{"id":"464fdcc1-cdde-4303-8302-1f7f593f368d","arxiv_id":"2506.16301","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"M-PSpliner adds multi-opponent tracking and Gaussian process trajectory prediction to the PSpliner planner, raising measured overtaking success from 81.53% to 91.65% on a 1:10 scale car.","lead":"This paper extends a single-opponent overtaking planner for 1:10 scale autonomous race cars to handle multiple opponents at once, using a Kalman filter tracker and Gaussian process predictions of each opponent's path and speed. The system reports an average overtaking success rate of 91.65%, about 10 percentage points higher than the previous single-opponent method at the same speed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M-PSpliner's 'arbitrary multi-opponent' claim is not supported by its single-RoC design: §III-B.2 predicts only the closest opponent's collision region and treats all others spatially, so overlapping or non-closest future risks are not planned for.","rationale":"The reader's weakest_assumption identified the same single-RoC simplification, and I agree that it is load-bearing because the paper's headline contribution is specifically arbitrary multi-opponent, spatiotemporal overtaking. The paper's own text in Sections III-B.2 and III-B.3 reduces all but the closest opponent to spatial obstacles, so the central claim covers a broader regime than the described algorithm can plan for. I see no internal inconsistency in the derivation; the gap is between the claim and the demonstrated domain. The placeholder code repository at github.com/ForzaETH/notavailableyet and the computational table measured on an AMD Ryzen 7 PRO 7840U rather than the stated Intel NUC i7-1165G7 are real reproducibility issues, but they are secondary to the algorithmic coverage gap. The proposed simulation check directly targets the overlapping-RoC regime where the assumption would break; if it passes, the remaining weaknesses are statistical reporting and platform verification, and if it fails, the claim must be narrowed to non-overlapping multi-opponent scenarios. In either case the reader's CONDITIONAL verdict remains appropriate, so I recommend no change.","tokens_in":9407,"tokens_out":8184,"duration_ms":96754,"concrete_test":"Re-run M-PSpliner in the ForzaETH stack or a faithful simulation with two opponents whose GP-predicted RoCs overlap: place Opp1 immediately ahead on the racing line at Smax around 77%, and Opp2 on the shortest-path line about 0.5 to 1 car length ahead so its predicted RoC intersects Opp1's RoC before the first overtake completes. Record collisions and trajectory adjustments, and compare against a control variant that computes one RoC per opponent. If the single-RoC version collides or cannot complete the overtake while the control succeeds, the Section III-B.2 assumption fails in exactly the regime the paper dismisses; if it succeeds, the simplification is adequate for that regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is 'arbitrary multi-opponent head-to-head racing' with spatiotemporal opponent information. That claim depends on the planner actually using each opponent's future intent. Section III-B.2 computes a single RoC from the closest opponent, justified only by an unspecified empirical analysis that multiple RoCs are expensive and unlikely to overlap; Section III-B.3 then handles every other opponent 'spatially if present.' Thus the spatiotemporal prediction applies to exactly one opponent. If two opponents' predicted RoCs overlap, or if the non-closest opponent is the one whose future motion constrains the overtake (for example, the closest opponent is slow and pulls wide while the second is on an intersecting line and will arrive at the same corner later), the planner has no mechanism to use that opponent's GP prediction. The two-opponent experiment in Table IV does not exclude this regime, because the reported scenario has a shortest-path front opponent and a centerline follower with sequential, non-overlapping encounters; only three trials are averaged and no error bars are given. The claimed 10.13-point safety improvement is therefore demonstrated only for single-opponent and non-overlapping multi-opponent situations, not for arbitrary multi-opponent racing. This is a correctness gap, not a disagreement with consensus: the described planner simply omits the temporal coupling between multiple RoCs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the single-opponent Predictive Spliner (PSpliner) overtaking planner to multiple opponents, naming the result M-PSpliner. The system uses a Kalman-filter-based multi-object tracker with re-identification to maintain opponent identities, fits Gaussian process models of each opponent's spatial trajectory and velocity, and then plans an overtake using a single future Region of Collision (RoC) computed from the closest opponent; other opponents are handled purely spatially. The method is evaluated on a physical 1:10 scale autonomous racing car, with four single-opponent behavior conditions and one two-opponent scenario. The reported results claim an overtaking success rate up to 91.65%, an average 10.13-percentage-point improvement over PSpliner, and nearly equal per-overtake time in the two-opponent case compared with the single-opponent case.","tokens_in":9686,"tokens_out":5618,"duration_ms":65078,"significance":"If the results hold, this is a useful step toward practical multi-opponent autonomous racing: the work is openly built on the ForzaETH stack, provides physical-vehicle validation rather than simulation only, and shows that multi-opponent tracking and simultaneous GP construction are feasible on embedded hardware. The CPU and latency measurements are also a useful comparison point for the community. However, the paper's central claim of 'arbitrary multi-opponent head-to-head racing while considering the opponents' future intent' is not actually implemented, because only the closest opponent receives a temporal RoC; all other opponents are reduced to spatial obstacles. In addition, the quantitative evidence for the headline safety improvement is thin: success rates are point estimates from experiments stopped after five successful overtakes, and the two-opponent experiment consists of three trials with no error bars. The contribution is therefore real but narrower than claimed, and the statistical support needs strengthening before the quantitative headline numbers can be accepted as stated.","major_comments":[{"comment":"The planner computes exactly one RoC, from the closest opponent, and Section III-B.3 states that all other opponents are accounted for 'spatially if present.' Therefore the abstract's claim of 'arbitrary multi-opponent head-to-head racing while considering the opponents' future intent' overstates the method: only one opponent's future intent enters the spatiotemporal planning. No experiment in Section IV-B includes overlapping RoCs or a case where the non-closest opponent is the limiting constraint; the Table IV scenario has sequential, non-overlapping encounters (Figure 2). Please either compute and plan with multiple RoCs, or restrict the claims to 'closest-opponent spatiotemporal planning with spatial avoidance of other opponents.'","section":"III-B.2, III-B.3; Abstract"},{"comment":"The 10.13-percentage-point improvement is based on experiments stopped after five successful overtakes per condition, and the displayed Rot/c values are coarse counts (e.g., 71.4 = 5/7, 83.3 = 5/6, 100 = 5/5). With these counts, the difference between 71.4% and 83.3% is a single crash, and the four-condition average is compatible with large sampling variability. No confidence intervals, raw crash counts, or significance tests are reported. Please report the raw counts with intervals, or soften the safety-improvement claim to reflect the small-sample nature of the measurement.","section":"IV-B.1, Table II"},{"comment":"The multi-opponent evaluation uses three trials per condition and reports only means, with no variance, error bars, or individual trial values. Moreover, the test scenario places Opp1 on the shortest path and Opp2 on the centerline, with the ego overtaking them one after another, so it does not exercise simultaneous or overlapping threats. This is insufficient to establish 'arbitrary' multi-opponent capability; additional scenarios that stress the planner with spatially or temporally overlapping opponents, or an explicit statement of this limitation, are needed.","section":"IV-B.2, Table IV"},{"comment":"The decision to use a single RoC is justified by an 'empirical analysis' that is neither described nor referenced. The reader cannot check the claims that computing multiple RoCs is too expensive or that overlapping RoCs are unlikely. Please provide the analysis, report the measured computational cost and overlap statistics, or remove the unsupported justification.","section":"III-B.2"}],"minor_comments":[{"comment":"The text says opponents 'must be trailed for one lap' to construct their GPs, but Section IV-B.2 says GP2 is learned while overtaking Opp1. Please clarify whether a full lap of observations per opponent is required or whether partial observations are sufficient.","section":"III-B.1"},{"comment":"There are typos in the text before Table III ('handle multiple oppnents') and in the table caption ('algorihtm'). These should be corrected.","section":"Table III and surrounding text"},{"comment":"The notation 'P T' and 'PPP T' is confusing; please define the total race time symbol cleanly and use it consistently in the table and text.","section":"IV-B.2, Table IV"},{"comment":"The code link github.com/ForzaETH/notavailableyet appears to be a placeholder; this contradicts the claim that 'everything in this project is fully open-sourced.' Please provide the actual repository URL or state clearly that the code will be released later.","section":"V, Open-source availability"},{"comment":"The CPU utilization is reported per core, which is noted in the table caption, but the high standard deviations (e.g., 30.57 for M-PSpliner vs 31.87 for PSpliner) make the reported CPU difference between M-PSpliner and PSpliner statistically unconvincing; please comment on this variability.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is already accepted at IROS 2025. If the intended venue is a journal, the current evidence level is closer to a conference contribution than to a journal paper, and the central 'arbitrary multi-opponent' claim needs either a methodological change (multiple RoCs) or a significant reframing. The placeholder repository URL should also be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know this paper is a competent engineering extension of PSpliner to multiple opponents, but the 'arbitrary multi-opponent' claim in the abstract is not supported by the design or the experiments. The planner computes a single Region of Collision from the closest opponent's GP and treats every other opponent spatially. The two-opponent experiment has sequential, non-overlapping encounters, so it never stresses the temporal coupling between opponents. That gap is real, and the stress-test note lands.\n\nWhat's genuinely new and good: the tracker is a solid piece of work. Adapting Norfair to spatial LiDAR tracking with reID, and building GPs for multiple opponents simultaneously even under occlusion, is a useful step. The physical experiments are more than a simulation: they run on a 1:10 car with a real opponent, and the improvement over PSpliner in the tested single-opponent conditions (10.13 pp success rate) is believable as an empirical result. The multi-opponent lap time staying flat is also a nice demonstration.\n\nThe soft spots are mostly about overclaiming and evidence thickness. The success rates are point estimates from experiments stopped after five successful overtakes; no confidence intervals. The multi-opponent table averages three trials, no error bars. There's no head-to-head comparison with another multi-opponent planner, so 'previous SotA' really means 'their own single-opponent PSpliner.' The code link is a placeholder. And the computational numbers were measured on an AMD Ryzen, while the vehicle's stated computer is an Intel NUC, which makes the real-time claim less direct.\n\nNone of these are fatal. The single-RoC design may be a sensible engineering trade-off, but calling it 'arbitrary multi-opponent' is simply not accurate. The paper would be stronger if it said 'up to two opponents' and explicitly discussed the overlap limitation. I'd want the authors to add error bars, run more multi-opponent trials with overlapping trajectories, and ideally compare against a spatial-only baseline that ignores GP predictions, to isolate the value of the temporal model.\n\nThis paper deserves a serious referee. It's a real system with real experiments and an honest comparison against their own prior work. The contribution is incremental but useful for the scaled-racing community and for anyone doing 2D LiDAR-based multi-agent tracking. I'd engage with it, but I'd push for revision before accepting.\n\nRecommendation: send it to peer review with a request for major revision.","headline":"Solid multi-opponent extension of PSpliner with real experiments, but the 'arbitrary' claim outruns a design that predicts only the closest opponent's collision region.","tokens_in":10214,"tokens_out":2765,"would_cite":false,"duration_ms":30913,"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 single racing planner can overtake multiple opponents by predicting each opponent's future trajectory and planning against the closest one.","keywords":["autonomous racing","multi-opponent overtaking","spatiotemporal planning","Gaussian process regression","Kalman filter tracking","re-identification","motion prediction"],"falsifier":"Run the planner with two opponents whose Gaussian-process trajectory predictions overlap at the same corner; if the ego vehicle cuts toward the second opponent's future position or the overtake success rate falls to the spatial-baseline level, the single-Region-of-Collision assumption is falsified.","tokens_in":9199,"feed_emoji":"🏎️","tokens_out":8183,"duration_ms":83163,"temperature":0.7,"pith_summary":"The paper sets out to remove the single-opponent restriction of Predictive Spliner, a spatiotemporal overtaking planner that uses Gaussian-process regression to predict an opponent's trajectory and computes a future Region of Collision before committing to a pass. It argues that a Kalman-filter tracker with re-identification can keep stable identities for several opponents at once, and that regressing each opponent's spatial path and velocity separately supplies enough predictive information to overtake them in sequence. On a physical 1:10 scale car, the method reports an overtaking success rate up to 91.65 percent against a single opponent and an average 10.13-percentage-point improvement in safety over the prior planner, while the total time to overtake two opponents is nearly identical to the single-opponent time. If the claims hold, predictive, intent-aware overtaking is no longer limited to one rival and can be run on computationally modest hardware.","feed_headline":"Overtaking planner uses trajectory predictions to pass multiple rivals","feed_subtitle":"On a 1:10 scale race car, it raised success to 91.65 percent while two-overtake time stayed flat","key_machinery":"The central object is the spatiotemporal Region of Collision (RoC): the planner integrates the closest opponent's Gaussian-process-regressed velocity along its spatial trajectory to predict where and when the paths will overlap, then optimizes the ego trajectory to pass within that region. The supporting mechanism is a multi-opponent tracker built on a Kalman filter with re-identification, which associates LiDAR detections to persistent tracklets and lets the method build a separate Gaussian process for each opponent's spatial path and velocity. A sequential quadratic programming problem then adjusts the ego vehicle's lateral position while respecting track boundaries, curvature limits, and a minimum clearance that keeps other opponents as spatial constraints.","core_discovery":"The central claim is that arbitrary multi-opponent overtaking can be handled by a single planner that computes one Region of Collision from the closest opponent's predicted trajectory and treats every other opponent as a spatial obstacle during the pass. To make this possible, the paper extends the perception stack with a Kalman-filter multi-opponent tracker that re-identifies cars across LiDAR detections, so each opponent can be trailed and modeled by its own spatial and velocity Gaussian process even when occluded by another car or track curvature. In experiments on a 1:10 scale autonomous vehicle, the planner matched or beat the single-opponent baseline in one-opponent races, with a 91.65 percent success rate and an average safety improvement of 10.13 percentage points, and it achieved successful overtakes at up to 76.75 percent of the ego vehicle's speed. With two opponents, the average lap time for an overtaking lap was 9.275 seconds versus 9.25 seconds for one opponent, which the paper presents as evidence that the second opponent's model is learned while the first overtake is underway.","pith_inferences":["The single-RoC approximation is the point most worth probing: the paper justifies it by saying overlapping RoCs were unlikely, so a natural stress test is a track layout where two opponents' predicted futures cross at the same corner.","The one-lap trailing requirement means the method assumes a warm-up period for every opponent; replacing that with a generic prior or online model would make the planner usable in race entries and after tracking failures.","The same architecture, detection, identity-preserving tracking, per-agent trajectory regression, and a single predicted collision region, transfers to any high-speed navigation task with several predictable dynamic agents, not just racing."],"forward_implications":["A planner that predicts opponent intent can pass multiple cars in sequence without allocating a separate predicted collision region to each one.","Re-identification lets the system keep modeling an opponent that disappears behind another car or around a bend, so the second overtake can be planned while the first is still being executed.","The reported timing shows the cost of the second overtake is almost invisible: two overtakes took 37.1 seconds total, against 37.0 seconds for two single-overtake races.","The method's CPU usage and latency stay in real-time range on embedded hardware, so multi-opponent prediction does not require a large compute budget."],"supporting_citations":[{"why":"defines the PSpliner baseline, the RoC computation, and the sequential quadratic programming overtaking optimization that M-PSpliner extends.","marker":"[4]"},{"why":"supplies the open-source 1:10 scale racing stack, LiDAR detection, track-boundary filtering, and the single-opponent tracker used as baseline.","marker":"[3]"},{"why":"provides the Kalman-filter-based tracker software whose data association the paper adapts for spatial multi-opponent re-identification.","marker":"[9]"},{"why":"contributes the tracking-by-detection structure, including Kalman updates and Hungarian association, that the re-identification tracker is built on.","marker":"[15]"},{"why":"shows how to adapt the image-based tracker to operate on real-world spatial LiDAR detections.","marker":"[16]"},{"why":"defines the reactive Follow-the-Gap behavior used as one of the opponent racing behaviors in the overtaking experiments.","marker":"[6]"}],"fun_headline_variants":["AI racer passes multiple opponents using predicted trajectories","Autonomous racer overtakes multiple rivals with intent predictions","Multi-opponent overtaking planner hits 91.65% success on 1:10 car","Predicting rivals' paths enables multi-car overtakes in racing","Racer planner uses trajectory forecasts to pass several cars"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The planner assumes that a single predicted collision region computed from the closest opponent is sufficient and that every other opponent can be treated as a static spatial obstacle, so the method can fail if two opponents' predicted regions overlap or if the second opponent is the one that actually threatens the overtake.","fun_headline_variants_meta":{"raw":{"variants":["AI racer passes multiple opponents using predicted trajectories","Autonomous racer overtakes multiple rivals with intent predictions","Multi-opponent overtaking planner hits 91.65% success on 1:10 car","Predicting rivals' paths enables multi-car overtakes in racing","Racer planner uses trajectory forecasts to pass several cars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1229,"prompt_tokens":941,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":200}},"tokens_in":557,"tokens_out":288,"duration_ms":3817,"temperature":1.0,"reasoning_tokens":200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:44:16.128515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the planner with two opponents whose Gaussian-process trajectory predictions overlap at the same corner; if the ego vehicle cuts toward the second opponent's future position or the overtake success rate falls to the spatial-baseline level, the single-Region-of-Collision assumption is falsified.","supporting_citations":[{"cited_title":"Predictive spliner: Data-driven overtaking in autonomous racing using opponent trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"defines the PSpliner baseline, the RoC computation, and the sequential quadratic programming overtaking optimization that M-PSpliner extends."},{"cited_title":"Forzaeth race stack—scaled autonomous head-to-head racing on fully commer- cial off-the-shelf hardware,","cited_arxiv_id":null,"evidence_quote":"supplies the open-source 1:10 scale racing stack, LiDAR detection, track-boundary filtering, and the single-opponent tracker used as baseline."},{"cited_title":"Simple online and realtime tracking,","cited_arxiv_id":null,"evidence_quote":"contributes the tracking-by-detection structure, including Kalman updates and Hungarian association, that the re-identification tracker is built on."},{"cited_title":"Autonomous navigation in dynamic human environments with an embedded 2d lidar-based person tracker,","cited_arxiv_id":null,"evidence_quote":"shows how to adapt the image-based tracker to operate on real-world spatial LiDAR detections."},{"cited_title":"A novel obstacle avoidance algorithm: “follow the gap method","cited_arxiv_id":null,"evidence_quote":"defines the reactive Follow-the-Gap behavior used as one of the opponent racing behaviors in the overtaking experiments."}],"review_version":1}