{"id":"8eaaea93-a6b0-4748-a209-2d859f5e88b8","arxiv_id":"2412.15000","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An embedded real-time 2D LiDAR person tracking pipeline, combining DR-SPAAM, Norfair, and TEB, achieves 85.45% average MOTA on three new datasets and improves collision avoidance on a quadruped robot.","lead":"This paper combines an existing 2D LiDAR person detector, an off-the-shelf tracker, and a local navigation planner, and tests the full pipeline on a quadruped robot with an embedded GPU. In three newly recorded motion-capture datasets, the preferred configuration achieved an average tracking accuracy of 85.45% while running at 20 Hz, and one avoidance experiment showed earlier collision avoidance than without the tracker.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MOTA values in Table II do not follow from Eq. (1) given the reported error counts; the headline 85.45% average is not internally consistent.","rationale":"After reading the paper, the most load-bearing issue is not the detector-transfer assumption on which the reader focused, but an internal inconsistency in the central reported metric. Table II is the only quantitative evidence for the headline 85.45% MOTA, yet the numbers in that table do not satisfy the paper's own MOTA definition. This is not a matter of generalization or missing baseline; it is a direct checkable arithmetic inconsistency. If the MOTA values are wrong, the central claim fails. If they are right, the table's labels or the 'Valid' column are wrong and must be corrected. Either way, the paper as submitted does not support its central number. I would keep the verdict as CONDITIONAL, so verdict_should_be is UNCHANGED relative to the reader, because the issue is resolvable: the authors can recompute and disclose their exact method. But the condition is now specific: the MOTA discrepancy must be resolved and the corrected numbers reported. I disagree with the reader's choice of weakest assumption as the primary concern; detector transfer is a secondary generalization concern.","tokens_in":9829,"tokens_out":8145,"duration_ms":65432,"concrete_test":"Recompute each row of Table II using Eq. (1) with the reported Valid/ID/Miss/FP counts. For MR2 Config-1, 1 − (47+1052+338)/5667 should equal 78.76% if the table is consistent; it instead equals 74.65%. Also recompute the Config-3 average from the corrected per-dataset MOTA values. If the discrepancy persists, request the authors' exact matching and aggregation script (ideally release data) to identify whether the error is in the counts, the 'Valid' denominator, or the MOTA computation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is the average MOTA of 85.45% for Config-3. Yet in Table II, each reported MOTA is higher than what Eq. (1) yields when applied to the 'Valid', ID Switch, Miss, and FP counts in the same row. For example, MR2 Config-1: 1 − (47+1052+338)/5667 = 74.65%, but the table reports 78.76% — a 4.11 percentage point gap. Even the most controlled row, SR Config-1, gives 94.07% instead of 94.40%. Recomputing the Config-3 average from the three rows using Eq. (1) gives (94.11+80.62+79.41)/3 ≈ 84.71%, not 85.45%. Unless the authors disclose a non-standard aggregation (e.g., per-frame averaging with zero-GT frames excluded), the numbers cannot be reproduced from the paper's own data. This matters because the abstract and conclusion rest on this specific MOTA figure; if the reported accuracy is inflated, the tracking claim — and the downstream avoidance benefit — lose their evidentiary base.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a modular 2D LiDAR person-tracking pipeline that combines the DR-SPAAM detector with a Norfair/SORT multi-object tracker, runs at 20 Hz on a Jetson Xavier NX, and is integrated with the TEB local planner on a Unitree A1 quadruped. Tracking accuracy is evaluated against Vicon motion-capture ground truth in three newly recorded indoor datasets (SR, MR1, MR2) for three tracker configurations; Config-3 is reported to achieve an average MOTA of 85.45%. The integrated navigation experiment compares TEB behavior with and without the tracker in a frontal-collision scenario. The central claims are real-time embedded operation, reliable person tracking, and improved collision avoidance.","tokens_in":9999,"tokens_out":6225,"duration_ms":53849,"significance":"If the numbers held, this would be a useful systems contribution: it demonstrates reuse of an open-source 2D LiDAR person detector without retraining on a new sensor, validates tracking against independent motion-capture ground truth, reports raw error counts and computation times, and integrates with a widely used local planner. The modular treatment of detection, tracking, and planning is a strength, and the detailed configuration table aids reproducibility. However, the headline MOTA figures are not internally consistent with the paper's own Eq. (1), and the avoidance claim rests on a single qualitative trial. These issues currently undermine the quantitative support for the main claims, although they appear addressable with corrections and additional experiments.","major_comments":[{"comment":"The MOTA values in Table II cannot be reproduced from the reported Valid/ID/Miss/FP counts using Eq. (1). For example, MR2 Config-1 gives 1−(47+1052+338)/5667 = 74.65%, not the reported 78.76%; even the most stable row, SR Config-1, gives 94.07%, not 94.40%. Recomputing the Config-3 average from the counts in the table gives approximately 84.72%, not the stated 85.45%, and the stated Config-2 average of 89.99% matches neither the average of the reported row values (88.77%) nor the count-based row values (87.46%). Unless a non-standard aggregation (e.g., per-frame MOTA averaging with special handling of zero-GT frames) is disclosed and justified, the headline MOTA figure is unsupported and must be corrected or explained.","section":"Table II, Eq. (1)"},{"comment":"The avoidance experiment is a single trial per condition and reports no quantitative outcome measures. The claim that the integrated tracker 'enhances collision avoidance' is supported only by two illustrative trajectories; there is no minimum-distance metric, no success rate over repeated trials, and no statistical comparison. This is load-bearing for the integrated-system contribution, and the paper should either add repeated runs with a quantitative safety or clearance metric or substantially soften the claim.","section":"Section IV-D, Fig. 4"},{"comment":"Each dataset is evaluated only once, so no error bars or run-to-run variability are reported. As a result, the configuration comparison—for instance, the statement that Config-1 and Config-2 have similar scores, or that Config-3's lower misses are worth its higher false positives—is not statistically grounded. At minimum, the authors should state explicitly that these are single-run observations and, ideally, report repeated trials or a sensitivity analysis over the hand-tuned parameters listed in Table I.","section":"Section IV-B, Table II"},{"comment":"The transfer of DR-SPAAM without retraining is a key enabling assumption for the claimed modularity and transferability, but it is not directly tested: there is no comparison against a detector fine-tuned on the Hokuyo sensor, and all evaluation is confined to one 4 m × 4 m arena with three participants. The Discussion acknowledges the open stride question for larger ranges, but the specific claim of 'resolution and distance independence' is only indirectly supported by the tracking benchmark. The authors should state this limitation more prominently and, if feasible, report detector-level results or a second, larger environment.","section":"Section III-B"}],"minor_comments":[{"comment":"The text says MOTP remains below 0.2 m/s, but MOTP is a distance and the table header correctly lists it in meters; please correct the units in the text.","section":"Section IV-B, Table II"},{"comment":"The sentence 'both Config-2 and Config-3 maintain an average processing time well under the scan period T_i_det of 50 ms' appears to use the wrong symbol: T_i_det is the detector inference time, not the scan period. Use T_scan or a clearly defined scan-period variable.","section":"Section IV-C"},{"comment":"The column headers of Table III are easy to misread: 'Tracker Detector T_i_det' is not a clear header. Please restructure the table so that Detector T_i_det, Tracker T_i_track, and Total T_i_lat are separate, unambiguous columns.","section":"Table III"},{"comment":"The description of the Norfair adaptation would benefit from a short pseudocode block or a precise parameter list (e.g., the matching threshold, Kalman process noise, and the exact logic for Cinit and Cdel), since these values are central to reproducing the benchmark.","section":"Section III-C"},{"comment":"Reference [32] contains a space in the URL ('2D lidar person detection'), which may break the link; please provide the correct URL.","section":"References"},{"comment":"There are several typographical issues, such as '20 Hzon' in the abstract and 'all configurations of our achieve' in Section IV-B; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The numerical inconsistency in Table II is the main blocker; it directly affects the abstract's headline result. The paper is otherwise a reasonable systems demonstration, and the missing baseline comparison and single-trial avoidance experiment are addressable within a revision. I do not think rejection is warranted if the MOTA numbers can be corrected or transparently explained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core here is a working, fully embedded stack: DR-SPAAM detection, Norfair/SORT tracking, and TEB planning on a Unitree A1 with a Hokuyo 2D LiDAR and Jetson Xavier NX. The three new datasets with Vicon ground truth are a real contribution, and the 20 Hz real-time claim is backed by measured latency. This is not a new algorithm paper, and it does not pretend to be; it is a clean combination of existing pieces, evaluated honestly in a close-range human environment.\n\nOne thing the stress-test note got wrong: the Table II MOTA values do follow from Eq. (1) if you use the correct denominator. The 'Valid' column is matched ground truth, not total ground truth; the denominator should be Valid + Miss. For SR Config-1, (9+424+7)/(7421+424) gives 94.39%, matching the reported 94.40%. The same check works for the other rows. So the 85.45% average for Config-3 is internally consistent, and I do not see inflated accuracy in the central tracking claim.\n\nThe real arithmetic problem is in the text: it says Config-2 has an average MOTA of 89.99%, but the three rows in Table II give (94.46+90.58+81.26)/3 = 88.77%. That is a visible inconsistency, probably a leftover number or a typo, and it needs to be corrected.\n\nThe genuine soft spots are the ones the reader flagged: no comparison against earlier 2D LiDAR trackers (the paper itself lists future work for this), a single avoidance experiment with no quantitative safety metric, and no released code or data. The detector transfer from JRDB to the new sensor geometry is plausible and is implicitly validated by the Vicon benchmark, but a dedicated sensor-transfer check would make it stronger. The lack of error bars or repeated trials limits how much weight you can put on small differences between configurations.\n\nOverall this is a solid systems paper, not a breakthrough. It would benefit from a serious referee who asks for baselines, more avoidance trials, and a release, but the main claim is backed by measured data and the logic is clear. I would engage with it, and I would cite it if I were working on embedded LiDAR person tracking.","headline":"A modest but genuine embedded-systems integration with a real Vicon benchmark; the headline MOTA is internally consistent, but one reported average in the text is wrong and the avoidance result is a single run.","tokens_in":10614,"tokens_out":3949,"would_cite":true,"duration_ms":31327,"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 modular 2D-LiDAR tracking pipeline lets a quadruped robot track people at 85.45% MOTA and avoid collisions in real time at 20 Hz, using a detector pretrained on a different sensor.","keywords":["2D LiDAR","person detection","multi-object tracking","collision avoidance","embedded GPU","TEB local planner","quadruped robot","DR-SPAAM"],"falsifier":"Record a new dataset with the same robot and sensor in a larger, more cluttered space where people are tracked at distances beyond the 4 m by 4 m arena and frequently leave the field of view; if Config-3's MOTA drops well below 85% or track initiation becomes too slow to support 0.5 m/s avoidance, the no-retraining transfer claim fails.","tokens_in":9575,"feed_emoji":"🤖","tokens_out":8033,"duration_ms":61133,"temperature":0.7,"pith_summary":"The paper sets out to show that a modular pipeline—detect people with a pretrained 2D-LiDAR network, track them with a lightweight multi-object tracker, and feed the predicted human motion into a local planner—can make a quadruped robot navigate safely around walking people using only an embedded GPU. It reports an average tracking accuracy (MOTA) of 85.45% across three newly recorded close-range datasets, with two configurations sustaining the full pipeline at 20 Hz on the Jetson Xavier NX. In navigation experiments, the tracker-equipped planner starts an avoidance maneuver early enough to prevent a frontal collision, while treating all obstacles as static reacts too late. The wider point is that a state-of-the-art person detector can transfer to a new sensor without retraining, so the expensive perception component can be reused while the tracking and planning pieces are swapped or tuned independently.","feed_headline":"A 2D-LiDAR tracker lets a quadruped dodge people at 20 Hz","feed_subtitle":"Pretrained detector transfers to a new 270-degree sensor, so person avoidance runs embedded in real time.","key_machinery":"The pipeline chains three components. Detection uses DR-SPAAM, a 1D convolutional network that processes a fixed one-meter window around each LiDAR point and votes on person locations; because the window is resampled to a fixed sample count, the detector is claimed to be independent of angular resolution and distance. Tracking uses a lightweight SORT-style multi-object tracker: a Kalman filter with a constant-velocity model for motion, Hungarian assignment with an Euclidean cost for data association, and counters that initiate tracks after enough consecutive matches and delete them after enough misses; detections are transformed into the robot's odometry frame first. Planning uses the TEB local planner, which optimizes trajectories and can incorporate dynamic obstacles with estimated velocities; the authors filter out static obstacles around the tracked person so the planner does not see the person as a fixed wall.","core_discovery":"On the paper's own terms, the central claim is that a detection-tracker-planner pipeline can be assembled from existing components—a pretrained 1D-convolutional person detector, a lightweight SORT-style tracker with constant-velocity Kalman filtering, and the TEB local planner—and run entirely on an embedded GPU at 20 Hz while tracking people accurately enough for collision avoidance. The benchmark results give an average MOTA of 85.45% for the favored configuration across three newly recorded close-range datasets, and the navigation experiments show that feeding the tracker's velocity estimates into TEB lets the robot begin an avoidance maneuver early enough to avoid a frontal collision, whereas treating all obstacles as static does not. The authors take the high scores on a new 270° LiDAR without retraining as experimental confirmation that the detector is resolution- and distance-independent.","pith_inferences":["If the detector really is sensor-independent as claimed, the same modular pipeline should transfer to other 2D LiDARs with different resolutions and mounting heights, but only for the close-range distances tested; longer-range behavior remains unverified.","The constant-velocity motion model is the paper's weakest motion assumption: sudden stops, sharp turns, or socially interactive movement could break both tracking association and the planner's predicted obstacle trajectories.","Because the planner removes static obstacles around a tracked person, a lost track during occlusion could leave the person effectively invisible to the costmap; identity maintenance under re-entry is therefore the practical bottleneck for safety.","A direct testable extension is to filter detections near walls with the local costmap, which the paper names as a limitation; doing so should reduce the false positives observed while the robot is moving and raise MOTA in the moving-robot datasets."],"forward_implications":["Two of the three configurations sustain the full detection-tracking pipeline at 20 Hz on a Jetson-class embedded GPU, so person-aware avoidance does not require off-board computation.","In the navigation test, the tracker-equipped robot starts its avoidance maneuver early enough to prevent collision with a person walking at roughly 1 m/s, while the static-obstacle version reacts too late.","Config-3's fast track initiation lets the system react when a person emerges from behind a wall, which is the critical case for close-range safety.","Config-1 and Config-2 achieve nearly equal MOTA, showing that skipping detection windows has little effect at close range, so the faster configuration loses almost no tracking quality."],"supporting_citations":[{"why":"Supplies DR-SPAAM, the pretrained 2D-LiDAR person detector whose resolution/distance independence lets the authors skip retraining for their new sensor.","marker":"[9]"},{"why":"JRDB is the dataset the detector was trained on and the transfer source for the new sensor configuration.","marker":"[16]"},{"why":"Provides the benchmark framework and matching pipeline used to compute MOTA/MOTP, and an earlier person tracking/following baseline.","marker":"[17]"},{"why":"Defines the SORT tracking approach (Kalman filter plus Hungarian association) that the adapted tracker generalizes.","marker":"[22]"},{"why":"The lightweight multi-object tracker library that is adapted to LiDAR person detections, supplying the tracking backbone.","marker":"[34]"},{"why":"The TEB local planner that consumes tracked dynamic obstacles with velocities for collision avoidance.","marker":"[26]"},{"why":"Defines the CLEAR MOT metrics (MOTA, MOTP) used for the quantitative benchmark.","marker":"[35]"}],"fun_headline_variants":["Quadruped dodges humans with 20 Hz 2D-LiDAR tracking","Embedded 2D-LiDAR person tracker lets robot avoid humans at 20 Hz","Quadruped uses LiDAR person tracking for real-time collision avoidance","Person tracker at 20 Hz on embedded LiDAR helps quadruped navigate","20 Hz 2D-LiDAR tracking lets a quadruped avoid pedestrians"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach stands on the assumption that the pretrained detector, trained on a different 360° LiDAR dataset, keeps detecting people accurately from a new 270° sensor mounted 45 cm high on a moving robot without retraining; the benchmarks only demonstrate this inside a 4 m by 4 m room.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped dodges humans with 20 Hz 2D-LiDAR tracking","Embedded 2D-LiDAR person tracker lets robot avoid humans at 20 Hz","Quadruped uses LiDAR person tracking for real-time collision avoidance","Person tracker at 20 Hz on embedded LiDAR helps quadruped navigate","20 Hz 2D-LiDAR tracking lets a quadruped avoid pedestrians"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3383,"prompt_tokens":959,"completion_tokens":2424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2321}},"tokens_in":575,"tokens_out":2424,"duration_ms":13462,"temperature":1.0,"reasoning_tokens":2321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:43:00.082162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a new dataset with the same robot and sensor in a larger, more cluttered space where people are tracked at distances beyond the 4 m by 4 m arena and frequently leave the field of view; if Config-3's MOTA drops well below 85% or track initiation becomes too slow to support 0.5 m/s avoidance, the no-retraining transfer claim fails.","supporting_citations":[{"cited_title":"Dr-spaam: A spatial-attention and auto-regressive model for person detection in 2d range data,","cited_arxiv_id":null,"evidence_quote":"Supplies DR-SPAAM, the pretrained 2D-LiDAR person detector whose resolution/distance independence lets the authors skip retraining for their new sensor."},{"cited_title":"Jrdb: A dataset and bench- mark of egocentric robot visual perception of humans in built environ- ments,","cited_arxiv_id":null,"evidence_quote":"JRDB is the dataset the detector was trained on and the transfer source for the new sensor configuration."},{"cited_title":"Integrated online trajectory planning and optimization in distinctive topologies,","cited_arxiv_id":null,"evidence_quote":"The TEB local planner that consumes tracked dynamic obstacles with velocities for collision avoidance."},{"cited_title":"Evaluating multiple object tracking performance: the clear mot metrics,","cited_arxiv_id":null,"evidence_quote":"Defines the CLEAR MOT metrics (MOTA, MOTP) used for the quantitative benchmark."}],"review_version":1}