariance for the endpoint state.","marker":"[18]"},{"why":"Gives the standard Bayesian filtering and Kalman filter background that the EKF implementation builds on.","marker":"[19]"},{"why":"LiDAR-inertial odometry used for localization and mapping on the real robot, providing the world-frame pose needed for state initialization.","marker":"[20]"}],"fun_headline_variants":["Bayesian split-state filter reconstructs stairs from sparse data","Split-state Bayesian model maps occluded staircases in real time","Infinite-line state EKF estimates unseen stair steps from clutter","Bayesian split-state filter fuses noisy scans to predict stair geometry","Model uses infinite lines and endpoints to reconstruct cluttered stairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes every step in a staircase has the same height, depth, width, orientation, and curvature, so it predicts unseen steps by repeating the geometry of the steps it can see.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian split-state filter reconstructs stairs from sparse data","Split-state Bayesian model maps occluded staircases in real time","Infinite-line state EKF estimates unseen stair steps from clutter","Bayesian split-state filter fuses noisy scans to predict stair geometry","Model uses infinite lines and endpoints to reconstruct cluttered stairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2843,"prompt_tokens":917,"completion_tokens":1926,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1841}},"tokens_in":533,"tokens_out":1926,"duration_ms":13635,"temperature":1.0,"reasoning_tokens":1841,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:39:25.366168+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the estimator's predicted step positions on a staircase whose step depth changes partway, such as five 30 cm treads followed by five 25 cm treads, with only the lower half observed; if the predicted upper steps show a systematic, growing offset from ground truth, the constant-parameter assumption is the cause.","supporting_citations":[{"cited_title":"Fast staircase detection and estimation using 3d point clouds with multi-detection merging for heterogeneous robots,","cited_arxiv_id":null,"evidence_quote":"Supplies the staircase detection algorithm that produces the measurements used by the estimator, and provides the averaging and maximizing baselines it is compared against."},{"cited_title":"An easy-to-use airborne lidar data filtering method based on cloth simulation,","cited_arxiv_id":null,"evidence_quote":"Provides the Cloth Simulation Filter baseline used for stair surface segmentation comparison."},{"cited_title":"Weighted line fitting algorithms for mobile robot map building and efficient data representation,","cited_arxiv_id":null,"evidence_quote":"Weighted line fitting algorithm used inside the detection module, which motivates the design choice of not maintaining a covariance for the endpoint state."},{"cited_title":"LiPO: LiDAR Inertial Odometry for ICP Comparison","cited_arxiv_id":"2410.08097","evidence_quote":"LiDAR-inertial odometry used for localization and mapping on the real robot, providing the world-frame pose needed for state initialization."}],"review_version":1}