{"id":"501192fc-c786-49b9-a662-db64a82a79a3","arxiv_id":"2603.02742","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Gate-corner reprojection errors fused directly in an Error-State Kalman Filter give robust monocular visual-inertial state estimation for drone racing, with an offline factor-graph smoother used as a reference.","lead":"A drone-racing state estimator that feeds gate-corner pixel errors directly into a filter, working with as few as two visible corners, plus an offline smoothing pipeline to score races without motion capture. It reports half-decimeter accuracy on a public dataset and survived real competition flights above 20 m/s.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 0.134 m RMS error is inconsistent with Table II; recomputation from per-sequence data yields 0.143 m.","rationale":"I read the paper in good faith. The estimator's core mechanism (ESKF with direct reprojection factors) is standard and plausible, and the two-corner ablation on Flight 18A supports the claim of robustness with partial gate visibility. The TII-RATM dataset provides independent MoCap ground truth, so the ADR-FGO self-reference does not affect that portion. However, the headline number in the abstract and the table's average are internally inconsistent: the per-sequence numbers imply 0.143 m, not 0.134 m. Since the abstract's primary accuracy claim is the most visible and load-bearing quantitative result, this discrepancy is a serious correctness risk. It is a simple arithmetic check, and it is concerning that this was not caught before submission. The reader's rationale already noted 'abstract/body accuracy numbers disagree,' so this is a partial agreement, but I elevate it to the primary concern because it directly undermines the headline result. The ADR-FGO independence issue is a valid limitation, but the TII-RATM MoCap results stand independently. I recommend the verdict remain CONDITIONAL: the paper should be accepted only after the numerical inconsistency is resolved and the accuracy numbers are re-stated with uncertainty estimates and a corrected table.","tokens_in":13923,"tokens_out":11458,"duration_ms":103265,"concrete_test":"Recompute the overall mean translation RMS error from the per-sequence ADR-VINS et column in Table II: sum = 0.130+0.134+0.132+0.169+0.132+0.132+0.107+0.099+0.191+0.208 = 1.434; mean = 0.1434. If the published 'Avg' row and abstract do not equal this value, the authors should correct the table/abstract and re-state any conclusions that depend on the 0.134 m figure. Also verify that the per-sequence et values themselves are correct by re-running the TII-RATM benchmark with the released code or by consulting the original dataset's ground-truth trajectories.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central accuracy claim, 'average RMS translation error of 0.134 m', does not match the paper's own data. From Table II, the ADR-VINS translation errors (et) are: 05P .130, 06P .134, 11P .132, 12P .169, 05A .132, 06A .132, 11A .107, 12A .099, 17A .191, 18A .208. Summing gives 1.434 m, whose arithmetic mean is 0.1434 m. The table's average row reports et=0.134 (while er=2.06 and ev=0.283 are consistent with the per-sequence means), and the abstract repeats 0.134. The paper's own sub-averages are 0.141 (piloted) and 0.145 (autonomous), which also average to 0.143, not 0.134. Thus the most prominently reported quantitative result is not reproducible from the presented data. If the per-sequence values are correct, the headline should be 0.143 m; if the headline is correct, the per-sequence table contains errors. Either way, the central empirical claim is currently unsupported. This is more load-bearing than the ADR-FGO self-reference because it affects the TII-RATM evaluation that is the primary quantitative evidence for the estimator's accuracy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADR-VINS, an Error-State Kalman Filter (ESKF) based monocular visual-inertial state estimator for autonomous drone racing. Instead of the conventional cascade of a general VIO with a PnP-based gate pose correction, ADR-VINS directly feeds gate-corner pixel reprojection errors into the filter, enabling updates with as few as two visible corners and using Huber reweighting in place of RANSAC. The paper also introduces ADR-FGO, an offline factor-graph optimizer that fuses IMU preintegration, gate-corner factors, ADR-VINS pose priors, and extrinsic refinement to produce a reference trajectory for post-flight evaluation. The system is evaluated on the TII-RATM dataset (MoCap ground truth) and deployed in the A2RL Drone Championship Season 2, where it is reported to run at 20.9 m/s with low reprojection error and sub-9 ms per-frame latency.","tokens_in":14380,"tokens_out":3999,"duration_ms":36514,"significance":"If the reported accuracy and robustness hold, the paper makes a useful practical contribution: a tightly coupled, PnP-free, low-latency estimator that degrades gracefully under partial gate visibility, plus an offline smoother for uninstrumented environments. The two-corner ablation (Fig. 5) and the runtime breakdown (Table V) support the central mechanism. However, the current empirical validation contains a serious internal inconsistency in the headline accuracy number, and the A2RL evaluation reference is partly self-referential. These issues must be resolved before the quantitative claims can be accepted.","major_comments":[{"comment":"The headline quantitative claim is not reproducible from the paper's own data. The per-sequence translation errors in Table II (05P .130, 06P .134, 11P .132, 12P .169, 05A .132, 06A .132, 11A .107, 12A .099, 17A .191, 18A .208) sum to 1.434 m, which averages to 0.1434 m, not the reported 0.134 m. The body text reports piloted and autonomous averages of 0.141 m and 0.145 m, which also average to 0.143 m. The abstract and the 'Avg' row state 0.134 m. Thus the most prominent accuracy figure is unsupported by the presented numbers; either the table or the headline must be corrected, and the entire set of derived claims (e.g., percentage improvements over baselines) should be recomputed accordingly.","section":"Abstract / Table II"},{"comment":"The A2RL evaluation uses ADR-FGO as the 'GT (ADR-FGO)' reference, but ADR-FGO is anchored by soft priors taken from ADR-VINS states (Eq. 20) and consumes the same gate-corner detections as ADR-VINS. The reported A2RL translation RMSE of 0.152 m is therefore a consistency measure with a smoothed version of the same system, not an independent accuracy measurement. This does not invalidate the TII-RATM MoCap results, but it weakens the competition-based accuracy claims in Sec. V-D and Table IV.","section":"Sec. IV-A3, Eq. (20); Sec. V-D"},{"comment":"All online methods are initialized with ground-truth states, as stated in Sec. V-C. This removes the initialization and convergence challenge that is often a major differentiator in practice, and it makes the reported improvement percentages (67–74% in translation) less generalizable to real deployment. In addition, no error bars or repeated trials are reported for any of the three online methods; Table II contains one run per sequence. Without this information, the statistical significance of the claimed improvements cannot be assessed.","section":"Sec. V-C, Initialization"}],"minor_comments":[{"comment":"The Huber weight formula appears to have a typo: w = min(1.0, τe/e) should presumably be w = min(1.0, τ/e). Please clarify.","section":"Eq. (16)"},{"comment":"The text refers to 'Sec.V.B' when comparing A2RL reprojection errors; the intended cross-reference is probably Sec. V-C (TII-RATM dataset experiments), not Sec. V-B (Metrics).","section":"Sec. V-D"},{"comment":"The abstract in the front matter reports 0.143 m average RMS translation error, while the abstract in the full text body reports 0.134 m, and the body text reports 0.141/0.145 m for the two subsets. These numbers should be made consistent throughout after the Table II issue is resolved.","section":"Abstract vs. full text"},{"comment":"Minor typographical issues such as 'MA Vs' in the introduction (should be 'MAVs') and the spacing in 'Kore an Goverment' (should be 'Korean Government') should be corrected.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The central ESKF formulation and the partial-gate ablation are promising, but the Table II inconsistency is a serious data-quality problem that must be fixed before any further acceptance decision. The self-referential ADR-FGO reference for the A2RL evaluation should be acknowledged explicitly as a consistency check rather than ground truth, or the claims should be rephrased. I do not see evidence of deliberate misreporting, but the paper needs a thorough data audit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You'll want to know two things about this paper: the core idea has a genuine niche — feeding gate-corner reprojection errors directly into an ESKF rather than through a PnP pose — and it performs well enough to be deployed at 20.9 m/s in A2RL Season 2. But the headline accuracy number in the abstract doesn't match the table, and the offline \"ground truth\" used for competition evaluation is partly derived from the estimator itself. Both problems need fixing before I'd trust the quantitative claims.\n\nWhat's actually new: the two-corner update mechanism is a clean way to stay in the filter loop when only part of a gate is visible, and the ablation in Fig. 5 (2 vs 4 vs 6 corners) supports the argument. The Huber reweighting instead of RANSAC is also sensible and the runtime table (under 9 ms) shows it's practical. The ADR-FGO smoother with visual-less keyframes and extrinsics refinement is a reasonable idea for uninstrumented arenas, though not revolutionary. The competition deployment is a strong point — real-world robustness at 20 m/s is not something every VIO paper has.\n\nThe soft spots. First, the abstract says 0.134 m average RMS translation error on TII-RATM, but Table II's per-sequence numbers sum to 1.434 m over 10 sequences, i.e., 0.1434 m. The piloted and autonomous sub-averages in the table (0.141 and 0.145) also average to 0.143, not 0.134. The paper's own numbers disagree with its headline. That's not a rounding nit; it's the central empirical claim. Second, the A2RL evaluation uses ADR-FGO as \"GT,\" but ADR-FGO is anchored by soft priors from ADR-VINS (Eq. 20) and consumes the same gate detections. So the 0.152 m error on the A2RL track is partly a consistency-with-self measure, not an independent accuracy assessment. The TII-RATM evaluation does have MoCap GT, so the 0.143 m claim is independent — but only after the discrepancy is resolved. Third, all online methods are initialized with ground truth, which is a common practical choice but inflates relative performance in a benchmark that doesn't report sensitivity to initialization. No error bars or repeated trials either.\n\nThe math itself is standard ESKF; I didn't see a load-bearing flaw there. The citation pattern is honest and covers the relevant ADR/VIO literature.\n\nWho gets value: people working on gate-based state estimation for racing or landmark-based VIO under high dynamics. It's a solid incremental system paper, but the evaluation needs to be cleaned up before publication. I'd send it to review — the idea and deployment are worth referees' time — but I'd expect a revision that fixes the number inconsistency and re-frames the A2RL numbers as consistency checks rather than independent errors.\n\nBest,","headline":"Gate-corner ESKF is a real ADR contribution, but the headline accuracy number contradicts the table and the A2RL reference is partly self-derived — fix those before trusting the numbers.","tokens_in":14839,"tokens_out":3271,"would_cite":false,"duration_ms":27488,"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 tightly-coupled filter that feeds raw gate-corner pixel errors directly into an error-state Kalman filter can keep a monocular racing drone localized with as few as two visible corners, and an offline factor-graph smoother supplies the re","keywords":["monocular visual-inertial state estimation","error-state Kalman filter","direct reprojection residual","gate-corner detection","autonomous drone racing","robust reweighting","factor-graph smoothing","GNSS-denied evaluation"],"falsifier":"Install a motion-capture system on a competition-style course and compare the offline smoother's trajectory to the motion-capture ground truth. If the smoother's deviation from motion capture is comparable to or larger than its reported gap from the online filter, the race reference is not independent and the 0.152 m RMSE is not an accuracy figure. A cheaper check: run the filter on a public dataset with motion-capture ground truth while varying the minimum visible corners; if two-corner updates do not beat four-corner updates there, the central claim fails.","tokens_in":13849,"feed_emoji":"🏁","tokens_out":6095,"duration_ms":54364,"temperature":0.7,"pith_summary":"The paper tries to establish that a drone-racing state estimator does not need the usual middle step of computing a camera pose from visible gate corners. Instead, raw pixel locations of gate corners can be fed directly into an error-state Kalman filter, so the filter stays valid even when only two corners of a gate are visible, and noisy outlier detections are handled by robust downweighting rather than a separate rejection stage. The same work introduces an offline factor-graph smoother that produces a high-fidelity reference trajectory after a flight, giving a way to score estimators in arenas without motion capture or GPS. If true, this makes high-speed monocular racing systems faster, more reliable during sharp maneuvers, and objectively comparable in uninstrumented venues.","feed_headline":"Two visible gate corners are enough to correct a racing drone's drift","feed_subtitle":"Pixel errors update the filter with a half-visible gate; a smoother scores flights without GPS or motion capture.","key_machinery":"The load-bearing object is the measurement residual r = u_tilde - h(x_bar, p_W_G): a direct pixel-space comparison between a detected gate corner and the projection of the known 3D gate corner through the current state. In the filter, this residual's Jacobian with respect to the error state replaces the pose-estimation layer, and its covariance is inflated by a robust weight w = min(1, tau/e) based on Mahalanobis distance. In the offline smoother, the same corner reprojection factors plus IMU preintegration and soft priors from the online filter are optimized in one batch to refine the trajectory and produce a reference for evaluation.","core_discovery":"ADR-VINS is an error-state Kalman filter that treats each 2D gate-corner detection as an innovation by projecting the known 3D corner into the image and computing a pixel residual. Because it bypasses the Perspective-n-Points solver and its RANSAC filtering, it needs only two visible corners for a valid update, and robust reweighting inflates the measurement covariance for outliers. ADR-FGO is a batch factor-graph optimization that combines IMU preintegration, gate-corner reprojection factors, soft priors from ADR-VINS, and a camera-IMU extrinsics factor to smooth the full trajectory and serve as a reference. On an open high-speed flight dataset with motion-capture ground truth, the filter r","pith_inferences":["The race-error numbers are consistency-with-a-smoothed-version rather than fully independent accuracy: because the smoother is anchored by soft priors from the online filter and consumes the same gate-corner detections, the reported 0.152 m may understate true error if the reference is pulled toward the estimator's drift.","With a two-corner minimum, the filter's accuracy becomes highly sensitive to systematic detection bias; biased corner labels could yield a self-consistent but offset trajectory, a failure mode worth testing by shifting detections in a motion-capture experiment.","The evaluation protocol could transfer to other structured environments—warehouse or agricultural flight, for instance—where a known map and distinctive corners replace racing gates.","The smoother's refined extrinsics factor could be fed back into the online estimator after a crash or hard impact, potentially removing a mechanical-shift failure mode."],"forward_implications":["Real-time systems no longer need four visible corners; partial gate views become usable updates, reducing drift between gates.","Eliminating the pose-solver and rejection stages cuts per-frame latency (reported under 9 ms, with detection dominating), making the pipeline fit tight camera deadlines.","The offline smoother gives a repeatable evaluation protocol for GNSS- and motion-capture-denied arenas, letting estimation, control, and total error be separated after a flight.","The same filter structure could generalize to any known-landmark task where distinctive corners can be detected in a monocular camera.","The smoother's refined trajectory could also be used to retune the online filter or to supervise learned motion models for visual outages."],"fun_headline_variants":["Drone racing filter needs only 2 visible gate corners","No PnP: ESKF drone estimator works with 2 gate corners","Racing drone state estimation: 2 corners beat 4-corner PnP","Monocular VIO for drone racing: robust to half-hidden gates","Autonomous drone racing: filter updates with just 2 corners"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation reference used for the uninstrumented race is an offline smoother that is anchored by soft priors taken from the same online estimator and consumes the same gate-corner detections, so if the smoother is dragged toward the estimator's drift, the race error numbers measure self-consistency rather than independent accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Drone racing filter needs only 2 visible gate corners","No PnP: ESKF drone estimator works with 2 gate corners","Racing drone state estimation: 2 corners beat 4-corner PnP","Monocular VIO for drone racing: robust to half-hidden gates","Autonomous drone racing: filter updates with just 2 corners"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1472,"prompt_tokens":883,"completion_tokens":589,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":627,"tokens_out":589,"duration_ms":5859,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:15:06.363878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Install a motion-capture system on a competition-style course and compare the offline smoother's trajectory to the motion-capture ground truth. If the smoother's deviation from motion capture is comparable to or larger than its reported gap from the online filter, the race reference is not independent and the 0.152 m RMSE is not an accuracy figure. A cheaper check: run the filter on a public dataset with motion-capture ground truth while varying the minimum visible corners; if two-corner updates do not beat four-corner updates there, the central claim fails.","supporting_citations":[],"review_version":1}