Pith. sign in

REVIEW 3 major objections 4 minor 42 references

A drone-recorded, multi-city dataset with semantic risk labels turns rare intersection near-misses into repeatable stress tests that expose failures in autonomous driving planners.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:27 UTC pith:YML7DRBJ

load-bearing objection A genuinely useful multi-city intersection dataset with SOTIF-oriented semantic tags, but the semantic-subset effectiveness claim rests on an unaudited label pipeline and a partly circular low-MprTTC evaluation; worth refereeing with targeted revision requests. the 3 major comments →

arxiv 2607.16943 v1 pith:YML7DRBJ submitted 2026-07-18 cs.RO

SinD 2.0: A Multi-City UAV Dataset with Semantic Risk Annotations for SOTIF-Oriented Safety Validation at Signalized Intersections

Yunwei Li , Shengjie Fu , Chunrong Chen , Chengxiang Zhao , Yuchen Fan , Mingyu Zhu , Yanchao Xu , Yuxin Zhang
show 8 more authors
Lan Yang Chuzhao Li Jie Ji Yi He Abhijit Sarkar Akash Sonth Hong Wang Jun Li
This is my paper
classification cs.RO
keywords autonomous drivingSOTIFdrone datasetsignalized intersectionsafety-critical eventsemantic annotationscenario-based testingcross-domain generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to close three gaps that limit SOTIF safety validation at signalized intersections: geographic homogeneity, sparsity of critical events, and missing semantic annotations. Its answer is SinD 2.0, a drone-captured dataset of six intersections across four Chinese cities with 32,682 safety-critical events and four families of risk labels. The authors argue that the semantic subsets—violations, high-risk interactions, visual shielding, narrow drivable areas—concentrate risk, so a few targeted runs reveal limitations that raw naturalistic data hide. They also show that the six sites form measurably different operational domains, from grid-like crossings to skewed, compressed mountain-city geometries. A sympathetic reader would care because the dataset and toolchain turn the long-tail problem of intersection safety into a queryable, replayable test process.

Core claim

On its own terms, SinD 2.0 is a large-scale drone-based intersection dataset built specifically for SOTIF-oriented validation. It provides about 53,000 tracks over 22.8 hours, spanning 4,188 km of travel, and 32,682 safety-critical events extracted with a surrogate-safety measure that combines time-to-collision and post-encroachment time. The paper's distinctive addition is the hierarchical semantic layer: 53,901 scenario-level records (31,005 high-risk MprTTC interactions, 1,286 visual-shielding cases, 21,610 narrow-feasible-area records) plus 25,966 rule-based violation events, all exposed through a shared queryable schema with stored evidence. Benchmark experiments report that these seman

What carries the argument

The load-bearing mechanism is the hierarchical semantic scenario labeling framework: a common schema (scenario ID, time window, ego ID, semantic tags, event-specific evidence) plus four deterministic rule-based extractors. High-risk interactions are mined with MprTTC, a predicted minimum time-to-collision from short-horizon motion rollouts; violations use map, signal-phase, and crosswalk geometry rules; visual shielding uses a visibility-sector occlusion check with converging-path prediction; narrow drivable areas use an ego-centric reachable-grid depth-first search. Together they convert raw trajectories into searchable SOTIF trigger cases that can be replayed in open-loop or reactive close

Load-bearing premise

The semantic risk labels are generated by deterministic rules and therefore inherit map, signal-binding, and trajectory-estimation errors; a manual precision/recall audit is deferred, so the reported scenario counts and the conclusion that semantic subsets expose ADS limitations rest on the assumption that those labels are accurate enough.

What would settle it

A manual audit of a random sample of the semantic labels against the drone video, or a threshold-sensitivity study that re-runs the open-loop benchmark with the MprTTC threshold varied from 1.5 to 4.0 seconds and the visual-shielding field-of-view angle varied; if the high collision rates on semantic subsets disappear at nearby thresholds, the paper's central stress-test claims would not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Scenario search becomes targeted: engineers can query specific semantic triggers and get a time window, ego/challenger IDs, and stored evidence for each case, instead of scanning raw recordings.
  • Cross-city benchmarking quantifies generalization: the paper documents domain shifts in speed-acceleration envelopes, accepted gaps, violation rates, and interaction durations across the six intersections, allowing a single policy to be evaluated in deliberately different operational domains.
  • The same scenario can be tested under open-loop replay and reactive closed-loop traffic, letting developers separate imitation error from failures caused by multi-agent interaction feedback.
  • If the semantic labels are valid, testing efficiency improves dramatically: reported collision rates jump from 3% on raw data to 30–98% on semantic subsets, so fewer runs are needed to find safety-relevant failures.
  • The dataset's VRU-heavy composition—403 km of vulnerable-road-user tracks—provides a distinctive testbed for prediction and planning around non-compliant pedestrians and two-wheelers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An obvious next step is a human audit of a random sample of the 53,901 semantic labels against the raw drone video; if precision and recall hold up, the deterministic pipeline could auto-generate SOTIF scenario libraries for any city with map and signal data.
  • The paper does not analyze how sensitive its stress-test conclusions are to the MprTTC horizon and the 3.0-second threshold; a threshold sweep could reveal whether the reported failure rates are robust or artifacts of the chosen cutoffs.
  • The finding that visual-shielding scenarios break even conservative planners suggests that trajectory-level conflict metrics alone are insufficient for SOTIF validation—perception-aware simulation that models occlusion may be necessary.
  • Because accepted gaps and violation rates differ widely across sites, SinD 2.0 could serve as calibration data for regional driving behavior in traffic simulators, making simulated background traffic more realistic per city.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. SinD 2.0 is a drone-based, bird's-eye-view trajectory dataset covering six signalized intersections in four Chinese cities, with 4,188 km of trajectories, 53,000 tracks, and 22.8 hours of recording. The paper's main claims are: (1) cross-domain diversity in topology, kinematics, and interaction structure; (2) high-density safety-critical-event mining (32,682 events by a PET/TTC criterion); (3) hierarchical semantic annotations for traffic violations, high-risk interactions, visual shielding, and narrow feasible areas; and (4) a full toolchain for scenario extraction, prediction-only evaluation, open-loop replay, reactive closed-loop testing, and photorealistic rendering. Benchmark experiments are used to support two further claims: significant domain shifts across cities, and that semantic risk subsets effectively expose ADS performance limitations. The dataset, annotations, and toolchain are released at a public repository.

Significance. If the dataset and annotations are reliable, SinD 2.0 is a useful resource for SOTIF-oriented intersection safety research. Its strengths are concrete: multi-city collection with heterogeneous intersection topologies and VRU behavior; a reproducible, rule-based semantic label layer; and an integrated open-loop/closed-loop testing toolchain built on trajdata. The descriptive analytics in Section IV (kinematic envelopes, lateral deviation, accepted gaps, occupancy density, interaction degree) are coherent and internally consistent. However, the central validation claim—that the semantic subsets effectively expose ADS limitations—is not yet independently established. The evaluation of the Low-MprTTC subset is partly circular, and the semantic labels have not been manually audited or subjected to threshold-sensitivity analysis, as the paper itself acknowledges in Section V-A. These issues are fixable with additional experiments and disclosures, but they are load-bearing for the paper's headline conclusion.

major comments (3)
  1. [§VI-B, Table X, Eq. (4)] The central claim that semantic subsets expose ADS limitations is weakened by a circular evaluation design. The Low-MprTTC subset is selected using the MprTTC criterion in Eq. (4), and MinTTC is then reported as computed 'using the MprTTC-style forward bicycle-model rollout' (Section VI-B). Selecting scenarios because they have low predicted-occupancy overlap under this rollout, and then scoring them with the same rollout metric, guarantees part of the observed degradation. The tenfold collision increase for RiskIDM and the 98.3% collision rate for QCNet therefore do not cleanly demonstrate semantic-trigger-specific weakness. Please re-evaluate the Low-MprTTC subset with metrics not used in selection (e.g., PET, DRAC, or independently annotated collision/near-collision outcomes), or compare against a control set matched on traffic density, route, and kinematic criticality.
  2. [§V-A, Eqs. (3)–(5), Algorithms 1–2] The paper's own auditability statement says all semantic labels are generated by deterministic rules and that 'a larger manual precision/recall audit and threshold-sensitivity study are left for the dataset release validation.' The labels therefore inherit map, signal-binding, and trajectory-estimation errors, as acknowledged. The main text does not report numerical values for key thresholds: tau_mpr, horizon H, pair-pruning radius d_max, visibility radius R and field-of-view theta, the path-intersection tolerance epsilon, and the feasible-depth threshold Dmin. Without these values and a sensitivity analysis, counts such as 31,005 MprTTC scenarios, 1,286 visual-shielding scenarios, and 21,610 narrow-feasible-area scenarios, as well as the benchmark ordering in Tables X–XI, could shift materially. Please disclose all thresholds in the appendix/repository and add a threshold-sensitivity st
  3. [§VI-D, Table X] Even setting the circularity issue aside, the raw baseline in Table X is itself very hard: QCNet has a 52.9% collision rate on raw naturalistic scenarios in dataset-open-loop mode, and ASAPRL has 58.4%. The increment to 98.3% on Low MprTTC is large, but the baseline is affected by well-known covariate shift in open-loop replay, and no confidence intervals or per-city/per-intersection breakdowns are provided. To support the claim that semantic subsets 'massively accelerate the discovery of algorithmic deficiencies,' a control experiment is needed: for example, sample equal numbers of random scenarios matched by intersection, ego maneuver, and participant density, and show that the semantic subsets are more efficient at exposing failures than the matched control. Without such a control, the added value of the semantic selector over plain kinematic criticality is not yet established.
minor comments (4)
  1. [Table I] The SCE counts and densities for external datasets are marked approximate, but the table presents them alongside SinD 2.0 as if directly comparable. Please add a more prominent caveat in the table caption stating that differences in collection protocol, viewpoint, and annotation make cross-dataset SCE density comparisons indicative only.
  2. [§IV-A, Figures 4–5] Spearman correlations computed from only six intersections are presented without confidence intervals or scatter plots. The text appropriately calls them descriptive, but the figures with |r| values may overstate the strength of the relationships. Adding scatter plots or bootstrap intervals would help.
  3. [Appendix A] The final representative measurement is defined as 'the 20% quantile of these valid consecutive values.' Please specify whether this quantile is taken over the valid consecutive frames of a single pair, and over what time window, so the SCE definition is unambiguous.
  4. [Table IX] Typo: 'A V2 val' should read 'Argoverse 2 val'. A pass through the references and figure captions for similar spacing/typos would be useful.

Circularity Check

1 steps flagged

Low-MprTTC semantic subset is selected and scored with the same forward-rollout TTC metric, so part of the claimed validation is built into the selection criterion.

specific steps
  1. self definitional [Section V-C (Eqs. 3–4) and Section VI-B (Table X)]
    "‘MprTTCij(t) = min Δ∈{1,...,H} Δδt, s.t. B i(t+ Δ)∩ B j(t+ Δ)≠∅’ and ‘IHR(i, j, t) =1[ MprTTCij(t)≤τ mpr ∧ d ij(t)≤d max ∧ (i∈ Vm ∨j∈ Vm)]’ (Section V-C); ‘MinTTC is computed using the MprTTC-style forward bicycle-model rollout’ (Section VI-B)."

    The 'Low MprTTC' scenario set in Table X is constructed by thresholding Eq. (4), which uses the MprTTC forward predicted-occupancy-overlap score of Eq. (3). The evaluation metric MinTTC is then 'computed using the MprTTC-style forward bicycle-model rollout,' i.e., the same forward-rollout overlap logic. Therefore, low MinTTC on this subset is guaranteed by the inclusion criterion, not discovered by the benchmark. The paper's statement that 'the minimum time-to-collision drops to 1.62 seconds' as validation of the kinematic mining pipeline is self-referential. Collision, off-road, and rule-violation rates are partially independent, and the visual-shielding and narrow-feasible-area subsets are not selected by TTC, so the circularity is partial rather than total.

full rationale

SinD 2.0 is a substantial, externally grounded dataset contribution: the collection spans multiple cities, the prediction benchmark uses public checkpoints (QCNet, Diffuser) against Argoverse 2 and nuScenes, and the closed-loop and visual-shielding/narrow-feasible-area evaluations do not reduce to the selection metric. The main circularity is confined to the Low-MprTTC validation loop: scenarios are selected by MprTTC (Eq. 4) and then scored with MinTTC defined as a 'MprTTC-style forward bicycle-model rollout,' so the headline MinTTC improvements are partly guaranteed by construction. The collision-rate increases are less forced because they depend on the policy's behavior, but they are still measured on a set already enriched for predicted occupancy overlap. The paper itself flags the other major validity concern: 'a larger manual precision/recall audit and threshold-sensitivity study are left for the dataset release validation,' meaning the label counts and subset compositions are not yet externally audited. That is a correctness/validation gap, not itself circularity. There is no load-bearing self-citation chain or imported uniqueness theorem; the SinD v1 citation is for pipeline inheritance, and the benchmark comparisons are computed on the released data. Overall, one central semantic-subset claim is partially circular, while enough independent evidence remains that the paper should not be scored as wholly circular.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 0 invented entities

The central claims rest mainly on hand-set thresholds, unvalidated tracking and map assumptions, and rule-based semantic labels without manual audit. The paper introduces no new physical entities; MprTTC and the semantic labels are computational constructs with documented procedures, not independently evidenced postulations.

free parameters (8)
  • SCE threshold tau = 3.0 s
    Eq. (1) classifies safety-critical events when min(PET, TTC_robust) < 3.0s. No calibration or sensitivity analysis is provided.
  • SCE candidate pair filters = travel distance >2.0m, max speed >0.5m/s, pair distance <20.0m, temporal overlap <15.0s, conflict-point distance <2.5m,
    Appendix A defines these thresholds to filter stationary, non-interacting, or routine trajectories. They are hand-set and not justified against external benchmarks.
  • MprTTC horizon and threshold = H, tau_mpr, d_max (unnumbered in text)
    Eq. (3)-(4) define high-risk interactions using a short-horizon rollout and thresholds tau_mpr and d_max. The paper does not report their numerical values or sensitivity.
  • Visual-shielding sector parameters = R, theta, time tolerance epsilon
    Algorithm 1 uses a visibility sector and path-intersection tolerance, but the specific values are omitted from the paper.
  • Narrow feasible-area threshold = Dmin and consecutive-frame duration
    Algorithm 2 emits narrow-area events when maximum reachable depth stays below Dmin for consecutive frames; the threshold values are not given.
  • Interaction component parameters = 15 m proximity, 5 s prediction horizon
    Section IV.C.1 constructs interaction groups using these parameters; their choice is not derived or sensitivity-tested.
  • Reactive closed-loop controller settings = 25 m neighborhood, 5 closest agents, 5-frame update interval
    Section VI.A fixes these settings for the Diffuser-based reactive background model, affecting all closed-loop results.
  • Miss-rate threshold = 2 m FDE
    Table IX defines miss rate with the 2m final displacement error threshold per standard convention.
axioms (7)
  • domain assumption Drone-based object detection, tracking, and calibration produce accurate trajectories and kinematic states.
    Section III-B describes trajectory extraction and smoothing, but the paper provides no ground-truth validation of the tracking output.
  • domain assumption HD maps and SPaT bindings are sufficiently accurate for violation detection and semantic annotation.
    Section V-B and the auditability paragraph in Section V-A state the labels inherit map, signal-binding, and trajectory-estimation errors.
  • domain assumption Rule-based semantic labels are usable as queryable test selectors without manual verification.
    Section V-A explicitly defers manual precision/recall audit to the release validation.
  • domain assumption Surrogate safety measures such as PET, TTC, and MprTTC are valid proxies for safety-criticality.
    Section II-B motivates SSMs from the literature, but the specific MprTTC formulation in Eq. (3) is not validated against crash outcomes.
  • domain assumption Six intersections in four Chinese cities are representative of cross-domain ODD variation.
    Section III-A describes the sites, but there is no external measure or sample-size justification establishing representativeness.
  • domain assumption The Diffuser-based reactive model reproduces realistic non-ego responses in closed-loop testing.
    Section VI-E uses this controller; the conclusion admits it is 'susceptible to behavioral approximations and semantic drifting under severe edge-case contentions.'
  • domain assumption External dataset SCE counts are comparable despite different sensors and approximate estimation.
    Table I notes 'External vehicle-view SCE metrics are approximate,' so the cross-dataset density comparison is only qualitative.

pith-pipeline@v1.3.0-alltime-deepseek · 23192 in / 9695 out tokens · 91218 ms · 2026-08-01T19:27:21.327238+00:00 · methodology

0 comments
read the original abstract

Safety validation at signalized intersections remains a critical bottleneck for the deployment of autonomous driving systems (ADS), as these scenarios involve dense heterogeneous traffic, contested right of way, and long-tail safety-critical interactions, posing significant challenges to the Safety of the Intended Functionality (SOTIF). Existing naturalistic driving datasets often suffer from geographical homogeneity, sparsity of safety-critical events, and lack of semantic risk annotations, which limit the evaluation of algorithmic generalizability and targeted SOTIF verification. To address these gaps, this paper introduces SinD 2.0, a large-scale drone-based intersection dataset dedicated to cross-domain ADS safety analysis. The main contributions of SinD 2.0 are: (1) Cross-domain diversity: It covers six signalized intersections across four Chinese cities, capturing distinct intersection topologies and regional driving behavior characteristics; (2) High-density risk interactions: A total of 32,682 safety-critical events are extracted via surrogate safety measures, significantly enriching the density of boundary test scenarios; (3) Hierarchical semantic annotations: Besides integration with high-definition (HD) maps and Signal Phase and Timing (SPaT) data, it provides multi-dimensional semantic labels including traffic violations, high-risk interactions, visual shielding, and narrow feasible areas; (4) Full-stack testing toolchain: It supports automated scenario extraction, prediction-only evaluation, open-loop replay, reactive closed-loop testing, and photorealistic rendering. Benchmark experiments demonstrate that SinD 2.0 exhibits significant domain shifts across cities, and the semantic risk subsets can effectively expose the performance limitations of ADS algorithms. The dataset, annotations, and testing toolchain are available at https://github.com/SOTIF-AVLab/SinD/tree/main.

Figures

Figures reproduced from arXiv: 2607.16943 by Abhijit Sarkar, Akash Sonth, Chengxiang Zhao, Chunrong Chen, Chuzhao Li, Hong Wang, Jie Ji, Jun Li, Lan Yang, Mingyu Zhu, Shengjie Fu, Yanchao Xu, Yi He, Yuchen Fan, Yunwei Li, Yuxin Zhang.

Figure 1
Figure 1. Figure 1: Overview of the SinD 2.0 dataset construction, semantic annotation, and toolchain. (A) Geographic distribution of the six intersections across four [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of traffic participant type distributions across SinD [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Traffic participant type composition across the six SinD 2.0 intersec [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Core descriptive correlations between static geometric descriptors and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of intersection angle skewness on crossing-conflict ratio and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-city speed-acceleration envelopes estimated from trajectory [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Accepted time-gap distributions for unprotected conflict passages. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Spatiotemporal vehicle occupancy density for all six intersections. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of interaction degree by intersection. [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Hierarchical semantic scenario labeling framework in SinD 2.0. A [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Structured violation and non-compliance rates across intersections. [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Representative structured violation and non-compliance examples. The montage shows motor-vehicle (MV) and non-motorized-vehicle (NMV) red [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Representative high-risk interactions mined by MprTTC. The green [PITH_FULL_IMAGE:figures/full_fig_p012_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Representative visual-shielding scenarios. The yellow occluding [PITH_FULL_IMAGE:figures/full_fig_p013_17.png] view at source ↗
Figure 19
Figure 19. Figure 19: Auxiliary 3DGS-based visual simulation pipeline. A BEV map [PITH_FULL_IMAGE:figures/full_fig_p016_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Egocentric visual simulation demonstrations generated by the 3DGS-based pipeline. The sequences illustrate continuous view rendering for an [PITH_FULL_IMAGE:figures/full_fig_p017_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 2 canonical work pages

  1. [1]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012, pp. 3354–3361

    2012

  2. [2]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 621–11 631

    2020

  3. [3]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caineet al., “Scalability in perception for autonomous driving: Waymo open dataset,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2446–2454. SIND 2.0 DRAFT MANUSCRIPT 18

    2020

  4. [4]

    Argoverse: 3d tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan, and J. Hays, “Argoverse: 3d tracking and forecasting with rich maps,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 8748–8757. [Online]. Available: https://arxiv.org/abs/1911. 02620

    2019

  5. [5]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

    S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou, Z. Yang, A. Chouard, P. Sun, J. Ngiam, V . Vasudevan, A. McCauley, J. Shlens, and D. Anguelov, “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” inProceedings of the IEEE/CVF International Conference on Com...

    Pith/arXiv arXiv 2021

  6. [6]

    Assessment of vehicle trajectory data accuracy and application to the ngsim trajectory data,

    V . Punzo, M. T. Borzacchiello, and B. Ciuffo, “Assessment of vehicle trajectory data accuracy and application to the ngsim trajectory data,” Transportation Research Part C: Emerging Technologies, vol. 19, no. 6, pp. 1243–1262, 2011

    2011

  7. [7]

    The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

    R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” in2018 21st International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018, pp. 2118–2125

    2018

  8. [8]

    The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,

    J. Bock, R. Krajewski, T. Moers, S. Runde, L. Vater, and L. Eckstein, “The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,” in2020 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2020, pp. 1929–1934

    2020

  9. [9]

    Interaction dataset: An inter- national, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,

    W. Zhan, L. Sun, D. Wang, H. Shi, A. Clausse, M. Naumann, J. Kum- merle, H. Konig, C. Dong, Y . Jianget al., “Interaction dataset: An inter- national, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,”arXiv preprint arXiv:1910.03088, 2019

    Pith/arXiv arXiv 1910

  10. [10]

    The round dataset: A drone dataset of road user trajectories at roundabouts in germany,

    R. Krajewski, T. Moers, J. Bock, L. Vater, and L. Eckstein, “The round dataset: A drone dataset of road user trajectories at roundabouts in germany,” in2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2020, pp. 1–6. [Online]. Available: https://www.round-dataset.com/

    2020

  11. [11]

    opendd: A large-scale roundabout drone dataset,

    A. Breuer, J.-A. Term ¨ohlen, S. Homoceanu, and T. Fingscheidt, “opendd: A large-scale roundabout drone dataset,” in2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2020, pp. 1–6. [Online]. Available: https://arxiv.org/abs/2007. 08463

    2020

  12. [12]

    Citysim: A drone-based vehicle trajectory dataset for safety oriented research and digital twins,

    O. Zheng, M. Abdel-Aty, L. Yue, A. Abdelraouf, Z. Wang, and N. Mahmoud, “Citysim: A drone-based vehicle trajectory dataset for safety oriented research and digital twins,”arXiv preprint arXiv:2208.11036, 2022. [Online]. Available: https://arxiv.org/abs/2208. 11036

    Pith/arXiv arXiv 2022

  13. [13]

    A9 intersection dataset: All you need for urban 3d camera-lidar roadside perception,

    W. Zimmer, C. Creß, H. T. Nguyen, and A. C. Knoll, “A9 intersection dataset: All you need for urban 3d camera-lidar roadside perception,”arXiv preprint arXiv:2306.09266, 2023. [Online]. Available: https://arxiv.org/abs/2306.09266

    Pith/arXiv arXiv 2023

  14. [14]

    Sind: A drone dataset at signalized intersection in china,

    Y . Xuet al., “Sind: A drone dataset at signalized intersection in china,” in 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2022, pp. 2471–2478

    2022

  15. [15]

    ISO/PAS 21448:2019,Road vehicles — Safety of the intended function- ality, International Organization for Standardization Std., 2019

    2019

  16. [16]

    A survey on scenario-based testing for automated driving systems in high-fidelity simulation,

    Z. Zhong, Y . Tang, Y . Zhou, V . de Oliveira Neves, Y . Liu, and B. Ray, “A survey on scenario-based testing for automated driving systems in high-fidelity simulation,”arXiv preprint arXiv:2112.00964, 2021. [Online]. Available: https://arxiv.org/abs/2112.00964

    Pith/arXiv arXiv 2021

  17. [17]

    Application of proximal surrogate indicators for safety evaluation: A review of recent developments and research needs,

    S. M. Mahmud, L. Ferreira, M. S. Hoque, and A. Tavassoli, “Application of proximal surrogate indicators for safety evaluation: A review of recent developments and research needs,”Iatss Research, vol. 41, no. 4, pp. 153–163, 2017

    2017

  18. [18]

    Curse of rarity for autonomous vehicles,

    H. X. Liu and S. Feng, “Curse of rarity for autonomous vehicles,”nature communications, vol. 15, no. 1, p. 4808, 2024

    2024

  19. [19]

    Accelerated evaluation of automated vehicles safety in lane-change scenarios based on importance sampling techniques,

    D. Zhao, H. Lam, H. Peng, S. Bao, D. J. LeBlanc, K. Nobukawa, and C. S. Pan, “Accelerated evaluation of automated vehicles safety in lane-change scenarios based on importance sampling techniques,” inIEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 3. IEEE, 2017, pp. 595–607. [Online]. Available: https://arxiv.org/abs/1605.04965

    Pith/arXiv arXiv 2017

  20. [20]

    Deep Probabilistic Accelerated Evaluation: A Robust Certifiable Rare-Event Simulation Methodology for Black-Box Safety- Critical Systems,

    M. Arief, Z. Huang, G. K. S. Kumar, Y . Bai, S. He, W. Ding, H. Lam, and D. Zhao, “Deep Probabilistic Accelerated Evaluation: A Robust Certifiable Rare-Event Simulation Methodology for Black-Box Safety- Critical Systems,” Mar. 2021

    2021

  21. [21]

    Adaptive stress testing for autonomous vehicles,

    M. Koren, S. Alsaif, R. Lee, and M. J. Kochenderfer, “Adaptive stress testing for autonomous vehicles,”2018 IEEE Intelligent Vehicles Symposium (IV), pp. 1–7, 2018. [Online]. Available: https://arxiv.org/abs/1902.01909

    Pith/arXiv arXiv 2018

  22. [22]

    Safety-critical scenario generation for automated driving systems: A review,

    D. Zhaoet al., “Safety-critical scenario generation for automated driving systems: A review,” inIEEE Transactions on Intelligent Transportation Systems, 2020

    2020

  23. [23]

    A survey on safety- critical driving scenario generation: A methodological perspective,

    W. Ding, B. Chen, M. Xu, and D. Zhao, “A survey on safety- critical driving scenario generation: A methodological perspective,” arXiv preprint arXiv:2202.02215, 2022. [Online]. Available: https: //arxiv.org/abs/2202.02215

    Pith/arXiv arXiv 2022

  24. [24]

    Testing sce- nario library generation for connected and automated vehicles, part i: Methodology,

    S. Feng, Y . Feng, C. Yu, Y . Zhang, and H. X. Liu, “Testing sce- nario library generation for connected and automated vehicles, part i: Methodology,”IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 3, pp. 1573–1582, 2021

    2021

  25. [25]

    Simnet: Learning reactive self-driving simulations from real-world observations,

    L. Bergamini, Y . Ye, O. Scheel, L. Chen, C. Hu, L. Del Pero, B. Osinski, H. Grimmett, and P. Ondruska, “Simnet: Learning reactive self-driving simulations from real-world observations,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 5119–5125. [Online]. Available: https://arxiv.org/abs/2105.12332

    Pith/arXiv arXiv 2021

  26. [26]

    Bits: Bi-level imitation for traffic simulation,

    D. Xu, Y . Lu, F. Shkurti, A. Garg, M. Tomizuka, and Y . Zhu, “Bits: Bi-level imitation for traffic simulation,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 2929–2936. [Online]. Available: https://arxiv.org/abs/2208.12403

    Pith/arXiv arXiv 2023

  27. [27]

    Trafficgen: Learning to generate diverse and realistic traffic scenarios,

    S. Feng, X. Yan, H. Sun, Y . Feng, and H. X. Liu, “Trafficgen: Learning to generate diverse and realistic traffic scenarios,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 3567–3575. [Online]. Available: https: //arxiv.org/abs/2210.06609

    Pith/arXiv arXiv 2023

  28. [28]

    Scenario diffusion: Controllable driving scenario generation with diffusion,

    E. Pronovost, M. Ganesina, N. Hendy, Z. Wang, A. Morales, K. Wang, and N. Roy, “Scenario diffusion: Controllable driving scenario generation with diffusion,”Advances in Neural Information Processing Systems, vol. 36, 2023. [Online]. Available: https: //arxiv.org/abs/2311.02738

    Pith/arXiv arXiv 2023

  29. [29]

    Generation of signalized intersection test scenarios based on traffic participant model,

    X. Gu, S. Wu, S. Zhao, T. Zhang, X. Li, X. Jiao, and H. Wang, “Generation of signalized intersection test scenarios based on traffic participant model,”Automotive Innovation, vol. 8, pp. 882–895, 2025. [Online]. Available: https://doi.org/10.1007/s42154-025-00370-y

    work page doi:10.1007/s42154-025-00370-y 2025

  30. [30]

    Text2scenario: Text-driven scenario generation for autonomous driving test,

    X. Cai, X. Bai, Z. Cui, D. Xie, D. Fu, H. Yu, and Y . Ren, “Text2scenario: Text-driven scenario generation for autonomous driving test,”Automotive Innovation, vol. 9, pp. 102–127, 2026. [Online]. Available: https://doi.org/10.1007/s42154-025-00374-8

    work page doi:10.1007/s42154-025-00374-8 2026

  31. [31]

    Recent development and applications of sumo: Simulation of urban mobility,

    D. Krajzewicz, J. Erdmann, M. Behrisch, and L. Bieker, “Recent development and applications of sumo: Simulation of urban mobility,” International Journal On Advances in Systems and Measurements, vol. 5, no. 3–4, pp. 128–138, 2012

    2012

  32. [32]

    Carla: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “Carla: An open urban driving simulator,” inProceedings of the 1st Annual Conference on Robot Learning, 2017, pp. 1–16. [Online]. Available: https://arxiv.org/abs/1711.03938

    Pith/arXiv arXiv 2017

  33. [33]

    Scenic: A language for scenario specification and scene generation,

    D. J. Fremont, T. Dreossi, S. Ghosh, X. Yue, A. L. Sangiovanni- Vincentelli, and S. A. Seshia, “Scenic: A language for scenario specification and scene generation,” inProceedings of the 40th ACM SIGPLAN Conference on Programming Language Design and Implementation, 2019, pp. 63–78. [Online]. Available: https: //arxiv.org/abs/1809.09310

    Pith/arXiv arXiv 2019

  34. [34]

    Commonroad: Composable benchmarks for motion planning on roads,

    M. Althoff, M. Koschi, and S. Manzinger, “Commonroad: Composable benchmarks for motion planning on roads,” in2017 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2017, pp. 719–726. [Online]. Available: https://commonroad.in.tum.de/

    2017

  35. [36]

    Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling,

    Q. Li, Z. Peng, L. Feng, Z. Liu, C. Duan, W. Mo, and B. Zhou, “Scenarionet: Open-source platform for large-scale traffic scenario simulation and modeling,”arXiv preprint arXiv:2306.12241, 2023. [Online]. Available: https://arxiv.org/abs/2306.12241

    Pith/arXiv arXiv 2023

  36. [37]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting,

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J. K. Pontes, D. Ramanan, P. Carr, and J. Hays, “Argoverse 2: Next generation datasets for self-driving perception and forecasting,” inProceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2023. [Online]. Available: https:...

    Pith/arXiv arXiv 2023

  37. [38]

    trajdata: A unified interface to multiple human trajectory datasets,

    B. Ivanovic, G. Song, I. Gilitschenski, and M. Pavone, “trajdata: A unified interface to multiple human trajectory datasets,” inProceedings of the Neural Information Processing Systems (NeurIPS) Track on SIND 2.0 DRAFT MANUSCRIPT 19 Datasets and Benchmarks, New Orleans, USA, Dec. 2023. [Online]. Available: https://arxiv.org/abs/2307.13924

    Pith/arXiv arXiv 2023

  38. [39]

    High- dimensional functional boundaries search for deviation-robust testing of autonomous driving system,

    Y . Li, S. Wu, A. Wang, L. Yang, H. Wang, J. Li, and C. Huang, “High- dimensional functional boundaries search for deviation-robust testing of autonomous driving system,”Accident Analysis & Prevention, vol. 221, p. 108156, Oct. 2025

    2025

  39. [40]

    Efficient Reinforcement Learning for Autonomous Driving with Parameterized Skills and Priors,

    L. Wang, J. Liu, H. Shao, W. Wang, R. Chen, Y . Liu, and S. Waslan- der, “Efficient Reinforcement Learning for Autonomous Driving with Parameterized Skills and Priors,” inRobotics: Science and Systems XIX. Robotics: Science and Systems Foundation, Jul. 2023

    2023

  40. [41]

    Query-centric trajectory prediction,

    Z. Zhou, J. Wang, Y .-H. Li, and Y .-K. Huang, “Query-centric trajectory prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 17 863–17 873. [On- line]. Available: https://openaccess.thecvf.com/content/CVPR2023/html/ Zhou Query-Centric Trajectory Prediction CVPR 2023 paper.html

    2023

  41. [42]

    Planning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” inInternational Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 162, 2022, pp. 9902–9915. [Online]. Available: https://proceedings.mlr.press/v162/janner22a.html

    2022

  42. [2024]

    Available: https://arxiv.org/abs/2403.04133

    [Online]. Available: https://arxiv.org/abs/2403.04133

    Pith/arXiv arXiv