{"id":"e35e64f5-47a1-44ab-a446-f2755fa4ad8e","arxiv_id":"2505.02272","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A low-cost quadruped SLAM stack that adds IMU-stabilized depth scans and leg-odometry velocity constraints achieves accurate 2D mapping, localization, and navigation in indoor environments.","lead":"This paper combines leg odometry, visual-inertial odometry, and an IMU-stabilized depth scan to help cheap quadruped robots build indoor maps and navigate autonomously. Real-robot and simulation tests show the combined system beats baselines that leave out the stabilization or the leg odometry.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's ablation data contradict the paper's claim that scan stabilization is the most critical component: the B+SS and B+LO rows contain impossible APE values, and the printed ATE values show leg odometry, not scan stabilization, driving the improvement.","rationale":"I examined the reader's weakest assumption about Eq. 5 in Section III-C. Despite the paper deriving the intercept for a single on-axis point, the resulting line model is not as fragile as it first appears: under a pinhole camera, a pure pitch of the body-horizontal plane shifts every point on that slice by the same f_y*tan(p), and a pure roll produces a line with slope approximately tan(r), up to f_y/f_x and 1/cos(r) corrections. Lens distortion and mounting offsets are calibration issues, and the real-world OptiTrack results give empirical support that the stabilization works in practice. I therefore do not consider the geometric approximation itself to be the single most load-bearing problem. The more concrete and load-bearing concern is the inconsistency in Table I, the only controlled ablation that isolates the components. The impossible APE values and the contradiction between the printed ATE values and the claim that scan stabilization is the most critical component directly affect the causal interpretation of the central claim. This is a substantive reporting/evidence problem, but it is fixable by recomputing the table, so it supports the existing CONDITIONAL verdict rather than moving it.","tokens_in":12609,"tokens_out":17790,"duration_ms":224319,"concrete_test":"Download the per-run trajectory logs from the project website (or rerun the four configurations in the AWS warehouse and house environments with the same seeds), then recompute ATE, ARE, and APE using evo for the B+SS and B+LO rows. If the recomputed values show B+SS having low ATE and B+LO high ATE (i.e., the printed columns were swapped), the current conclusion can be restored by correcting the table. If the recomputed values match the printed ATE/ARE values, then the paper's claim that scan stabilization is the most critical component is not supported by its own ablation data and the conclusion must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the ablation study in Table I, which is supposed to show that the full system outperforms alternatives and that scan stabilization is the most critical component (Section IV-C). As printed, the table cannot be trusted: for SB in the warehouse, B+SS has ATE 5.19±2.24 and ARE 0.75±0.21 but APE 0.45±0.11, an impossibility if APE is a combined pose error that includes translation. The B+LO row shows the reverse problem (ATE 0.44, APE 7.90), indicating the APE columns for B+SS and B+LO are swapped. More importantly, reading the ATE/ARE columns literally, B+LO (baseline plus leg odometry, without scan stabilization) improves SB warehouse ATE from 5.00 to 0.44, while B+SS (baseline plus scan stabilization, without leg odometry) leaves ATE at 5.19. This directly contradicts the paper's statement that scan stabilization is the most critical component, and it undercuts the causal story behind the central claim: the system's performance could be driven primarily by leg odometry rather than by the depth-stabilized scans that the paper highlights as the key novelty. The real-world Table III provides some independent support for the full system, but it does not resolve which component should receive credit, because the simulation ablation is the only controlled comparison that isolates the components.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents an integrated localization, mapping, and navigation system for low-cost quadruped robots equipped with an RGB-D camera and an IMU. It augments a standard RTAB-Map/SLAM Toolbox 2D pipeline with three components: contact estimation from generalized momentum, legged odometry via least-squares twist estimation, and IMU-stabilized scan extraction from depth images. The authors claim that the full system produces accurate 2D maps, robust localization, and autonomous navigation, and they identify scan stabilization as the most critical component. The evaluation uses Gazebo simulation (AWS small house and warehouse), real-world Silver Badger experiments with OptiTrack ground truth, navigation success tests, and autonomous exploration, with five runs per condition and ablations of leg odometry and scan stabilization. Code, videos, and additional material are released on a project website.","tokens_in":12960,"tokens_out":9442,"duration_ms":133270,"significance":"If the claims survive the required corrections, the paper makes a useful practical contribution: it targets low-cost sensors, evaluates against external ground truth, reports real-hardware results, and releases code and supplementary material. The ablation design is methodologically sound in principle, and the real-world full-system results are encouraging. However, the printed simulation ablation in Table I contains internally inconsistent values that undermine the stated causal attribution of the improvements, the scan-stabilization derivation is not fully justified for the implemented slice extraction, and the contact-estimation module is not isolated in any controlled ablation. These issues must be fixed before the central claims can be accepted.","major_comments":[{"comment":"Table I contains internally inconsistent APE values for the Silver Badger rows. In the AWS Small Warehouse, B+SS is reported with ATE 5.19±2.24 and APE 0.45±0.11, while B+LO is reported with ATE 0.44±0.11 and APE 7.90±0.83. If APE is a combined pose error that includes translation, these entries cannot both be correct, and the B+SS and B+LO APE columns appear to be swapped. More importantly, reading the ATE/ARE columns literally, B+LO reduces the warehouse ATE from 5.00 to 0.44, whereas B+SS leaves it at 5.19; this contradicts the Section IV-C statement that \"scan stabilization proves to be the most critical component\" and instead credits leg odometry with the main translation improvement. The same pattern appears in the AWS Small House SB rows. Even after correcting the swap, the claim needs to be metric-specific: B+SS improves RPE locally (warehouse RPE 2m: 0.51 to 0.18) while B+LO improves ATE, so \"most critical\" is too broad as printed. Please correct Table I, re-analyze the attribution, and revise Sections IV-C and V accordingly.","section":"Table I and Section IV-C"},{"comment":"The derivation of the scan-line equation is carried out for a single point on the optical axis, and Eq. (5) is then applied as a global row offset and slope to the entire depth slice. For an ideal pinhole camera the offset is exact for points on the horizontal plane through the camera center, but the manuscript does not state this assumption, and the implementation also ignores lens distortion and any camera-IMU extrinsic misalignment. Since Section IV-C identifies scan stabilization as the most critical component, this is not merely a presentation issue. Please either extend the derivation to the full projection model, quantify the approximation error on the actual D435i (for example, by comparing Eq. (5) against ray-plane intersection for the depth image), or otherwise validate that the fixed line and fixed slice width produce a scan that corresponds to a consistent world plane.","section":"Section III-C, Eq. (5)"},{"comment":"Contact estimation is listed as a contribution, but it is never ablated. In the simulation experiments the authors state they \"use the ground-truth contact instead\" (Section IV-A), and the real-world experiments in Table III compare configurations with and without leg odometry and scan stabilization but do not isolate the contact estimator. Thus there is no controlled evidence that the generalized-momentum observer with feet-specific thresholds performs comparably to true contact sensing. Please add an evaluation that compares estimated contacts against ground-truth contacts (for example, B+LO with estimated versus true contacts in simulation, or a real-world sequence where contacts are also measured by an instrumented foot), or explicitly narrow the contribution claim to the integration rather than to the contact-estimation module itself.","section":"Section IV-A and contribution (i)(b)"}],"minor_comments":[{"comment":"The text contains a typo: \"RBGD camera\" should be \"RGBD camera\".","section":"Section I"},{"comment":"The units of ARE are not stated. Please specify whether the reported angular errors are in degrees or radians, since this is necessary for interpreting the magnitudes.","section":"Tables I and III"},{"comment":"The navigation success results report 0% and 100% values but do not state the number of independent trials. Please clarify whether these are five navigation runs per condition or five goals within a single run, and report confidence intervals or per-goal counts.","section":"Section IV-D / Table II"},{"comment":"The sign convention in Eq. (5) is not defined: it should be stated whether a positive roll/pitch corresponds to a clockwise or counterclockwise rotation in the image frame, and how the camera frame is oriented relative to the robot body frame.","section":"Section III-C"},{"comment":"Several tuning parameters that affect the results are not given numerically in the paper: the observer gains L1, L2 and L, the feet-specific contact thresholds, the slice width, and the factor-graph velocity constraint weights. The project website is a good resource, but a complete reproducibility table in the paper would be preferable.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The Table I inconsistency is the main barrier to acceptance. If the corrected numbers confirm the printed ATE columns, the authors should reframe the causal claim: leg odometry appears responsible for the large ATE reduction, while scan stabilization contributes to local consistency and map quality. The real-world Table III is more favorable and suggests the integrated system is effective, so I do not regard the issue as unfixable. I would also ask the authors to verify that the B+LO and B+SS APE entries are swapped in both simulation environments before resubmitting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a practical system integration for low-cost quadruped navigation, and the full system appears to work. The real-world results are the strongest part: with only a depth camera and IMU, the full pipeline reduces ATE from 2.51 m to 0.22 m in one environment and from 1.50 m to 0.65 m in another, all against OptiTrack ground truth. That is real, reproducible evidence, and the code and videos are promised on the project site. The scan-stabilization line-slice idea is simple and likely useful, and the mixed-mode contact observer is a nice pragmatic touch for platforms without contact sensors.\n\nBut the stress-test note is on target, and the problem sits at the center of the paper's argument. Table I is internally inconsistent: for the Silver Badger in the warehouse, B+SS shows ATE 5.19 and APE 0.45, which is impossible if APE combines translation and rotation; B+LO shows the reverse problem. The APE columns for those two rows are clearly swapped. And when you read the ATE columns literally, leg odometry (B+LO) improves the baseline from 5.00 to 0.44 while scan stabilization (B+SS) leaves it at 5.19. That directly contradicts the repeated claim that scan stabilization is the most critical component. The real-world Table III does not rescue the story either: in Environment 2, neither ablation component alone beats the baseline, only the full system does. So the causal attribution is murky even after ignoring the typo.\n\nThe scan-stabilization derivation is also an approximation: Eq. 5 is derived for a point on the optical axis and then applied as a uniform pixel line across all depths. That could be fine with a wide enough slice, but the paper does not discuss the limits, and since the ablation is the main evidence for its importance, the approximation deserves more scrutiny.\n\nThis paper deserves a serious referee, however, because the full system works, the evaluation design is otherwise above average (external ground truth, multiple runs, real-world validation), and the scan-stabilization idea is worth testing. But it needs a corrected table, a re-analysis of which component actually contributes what, and some discussion of the depth-slice approximation. As it stands, the central claim is not supported by the printed data. I would send it to review with instructions to require those fixes before acceptance.","headline":"Useful system integration for low-cost quadrupeds, but the central ablation table is internally inconsistent and contradicts the paper's own causal story.","tokens_in":13477,"tokens_out":3768,"would_cite":false,"duration_ms":49267,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A low-cost quadruped robot can localize, map, and navigate accurately by fusing contact-aided leg odometry, visual-inertial odometry, and IMU-stabilized depth scans.","keywords":["quadruped robots","2D SLAM","leg odometry","visual-inertial odometry","scan stabilization","contact estimation","autonomous navigation","low-cost sensors"],"falsifier":"Tilt a mounted depth camera by a known pitch angle, point it at a flat wall at a known distance, and read off the image row where the wall appears. Compare that row with the prediction of Eq. 5 using the camera's focal length; a discrepancy of more than a few pixels across the usable width would show that the single-point pinhole derivation does not describe the full row, and because the paper's ablation identifies scan stabilization as the most critical component, that failure would undermine the central claim.","tokens_in":1563,"feed_emoji":"🤖","tokens_out":1839,"duration_ms":95217,"temperature":0.7,"pith_summary":"The paper sets out to show that a cheap quadruped robot—equipped only with a depth camera, an IMU, and joint encoders—can map a room, know where it is, and navigate to goals, tasks usually reserved for expensive platforms with long-range lidars. It argues that the obstacle is not the SLAM algorithms themselves but the instability of depth-derived scans when a legged robot pitches and rolls with every step. The proposed fix selects a slanting row of depth pixels determined by the IMU's roll and pitch, producing a stabilized 2D scan, and adds leg odometry as a velocity constraint in the pose graph. In the paper's experiments, the full system outperforms every ablated variant, reaches 100 percent navigation success in the simulated house and warehouse scenarios tested, and the real-robot experiments confirm the same ordering of ablations. The takeaway is that one small sensor-fusion detail, not a new SLAM paradigm, unlocks robust autonomy for low-cost legged robots.","feed_headline":"Budget quadrupeds map and navigate with one stabilized-scan fix","feed_subtitle":"IMU-stabilized scans plus leg odometry push goal-reaching to 100 percent in tested homes and warehouses.","key_machinery":"The load-bearing object is an IMU-stabilized scan: from a depth image, choose the row whose slope is set by the roll angle and whose intercept is $-f_y\\tan(p)$, where $p$ is pitch and $f_y$ is the focal length, and combine pixels above and below that row into a lidar-like 2D scan. Around this, the paper stacks a least-squares leg odometry that solves for the body twist $\\hat{V}_b$ from foot-contact constraints plus IMU angular velocity measurements, and a contact-state observer based on generalized momentum that estimates foot contacts from joint torques alone. The leg twist is used to reinitialize visual-inertial odometry after tracking loss and to impose velocity constraints between consecutive poses in the 2D SLAM factor graph, preventing scan mismatches from corrupting the map.","core_discovery":"The central claim, stated on the paper's own terms, is that contact-aided kinematic leg odometry, visual-inertial odometry, and depth-stabilized vision can be fused into a single 2D localization, mapping, and navigation stack that works on low-cost quadruped platforms. The discovery is that scan stabilization is the load-bearing component: without it, localization error grows by roughly an order of magnitude in several simulated and real configurations and navigation often fails; with it, the robot localizes on a pre-built map and reaches every commanded pose in the tested scenarios. Leg odometry alone does not rescue the baseline, and in some cases degrades it when scans are badly aligned; it becomes valuable only in combination with stable scans, where it reinitializes lost visual odometry and adds velocity edges to the factor graph.","pith_inferences":["The IMU-row selection trick is not inherently tied to legged platforms; the same formula should stabilize depth slices on any tilt-prone base, so a direct test on a wheeled rover with suspension or a handheld depth camera would check how far the mechanism generalizes.","Because the stabilized scan is a single horizontal slice of the world, an obvious extension is to sample several parallel slices and stack them into a 2.5D or multi-floor map; the paper only commits to 2D.","The strong ablation result suggests a practical rule of thumb the authors do not state: before improving hardware, check camera calibration and mounting stiffness, since a distorted or loose camera would break the exact row-to-plane correspondence the method relies on.","A testable refinement of the contact estimator would be to compare its torque-based contact states against a platform with real foot contact sensors during the same gaits; the paper validates the estimator mainly through downstream localization accuracy."],"forward_implications":["With the full pipeline, navigation on a pre-built map succeeds 100 percent of the time in the tested house and warehouse scenarios for both robots, compared with 0 to 80 percent for the baseline.","The ablation shows scan stabilization is the single most critical module; removing it inflates absolute trajectory error by roughly an order of magnitude in several configurations.","Leg odometry contributes by reinitializing visual odometry after tracking loss and by adding velocity constraints to the 2D factor graph, but only when scans are already stabilized.","The same pipeline, including torque-based contact estimation, works on a real quadruped with an actuated spine and in cluttered indoor environments, not just in simulation.","Because the system uses 2D grid maps, it plugs into standard 2D navigation stacks rather than requiring expensive 3D mapping hardware or algorithms."],"supporting_citations":[{"why":"Supplies the learned locomotion controllers whose fast, jerky motion creates the scan instability the system must overcome.","marker":"[2]"},{"why":"Provides the visual-inertial odometry front-end that the full system wraps.","marker":"[11]"},{"why":"Supplies the 2D pose-graph SLAM back-end into which leg odometry constraints are injected.","marker":"[38]"},{"why":"Provides the leg-odometry formulation of foot-contact twist constraints used in the least-squares body velocity estimate.","marker":"[35]"},{"why":"Provides the generalized-momentum disturbance-observer basis for contact estimation.","marker":"[33]"},{"why":"Prior quadruped visual-SLAM that combines leg odometry with visual odometry, motivating the integration approach.","marker":"[12]"},{"why":"Defines the ATE, ARE, and RPE metrics used to evaluate localization accuracy.","marker":"[37]"}],"fun_headline_variants":["Stabilized scans turn budget quadrupeds into reliable navigators","Scan stabilization is the linchpin for low-cost quadruped SLAM","Cheap quadrupeds navigate reliably when depth scans are stabilized","Scan stabilization makes or breaks budget quadruped SLAM","Contact-aided odometry only helps with stabilized depth scans"],"cache_read_input_tokens":15616,"weakest_assumption_plain":"The system stands on the assumption that one slanted row of depth pixels always samples the same horizontal plane in the world, no matter how the robot is tilted; camera distortion or a tilted mounting would break the stabilized scans, and the paper's own ablation says those scans are the most critical part.","fun_headline_variants_meta":{"raw":{"variants":["Stabilized scans turn budget quadrupeds into reliable navigators","Scan stabilization is the linchpin for low-cost quadruped SLAM","Cheap quadrupeds navigate reliably when depth scans are stabilized","Scan stabilization makes or breaks budget quadruped SLAM","Contact-aided odometry only helps with stabilized depth scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000786,"raw_usage":{"total_tokens":3439,"prompt_tokens":890,"completion_tokens":2549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2463}},"tokens_in":506,"tokens_out":2549,"duration_ms":18450,"temperature":1.0,"reasoning_tokens":2463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:56:44.557201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Tilt a mounted depth camera by a known pitch angle, point it at a flat wall at a known distance, and read off the image row where the wall appears. Compare that row with the prediction of Eq. 5 using the camera's focal length; a discrepancy of more than a few pixels across the usable width would show that the single-point pinhole derivation does not describe the full row, and because the paper's ablation identifies scan stabilization as the most critical component, that failure would undermine the central claim.","supporting_citations":[{"cited_title":"Slam toolbox: Slam for the dynamic world,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D pose-graph SLAM back-end into which leg odometry constraints are injected."},{"cited_title":"Leg odometry for slam,","cited_arxiv_id":null,"evidence_quote":"Provides the leg-odometry formulation of foot-contact twist constraints used in the least-squares body velocity estimate."},{"cited_title":"Contact model fusion for event-based locomotion in unstructured terrains,","cited_arxiv_id":null,"evidence_quote":"Provides the generalized-momentum disturbance-observer basis for contact estimation."},{"cited_title":"A robust visual slam system for small-scale quadruped robots in dynamic environments,","cited_arxiv_id":null,"evidence_quote":"Prior quadruped visual-SLAM that combines leg odometry with visual odometry, motivating the integration approach."},{"cited_title":"A benchmark for the evaluation of rgb-d slam systems,","cited_arxiv_id":null,"evidence_quote":"Defines the ATE, ARE, and RPE metrics used to evaluate localization accuracy."}],"review_version":1}