{"id":"4c0a8a1d-f8a6-404e-a6d3-ab0d49258fae","arxiv_id":"2501.12799","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Int2Planner samples intention points from the ego vehicle's route path to generate multi-modal planning trajectories, improving integrated prediction and planning on nuPlan and a private dataset.","lead":"The paper presents Int2Planner, a transformer-based model that combines trajectory prediction and motion planning for autonomous vehicles by sampling 'intention points' from the route ahead. The authors report improved planning accuracy over several baselines and say they have driven hundreds of kilometers with it in real urban traffic.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4 m route-intention grid (dr=4, Nq=64) is assumed to cover the optimal plan endpoint, but this coverage is never measured against GT endpoints; an uncovered GT endpoint cannot be represented or selected, undermining the central claim.","rationale":"Good-faith reading: the paper's strongest, most novel claim is that route-conditioned intention queries are a better inductive bias for integrated planning than static or clustered anchors. The internal ablations (Tables 3 and 2) genuinely support this direction: replacing cluster anchors with route anchors improves both planning and prediction metrics, and the real-world appendix is unusually candid about safety. But the mechanism by which route points help is exactly their ability to cover the space of plausible plans. The architecture's loss assigns each GT plan to the nearest sampled intention point, so if no point lies near the GT endpoint, that GT plan is simply unrepresentable; no amount of decoder refinement can be trained to reach it. The sampling interval dr=4 m and query count Nq=64 are chosen without reporting the induced distance-to-nearest-point distribution, making the coverage assumption the softest load-bearing step. It is directly testable and likely fixable (sample more densely or add learned offsets), so this does not invalidate the paper; it does make full acceptance premature until the check is reported. Secondary issues (no error bars, unreleased code and dataset, the Val14 SOTA overstatement) reinforce the conditional verdict but are not the single load-bearing premise. The reader's weakest_assumption matched this concern, so agreement is 'agree'. The verdict should remain conditional: accept the architectural idea as promising, but require the coverage validation and the availability fixes before treating the SOTA claim as established.","tokens_in":13297,"tokens_out":7394,"duration_ms":80266,"concrete_test":"On the private validation set (54,505 scenarios), compute for each GT planning trajectory the minimum Euclidean distance from the GT endpoint to the route intention point set used at inference (dr=4 m, Nq=64, primary+secondary); report the fraction of scenes with this distance above 2 m and above 4 m. Then train/evaluate Int2Planner with dr=2 m (Nq≈128) and dr=8 m (Nq=32) on the same split, keeping all else fixed. If the dr=2 m model does not improve plan ADE/FDE and the above-4 m fraction is negligible, the 4 m grid is adequate; if either condition fails, the central claim must be qualified because the planner's representable mode set is too coarse.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the 64 route intention points sampled every 4 m along primary and secondary routes (Section 3, Eq. 4; implementation details) cover all plausible ego-planning endpoints. This is never checked. The loss assigns exactly one intention point as positive: the one closest to the GT endpoint (Loss Function). If the GT endpoint lies more than a few meters from every sampled point, no mode is responsible, the decoder is never trained to produce that trajectory, and the closed-loop/real-vehicle policy of taking the highest-confidence mode can commit to a trajectory whose endpoint is nowhere near the intended one. The authors' own limitation paragraph concedes that the highest-confidence trajectory is not necessarily optimal, which makes the coverage assumption more, not less, critical. Table 2 and Table 3 compare route versus K-means cluster intention points, but the K-means anchors are global and context-free; the observed gain could come from being scene-conditioned and route-constrained rather than from the specific 4 m spacing. No analytic or empirical bound links dr/Nq to plan ADE/FDE, and no distribution of GT-to-nearest-intention-point distances is reported. Until this is measured, the central architectural claim is conditional on an unvalidated quantization of the plan space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Int2Planner, a transformer-based integrated motion prediction and planning model for autonomous driving. For the ego vehicle, a set of 'route intention points' is sampled at a fixed interval along the primary and secondary route polylines; these points initialize query embeddings in a multi-modal trajectory decoder that outputs one planning trajectory per intention point, along with confidence scores. The same decoder is shared for multi-agent trajectory prediction. The method is evaluated on the closed-loop nuPlan benchmark (Val14 and Test14-hard) and on a large private dataset, with ablations comparing route intention points against K-means-derived cluster intention points and isolating the effect of integrated prediction. The authors also report real-vehicle tests covering hundreds of kilometers in urban areas.","tokens_in":13650,"tokens_out":9268,"duration_ms":88501,"significance":"The paper makes a credible case that route-constrained intention points are a better inductive bias for ego planning than static, globally clustered anchors, which is of practical relevance for learning-based planners. The planned release of a large private dataset and code is a positive step for reproducibility. However, the empirical evidence as presented is not yet sufficient to support the strongest claims: the sampling-grid coverage is unexamined, the open-loop planning metrics are under-specified, the ablation for intention-point type is confounded by scene-conditioning, and the claimed state-of-the-art results require qualification. The manuscript would benefit from targeted additional experiments and clarifications.","major_comments":[{"comment":"The route intention points are sampled at a fixed interval dr = 4 m with Nq = 64, but the paper never reports how far ground-truth planning endpoints are from the nearest sampled intention point on the validation set, nor does it ablate dr or Nq. Since the loss assigns the positive mode as the intention point closest to the GT endpoint (Loss Function paragraph), any GT endpoint that is far from every grid point is assigned to a poorly conditioned mode and the decoder is forced to extrapolate beyond its query. The closed-loop and real-vehicle deployment then selects the highest-confidence trajectory (Section 5), so an uncovered endpoint can directly cause a poor commitment. Please report the GT-to-nearest-intention-point distance distribution on the validation set and include ablations over dr (e.g., 2, 4, 8 m) and Nq (e.g., 32, 64, 128) to establish that the reported gains are not an artifact of the specific grid.","section":"§3, Eq. (4); §4 Implementation Details"},{"comment":"The planning ADE/FDE metrics are not defined. It is unclear whether they are measured on the highest-confidence mode, the best-of-N mode (minimum over the 64 outputs), or the mode corresponding to the intention point closest to the GT endpoint. This matters because the closed-loop and real-vehicle policies use the highest-confidence trajectory, while the open-loop numbers may reflect a different selection rule. Please specify the metric definitions and report both best-mode and confidence-selected planning ADE/FDE.","section":"§4 Datasets and Metrics; Tables 2 and A2"},{"comment":"The comparison between cluster intention (CI) and route intention (RI) does not control for scene-conditioning. CI uses global K-means centers of GT endpoints, which are identical for all scenes, whereas RI anchors are re-sampled from the per-scene route. The observed improvement could therefore be attributed to the anchors being scene-conditional and route-constrained rather than to the specific route-sampling scheme. Please add a baseline with a matched number of scene-conditioned anchors (e.g., lane-center samples or per-scene projected cluster centers) to isolate the contribution of the route-based sampling.","section":"§4 Ablation Study; Table 3"},{"comment":"The abstract's claim that 'Int2Planner achieves state-of-the-art performance' is not supported by the full results. On the Val14 benchmark in Table 1, the best Int2Planner variant (0.8385 overall) is below PDM-Hybrid (0.8967) and only on par with PlanTF (0.8360); the top score is achieved only on Test14-hard. Moreover, the comparison in Table 1 mixes models with different training data sizes and post-processing, and the private-dataset evaluation in Table 2 includes a single baseline (GameFormer). Please qualify the SOTA claim to the specific benchmark (Test14-hard relative to the compared planners) and expand the number of baselines on the private dataset.","section":"§4 Main Results; Table 1; Abstract"},{"comment":"No error bars or multiple-seed runs are reported for any table. The differences in Tables 4-6 are small (e.g., NR-CL 0.6784 vs 0.6971 for integrated prediction), so it is possible that the reported improvements are within run-to-run noise. Please report mean and standard deviation over at least three random seeds for the main results and ablations, and state whether the differences are statistically significant.","section":"All experimental tables; especially Tables 4-6"}],"minor_comments":[{"comment":"The abstract contains a typo: 'avaliable' should be 'available'; additionally, the sentence 'we construct Int2Planner, an Intention-based Integrated motion Planner achieves multi-modal planning' is grammatically incomplete and should be reworded.","section":"Abstract"},{"comment":"The references Hu et al. 2023a and Hu et al. 2023b appear to be the same paper (identical title and venue) and should be merged into a single citation.","section":"References"},{"comment":"There are missing spaces in several places, e.g., 'useth = 15' and 'tf = 50future'; these should be corrected for readability.","section":"§4 Implementation Details"},{"comment":"The caption lists subfigures (a)-(d), but the data flow between the modules is not explained in the text; a short walk-through of the figure would greatly improve readability.","section":"Figure 1"},{"comment":"The qualitative claims of 'safe and reasonable planning trajectories' would be more convincing with quantitative safety indicators, such as the number of take-overs or a breakdown by scenario type.","section":"Real-world Vehicle Test"},{"comment":"There is a typo in the table: 'PDM-Hybird' should be 'PDM-Hybrid'.","section":"Table A1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central idea is interesting, but the current evaluation does not fully isolate the proposed route-intention mechanism from other factors, and the sampling-grid coverage is unvalidated. I believe the paper can be made publishable with the requested revisions, particularly the coverage analysis, metric clarification, and ablation controls. The overbroad SOTA claim in the abstract should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is simple and defensible: instead of using global cluster centers as intention queries for the ego planner, sample points along the primary and secondary route at a fixed spacing. That gives the decoder a route-aware, scene-conditioned set of modes. The ablations back this up—route intention beats cluster intention on both the private set and Test14-hard, and integrated prediction helps. The real-vehicle deployment, with operator oversight and a candid safety caveat, is more than most planning papers bother to include. Credit where due: the writing is clear, the experiments are reasonably thorough, and the mechanism is concrete enough to adopt or test in other work.\n\nThe main soft spots are three. First, the 'state-of-the-art' claim in the abstract does not hold on Val14: Int2Planner scores 0.8226 overall versus PlanTF's 0.8360. They are honest in the text ('competitive with PlanTF') but the abstract overreaches. Second, there are no error bars or multiple seeds anywhere, which matters for a paper whose gains on the private dataset are a few hundredths of ADE. That is a minor issue but should be fixed. Third, and more substantively, the stress-test concern about coverage of the dr=4 m route grid is real but overstated. The intention points are learnable query embeddings, not hard output constraints; the decoder can in principle emit a trajectory whose endpoint sits far from every sampled point, and the closest-point loss assignment will still train that query to hit the GT endpoint. So an 'uncovered endpoint cannot be represented' is too strong. What is true, and important, is that the query set defines the mode structure and the paper itself argues that endpoints land near intention points—so a missing empirical distribution of GT-to-nearest-query distances and ablations over dr/Nq leaves the inductive-bias claim less supported than it should be. A few sentences and one figure would settle it.\n\nThe paper is a solid workshop-to-conference contribution, not a field reshaper. It deserves a serious referee: the route-intention mechanism is novel relative to MTR, GameFormer, and PlanTF, and the negative result on Val14 plus the unresolved coverage question make it a good candidate for constructive revision. I would send it to review and ask for the distance distribution, a corrected SOTA claim, and ideally one more seed.\n\nFor a reading group, it is a reasonable 'maybe'—good for discussing what counts as a mode in goal-conditioned planning.","headline":"Route intention points are a sensible, well-ablated inductive bias for integrated planning, but the SOTA claim is too broad and the 4 m query-grid coverage is never validated.","tokens_in":14059,"tokens_out":2077,"would_cite":true,"duration_ms":24522,"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":"Int2Planner claims that route-sampled intention points, rather than static clustered anchors, improve multi-modal planning in an integrated prediction-and-planning model, and reports top nuPlan scores plus hundreds of urban driving…","keywords":["autonomous driving","motion planning","integrated prediction and planning","route intention points","multi-modal planning","transformer decoder","nuPlan benchmark","real-world deployment"],"falsifier":"In the released validation set, compute for every ground-truth ego trajectory endpoint the distance to the nearest sampled route intention point; a substantial share of endpoints farther than $4$ m from all intention points would show that the fixed sampling grid cannot cover the trajectories the planner is asked to produce.","tokens_in":13073,"feed_emoji":"🚗","tokens_out":7839,"duration_ms":73277,"temperature":0.7,"pith_summary":"Int2Planner's central claim is that the uncertainty in ego motion planning is best constrained by intention points sampled from the route path, not by static or clustered anchor points. The paper argues that because the ego vehicle always follows a route to a destination, its possible short-term goals lie on that route, so sampling candidate goals every 4 meters along primary and secondary routes yields a more useful multi-modal planning space. In an integrated prediction-and-planning transformer, each route intention point initializes a query that produces one candidate trajectory with a confidence score. On the private dataset and the nuPlan Val14 and Test14-hard benchmarks, the route-intention design outperforms clustered intentions and reaches state-of-the-art planning scores, and a real-vehicle deployment drove hundreds of kilometers in urban areas. A sympathetic reader would care because route-conditioned goal sampling is a simple, cheap inductive bias that could apply to any route-following planner.","feed_headline":"Route-sampled goals beat clustered anchors in self-driving planners","feed_subtitle":"Route-sampled goals let one transformer propose many trajectories and top nuPlan's Test14-hard scores.","key_machinery":"The central object is the route intention point set $G_{EA}$, containing $N_q$ points sampled at equal distance intervals along primary and secondary route polylines. Each intention point is embedded by an MLP into a planning intention query, used as position embedding in a transformer decoder that performs self-attention and cross-attention over context and route embeddings. The decoder's output is concatenated with route content and mapped by an MLP into a planning trajectory and confidence score per intention point, with K iterative refinements. This machinery lets route information enter twice: as route embedding through a Route Attention module and as goal queries that anchor multi-modal trajectory proposals.","core_discovery":"The discovery is that replacing static or clustered intention anchors with route intention points improves multi-modal motion planning in a joint prediction-and-planning model. The ego vehicle's route provides stable short-term destinations, so intention points are sampled from primary and secondary route polylines at a fixed distance interval $d_r = 4$ m with $N_q = 64$, rather than computed by K-means on ground-truth endpoints. Each point initializes a planning intention query in a transformer decoder; the decoder refines trajectories over K iterations and outputs a candidate trajectory and confidence per intention point. The paper's experiments show that route intention points beat cluster intention points on both planning and prediction metrics, that including secondary routes adds useful alternatives when the primary route is blocked, and that the full model achieves the best overall score among purely learning-based planners on nuPlan Test14-hard.","pith_inferences":["Editorial inference: the same route-intention idea could carry over to predicting other agents whose routes are known, such as vehicles with active navigation, shrinking the prediction query set to route-relevant goals rather than scene-wide clusters.","Editorial inference: the fixed 4-meter sampling could be made adaptive by spacing intention points according to local curvature, traffic-light positions, or reachable-area boundaries, potentially reducing $N_q$ while keeping coverage.","Editorial inference: the confidence distribution over route intention points is itself a compact explainable signal, because where the mass concentrates tells a human operator which route-level behavior the planner is committing to, and it could be exposed in a monitoring interface.","Editorial inference: because the authors acknowledge that the highest-confidence trajectory is not always optimal, a natural testable extension is a second-stage selector that scores candidate trajectories by interaction-aware metrics rather than learned confidence alone."],"forward_implications":["Route-conditioned goal sampling can replace K-means cluster anchors in transformer-based planners without losing multi-modal coverage, while removing irrelevant targets.","Sampling from secondary as well as primary routes gives the planner explicit alternatives for blocked or congested primary-route maneuvers.","Jointly training prediction and planning in one decoder improves planning metrics compared with planning-only training, because surrounding-agent predictions inform ego trajectory selection.","Adding light rule-based post-processing to the learned multi-modal outputs further lifts closed-loop scores, so the learned planner and classical safety checks are complementary.","The planner can be deployed in real vehicles using the highest-confidence trajectory as the control reference, with the caveat the authors state that top confidence is not always optimal in complex scenes."],"supporting_citations":[{"why":"Supplies the static/clustered intention-query design and the positive-query loss assignment that Int2Planner replaces for the ego vehicle and keeps for other agents.","marker":"Shi et al. 2022"},{"why":"Provides the nuPlan benchmark and closed-loop simulator used for the paper's planning evaluation.","marker":"Caesar et al. 2021"},{"why":"Defines the Val14 benchmark and the PDM-Hybrid hybrid planner, and supplies comparison results for the closed-loop simulations.","marker":"Dauner et al. 2023"},{"why":"Defines the Test14-hard benchmark and the PlanTF learning-based planner, and supplies the baseline scores for the main comparison.","marker":"Cheng et al. 2023"},{"why":"Provides the GameFormer integrated prediction-and-planning baseline that the paper trains and compares against on both nuPlan and the private dataset.","marker":"Huang, Liu, and Lv 2023"}],"fun_headline_variants":["Route intention points outperform clustered anchors for planning","Road-based goals enable multi-modal motion planning in nuPlan","Autonomous planner samples route goals to beat nuPlan baselines","Int2Planner leverages route goals for real-world autonomous driving","Secondary-route goals improve planning when primary route is blocked"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The route intention points, sampled every 4 meters from the route, are assumed to lie close enough to every reasonable ego path that the model can still represent and select the right trajectory.","fun_headline_variants_meta":{"raw":{"variants":["Route intention points outperform clustered anchors for planning","Road-based goals enable multi-modal motion planning in nuPlan","Autonomous planner samples route goals to beat nuPlan baselines","Int2Planner leverages route goals for real-world autonomous driving","Secondary-route goals improve planning when primary route is blocked"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2712,"prompt_tokens":921,"completion_tokens":1791,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":1711}},"tokens_in":537,"tokens_out":1791,"duration_ms":15755,"temperature":1.0,"reasoning_tokens":1711,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:46:02.184214+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the released validation set, compute for every ground-truth ego trajectory endpoint the distance to the nearest sampled route intention point; a substantial share of endpoints farther than $4$ m from all intention points would show that the fixed sampling grid cannot cover the trajectories the planner is asked to produce.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the static/clustered intention-query design and the positive-query loss assignment that Int2Planner replaces for the ego vehicle and keeps for other agents."}],"review_version":1}