Pith. sign in

REVIEW 5 major objections 4 minor 62 references

Towards Human-Like Trajectory Prediction for Autonomous Driving: A Behavior-Centric Approach

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read HiT claims that scoring each driver's aggressiveness with a fuzzy inference system and encoding it through a hypergraph yields more accurate trajectory predictions than static interaction graphs, with gains concentrated in…

desk verdict A behavior-aware trajectory predictor with plausible gains and real engineering effort, but the missing BAT baseline and self-labeled aggression subsets mean the 'consistently outperforms' claim is not yet established. read the letter →

arxiv 2505.21565 v1 pith:U2CYMD7S submitted 2025-05-27 cs.RO cs.AI

classification cs.ROcs.AI
keywords trajectorypredictionautonomousdrivingbehaviorfuzzyinferencesystemhypergraphneuralnetworkcentralitymeasurespolarcoordinatesmultimodal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that trajectory prediction for autonomous driving improves when the model explicitly estimates each driver's aggressiveness instead of relying only on static spatial graphs. Its evidence is a model called HiT, which pairs a behavior-aware module with a polar-coordinate interaction module and a multimodal decoder. Across the NGSIM, HighD, RounD, MoCAD++, and ApolloScape datasets, HiT reports lower prediction error than the compared baselines, particularly at long horizons and for aggressive driving, while using fewer parameters and still performing well when trained on only 25% of the data. A sympathetic reader would care because the result suggests that behavior-centric cues, not just geometry, are a practical route to safer and more data-efficient autonomous driving.

What carries the argument

The load-bearing mechanism is the behavior-aware module, which builds dynamic geometric graphs, computes three centrality measures (dynamic degree centrality, dynamic closeness centrality, and eigenvector centrality), and converts their first and second time derivatives into Behavior Intensity Estimate and Behavior Fluctuation Estimate criteria. Those criteria feed a fuzzy inference system that uses data-fitted membership functions, a hand-coded rule table, and the q-rung orthopair fuzzy weighted Einstein Bonferroni mean operator to output an aggressiveness score for every agent. The aggressiveness scores then define hyperedges in a hypergraph, so vehicles with similar behavior are grouped together and processed by a hypergraph convolution to produce the behavior vector used by the decoder.

What would settle it

Replace the fuzzy aggressiveness scores in the behavior encoder with shuffled or random scores while keeping the rest of HiT fixed, and compare RMSE on NGSIM and HighD; if performance does not degrade, the reported gains do not come from behavior information. A second check would swap the fuzzy scores for human-annotated aggressiveness labels as an oracle and see whether the predicted scores capture the same signal.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a continuous, interpretable score of driving aggressiveness, computed from dynamic centrality measures and then refined by a hypergraph behavior encoder, makes vehicle trajectory prediction more accurate and more human-like than prior interaction models. The paper reports that HiT achieves the lowest average RMSE on NGSIM, HighD, RounD, and MoCAD++, and the lowest weighted ADE and FDE on ApolloScape, with the largest margins appearing at 4- to 5-second horizons and in subsets classified as high- and low-aggression driving. It also reports that HiT maintains competitive accuracy trained on a quarter of the data and uses roughly 40-50% fewer parameters than two strong baselines, supporting the claim that behavior awareness substitutes for some of the usual data and capacity requirements.

Load-bearing premise

The load-bearing premise is that the fuzzy inference system's hand-written rules and data-fitted membership functions produce an aggressiveness score that genuinely reflects driving behavior, and that grouping vehicles by that score is what improves the predictions.

Editorial extensions

If this is right

  • If HiT is right, behavior-aware scoring lowers trajectory prediction error across highway, roundabout, campus, and urban scenes, with the largest gains at long prediction horizons.
  • The reported 25%-data results imply that explicitly modeling driving behavior can reduce the amount of annotated trajectory data needed to train a competitive predictor.
  • The efficiency results, roughly 182K parameters and 22 ms inference for 24 agents, suggest the behavior-aware design can fit within real-time onboard computation budgets.
  • The case-study results imply that the model's advantage is strongest for low- and high-aggression driving, where predicting extreme behavior matters most for safety.
  • Because the pattern holds across five datasets, the paper claims the approach transfers across road types and driving cultures rather than overfitting one scene.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test whether the fuzzy aggressiveness score carries information beyond what a simple speed-acceleration heuristic could provide; randomizing or shuffling the scores in the behavior encoder would settle that.
  • If aggressiveness scores reflect stable per-driver tendencies, a natural extension the paper leaves implicit is personalizing predictions to individual drivers when identity or driver-history data is available.
  • The dynamic-centrality and hypergraph behavior encoding are not vehicle-specific, so the same module could plausibly transfer to pedestrians or mixed traffic, where observable 'behavior' is noisier.
  • A practical downstream use the paper does not develop is coupling the aggressiveness score with risk estimation, so an autonomous system could react early to a high-aggression neighbor rather than only predicting its path.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes HiT, a trajectory prediction model for autonomous driving that combines a behavior-aware module, an interaction-aware module, and a multimodal decoder. The behavior-aware module uses dynamic geometric graphs, three centrality measures, a fuzzy inference system with a q-ROFWEBM aggregation operator, and a hypergraph encoder to produce a continuous aggressiveness score for each agent. The interaction-aware module uses a polar-coordinate pooling mechanism and a transformer-style encoder, and the decoder outputs a Gaussian mixture over maneuvers. The authors evaluate HiT on NGSIM, HighD, RounD, MoCAD++, and ApolloScape, reporting lower RMSE/ADE/FDE than several baselines, including configurations trained on 25% of the data and a smaller 'HiT (S)' model. The central advertised claim is that HiT 'consistently outperforms other top models across multiple metrics, particularly excelling in scenarios involving aggressive driving behaviors.'

Significance. If the empirical claims are correct, the paper would make a meaningful contribution to behavior-aware trajectory prediction by demonstrating that a continuous, fuzzy-logic-based behavior representation, combined with dynamic centrality measures and hypergraph interaction modeling, improves accuracy over strong baselines while reducing parameter count and training data needs. The efficiency results, if reproducible, would be practically valuable for real-time autonomous driving systems. However, the current manuscript does not establish these claims convincingly: there are no error bars or significance tests, the most relevant predecessor (BAT, the authors' own prior behavior-aware model) is absent from all comparisons, the aggression-subset evaluation is circular because subsets are labeled by the model itself, and there are arithmetic and methodological errors in the tables and proofs. These issues must be resolved before the contribution can be assessed.

major comments (5)
  1. [§2, Tables 3–5 and 7] The most relevant behavior-aware baseline, BAT (Liao et al. 2024c), is cited in Related Work but is never included in any performance comparison. Because BAT is the authors' own closest prior model and HiT is positioned as a behavior-centric successor, the abstract's claim that HiT 'consistently outperforms other top models' cannot be evaluated without comparing against BAT. The authors should add BAT to all benchmark tables or explicitly justify its exclusion; otherwise the central claim is unsupported.
  2. [§5.2, Tables 3–5] All reported metrics are point estimates with no standard deviations, confidence intervals, or multiple-seed experiments, and Section 5.1.2 says only that data were 'partitioned using standard sampling' without specifying the split. Several reported differences are small (e.g., Table 3 HighD at 3 s: HiT 0.22 vs iNATran 0.21; NGSIM at 3 s: HiT 1.42 vs GaVa/HLTP++ 1.52), so the claimed superiority is not statistically established. Additionally, Table 3 shows iNATran outperforming HiT at HighD horizons 1–3 s, which conflicts with the abstract's 'consistently outperforms' wording; Appendix G.1 acknowledges this only in passing, and the inconsistency should be reconciled.
  3. [Table 3, Table 8, Appendix G.3] There are concrete arithmetic inconsistencies in the numerical comparisons. In Table 3, the NGSIM row for HiT (25%) lists AVG = 2.01, but the five horizon values (0.46, 0.99, 1.54, 2.18, 3.04) average to 1.64. Table 8 reports HiT NGSIM AVG = 1.64, whereas Table 3 reports HiT NGSIM AVG = 1.53. Table 8 also reports GaVa HighD AVG = 0.39, but Table 3's GaVa HighD values (0.17, 0.24, 0.42, 0.86, 1.31) average to 0.60, which is instead the value listed for Wave. These errors directly affect the efficiency and data-efficiency claims in Appendix G.3 and must be corrected and re-verified.
  4. [§4.2.2, Appendix C; §4.2.4, Eqs. (30)–(33)] The proof of Theorem 2 is not a proof of IIC violation: showing that the centrality values change when an edge is removed does not establish that the relative ordering of two unaffected nodes flips; a concrete graph with x_i(ε) ≥ x_j(ε) in one graph and the reverse in the modified graph is required. The proof of Theorem 3 is also invalid: in Eqs. (30)–(33), the nonlinear function g^{-1} is moved inside the summation over i,j after first being applied to each summand, an operation that is not allowed for nonlinear Archimedean generators. Since the closed form in Eq. (19) and the subsequent q-ROFWEBM defuzzification formulas (36)–(44) depend on this derivation, the behavior-aware module's theoretical foundation is not sound as written.
  5. [§5.3, Tables 5 and 6] The evaluation of performance on 'Low', 'Medium', and 'High' aggression subsets is circular: the subsets are defined by the model's own fuzzy inference system, and the claim that HiT 'excels in aggressive driving behaviors' is then measured only on those self-labeled groups. Table 6 merely describes the speed, acceleration, and time-headway statistics of these self-labeled subsets; it does not validate the aggressiveness score against an independent ground-truth behavior label. The authors should either use externally annotated behavior labels, evaluate with a perturbation/sensitivity analysis of the fuzzy rule boundaries, or clearly reframe the claims as conditional on the proposed FIS.
minor comments (4)
  1. [Throughout Section 4.2.4, Appendix E, Section 5.1.2] There are several typos and inconsistent terms: 'denazify' should be 'defuzzify' in Step-3 of Section 4.2.4; 'NGISM' appears instead of 'NGSIM' in Section 5.1.2; 'BLE' appears instead of 'BFE' in the description of the weight hyperparameters in Section 4.2.4; 'Activaiton' appears in Appendix E; and 'Q-order pair fuzzy numbers' is used interchangeably with 'q-rung ortho-pair fuzzy numbers' in Definition 6 and Theorem 4, which is confusing because Q is also used for a fuzzy number set.
  2. [Table 2] Table 2 is titled as a comparison of HiT with baseline models, but it omits MHA-LSTM, EA-Net, iNATran, and Wave, all of which appear in Table 3's performance comparisons. The table should either include these models or state that it lists only a representative subset of comparison methods.
  3. [Eq. (50), §5.1.4] Equation (50) is described as the maneuver loss Lman, but it is the negative log-likelihood of a bivariate Gaussian distribution. The text should state this explicitly and clarify that σ_tk,ρ and σ_tk,θ denote variances (or standard deviations, depending on the parameterization), since the notation in the surrounding text is ambiguous.
  4. [Appendix H.2, Table 12] The descriptions of ablation models G–L are repetitive and partly inconsistent: Model I is said to use 'the classical closeness centrality and classical degree centrality,' while Model J is described as 'using the classical degree centrality, instead of the dynamic degree centrality,' and Model K replaces the hypergraph with graph behavior learning. The table caption and the text should be rewritten so that each row is defined once, without overlapping descriptions.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity in the core derivation: the headline RMSE results are measured against external trajectory ground truth; the self-generated aggression strata are a validity concern, not a derivation circularity.

full rationale

HiT's central claim is empirical superiority on five externally benchmarked datasets. The prediction losses (Eq. 50) and evaluation protocols compare against recorded ground-truth trajectories, so the main Tables 3, 4, 7, and 8 cannot reduce to the model's own fitted quantities. The behavior-aware module is derived from centrality measures, a K-means-fitted fuzzy inference system, and hypergraph convolution (Eqs. 5-49), but none of these steps defines the target trajectory error; the trajectory loss is an independent MSE/NLL objective. The only self-referential episode is Section 5.3, where the Low/Medium/High 'aggressive' subsets are labeled by the paper's own FIS and then used to show the behavior module's benefit and the abstract's claim of excelling in aggressive scenarios. This is a statistical-validity limitation (the test strata are not independent of HiT's input features), and it is partially mitigated by Table 6, which confirms that the high-aggression subset has higher speed, acceleration, and shorter headway. It is not a circular derivation because the reported RMSE values are still external. Separately, the paper's closest prior behavior-aware model BAT (Liao et al. 2024c) is cited but never included as a baseline, which weakens the 'consistently outperforms other top models' claim as a completeness matter, but omission is not circularity. The mathematical theorems (IIC, q-ROFWEBM closure) are internal consistency results, not imported from the authors' prior work. Accordingly, no specific circular reduction can be exhibited, and the score is a low 2 reflecting only minor self-referential framing rather than a circular derivation.

Assumptions & free parameters 8 free parameters · 5 assumptions · 3 invented entities

HiT introduces a large number of hand-set and data-fitted parameters (kernel widths, mixing parameters, weighting factors, fuzzy boundaries) and rests on domain assumptions about interaction radius and GMM output. The behavior constructs (BIE, BFE, aggressiveness score) are internally defined and not validated against external labels. This makes the source of the reported gains hard to attribute.

free parameters (8)
  • sigma_k (kernel widths in multimodal adjacency) = not reported
    Eq. (5) uses sigma_k to scale each feature in the adjacency kernel; no values or sensitivity analysis are given.
  • epsilon (dynamic degree centrality mixing) = not reported
    Eq. (6) controls how much indirect connections matter; hand-set.
  • gamma_1..gamma_4 (closeness weighting) = not reported
    Eq. (7) sets the weight of position, speed, and acceleration differences; hand-set.
  • eta_1..eta_4 (distance metric weights) = not reported
    Eq. (8) defines the dynamic distance; hand-set.
  • q (q-ROFS control parameter) = not reported
    Eqs. (15) and (16) require q >= 1; the value affects non-membership degrees and is not justified.
  • o, h (q-ROFWEBM smoothness parameters) = not reported
    Used in Definitions 5-6 and Eq. (40); hand-set.
  • beta (time decay rate) = not reported
    Weights fuzzy pairs by recency; no value or sensitivity analysis is given.
  • Fuzzy set boundaries (BIE, BFE, Score) = e.g., {0.3, 0.8}, {0.35, 0.75}, {0.4, 0.7}
    Eqs. (12)-(14): boundaries are chosen from K-means clustering on the training data, i.e., fitted to the data.
assumptions (5)
  • standard math Perron-Frobenius theorem ensures eigenvector centrality is well-defined and unique.
    Invoked in Section 4.2.2 for eigenvector centrality; standard result.
  • domain assumption Interaction edges exist only for vehicle pairs within a distance threshold r.
    Section 4.2.1, Eq. (2): all interaction modeling rests on this threshold; r is not reported.
  • domain assumption Future trajectories follow a Gaussian mixture model conditioned on maneuvers.
    Appendix A and Section 5.1.4: the decoder outputs GMM parameters; this distributional assumption is standard in trajectory prediction.
  • ad hoc to paper The fuzzy rules in Table 1 correctly map BIE and BFE to aggressiveness levels.
    Table 1: the 3 by 3 rule set is hand-defined, with no learning or external validation.
  • ad hoc to paper The Independence of Irrelevant Connections (IIC) criterion is a sensible normative property for centrality measures.
    Definition 2 introduces a new criterion; the paper's theoretical narrative depends on it, but it is not an established metric.
invented entities (3)
  • Behavior Intensity Estimate (BIE)
    purpose: Quantify how fast a vehicle's centrality changes, as a proxy for sudden maneuvers.
    Defined in Eq. (10); used only within HiT; no external validation.
  • Behavior Fluctuation Estimate (BFE)
    purpose: Quantify variability of centrality change to detect erratic driving.
    Defined in Eq. (11); same internal-only status as BIE.
  • Aggressiveness score S
    purpose: Classify drivers into low, medium, and high aggressive groups in the fuzzy inference and hypergraph.
    Produced by the q-ROFWEBM defuzzification; the only supporting evidence is internal correlation with speed, acceleration, and headway in Table 6, not an independent ground-truth label.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Human-Like Trajectory Prediction for Autonomous Driving: A Behavior-Centric Approach." pith.science (2026). https://pith.science/paper/U2CYMD7S

@misc{pith2026250521565,
  author       = {Pith},
  title        = {Pith review of: Towards Human-Like Trajectory Prediction for Autonomous Driving: A Behavior-Centric Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2CYMD7S}},
  note         = {Machine review of arXiv:2505.21565}
}
read the original abstract

Predicting the trajectories of vehicles is crucial for the development of autonomous driving (AD) systems, particularly in complex and dynamic traffic environments. In this study, we introduce HiT (Human-like Trajectory Prediction), a novel model designed to enhance trajectory prediction by incorporating behavior-aware modules and dynamic centrality measures. Unlike traditional methods that primarily rely on static graph structures, HiT leverages a dynamic framework that accounts for both direct and indirect interactions among traffic participants. This allows the model to capture the subtle yet significant influences of surrounding vehicles, enabling more accurate and human-like predictions. To evaluate HiT's performance, we conducted extensive experiments using diverse and challenging real-world datasets, including NGSIM, HighD, RounD, ApolloScape, and MoCAD++. The results demonstrate that HiT consistently outperforms other top models across multiple metrics, particularly excelling in scenarios involving aggressive driving behaviors. This research presents a significant step forward in trajectory prediction, offering a more reliable and interpretable approach for enhancing the safety and efficiency of fully autonomous driving systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 55 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in "" FUNCTION format.date year ...

  3. [3]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  4. [4]

    2015 IEEE International Conference on Systems, Man, and Cybernetics, 2174--2179, ://dx.doi.org/10.1109/SMC.2015.380

    Ahmed K, Tazi NE, Hossny AH, 2015 Sentiment analysis over social networks: An overview. 2015 IEEE International Conference on Systems, Man, and Cybernetics, 2174--2179, ://dx.doi.org/10.1109/SMC.2015.380

  5. [5]

    Proceedings of the IEEE CVPR, 961--971

    Alahi A, Goel K, Ramanathan V, Robicquet A, Fei-Fei L, Savarese S, 2016 Social lstm: Human trajectory prediction in crowded spaces. Proceedings of the IEEE CVPR, 961--971

  6. [6]

    Baron J, 2000 Thinking and deciding (Cambridge University Press)

  7. [7]

    2009 IEEE Intelligent Vehicles Symposium, 907--912, ://dx.doi.org/10.1109/IVS.2009.5164400

    Batz T, Watson K, Beyerer J, 2009 Recognition of dangerous situations within a cooperative group of vehicles. 2009 IEEE Intelligent Vehicles Symposium, 907--912, ://dx.doi.org/10.1109/IVS.2009.5164400

  8. [8]

    IEEE Transactions on Vehicular Technology 70(11):11216--11227

    Cai Y, Wang Z, Wang H, Chen L, Li Y, Sotelo MA, Li Z, 2021 Environment-attention network for vehicle trajectory prediction. IEEE Transactions on Vehicular Technology 70(11):11216--11227

Show all 62 references
  1. [9]

    2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2035–2042 (IEEE Press), ://dx.doi.org/10.1109/IROS45743.2020.9341720

    Chandra R, Bhattacharya U, Mittal T, Bera A, Manocha D, 2020 Cmetric: A driving behavior measure using centrality functions. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2035–2042 (IEEE Press), ://dx.doi.org/10.1109/IROS45743.2020.9341720

  2. [10]

    Chen W, Wang F, Sun H, 2021 S2tnet: Spatio-temporal transformer networks for trajectory prediction in autonomous driving. Balasubramanian VN, Tsang I, eds., Proceedings of The 13th Asian Conference on Machine Learning, volume 157 of Proceedings of Machine Learning Research, 45...

  3. [11]

    IEEE Transactions on Instrumentation and Measurement 71:1--12, ://dx.doi.org/10.1109/TIM.2022.3192056

    Chen X, Zhang H, Zhao F, Cai Y, Wang H, Ye Q, 2022 a Vehicle trajectory prediction based on intention-aware non-autoregressive transformer with multi-attention learning for internet of vehicles. IEEE Transactions on Instrumentation and Measurement 71:1--12, ://dx.doi.org/10.11...

  4. [12]

    IEEE TITS 23(10):19471--19483

    Chen X, Zhang H, Zhao F, Hu Y, Tan C, Yang J, 2022 b Intention-aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles. IEEE TITS 23(10):19471--19483

  5. [13]

    Physical review E 66(1):016121

    Dall J, Christensen M, 2002 Random geometric graphs. Physical review E 66(1):016121

  6. [14]

    Proceedings of the IEEE CVPR Workshops, 1468--1476

    Deo N, Trivedi MM, 2018 a Convolutional social pooling for vehicle trajectory prediction. Proceedings of the IEEE CVPR Workshops, 1468--1476

  7. [15]

    2018 IEEE IV, 1179--1184 (IEEE)

    Deo N, Trivedi MM, 2018 b Multi-modal trajectory prediction of surrounding vehicles with maneuver based lstms. 2018 IEEE IV, 1179--1184 (IEEE)

  8. [16]

    2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6796--6805, ://dx.doi.org/10.1109/CVPR42600.2020.00683

    Fang L, Jiang Q, Shi J, Zhou B, 2020 Tpnet: Trajectory proposal network for motion prediction. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6796--6805, ://dx.doi.org/10.1109/CVPR42600.2020.00683

  9. [17]

    Nature 615(7953):620--627

    Feng S, Sun H, Yan X, Zhu H, Zou Z, Shen S, Liu HX, 2023 Dense reinforcement learning for safety validation of autonomous vehicles. Nature 615(7953):620--627

  10. [18]

    Proceedings of the IEEE/CVF ICCV Workshops, 0--0

    Fernando T, Denman S, Sridharan S, Fookes C, 2019 Neighbourhood context embeddings in deep inverse reinforcement learning for predicting pedestrian motion over long time horizons. Proceedings of the IEEE/CVF ICCV Workshops, 0--0

  11. [19]

    IEEE TITS

    Gao K, Li X, Chen B, Hu L, Liu J, Du R, Li Y, 2023 Dual transformer based prediction for lane change intentions and trajectories in mixed traffic environment. IEEE TITS

  12. [20]

    2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2255--2264, ://dx.doi.org/10.1109/CVPR.2018.00240

    Gupta A, Johnson J, Fei-Fei L, Savarese S, Alahi A, 2018 Social gan: Socially acceptable trajectories with generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2255--2264, ://dx.doi.org/10.1109/CVPR.2018.00240

  13. [21]

    Thrun S, Saul L, Sch\" o lkopf B, eds., Advances in Neural Information Processing Systems, volume 16 (MIT Press)

    Hamerly G, Elkan C, 2003 Learning the k in k-means. Thrun S, Saul L, Sch\" o lkopf B, eds., Advances in Neural Information Processing Systems, volume 16 (MIT Press)

  14. [22]

    arXiv preprint arXiv:2104.14079

    Hasan M, Solernou A, Paschalidis E, Wang H, Markkula G, Romano R, 2021 Maneuver-aware pooling for vehicle trajectory prediction. arXiv preprint arXiv:2104.14079

  15. [23]

    Proceedings of the IEEE CVPR workshops, 954--960

    Huang X, Cheng X, Geng Q, Cao B, Zhou D, Wang P, Lin Y, Yang R, 2018 The apolloscape dataset for autonomous driving. Proceedings of the IEEE CVPR workshops, 954--960

  16. [24]

    Proceedings of the IEEE CVPR, 7482--7491

    Kendall A, Gal Y, Cipolla R, 2018 Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. Proceedings of the IEEE CVPR, 7482--7491

  17. [25]

    arXiv preprint arXiv:1907.07792

    Li X, Ying X, Chuah MC, 2020 Grip++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving. arXiv preprint arXiv:1907.07792

  18. [26]

    Computer-Aided Civil and Infrastructure Engineering 39(1):120--135

    Li Z, Chen Z, Li Y, Xu C, 2024 a Context-aware trajectory prediction for autonomous driving in heterogeneous environments. Computer-Aided Civil and Infrastructure Engineering 39(1):120--135

  19. [27]

    CHAIN 1(1):46--53

    Li Z, Cui Z, Liao H, Ash J, Zhang G, Xu C, Wang Y, 2024 b Steering the future: Redefining intelligent transportation systems with foundation models. CHAIN 1(1):46--53

  20. [28]

    IEEE Transactions on Artificial Intelligence

    Liao H, Kong H, Wang B, Wang C, Ye W, He Z, Xu C, Li Z, 2025 a Cot-drive: Efficient motion forecasting for autonomous driving with llms and chain-of-thought prompting. IEEE Transactions on Artificial Intelligence

  21. [29]

    Liao H, Li X, Li Y, Kong H, Wang C, Wang B, Guan Y, Tam K, Li Z, 2024 a Cdstraj: Characterized diffusion and spatial-temporal interaction network for trajectory prediction in autonomous driving. IJCAI

  22. [30]

    arXiv preprint arXiv:2407.07020

    Liao H, Li Y, Li Z, et al., 2024 b Less is more: Efficient brain-inspired learning for autonomous driving trajectory prediction. arXiv preprint arXiv:2407.07020

  23. [31]

    AAAI, volume 38, 10332--10340

    Liao H, Li Z, Shen H, Zeng W, Liao D, Guofa, Li SE, Xu C, 2024 c Bat: Behavior-aware human-like trajectory prediction for autonomous driving. AAAI, volume 38, 10332--10340

  24. [32]

    IJCAI 2024

    Liao H, Li Z, Wang C, Shen H, Wang B, Liao D, Li G, Xu C, 2024 d Mftraj: Map-free, behavior-driven trajectory prediction for autonomous driving. IJCAI 2024

  25. [33]

    ICRA (IEEE)

    Liao H, Liu S, Li Y, Li Z, Wang C, Li Y, Li SE, Xu C, 2024 e Human observation-inspired trajectory prediction for autonomous driving in mixed-autonomy traffic environments. ICRA (IEEE)

  26. [34]

    Communications in Transportation Research 4:100116

    Liao H, Shen H, Li Z, Wang C, Li G, Bie Y, Xu C, 2024 f Gpt-4 enhanced multimodal grounding for autonomous driving: Leveraging cross-modal attention with large language models. Communications in Transportation Research 4:100116

  27. [35]

    Proceedings of the IJCAI-24

    Liao H, Wang C, Li Z, Li Y, Wang B, Li G, Xu C, 2024 g Physics-informed trajectory prediction for autonomous driving under missing observation. Proceedings of the IJCAI-24

  28. [36]

    IEEE Transactions on Intelligent Transportation Systems 1--15

    Liao H, Wang C, Zhu K, Ren Y, Gao B, Li SE, Xu C, Li Z, 2025 b Minds on the move: Decoding trajectory prediction in autonomous driving with cognitive insights. IEEE Transactions on Intelligent Transportation Systems 1--15

  29. [37]

    Proceedings of the AAAI, volume 33, 6120--6127

    Ma Y, Zhu X, Zhang S, Yang R, Wang W, Manocha D, 2019 Trafficpredict: Trajectory prediction for heterogeneous traffic-agents. Proceedings of the AAAI, volume 33, 6120--6127

  30. [38]

    Theoretical Issues in Ergonomics Science 21(6):728--752

    Markkula G, Madigan R, Nathanael D, Portouli E, Lee YM, Dietrich A, Billington J, Schieben A, Merat N, 2020 Defining interactions: A conceptual framework for understanding interactive behaviour in human and automated road traffic. Theoretical Issues in Ergonomics Science 21(6)...

  31. [39]

    Proceedings of the IEEE

    Mendel JM, 1995 Fuzzy logic systems for engineering: a tutorial. Proceedings of the IEEE

  32. [40]

    2019 IEEE IV, 975--980 (IEEE)

    Messaoud K, Yahiaoui I, Verroust-Blondet A, Nashashibi F, 2019 Non-local social pooling for vehicle trajectory prediction. 2019 IEEE IV, 975--980 (IEEE)

  33. [41]

    IEEE Transactions on Intelligent Vehicles 6(1):175--185

    Messaoud K, Yahiaoui I, Verroust-Blondet A, Nashashibi F, 2021 Attention based vehicle trajectory prediction. IEEE Transactions on Intelligent Vehicles 6(1):175--185

  34. [42]

    Artificial Intelligence Review 54:3361--3430

    Peng X, Luo Z, 2021 A review of q-rung orthopair fuzzy information: bibliometrics and future directions. Artificial Intelligence Review 54:3361--3430

  35. [43]

    IEEE Signal Processing Magazine

    Pillai SU, Suel T, Cha S, 2005 The perron-frobenius theorem: some of its applications. IEEE Signal Processing Magazine

  36. [44]

    2008 11th international conference on information fusion, 1--6 (IEEE)

    Schubert R, Richter E, Wanielik G, 2008 Comparison and evaluation of advanced motion models for vehicle tracking. 2008 11th international conference on information fusion, 1--6 (IEEE)

  37. [45]

    Proceedings of the National Academy of Sciences 116(50):24972--24978

    Schwarting W, Pierson A, Alonso-Mora J, Karaman S, Rus D, 2019 Social behavior for autonomous vehicles. Proceedings of the National Academy of Sciences 116(50):24972--24978

  38. [46]

    Advances in Nips 32

    Tang C, Salakhutdinov RR, 2019 Multiple futures prediction. Advances in Nips 32

  39. [47]

    IET Intelligent Transport Systems 17(2):386--399

    Tang L, Yan F, Zou B, Li W, Lv C, Wang K, 2023 Trajectory prediction for autonomous driving based on multiscale spatial-temporal graph. IET Intelligent Transport Systems 17(2):386--399

  40. [48]

    Advances in Nips 30

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser , Polosukhin I, 2017 Attention is all you need. Advances in Nips 30

  41. [49]

    IEEE TPAMI

    Wang C, Liao H, Li Z, Xu C, 2025 a Wake: Towards robust and physically feasible trajectory prediction for autonomous vehicles with wavelet and kinematics synergy. IEEE TPAMI

  42. [50]

    Wang C, Liao H, Wang B, Guan Y, Rao B, Pu Z, Cui Z, Xu C, Li Z, 2025 b Nest: A neuromodulated small-world hypergraph trajectory prediction model for autonomous driving. AAAI

  43. [51]

    Information Fusion 102924

    Wang C, Liao H, Zhu K, Zhang G, Li Z, 2025 c A dynamics-enhanced learning model for multi-horizon trajectory prediction in autonomous vehicles. Information Fusion 102924

  44. [52]

    Proceedings of the AAAI, 4685--4692

    Wang R, Wang S, Yan H, Wang X, 2023 Wsip: Wave superposition inspired pooling for dynamic interactions-aware trajectory prediction. Proceedings of the AAAI, 4685--4692

  45. [53]

    Foundations and Trends in Robotics 10(3-4):198--376

    Wang W, Wang L, Zhang C, Liu C, Sun L, et al., 2022 Social interactions for autonomous driving: A review and perspectives. Foundations and Trends in Robotics 10(3-4):198--376

  46. [54]

    IEEE Transactions on Vehicular Technology 70(1):107--120

    Wang Y, Wang C, Zhao W, Xu C, 2021 Decision-making and planning method for autonomous vehicles based on motivation and risk assessment. IEEE Transactions on Vehicular Technology 70(1):107--120

  47. [55]

    Wen L, Fu D, Li X, Cai X, MA T, Cai P, Dou M, Shi B, He L, Qiao Y, 2024 Dilu: A knowledge-driven approach to autonomous driving with large language models. ICLR

  48. [56]

    IEEE Transactions on Industrial Electronics 65(7):5999--6008

    Xie G, Gao H, Qian L, Huang B, Li K, Wang J, 2017 Vehicle trajectory prediction by integrating physics-and maneuver-based approaches using interactive multiple models. IEEE Transactions on Industrial Electronics 65(7):5999--6008

  49. [57]

    2021 ICRA, 13693--13700 (IEEE)

    Xie X, Zhang C, Zhu Y, Wu YN, Zhu SC, 2021 Congestion-aware multi-agent trajectory prediction for collision avoidance. 2021 ICRA, 13693--13700 (IEEE)

  50. [58]

    Advances in Nips 32

    Yadati N, Nimishakavi M, Yadav P, Nitin V, Louis A, Talukdar P, 2019 Hypergcn: A new method for training graph convolutional networks on hypergraphs. Advances in Nips 32

  51. [59]

    IEEE TITS

    Yang B, Fan F, Ni R, Wang H, Jafaripournimchahi A, Hu H, 2024 A multi-task learning network with a collision-aware graph transformer for traffic-agents trajectory prediction. IEEE TITS

  52. [60]

    Nature communications 15(1):408

    Yu W, Zhao C, Wang H, Liu J, Ma X, Yang Y, Li J, Wang W, Hu X, Zhao D, 2024 Online legal driving behavior monitoring for self-driving vehicles. Nature communications 15(1):408

  53. [61]

    IEEE Transactions on Intelligent Vehicles 8(1):73--83

    Zhang K, Zhao L, Dong C, Wu L, Zheng L, 2022 Ai-tp: Attention-based interaction-aware trajectory prediction for autonomous driving. IEEE Transactions on Intelligent Vehicles 8(1):73--83

  54. [62]

    IEEE TITS

    Zhao C, Chu D, Deng Z, Lu L, 2024 Human-like decision making for autonomous driving with social skills. IEEE TITS

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.