Pith. sign in

REVIEW 3 major objections 6 minor 68 references

ILNet: Trajectory Prediction with Inverse Learning Attention for Enhancing Intention Capture

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read ILNet claims that conditioning agent interactions on the immediately-next historical state, then reasoning backwards through history, yields sharper intention capture and state-of-the-art joint trajectory prediction on INTERACTION…

desk verdict The DAS module is solid and the inverse attention idea is fresh, but the paper's own description of the IL attention may peek at future ground truth at the last history step—resolve that before trusting the INTERACTION gains. read the letter →

arxiv 2507.06531 v1 pith:AAIQ5KD5 submitted 2025-07-09 cs.CV

classification cs.CV
keywords trajectorypredictioninverselearningattentionintentioncapturedynamicanchorselectionmulti-agentinteractionmultimodalmotionforecastingautonomousdriving
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

This paper claims that trajectory prediction improves when the model is allowed to reason backwards through the observed history. ILNet's inverse learning attention lets each agent first attend to the interacting agent's state one step into the future, still inside the observed window, and then rolls that information back to earlier timestamps to form a "proposed intention" that conditions the current interaction. On top of that, a dynamic anchor selection module picks a keypoint of each proposed trajectory to anchor a cheap refinement pass. On the INTERACTION benchmark the method reports the best joint prediction error on the leaderboard (minJointFDE 0.777, minJointADE 0.248), and on Argoverse it reports top minFDE (1.099) and miss rate (0.106) among the single-model baselines listed, with fewer parameters than several competitors and only about 3 ms added inference latency. If these results hold, inverse temporal reasoning is a stronger inductive bias for intention capture than the forward temporal attention used by prior factorized-attention models.

What carries the argument

Inverse Learning (IL) attention is the load-bearing object: three graph-attention layers, namely Ego-Agent Temporal attention over the ego's own history, Agent Future attention from ego at time t to interacting agents at time t+1, and Agent History attention from that prior embedding back to time t-1. The t+1 state is treated as prior information and the t-1 state as historical context, and the resulting inverse learning embedding is added to the spatial and temporal embeddings before factorized attention. The Dynamic Anchor Selection (DAS) module converts historical and proposed trajectories to polar coordinates, runs two 2D convolutions over history and mode dimensions, concatenates the features, and decodes a normalized keypoint position that becomes the refinement anchor.

What would settle it

Run the INTERACTION experiments with the Agent Future edge at the final historical timestamp strictly masked so that t+1 never points into the first prediction frame, and compare the inverse (FA+HA) ordering against the forward ordering; if the validation gap collapses, the reported gain is leakage of the future outcome rather than inverse reasoning. A complementary check is to train with the t+1 query offset by two timesteps and see whether the claimed intention-capture effect scales with lookahead.

Watch

Extended reading notes

Core claim

The central claim is that inverse temporal reasoning within the observed history is a better inductive bias for interactive intention capture than forward temporal attention. The method's Agent Future attention links the ego agent at timestamp t to interacting agents at t+1, which is still inside the observed window, and its Agent History attention then reverses that message direction back to t-1, producing a proposed intention that is summed with the ego's spatial and temporal embeddings to drive the mode queries. The authors report that this inverse ordering outperforms the forward ordering on the INTERACTION validation set, and that the full model reaches minJointFDE 0.777 and minJointADE 0.248 on the INTERACTION test leaderboard and minFDE 1.099 with miss rate 0.106 on Argoverse. They further claim that the dynamic anchor selection module improves accuracy with almost no parameter increase and about 3 ms added latency.

Load-bearing premise

The mechanism assumes that for every historical timestamp t the interacting agent's state at t+1 is available as a clean training signal, and the paper does not state how the final historical timestamp's t+1 is handled, so it is unclear whether the last step is masked or already belongs to the future prediction window.

Editorial extensions

If this is right

  • The full pipeline (inverse attention plus DAS) reports a 4.4% minJointFDE and 6.7% minJointADE reduction on INTERACTION validation relative to the temporal-attention baseline.
  • ILNet's reported INTERACTION test numbers (minJointFDE 0.777, minJointADE 0.248) place it first on that leaderboard at the time of submission, against single-model baselines that include HPNet and FJMP.
  • On Argoverse, the reported minFDE 1.099 and miss rate 0.106 are the best among the listed single-model methods, with 0.7 to 0.8 million fewer parameters than the closest competitor (DGFNet).
  • The inverse attention pair adds only about 3 ms of inference latency on challenging Argoverse scenarios while improving every accuracy and diversity metric except miss rate and drivable-area occupancy relative to the baseline.
  • Ablation of the DAS component shows the 2D CNN features are needed for its gain, and a two-anchor variant costs 8 ms for a 0.001 minJointFDE improvement, so the single-anchor design is the claimed efficiency sweet spot.

Reading between the lines

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

  • The authors leave implicit that the t+1 cue acts as a short-horizon bootstrapping signal; a natural test is to swap in t+2 or a learned offset to see whether the gain scales with the lookahead distance.
  • A separate testable consequence is that the inverse ordering could transfer to other sequence prediction tasks with partially observed futures, such as pedestrian intent or general behavior prediction; the paper does not report such experiments.
  • Because the DAS anchor is a continuous keypoint, an obvious extension is to regress multiple keypoints per mode to split the trajectory into segments, trading the 8 ms penalty the paper measured for a second anchor against finer refinement.
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

3 major / 6 minor

Summary. The paper proposes ILNet, a two-stage trajectory prediction model built on a query-centric architecture. The first contribution is an Inverse Learning (IL) attention mechanism that models historical interactions by connecting an ego agent at time t to interacting agents at time t+1, then reasoning backward from t+1 to t−1 to form 'proposed intentions.' The second contribution is a Dynamic Anchor Selection (DAS) module that predicts a keypoint on each proposed trajectory and uses it as the refinement anchor. The method is evaluated on the INTERACTION joint-prediction benchmark and the Argoverse marginal-prediction benchmark, with additional diversity metrics, latency measurements, and ablation studies. The paper claims state-of-the-art results on both benchmarks, attributing most of the gains to the IL attention and showing that DAS adds little latency or parameter overhead.

Significance. If the IL attention mechanism operates without leakage, the paper makes a plausible and useful architectural proposal: adding a temporally shifted interaction edge and a learnable anchor-selection module to a QCNet-style backbone yields consistent improvements on INTERACTION, and the release of code would support reproducibility. The ablation studies (Table III and Table IV) are reasonably thorough, the diversity metrics and inference-time measurements are a strength, and the parameter counts are reported for most comparisons. The central significance claim, however, is currently undercut by an unspecified temporal boundary in the core mechanism and by an overbroad state-of-the-art statement on Argoverse, so the contribution cannot yet be fully assessed.

major comments (3)
  1. [Section III-C.2, Eqs. (3)-(4)] The temporal boundary of the Agent Future attention is unstated, and this is load-bearing for the paper's central claim. The text defines edges between ego agents at timestamp t in [0,T-1] and interacting agents at t+1, and explicitly calls the t+1 states 'knowable' and 'visible in future spacetime.' With the standard data splits used in Section IV-A (2 s history for Argoverse and 1 s history for INTERACTION), the last historical timestamp is t=T-1, so t+1 is the first ground-truth future frame. The manuscript nowhere states that this edge is masked or index-shifted. If it is not masked, then q_f^a in Eq. (3) and q_i^a in Eq. (4) are conditioned on labels unavailable at inference, and the ablation gains in Table III (ID-5 vs. ID-6, ID-7 vs. ID-1) would be leakage artifacts rather than evidence for a causal inverse-learning mechanism. If the implementation masks this edge, the text should state the masking rule explicitly and also address the agent-history attention boundary at t=0, where t-1 is outside the observed interval.
  2. [Section IV-B, Table I] The abstract and Section IV-B claim state-of-the-art performance on Argoverse, but this is not supported by the official ranking metric. Table I reports ILNet B-minFDE6 = 1.745, which is worse than DGFNet (1.742) and DyMap (1.729), even though the table caption states that B-minFDE6 is the official ranking metric. The per-metric best minFDE6 (1.099) and MR6 (0.106) are genuine results, but the paper should either use the official ranking metric when making state-of-the-art claims or explicitly limit the claim to those two metrics. As written, the headline claim in the abstract and the conclusion overstate the empirical result.
  3. [Section III.E, Eq. (15)] The training loss is not precisely specified, which obstructs reproduction of the main empirical claims. The sum in Eq. (15) runs over t = -F+1 to 0, but the problem formulation in Section III.A defines predictions as future states S_1,...,S_F; if t=0 is included, a regression target at the final observed timestamp is used, and the normalization 1/(T N) refers to an undefined T. In addition, Eq. (12) uses n both as the agent index in the left-hand side and inside the argmin over modes, making the joint-mode selection rule formally ambiguous. These details should be clarified before the reported numbers can be independently reproduced.
minor comments (6)
  1. [Abstract and Section I] The word 'foward' appears twice and should be 'forward'; Section IV-B also contains 'Agroverse' and 'valuating' elsewhere in Section IV.
  2. [Table VIII] The table formatting is garbled: strings such as 's 37.6w/o DAS 4.3 p 91.4 0.540 0.170' merge the parameter column, latency column, and metric columns. Please restore a clean table layout with separate columns for 'Method', 'Param(M)', 'T Inf(ms)', and the metrics.
  3. [Table VI] The caption does not state that the results are on the validation set, and the Argoverse minFDE values (around 0.867) differ markedly from the test-set values in Table I (1.099). Please state the split explicitly in the caption and discuss the relationship between validation and test numbers.
  4. [Section IV-B] In the text comparing with LaneGCN, the method name is written 'LanGCN', while the table lists 'LaneGCN'; please use a consistent name.
  5. [Figure 2 and Figure 3 captions] The Figure 3 caption appears to concatenate method names ('HPNet ILNetFJMP', 'DGFNet HPNet SIMPL LAformer ILNet') without separators, making the qualitative panels hard to parse; please align the labels with the sub-panels.
  6. [Section IV-C, Table V] The 'challenging scenario' subset is constructed by the authors using four custom thresholds (Section IV-C). Since this subset is not a standardized benchmark, the comparison in Table V should be accompanied by a clear statement that the subset is non-official and that the selection criteria are part of the evaluation protocol, which the paper currently mentions only implicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ILNet's central claims rest on external benchmark evaluations and an independently trainable architecture; no reported prediction reduces by construction to a fitted constant or to a self-citation chain.

full rationale

The paper's claimed derivation chain is self-contained with respect to the reported results. The Inverse Learning attention is an architectural module that operates on historical state embeddings, and the Dynamic Anchor Selection module selects refine-ment anchors from proposed trajectories; neither module defines the final evaluation metrics (minJointFDE, minJointADE, minFDE, MR) in terms of its own fitted parameters. The reported numbers are test-set results on the INTERACTION and Argoverse benchmarks, not quantities that equal the training or validation losses by construction. The ablation results compare model variants on held-out validation data, so the gains attributed to FA+HA and DAS are empirical rather than definitional. Citations to prior work such as QCNet and HPNet are contextual related-work comparisons and are not used as load-bearing external theorems. The reader-raised concern about the temporal boundary of the Agent Future attention (whether t+1 at the last historical timestamp reaches into the ground-truth future) is a potential label-leakage or correctness issue, not a circularity issue: even if leakage occurred, it would invalidate the empirical comparison but would not make the prediction equivalent to the input by definition. No fitted input is renamed as a prediction, no uniqueness theorem is imported from the authors' own prior work, and no known result is repackaged under new coordinates. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 2 invented entities

The central claim rests on standard deep-learning training plus a data-availability assumption: next-timestamp states inside the history window are usable without leakage. Hand-set hyperparameters (K, Rf/Rh, Rm, hidden dimension, challenged-scenario thresholds) are tuned on validation data. The 'proposed intention' and 'anchor keypoint' are internal learned constructs, not physical entities with independent evidence.

free parameters (5)
  • Number of trajectory modes K = 6
    Set to 6 for both datasets and all metric evaluations, as stated in Section IV.A; mode count affects multimodal output and is a hand-chosen hyperparameter.
  • IL attention search radii Rf/Rh = 100 on Argoverse, 80 on INTERACTION
    Chosen via validation scan in Table VI; the paper reports best performance at 2x radius on Argoverse and 1x on INTERACTION, so these are tuned per dataset.
  • Local area radius Rm = 50 on Argoverse, 80 on INTERACTION
    Follows the HPNet setup (Section IV.A); affects which lane and agent neighbors are attended.
  • Hidden dimension = 128
    Model width chosen in implementation details; no sweep is reported.
  • Challenging-scenario thresholds = FDE > 5 m, proximity < 5 m, interaction time >= 25 timestamps, absolute angle >= 10 degrees
    Hand-defined filters selecting 429 Argoverse validation scenarios for the sub-analysis in Table V; this is a post hoc evaluation subset, not part of the main architecture.
assumptions (3)
  • domain assumption For every timestamp t used in Agent Future attention, the interacting agent's state at t+1 lies inside the observed history and is usable without leaking future ground truth.
    Invoked in Section III.C.2 when the paper states t in [0,T-1] and uses t+1; the boundary at the last historical timestamp is not specified.
  • domain assumption Multi-head graph cross-attention over mode queries can propagate the proposed intention embeddings into better trajectory proposals and refinements.
    Section III.C.3 assumes attention layers refine embeddings; no formal guarantee is given.
  • domain assumption The reported leaderboard and validation numbers come from a faithful single run of the architecture described.
    No multiple seeds or code verification are provided; the SOTA claim depends on this.
invented entities (2)
  • Proposed intention embedding (q_i^a)
    purpose: Bridges interacting agents' t+1 state to ego's t-1 history to modulate intention-aware attention.
    Internal latent vector; no external falsifiable handle beyond the final trajectory metrics.
  • Trajectory change keypoint anchor
    purpose: Selected point on each proposed trajectory that focuses refinement on uncertain future context.
    Construct of the Dynamic Anchor Selection module; no independent measurement outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ILNet: Trajectory Prediction with Inverse Learning Attention for Enhancing Intention Capture." pith.science (2026). https://pith.science/paper/AAIQ5KD5

@misc{pith2026250706531,
  author       = {Pith},
  title        = {Pith review of: ILNet: Trajectory Prediction with Inverse Learning Attention for Enhancing Intention Capture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AAIQ5KD5}},
  note         = {Machine review of arXiv:2507.06531}
}
read the original abstract

Trajectory prediction for multi-agent interaction scenarios is a crucial challenge. Most advanced methods model agent interactions by efficiently factorized attention based on the temporal and agent axes. However, this static and foward modeling lacks explicit interactive spatio-temporal coordination, capturing only obvious and immediate behavioral intentions. Alternatively, the modern trajectory prediction framework refines the successive predictions by a fixed-anchor selection strategy, which is difficult to adapt in different future environments. It is acknowledged that human drivers dynamically adjust initial driving decisions based on further assumptions about the intentions of surrounding vehicles. Motivated by human driving behaviors, this paper proposes ILNet, a multi-agent trajectory prediction method with Inverse Learning (IL) attention and Dynamic Anchor Selection (DAS) module. IL Attention employs an inverse learning paradigm to model interactions at neighboring moments, introducing proposed intentions to dynamically encode the spatio-temporal coordination of interactions, thereby enhancing the model's ability to capture complex interaction patterns. Then, the learnable DAS module is proposed to extract multiple trajectory change keypoints as anchors in parallel with almost no increase in parameters. Experimental results show that the ILNet achieves state-of-the-art performance on the INTERACTION and Argoverse motion forecasting datasets. Particularly, in challenged interaction scenarios, ILNet achieves higher accuracy and more multimodal distributions of trajectories over fewer parameters. Our codes are available at https://github.com/mjZeng11/ILNet.

Figures

Figures reproduced from arXiv: 2507.06531 by the authors.

Figure 1
Figure 1. The difference between previous agent-agent interaction methods and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of ILNet. The proposed ILNet consists of Trajectory Proposal stage and Trajectory Refinement stage. The Inverse Learning Attention [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparisons with recent works, (a) and (b) are in Argoverse, (c) and (d) are in INTERACTION. Predictions are indicated by red lines, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparisons of predictive accuracy (minFDE [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on different validation sets, (a) and (b) are in INTERACTION, (c) and (d) are in Argoverse. Predictions are indicated by red lines, [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Performance curves on the Argoverse validation set, evaluated at the [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 63 canonical work pages

  1. [1]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y . Chen, R. Liao, S. Feng, and R. Urtasun, “Learning lane graph representations for motion forecasting,” in Proc. Eur. Conf. Comput. Vis. (ECCV) , Glasgow, UK, 2020, pp. 541–556

  2. [2]

    Deep learning-based vehicle behavior prediction for autonomous driving applications: A review,

    S. Mozaffari, O. Y . Al-Jarrah, M. Dianati, P. Jennings, and A. Mouzakitis, “Deep learning-based vehicle behavior prediction for autonomous driving applications: A review,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 1, pp. 33–47, 2020

  3. [3]

    Social predictive intelligent driver model for autonomous driving simulation,

    Z. Deng, W. Hu, T. Huang, C. Sun, J. Zhong, and A. Khajepour, “Social predictive intelligent driver model for autonomous driving simulation,” Automot. Innov., vol. None, pp. 1–12, 2025. 12

  4. [4]

    Covernet: Multimodal behavior prediction using trajectory sets,

    T. Phan-Minh, E. C. Grigore, F. A. Boulton, O. Beijbom, and E. M. Wolff, “Covernet: Multimodal behavior prediction using trajectory sets,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2020, pp. 14074–14083

  5. [5]

    Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction,

    Y . Chai, B. Sapp, M. Bansal, and D. Anguelov, “Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction,” arXiv preprint arXiv:1910.05449, 2019

  6. [6]

    Rules of the road: Predicting driving behavior with a convolutional model of semantic interactions,

    J. Hong, B. Sapp, and J. Philbin, “Rules of the road: Predicting driving behavior with a convolutional model of semantic interactions,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 8454– 8462

  7. [7]

    Ganet: Goal area network for motion forecasting,

    M. Wang, X. Zhu, C. Yu, W. Li, Y . Ma, R. Jin, X. Ren, D. Ren, M. Wang, and W. Yang, “Ganet: Goal area network for motion forecasting,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA) , 2023, pp. 1609–1615

  8. [8]

    Lanercnn: Distributed representations for graph-centric motion forecasting,

    W. Zeng, M. Liang, R. Liao, and R. Urtasun, “Lanercnn: Distributed representations for graph-centric motion forecasting,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) , 2021, pp. 532–539

Show all 68 references
  1. [9]

    Vectornet: Encoding hd maps and agent dynamics from vectorized representation,

    J. Gao, C. Sun, H. Zhao, Y . Shen, D. Anguelov, C. Li, and C. Schmid, “Vectornet: Encoding hd maps and agent dynamics from vectorized representation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 11525–11533

  2. [10]

    Tpcn: Temporal point cloud networks for motion forecasting,

    M. Ye, T. Cao, and Q. Chen, “Tpcn: Temporal point cloud networks for motion forecasting,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 11318–11327

  3. [11]

    Convolutional neural network for tra- jectory prediction,

    N. Nikhil and B. Tran Morris, “Convolutional neural network for tra- jectory prediction,” in Proc. Eur. Conf. Comput. Vis. (ECCV) Workshops, 2018, pp. 0–0

  4. [12]

    Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction,

    A. Mohamed, K. Qian, M. Elhoseiny, and C. Claudel, “Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2020, pp. 14424–14432

  5. [13]

    SR-LSTM: State refinement for LSTM towards pedestrian trajectory prediction,

    P. Zhang, W. Ouyang, P. Zhang, J. Xue, and N. Zheng, “SR-LSTM: State refinement for LSTM towards pedestrian trajectory prediction,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition , 2019, pp. 12085–12094

  6. [14]

    Spatio-temporal graph trans- former networks for pedestrian trajectory prediction,

    C. Yu, X. Ma, J. Ren, H. Zhao, and S. Yi, “Spatio-temporal graph trans- former networks for pedestrian trajectory prediction,” in Proc. Computer Vision–ECCV 2020: 16th European Conf. , Glasgow, UK, Aug. 2020, pp. 507–523

  7. [15]

    TNT: Target-driven trajectory prediction,

    H. Zhao, J. Gao, T. Lan, C. Sun, B. Sapp, B. Varadarajan, Y . Shen, Y . Shen, Y . Chai, C. Schmid, and others, “TNT: Target-driven trajectory prediction,” in Proc. Conf. Robot. Learn. , 2021, pp. 895–904

  8. [16]

    Densetnt: End-to-end trajectory prediction from dense goal sets,

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory prediction from dense goal sets,” in Proc. IEEE/CVF Int. Conf. Computer Vision , 2021, pp. 15303–15312

  9. [17]

    GoHome: Graph-oriented heatmap output for future motion estimation,

    T. Gilles, S. Sabatini, D. Tsishkou, B. Stanciulescu, and F. Moutarde, “GoHome: Graph-oriented heatmap output for future motion estimation,” in Proc. 2022 Int. Conf. Robotics and Automation (ICRA) , 2022, pp. 9107–9114

  10. [18]

    Wayformer: Motion forecasting via simple & efficient attention networks,

    N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat, and B. Sapp, “Wayformer: Motion forecasting via simple & efficient attention networks,” in Proc. 2023 IEEE Int. Conf. Robotics and Automation (ICRA), 2023, pp. 2980–2987

  11. [19]

    A hierarchical hybrid learning framework for multi-agent trajectory prediction,

    Y . Jiao, M. Miao, Z. Yin, C. Lei, X. Zhu, X. Zhao, L. Nie, and B. Tao, “A hierarchical hybrid learning framework for multi-agent trajectory prediction,” IEEE Trans. Intell. Transp. Syst. , vol. 25, no. 8, pp. 10344– 10354, 2024

  12. [20]

    Scene trans- former: A unified architecture for predicting multiple agent trajectories,

    J. Ngiam, B. Caine, V . Vasudevan, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal, and others, “Scene trans- former: A unified architecture for predicting multiple agent trajectories,” arXiv preprint arXiv:2106.08417 , 2021

  13. [21]

    Query-centric trajectory prediction,

    Z. Zhou, J. Wang, Y .-H. Li, and Y .-K. Huang, “Query-centric trajectory prediction,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recogni- tion, 2023, pp. 17863–17873

  14. [22]

    HPNet: Dynamic trajectory forecasting with historical prediction attention,

    X. Tang, M. Kan, S. Shan, Z. Ji, J. Bai, and X. Chen, “HPNet: Dynamic trajectory forecasting with historical prediction attention,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition , 2024, pp. 15261–15270

  15. [23]

    QCNext: A next- generation framework for joint multi-agent trajectory prediction,

    Z. Zhou, Z. Wen, J. Wang, Y .-H. Li, and Y .-K. Huang, “QCNext: A next- generation framework for joint multi-agent trajectory prediction,” arXiv preprint arXiv:2306.10508, 2023

  16. [24]

    DiffusionDrive: Truncated diffusion model for end-to-end autonomous driving,

    B. Liao, S. Chen, H. Yin, B. Jiang, C. Wang, S. Yan, X. Zhang, X. Li, Y . Zhang, Q. Zhang, et al., “DiffusionDrive: Truncated diffusion model for end-to-end autonomous driving,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2025, pp. 12037–12047

  17. [25]

    Eliminating uncertainty of driver’s social preferences for lane change decision-making in realistic simulation environment,

    Z. Deng, W. Hu, C. Sun, D. Chu, T. Huang, W. Li, C. Yu, M. Pirani, D. Cao, A. Khajepour, and Y . Chao, “Eliminating uncertainty of driver’s social preferences for lane change decision-making in realistic simulation environment,” IEEE Trans. Intell. Transp. Syst. , 2024

  18. [26]

    Hyper-relational Interaction Modeling in Multi-modal Trajectory Prediction for Intelligent Connected Vehicles in Smart Cities,

    Y . Lu, W. Wang, R. Bai, S. Zhou, L. Garg, A. K. Bashir, W. Jiang, and X. Hu, “Hyper-relational Interaction Modeling in Multi-modal Trajectory Prediction for Intelligent Connected Vehicles in Smart Cities,”Inf. Fusion, vol. 114, pp. 102682, 2025

  19. [27]

    Probabilistic prediction of inter- active driving behavior via hierarchical inverse reinforcement learning,

    L. Sun, W. Zhan, and M. Tomizuka, “Probabilistic prediction of inter- active driving behavior via hierarchical inverse reinforcement learning,” in Proc. 2018 21st Int. Conf. Intelligent Transportation Systems (ITSC) , 2018, pp. 2111–2117

  20. [28]

    MTR++: Multi-agent motion prediction with symmetric scene modeling and guided intention query- ing,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “MTR++: Multi-agent motion prediction with symmetric scene modeling and guided intention query- ing,” IEEE Trans. Pattern Anal. Machine Intell. , 2024

  21. [29]

    GAMDTP: Dynamic trajectory prediction with graph attention Mamba network,

    Y . Liu, H. Niu, and J. Zhu, “GAMDTP: Dynamic trajectory prediction with graph attention Mamba network,” arXiv:2504.04862, 2025

  22. [30]

    ProphNet: Efficient agent-centric motion forecasting with anchor-informed proposals,

    X. Wang, T. Su, F. Da, and X. Yang, “ProphNet: Efficient agent-centric motion forecasting with anchor-informed proposals,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition , 2023, pp. 21995–22003

  23. [31]

    SmartRefine: A scenario-adaptive refinement framework for efficient motion prediction,

    Y . Zhou, H. Shao, L. Wang, S. L. Waslander, H. Li, and Y . Liu, “SmartRefine: A scenario-adaptive refinement framework for efficient motion prediction,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2024, pp. 15281–15290

  24. [32]

    Criteria: A new bench- marking paradigm for evaluating trajectory prediction models for au- tonomous driving,

    C. Chen, M. Pourkeshavarz, and A. Rasouli, “Criteria: A new bench- marking paradigm for evaluating trajectory prediction models for au- tonomous driving,” in Proc. IEEE Int. Conf. Robotics Autom. (ICRA) , 2024, pp. 8265–8271

  25. [33]

    Diverse and admissible trajectory forecasting through multimodal context understanding,

    S. H. Park, G. Lee, J. Seo, M. Bhat, M. Kang, J. Francis, A. Jadhav, P. P. Liang, and L. P. Morency, “Diverse and admissible trajectory forecasting through multimodal context understanding,” in Proc. Eur. Conf. Comput. Vis. (ECCV), Glasgow, UK, 2020, vol. 16, pp. 282–298

  26. [34]

    Reliable trajectory prediction in scene fusion based on spatio-temporal structure causal model,

    J. Liu, H. Lin, X. Wang, L. Wu, S. Garg, and M. M. Hassan, “Reliable trajectory prediction in scene fusion based on spatio-temporal structure causal model,” Inf. Fusion, vol. 107, p. 102309, 2024

  27. [35]

    Trajectory prediction for safety critical maneuvers in automated highway driving,

    C. Wissing, T. Nattermann, K.-H. Glander, and T. Bertram, “Trajectory prediction for safety critical maneuvers in automated highway driving,” in Proc. 2018 21st Int. Conf. Intelligent Transportation Systems (ITSC) , 2018, pp. 131–136

  28. [36]

    When will it change the lane? A probabilistic regression approach for rarely occurring events,

    J. Schlechtriemen, F. Wirthmueller, A. Wedel, G. Breuel, and K.-D. Kuhnert, “When will it change the lane? A probabilistic regression approach for rarely occurring events,” in Proc. 2015 IEEE Intelligent Vehicles Symp. (IV), 2015, pp. 1373–1379

  29. [37]

    Social-W AGDAT: Interaction-aware trajectory prediction via Wasserstein graph double- attention network,

    J. Li, H. Ma, Z. Zhang, and M. Tomizuka, “Social-W AGDAT: Interaction-aware trajectory prediction via Wasserstein graph double- attention network,” arXiv preprint arXiv:2002.06241 , 2020

  30. [38]

    DifTraj: Diffusion Inspired by Intrinsic Intention and Extrinsic Interaction for Multi-Modal Trajectory Prediction,

    Y . Liu, X. Dong, Y . Lin, and M. Ye, “DifTraj: Diffusion Inspired by Intrinsic Intention and Extrinsic Interaction for Multi-Modal Trajectory Prediction,” in Proc. Int. Joint Conf. Artif. Intell. , 2024

  31. [39]

    Social LSTM: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social LSTM: Human trajectory prediction in crowded spaces,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition , 2016, pp. 961–971

  32. [40]

    Multimodal motion prediction with stacked transformers,

    Y . Liu, J. Zhang, L. Fang, Q. Jiang, and B. Zhou, “Multimodal motion prediction with stacked transformers,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition , 2021, pp. 7577–7586

  33. [41]

    Scenario-transferable semantic graph reasoning for interaction-aware probabilistic prediction,

    Y . Hu, W. Zhan, and M. Tomizuka, “Scenario-transferable semantic graph reasoning for interaction-aware probabilistic prediction,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 12, pp. 23212–23230, 2022

  34. [42]

    Intention- aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles,

    X. Chen, H. Zhang, F. Zhao, Y . Hu, C. Tan, and J. Yang, “Intention- aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 10, pp. 19471–19483, 2022

  35. [43]

    GRIN: Generative relation and intention network for multi-agent trajectory prediction,

    L. Li, J. Yao, W. Li, T. He, T. Xiao, J. Yan, D. Wipf, and Z. Zhang, “GRIN: Generative relation and intention network for multi-agent trajectory prediction,” Adv. Neural Inf. Process. Syst., vol. 34, pp. 27107– 27118, 2021

  36. [44]

    Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network,

    X. Mo, Z. Huang, Y . Xing, and C. Lv, “Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 7, pp. 9554–9567, 2022

  37. [45]

    Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning,

    J. Li, F. Yang, M. Tomizuka, and C. Choi, “Evolvegraph: Multi-agent trajectory prediction with dynamic relational reasoning,” Adv. Neural Inf. Process. Syst., vol. 33, pp. 19783–19794, 2020

  38. [46]

    Multi-agent trajectory prediction with difficulty-guided feature enhancement network,

    G. Xin, D. Chu, L. Lu, Z. Deng, Y . Lu, and X. Wu, “Multi-agent trajectory prediction with difficulty-guided feature enhancement network,” IEEE Robotics and Automation Letters , 2025

  39. [47]

    Laformer: Trajectory prediction for autonomous driving 13 with lane-aware scene constraints,

    M. Liu, H. Cheng, L. Chen, H. Broszio, J. Li, R. Zhao, M. Sester, and M. Y . Yang, “Laformer: Trajectory prediction for autonomous driving 13 with lane-aware scene constraints,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2024, pp. 2039–2049

  40. [48]

    GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction,

    M. Pei, S. Shi, L. Zhang, P. Li, and S. Shen, “GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction,” in Proc. Forty-second International Conference on Machine Learning , 2025

  41. [49]

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone, “Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,” in Computer Vision–ECCV 2020: 16th European Conf., Glasgow, UK, Aug. 23–28, 2020, Proc. Part XVIII , 2020, pp. 683–700

  42. [50]

    Agentformer: Agent- aware transformers for socio-temporal multi-agent forecasting,

    Y . Yuan, X. Weng, Y . Ou, and K. M. Kitani, “Agentformer: Agent- aware transformers for socio-temporal multi-agent forecasting,” in Proc. IEEE/CVF Int. Conf. Computer Vision , 2021, pp. 9813–9823

  43. [51]

    HIVT: Hierarchical vector transformer for multi-agent motion prediction,

    Z. Zhou, L. Ye, J. Wang, K. Wu, and K. Lu, “HIVT: Hierarchical vector transformer for multi-agent motion prediction,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition , 2022, pp. 8823–8833

  44. [52]

    Bidirectional agent-map interaction feature learning leveraged by map-related tasks for trajectory prediction in autonomous driving,

    B. Fan, H. Yuan, Y . Dong, Z. Zhu, and H. Liu, “Bidirectional agent-map interaction feature learning leveraged by map-related tasks for trajectory prediction in autonomous driving,” IEEE Transactions on Automation Science and Engineering , vol. 2025, doi: 10.1109/TASE.2025.3529736

  45. [53]

    SOPHIE: An attentive GAN for predicting paths compliant to social and physical constraints,

    A. Sadeghian, V . Kosaraju, A. Sadeghian, N. Hirose, H. Rezatofighi, and S. Savarese, “SOPHIE: An attentive GAN for predicting paths compliant to social and physical constraints,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2019, pp. 1349–1358

  46. [54]

    Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,

    B. Varadarajan, A. Hefny, A. Srivastava, K. S. Refaat, N. Nayakanti, A. Cornman, K. Chen, B. Douillard, C. P. Lam, D. Anguelov, and others, “Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,” in Proc. 2022 Int. Conf. Robotics and Aut...

  47. [55]

    Trajectory prediction with graph-based dual-scale context fusion,

    L. Zhang, P. Li, J. Chen, and S. Shaojie, “Trajectory prediction with graph-based dual-scale context fusion,” in Proc. 2022 IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS) , 2022, pp. 11374–11381

  48. [56]

    R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement,

    S. Choi, J. Kim, J. Yun, and J. W. Choi, “R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement,” in Proc. IEEE/CVF Int. Conf. Computer Vision , 2023, pp. 8525–8535

  49. [57]

    Motion transformer with global intention localization and local movement refinement,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Motion transformer with global intention localization and local movement refinement,” Adv. Neural Inf. Process. Syst., vol. 35, pp. 6531–6543, 2022

  50. [58]

    Bootstrap motion forecasting with self-consistent constraints,

    M. Ye, J. Xu, X. Xu, T. Wang, T. Cao, and Q. Chen, “Bootstrap motion forecasting with self-consistent constraints,” in Proc. IEEE/CVF Int. Conf. Computer Vision, 2023, pp. 8504–8514

  51. [59]

    Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction,

    C. Feng, H. Zhou, H. Lin, Z. Zhang, Z. Xu, C. Zhang, B. Zhou, and S. Shaojie, “Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction,” IEEE Robotics and Automation Letters , 2023

  52. [60]

    SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving,

    L. Zhang, P. Li, S. Liu, and S. Shaojie, “SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving,” IEEE Robotics and Automation Letters , 2024

  53. [61]

    FJMP: Fac- torized joint multi-agent motion prediction over learned directed acyclic interaction graphs,

    L. Rowe, M. Ethier, E.-H. Dykhne, and K. Czarnecki, “FJMP: Fac- torized joint multi-agent motion prediction over learned directed acyclic interaction graphs,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2023, pp. 13745–13755

  54. [62]

    Traj-MAE: Masked autoencoders for trajectory prediction,

    H. Chen, J. Wang, K. Shao, F. Liu, J. Hao, C. Guan, G. Chen, and P.-A. Heng, “Traj-MAE: Masked autoencoders for trajectory prediction,” in Proc. IEEE/CVF Int. Conf. Computer Vision , 2023, pp. 8351–8362

  55. [63]

    HDGT: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding,

    X. Jia, P. Wu, L. Chen, Y . Liu, H. Li, and J. Yan, “HDGT: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding,” IEEE Trans. Pattern Anal. Mach. Intell. , 2023

  56. [64]

    THOMAS: Trajectory heatmap output with learned multi-agent sam- pling,

    T. Gilles, S. Sabatini, D. Tsishkou, B. Stanciulescu, and F. Moutarde, “THOMAS: Trajectory heatmap output with learned multi-agent sam- pling,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022

  57. [65]

    Latent variable sequential set transformers for joint multi-agent motion prediction,

    R. Girgis, F. Golemo, F. Codevilla, M. Weiss, J. A. D’Souza, S. E. Kahou, F. Heide, and C. Pal, “Latent variable sequential set transformers for joint multi-agent motion prediction,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2022

  58. [66]

    Stochastic multiple choice learning for training diverse deep ensembles,

    S. Lee, S. P. Purushwalkam, M. Cogswell, V . Ranjan, D. Crandall, and D. Batra, “Stochastic multiple choice learning for training diverse deep ensembles,” Adv. Neural Inf. Process. Syst. , vol. 29, 2016

  59. [67]

    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 others, “Argoverse: 3D tracking and forecasting with rich maps,” in Proc. IEEE/CVF Conf. Computer Vision Pattern Recognition, 2019, pp. 8748–8757

  60. [68]

    Interaction dataset: An international, 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. Kummerle, H. Konigshof, C. Stiller, A. de La Fortelle, and others, “Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps,” arXiv prepri...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.