Pith. sign in

REVIEW 4 major objections 6 minor 46 references

SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that injecting explicit soft-braid topological features into attention—closest-point pairs plus motion states at those points—improves multi-agent trajectory refinement across four baselines and two datasets.

desk verdict A solid, incremental refinement module with consistent ablation support; the 'new SOTA' claim is the soft spot, not the technique. read the letter →

arxiv 2507.04263 v1 pith:YSIXASL4 submitted 2025-07-06 cs.RO

Liwen Xiao , Zhiyu Pan , Zhicheng Wang , Zhiguo Cao , Wei Li This is my paper
classification cs.RO
keywords trajectorypredictionrefinementmulti-agentmotionforecastingsoft-braidtopologybraidtheoryattentionmechanismautonomousdrivingjoint
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

Trajectory refinement takes a coarse first-stage prediction and makes it more accurate, but existing refiners model interactions only implicitly and miss the topological structure of how predicted paths relate. This paper claims that injecting an explicit "soft-braid" topology—the closest pair of points on two trajectories at the same moment, together with the distance, angle, and each vehicle's velocity and acceleration there—into cross-attention guides the refinement and improves accuracy. The same idea is extended to the relation between trajectories and lane geometry. On Argoverse 2 and INTERACTION, the resulting multi-iteration refiner, SRefiner, is reported to improve all four tested baselines and to outperform previous refinement methods. If right, this gives a simple, interpretable prior that any future predictor can plug into its refinement stage.

What carries the argument

The load-bearing object is the soft intersection point: for any two predicted trajectories $y_i$ and $y_j$, the time $t_{ij} = \arg\min_t \|y_i(t) - y_j(t)\|$ and the closest-point pair $(P_{ij}, P_{ji})$ at that time. The topology feature $\tilde{\sigma}_{i\leftarrow j} = [\dot{y}_i^{(i)}(t_{ij}), \dot{y}_j^{(i)}(t_{ij}), \ddot{y}_i^{(i)}(t_{ij}), \ddot{y}_j^{(i)}(t_{ij}), d_{ij}, \theta_{ij}^{(i)}]$ bundles the distance, the connecting-line angle, and both agents' motion states in agent $i$'s local frame, with the symmetric $\tilde{\sigma}_{j\leftarrow i}$ for agent $j$. A parallel feature $\tilde{\lambda}_{i\leftarrow k}$ describes the closest trajectory-lane pair. A three-layer MLP maps these into additive bias terms that condition standard multi-head cross-attention, letting trajectories exchange information only through physically and temporally local topological links. Recomputing the topology each iteration is what lets refinement follow the improving prediction.

What would settle it

Run the original released code of the three closed-source refinement baselines reproduced in Table 2 on the same backbones and evaluation splits; if their reported errors are lower than the reproduced numbers, the claim that SRefiner establishes a new state of the art over prior refinement methods does not survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that explicit spatio-temporal topological relationships between predicted future trajectories are a useful, previously underused signal for trajectory refinement. It defines, for every pair of trajectories, the two points closest to each other at a common time (soft intersection points), and encodes their relative distance and angle together with each agent's velocity and acceleration in the corresponding local coordinate frame. These features, called soft-braid topology, are injected as additive biases in multi-head cross-attention, first among trajectories and then between trajectories and lane centerlines. A progressive strategy recomputes the topology from the previous iteration's refined trajectories, and all iterations are supervised with a joint winner-takes-all Huber loss. The paper reports consistent error reductions across four baselines on two datasets and state-of-the-art numbers among trajectory refinement methods.

Load-bearing premise

The claimed state of the art rests on the assumption that the paper's re-implementations of the three closed-source comparison methods match those methods' original performance; if the re-implementations are weaker, the reported margin over prior refinement may shrink.

Editorial extensions

If this is right

  • Adding soft-braid topology to the refinement stage improves average min-FDE, min-ADE, and miss rate on both datasets for every baseline tested, so refinement modules need not rely only on implicit latent interactions.
  • Because the representation also covers non-crossing, logically coupled behaviors, such as one car slowing while another passes, the method captures interactions that binary over/under-crossing braid attention misses.
  • Trajectory-lane soft-braid attention pulls off-road or opposing-lane predictions back into drivable, compliant regions, which directly targets the safety-relevant failures that two-stage refinement aims to correct.
  • Iterating with progressively recomputed topology gives the best accuracy at three iterations, with the gain flattening and latency growing beyond that.
  • SRefiner refines all agents jointly in one pass, making it cheaper per scenario than refinement methods that process each agent's trajectory separately.

Reading between the lines

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

  • Because the paper only wires soft-braid features into the refinement stage, the same features could be moved upstream into the initial prediction network; that is an extension the paper does not run.
  • A natural stress test the paper leaves implicit is whether soft-braid attention still helps when the refinement module is attached to a stronger, more recent predictor than the four baselines tested.
  • If the components of $\tilde{\sigma}_{i\leftarrow j}$ (distance, angle, velocity, acceleration) were ablated one at a time, the results would show which part of the topology is actually carrying the gain; the paper does not report that breakdown.
  • The soft intersection point could be computed for heterogeneous agents or for agent-map relations beyond lanes, which the paper does not evaluate.
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

4 major / 6 minor

Summary. The manuscript proposes SRefiner, a multi-iteration, multi-agent trajectory refinement module that augments multi-head cross-attention with geometric features derived from pairwise trajectory pairs and trajectory-lane pairs. For each pair, the 'soft intersection point' is defined as the time of minimum distance between the two curves, and the feature vector combines the velocity and acceleration of both agents (or of the agent and the lane) at that time together with the distance and orientation of the connecting line. These features are recomputed at each of three refinement iterations and injected as additive keys and values in cross-attention. The module is trained on top of motion forecasting baselines with a joint winner-takes-all Huber loss over all iterations. The paper reports consistent improvements on Argoverse 2 (FJMP, Forecast-MAE) and INTERACTION (AutoBots, FJMP, HPNet), presents ablations, and compares against DCMS, R-Pred, QCNet, SmartRefine, and MTR++, claiming state-of-the-art refinement performance.

Significance. If the reported gains are reproducible, SRefiner is a useful and simple contribution: it shows that explicit nearest-approach geometric features with relative motion states can improve multi-agent trajectory refinement on two benchmarks, and the paper releases code and reports extensive ablations. The internal ablations are a strength: component ablations (Trajectory-Trajectory vs Trajectory-Lane vs topology update), a comparison against BeTop-style hard braid attention, and radius/iteration sensitivity are all reported, and the improvements are consistent across baselines and datasets. I do not see a circularity problem: the soft-braid features are computed from the trajectories being refined, which are outputs of a baseline, and the evaluation is against held-out ground truth; no target metric enters the feature construction. The principal weakness is that the state-of-the-art claim in Table 2 depends on reproductions of non-open-source baselines without fidelity checks or variance estimates, and it uses an asymmetric deployment for SmartRefine; this must be resolved before the claim can be accepted.

major comments (4)
  1. [4.2.2, Table 2] The state-of-the-art claim is not yet supported because three of the six compared methods (DCMS, R-Pred, MTR++) are re-implemented 'within our framework following their described approaches' with no validation against the original papers' reported numbers, no release of the reproduction code, and no multi-seed variance. If the reproductions are weaker than the original implementations, the margin of SRefiner over these methods is an artifact of the reproduction. Please report, for each reproduced method, the original paper's official result on the same split and metric where available, or provide a fidelity table showing that the reproduced model matches the published numbers, and release the reproduction configurations. At minimum, report mean and standard deviation over several seeds.
  2. [4.2.2, Table 2] The comparison with SmartRefine is asymmetric. The paper states that SmartRefine's official code refines one agent per forward pass and that the authors therefore refine all agents independently. This removes joint, scene-level context from SmartRefine and drives its latency to 24×N ms, while SRefiner refines all agents simultaneously. The accuracy comparison may still favor SRefiner, but the protocol as described does not measure SmartRefine in the multi-agent joint-refinement regime, so it cannot be used to claim state-of-the-art over SmartRefine. Please either run SmartRefine in a joint multi-agent mode (if its code permits scene-level context sharing), or clearly present the result as a 'per-agent refinement deployment' and avoid the unqualified state-of-the-art statement.
  3. [3.2, Eq. (3)] The formal definition of braid topology is not the braid-theoretic one used in the motivation. A braid crossing in (x,y,t) requires two strands to coincide in the projected plane at the same time t, and the over/under relation σ_{i←j} is defined by the ordering along a chosen transverse spatial coordinate at that crossing. Equation (3), however, quantifies over two different times t_i < t_j and a distance threshold ε, with no requirement of same-time spatial coincidence and no rule for assigning over/under ordering. As written, the formula does not encode the 'below' condition stated in the text. Since the paper's conceptual novelty is the transition from hard braid topology to soft-braid topology, the hard-braid baseline should be defined correctly; alternatively, the paper should explicitly state that braid theory is used only as loose inspiration and that both the baseline and the proposed method are proximity-based interaction encoders rather than topological invariants.
  4. [4.2.1 and 4.2.2, Tables 1–2] The word 'significant' is used without statistical evidence. The tables show a single run per configuration, and several margins are small (for example, HPNet validation minJointFDE improves from 0.558 to 0.548 in Table 1, and the MTR++ gap in Table 2 is 0.018 in avgMinFDE). Please add means and standard deviations over at least three training seeds (or otherwise demonstrate that the differences exceed run-to-run variation), and provide seeds and configurations in the released code so that the variance estimates can be reproduced.
minor comments (6)
  1. [3.2, Eqs. (7), (8), (14)] The paper should specify how velocity ẏ and acceleration ÿ are computed from discrete trajectory points (for example, finite differences) and how the orientation angle of the connecting vector is handled for angle wrap-around; otherwise the feature definitions are incomplete.
  2. [3.3, Eq. (9)] The rotation formula is written as post-multiplying the offset vector by a rotation matrix; please clarify the row/column convention and define θ_i precisely (for example, the global heading at the end of the historical trajectory).
  3. [Throughout] The dataset name is inconsistent: 'INTERACTIONS' appears in Table 1 and Section 4.2.1, while 'INTERACTION' is used elsewhere; please use the official name 'INTERACTION' consistently.
  4. [Throughout] There are several typos and grammatical errors, including 'Furthremore' (§3.2), 'effecient' (§1), 'birad' (§3.4, near Eq. (17)), and 'Inplementation' (§8 of the supplement); please copyedit the manuscript.
  5. [4.2.1, Table 1] The text refers to 'AutoBot' while Table 1 and the reference list use 'AutoBots'; please unify the naming.
  6. [3.5, Eq. (20)] The notation for the ground-truth trajectory is not consistent (Y_{gt,i} versus Y^{i,gt}); please define the symbol once and use it consistently.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: soft-braid features are derived from the input trajectories and lanes, not from the target metrics, and the claimed gains are evaluated on held-out ground truth.

full rationale

The derivation chain is self-contained. The soft-braid topology features (Eqs. 4-8 and 11-14) are deterministic functions of the predicted trajectories and lane polylines, used as attention biases or residual feature augmentations; they are not solved from or fitted to the evaluation metrics, and the supervised objective (Eqs. 20-22) is a standard winner-takes-all Huber loss against ground truth on the training split. The reported improvements are therefore empirical and falsifiable on held-out validation/test data, not identities imposed by construction. The only name-matched self-citation is GANet [26] in Related Work, used as background for joint prediction and not load-bearing for SRefiner's mechanism or its SOTA claim. The main caveat is a verification/fairness concern rather than circularity: DCMS, R-Pred, and MTR++ are reproduced in-house 'following their described approaches' with no official-number fidelity checks, and SmartRefine is handicapped by its single-agent official code; these are reproducibility risks for Table 2, but they do not make SRefiner's output equal to its input by construction. No circular step was found.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities; its adjustable choices are validation-set hyperparameters. The 'soft-braid' construct is a feature engineering choice, not an independently evidenced entity.

free parameters (5)
  • T-T local radius tau_a = 50 m
    Set by ablation on Argoverse v2 validation; controls which agents are allowed to interact.
  • T-L local radius tau_l = 10 m
    Set by ablation; controls which lanes are included in trajectory-lane attention.
  • Number of refinement iterations I = 3
    Chosen as a trade-off between accuracy and latency, shown in Figure 5.
  • Embedding dimension D = 64
    Model capacity hyperparameter, stated in implementation details.
  • Initial learning rate = 1e-4 (Argoverse), 3e-4 (INTERACTION)
    Optimization hyperparameter, stated in implementation details.
assumptions (3)
  • domain assumption The hand-crafted nearest-point features (distance, angle, velocity, acceleration) capture useful interaction information for refinement.
    The entire method rests on the validity of these features as an interaction prior; no theoretical guarantee is given.
  • domain assumption The predicted trajectories Y0 and HD map are accurate enough that soft intersection points computed from them are meaningful.
    The topology is computed from the possibly noisy baseline predictions and then used for refinement, so errors in Y0 propagate into the attention weights.
  • standard math Standard multi-head cross-attention, MLPs, and Huber loss work as expected in this refinement setting.
    These are standard building blocks used without modification or formal analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement." pith.science (2026). https://pith.science/paper/YSIXASL4

@misc{pith2026250704263,
  author       = {Pith},
  title        = {Pith review of: SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YSIXASL4}},
  note         = {Machine review of arXiv:2507.04263}
}
read the original abstract

Accurate prediction of multi-agent future trajectories is crucial for autonomous driving systems to make safe and efficient decisions. Trajectory refinement has emerged as a key strategy to enhance prediction accuracy. However, existing refinement methods often overlook the topological relationships between trajectories, which are vital for improving prediction precision. Inspired by braid theory, we propose a novel trajectory refinement approach, Soft-Braid Refiner (SRefiner), guided by the soft-braid topological structure of trajectories using Soft-Braid Attention. Soft-Braid Attention captures spatio-temporal topological relationships between trajectories by considering both spatial proximity and vehicle motion states at ``soft intersection points". Additionally, we extend this approach to model interactions between trajectories and lanes, further improving the prediction accuracy. SRefiner is a multi-iteration, multi-agent framework that iteratively refines trajectories, incorporating topological information to enhance interactions within traffic scenarios. SRefiner achieves significant performance improvements over four baseline methods across two datasets, establishing a new state-of-the-art in trajectory refinement. Code is here https://github.com/Liwen-Xiao/SRefiner.

Figures

Figures reproduced from arXiv: 2507.04263 by the authors.

Figure 1
Figure 1. (a) Traffic scene with multi-agent future interaction. (b) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) The two points that are closest to each other at the same moment on trajectory [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overall pipeline of Soft-Braid Refiner (SRefiner). We first input historical trajectory information and HD maps into a motion forecasting baseline model. The predicted trajectories from this model are then encoded as initial inputs for SRefiner. SRefiner com￾prises a Trajectory-Trajectory Soft-Braid Attention Module, a Trajectory-Lane Soft-Braid Attention Module, and a predicting head. The Trajectory-Trajectory modu… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of refinement results. Red lines indicate predicted future trajectories, with red dots marking their starting points, while green lines and pentagrams represent ground truth (GT) future trajectories and endpoints. The first row demonstrates that SRefiner …
Figure 5
Figure 5. Figure 5: Ablation study of the number of iteration. To balance [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [1]

    Theory of braids.Annals of Mathematics, 48(1): 101–126, 1947

    Emil Artin. Theory of braids.Annals of Mathematics, 48(1): 101–126, 1947. 1, 2, 3, 8

    work page 1947

  2. [2]

    Adapt: Efficient multi-agent trajectory prediction with adaptation

    G ¨orkay Aydemir, Adil Kaan Akan, and Fatma G¨uney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295–8305, 2023. 1, 2

    work page 2023

  3. [3]

    Path planning for multiple tethered robots using topological braids

    M Cao, K Cao, S Yuan, K Liu, Loi Wong, and L Xie. Path planning for multiple tethered robots using topological braids. InProceedings of Robotics: Science and Systems,

    work page

  4. [4]

    Neptune: Nonentangling trajec- tory planning for multiple tethered unmanned vehicles.IEEE Transactions on Robotics, 39(4):2786–2804, 2023

    M Cao, K Cao, S Yuan, et al. Neptune: Nonentangling trajec- tory planning for multiple tethered unmanned vehicles.IEEE Transactions on Robotics, 39(4):2786–2804, 2023. 2

    work page 2023

  5. [5]

    Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449, 2019

    Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449, 2019. 2

    arXiv 1910

  6. [6]

    Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders

    Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 8679– 8689, 2023. 6, 7, 2

    work page 2023

  7. [7]

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

    Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525–8535, 2023. 1, 2, 3, 6

    work page 2023

  8. [8]

    Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation

    Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525– 11533, 2020. 2

    2020

Show all 46 references
  1. [9]

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

    Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D’Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. InIn- ternational Conference on Learning Representatio...

    2022

  2. [10]

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

    Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 15303–15312, 2021. 1

    2021

  3. [11]

    Optimal maneuver for mul- tiple aircraft conflict resolution: a braid point of view

    J Hu, M Prandini, and S Sastry. Optimal maneuver for mul- tiple aircraft conflict resolution: a braid point of view. In Proceedings of the IEEE Conference on Decision and Con- trol (CDC), pages 4164–4169. IEEE, 2000. 2

    2000

  4. [12]

    Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving

    Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903–3913, 2023. 1

    2023

  5. [13]

    Learning lane graph represen- tations for motion forecasting

    Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph represen- tations for motion forecasting. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part II 16, pages 541–556. Springer,

    2020

  6. [14]

    Reasoning multi-agent behavioral topology for interac- tive autonomous driving

    Haochen Liu, Li Chen, Yu Qiao, Chen Lv, and Hongyang Li. Reasoning multi-agent behavioral topology for interac- tive autonomous driving. InConference on Neural Informa- tion Processing Systems, 2024. 2, 8, 1

    2024

  7. [15]

    Christoforos Mavrogiannis, Jonathan A DeCastro, and Sid- dhartha S Srinivasa. Abstracting road traffic via topological braids: Applications to traffic flow analysis and distributed control.The International Journal of Robotics Research, 43 (9):1299–1321, 2024. 2

    2024

  8. [16]

    Scene transformer: A unified architecture for predicting fu- ture trajectories of multiple agents

    Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zheng- dong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting fu- ture trajectories of multiple agents. InIntern...

    2021

  9. [17]

    Covernet: Multimodal behavior prediction using trajectory sets

    Tung Phan-Minh, Elena Corina Grigore, Freddy A Boulton, Oscar Beijbom, and Eric M Wolff. Covernet: Multimodal behavior prediction using trajectory sets. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14074–14083, 2020. 2

    2020

  10. [18]

    From caging to grasp- ing.The International Journal of Robotics Research, 31(7): 886–900, 2012

    A Rodriguez, MT Mason, and S Ferry. From caging to grasp- ing.The International Journal of Robotics Research, 31(7): 886–900, 2012. 2

    2012

  11. [19]

    Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs

    Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745– 13755, ...

    2023

  12. [20]

    Motion transformer with global intention localization and lo- cal movement refinement.Advances in Neural Information Processing Systems, 35:6531–6543, 2022

    Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and lo- cal movement refinement.Advances in Neural Information Processing Systems, 35:6531–6543, 2022. 3

    2022

  13. [21]

    Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3, 6

    2024

  14. [22]

    M2i: From factored marginal trajectory pre- diction to interactive prediction

    Qiao Sun, Xin Huang, Junru Gu, Brian C Williams, and Hang Zhao. M2i: From factored marginal trajectory pre- diction to interactive prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6543–6552, 2022. 2

    2022

  15. [23]

    Hpnet: Dynamic trajectory fore- casting with historical prediction attention

    Xiaolong Tang, Meina Kan, Shiguang Shan, Zhilong Ji, Jin- feng Bai, and Xilin Chen. Hpnet: Dynamic trajectory fore- casting with historical prediction attention. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15261–15270, 2024. 6, 7

    2024

  16. [24]

    Identifying driver interactions via conditional be- havior prediction

    Ekaterina Tolstaya, Reza Mahjourian, Carlton Downey, Balakrishnan Vadarajan, Benjamin Sapp, and Dragomir Anguelov. Identifying driver interactions via conditional be- havior prediction. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3473–3479. IEEE,

  17. [25]

    Multipath++: Efficient information fu- sion and trajectory aggregation for behavior prediction

    Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivas- tava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fu- sion and trajectory aggregation for behavior prediction....

    2022

  18. [26]

    Ganet: Goal area network for mo- tion forecasting

    Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for mo- tion forecasting. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609–1615. IEEE,

  19. [27]

    Futurenet-lof: Joint trajectory prediction and lane occupancy field prediction with future context encod- ing.arXiv preprint arXiv:2406.14422, 2024

    Mingkun Wang, Xiaoguang Ren, Ruochun Jin, Minglong Li, Xiaochuan Zhang, Changqian Yu, Mingxu Wang, and Wen- jing Yang. Futurenet-lof: Joint trajectory prediction and lane occupancy field prediction with future context encod- ing.arXiv preprint arXiv:2406.14422, 2024. 2

    2024 arXiv

  20. [28]

    Coordination-free multi- robot path planning for congestion reduction using topologi- cal reasoning.arXiv preprint arXiv:2205.00955, 2022

    Xiaolong Wang and Alp Sahin. Coordination-free multi- robot path planning for congestion reduction using topologi- cal reasoning.arXiv preprint arXiv:2205.00955, 2022. 2

    2022 arXiv

  21. [29]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting.arXiv preprint arXiv:2301.00493, 2023

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting.arXiv preprint arXiv:2301...

    2023 arXiv

  22. [30]

    Fimp: Future interaction mod- eling for multi-agent motion prediction.arXiv preprint arXiv:2401.16189, 2024

    Sungmin Woo, Minjung Kim, Donghyeong Kim, Sungjun Jang, and Sangyoun Lee. Fimp: Future interaction mod- eling for multi-agent motion prediction.arXiv preprint arXiv:2401.16189, 2024. 2

    2024 arXiv

  23. [31]

    Tpcn: Tempo- ral point cloud networks for motion forecasting

    Maosheng Ye, Tongyi Cao, and Qifeng Chen. Tpcn: Tempo- ral point cloud networks for motion forecasting. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11318–11327, 2021. 2

    2021

  24. [32]

    Bootstrap motion forecast- ing with self-consistent constraints

    Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecast- ing with self-consistent constraints. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8504–8514, 2023. 2, 3, 6

    2023

  25. [33]

    Lanercnn: Distributed representations for graph-centric motion forecasting

    Wenyuan Zeng, Ming Liang, Renjie Liao, and Raquel Urta- sun. Lanercnn: Distributed representations for graph-centric motion forecasting. In2021 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), pages 532–

  26. [34]

    Interaction dataset: An international, adversarial and coop- erative motion dataset in interactive driving scenarios with semantic maps.arXiv preprint arXiv:1910.03088, 2019

    Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and coop- erative motion dataset in interactive driving scenar...

    1910 arXiv

  27. [35]

    Decoupling motion forecasting into directional intentions and dynamic states

    Bozhou Zhang, Nan Song, and Li Zhang. Decoupling motion forecasting into directional intentions and dynamic states. arXiv preprint arXiv:2410.05982, 2024. 1, 2

    2024 arXiv

  28. [36]

    Simpl: A simple and efficient multi-agent motion prediction base- line for autonomous driving.IEEE Robotics and Automation Letters, 2024

    Lu Zhang, Peiliang Li, Sikang Liu, and Shaojie Shen. Simpl: A simple and efficient multi-agent motion prediction base- line for autonomous driving.IEEE Robotics and Automation Letters, 2024. 2

    2024

  29. [37]

    Tnt: Target-driven trajectory pre- diction

    Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Ben Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, et al. Tnt: Target-driven trajectory pre- diction. InConference on Robot Learning, pages 895–904. PMLR, 2021. 1, 2

    2021

  30. [38]

    Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction

    Yang Zhou, Hao Shao, Letian Wang, Steven L Waslander, Hongsheng Li, and Yu Liu. Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 15281–15290, 202...

    2024

  31. [39]

    Hivt: Hierarchical vector transformer for multi-agent motion prediction

    Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Ke- jie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8823–8833, 2022. 1, 2

    2022

  32. [40]

    Query-centric trajectory prediction

    Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17863–17873, 2023. 1, 2, 3, 6

    2023

  33. [41]

    Biff: Bi-level future fusion with polyline-based coordinate for interactive trajec- tory prediction

    Yiyao Zhu, Di Luan, and Shaojie Shen. Biff: Bi-level future fusion with polyline-based coordinate for interactive trajec- tory prediction. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 8260–8271,

  34. [43]

    missed”(>2m FDE) in the the “best

    Evaluation metrics We first introduce three evaluation metrics used in the Ar- goverse v2 dataset [29]: avgMinFDE, avgMinADE and ac- torMR. avgMinFDE.The average Final Displacement Error (FDE) associated with the predicted world that has the lowest mean FDE among all ( K ) pre...

  35. [44]

    It pre- dicts the intersection relationships between future trajec- tories and uses ground truth (GT) to supervise these rela- tionships

    Discussion with BeTop BeTop [14] focuses on the integration of trajectory pre- diction and planning (IPP) in autonomous driving. It pre- dicts the intersection relationships between future trajec- tories and uses ground truth (GT) to supervise these rela- tionships. It is a di...

  36. [45]

    Ablation study of the effect of each components of SRefiner with FJMP [19] on Argoverse v2 [29]

    Inplementation details For both Argoverse v2 [29] and INTERACTIONS [34] dataset, we train the model for64epochs with a batch size Table 6. Ablation study of the effect of each components of SRefiner with FJMP [19] on Argoverse v2 [29]. Method Traj-Traj Soft-Braid Traj-Lane Sof...

  37. [46]

    Here, we present ablation studies using FJMP [19] on both two datasets

    More ablation study In the main paper, we report ablation studies using Forecast- MAE [6] on Argoverse v2 dataset [29] due to the page limit. Here, we present ablation studies using FJMP [19] on both two datasets. Tables 6 and 7 illustrate the ab- lation study of the impact of...

  38. [2023]

    6:Details of evaluation metrics

    1 SRefiner: Soft-Braid Attention for Multi-Agent Trajectory Refinement Supplementary Material The overall structure of the supplementary material is listed as follows: ▷Sec. 6:Details of evaluation metrics. ▷Sec. 7:Discussion with BeTop [14]. ▷Sec. 8:Implementation details. ▷S...

Pith tools

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