Pith. sign in

REVIEW 3 major objections 8 minor 49 references

ProgD: Progressive Multi-scale Decoding with Dynamic Graphs for Joint Multi-agent Motion Forecasting

T0 review · 3 major / 8 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A decoder that rebuilds future interaction graphs from its own predictions achieves top joint-motion forecasting results on INTERACTION and Argoverse 2.

desk verdict A genuinely new progressive-dynamic-graph decoder for joint motion forecasting with strong benchmark numbers, but the Argoverse 2 SOTA claim needs more baselines and error bars before I'd trust it fully. read the letter →

arxiv 2509.09210 v1 pith:54HVTO4K submitted 2025-09-11 cs.AI cs.RO

Xing Gao , Zherui Huang , Weiyao Lin , Xiao Sun This is my paper
classification cs.AIcs.RO
keywords jointmulti-agentmotionpredictiondynamicheterogeneousgraphsforecastingtrajectoryautonomousdrivingmulti-scaledecodinggraphneuralnetworksinteractionmodeling
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

ProgD argues that joint multi-agent motion prediction fails when future interactions are treated as static. The paper's central claim is that explicitly modeling how interactions among agents and roads evolve over time—by rebuilding a heterogeneous graph snapshot from each intermediate prediction—produces more accurate and more consistent joint forecasts. On the INTERACTION benchmark this claim is supported by ranking 1st among all submitted methods, reducing minJFDE from 0.9218 to 0.8620 and egoCR from 0.0044 to 0.0011; on Argoverse 2 it reduces B-minJFDE from 2.23 to 1.98. A multi-scale coarse-to-fine decoding loop, where a quick position estimate updates the graph before detailed states are predicted, is introduced to keep error from accumulating through the progressive construction. A sympathetic reader would take the paper as establishing that future interaction dynamics, not just observed ones, are worth modeling explicitly in motion forecasting.

What carries the argument

The central mechanism is the dynamic heterogeneous graph whose snapshots are rebuilt from the model's own intermediate predictions. In ProgD, the future is not a static interaction structure; each τ-second interval gets a new graph whose agent-lane edges are computed from the ℓ1 distance between predicted agent positions and lane centerlines (Eq. 2), and whose agent-agent edges form a complete graph. Two heterogeneous graph convolution modules (one on the constructed snapshot, one after a coarse prediction updates the snapshot) combined with a factorized temporal model carry the spatio-temporal message passing, and a coarse-to-fine 'predict midpoint/last position, update graph, predict full

What would settle it

On the Argoverse 2 validation split, run ProgD twice at inference: once with the coarse-predicted positions used to build graph snapshots, and once with ground-truth positions substituted at the same point. If the ground-truth version does not substantially improve B-minJFDE (or improves it only marginally), the progressive graph construction is not the source of the reported gains; if the standard model degrades when coarse predictions are deliberately corrupted, the feedback loop is a real risk.

Watch

Extended reading notes

Core claim

ProgD models future scenes as discrete-time dynamic heterogeneous graphs with agent nodes, lane nodes, agent-agent and agent-lane edges; agent-lane edges are drawn to lane centerlines within a threshold of the agents' previously predicted positions (Eq. 2). The decoder factorizes spatio-temporal processing: a cross-time attention module produces per-agent future feature sequences, then two heterogeneous graph convolution modules handle interactions—first on the constructed snapshot, and again after a coarse prediction of each interval's midpoint and final positions refreshes it. A joint header produces full multi-agent states, and the loop repeats. ProgD claims this progressive unfolding cap

Load-bearing premise

The load-bearing premise is that the coarse predicted positions are accurate enough that the dynamic graph rebuilt from them reflects the true future interactions; if that premise fails, ProgD conditions on its own errors and the progressive construction becomes a self-confirming artifact.

Editorial extensions

If this is right

  • If the claims hold, treating future interactions as static is a bottleneck; modeling their evolution is a usable path to more accurate joint forecasts.
  • Explicit coarse-to-fine decoding inside each interval is an effective mitigation for error accumulation in autoregressive graph construction.
  • The method scales to arbitrary numbers of agents with inference time (0.032 s per INTERACTION scenario) close to a static-graph baseline (0.024 s), so the dynamic machinery is cheap.
  • Lower egoCR (0.0011 vs 0.0044) and actorCR (0.01) imply that consistency metrics, not just point accuracy, improve when future graphs are regenerated.
  • Auxiliary marginal prediction and interval resolution τ are tunable knobs: finer snapshots (τ=1s) and marginal supervision both help, suggesting the decoder benefits from dense temporal supervision.

Reading between the lines

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

  • The paper does not test what happens if the coarse predictions that drive graph construction are corrupted or replaced by ground truth; a natural experiment is to feed ground-truth positions at construction time and measure B-minJFDE on Argoverse 2 validation, which would isolate how much of the gain comes from the graph dynamics versus the decoder's other components.
  • If the progressive graph is the real source of improvement, the same idea could transfer to other structured multi-agent problems such as traffic simulation or interactive planning, where future context is also generated step by step.
  • The ablation shows removing agent-lane edges hurts more than removing agent-agent edges; this suggests map compliance is a large part of the benefit, and a simpler model with only dynamic lane graphs might capture much of the gain.
  • Because the graph is rebuilt from the model's own outputs, the framework is self-confirming by design; if coarse predictions are biased, the graph encodes those biases. The paper acknowledges error accumulation (Sec. 3.3.2) but offers no analysis separating the feedback loop from the multi-scale mitigation, so the robustness of the loop is an open question.
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 / 8 minor

Summary. ProgD proposes a joint multi-agent motion forecasting method in which future interactions are explicitly represented by a discrete-time dynamic heterogeneous graph. The decoder progressively constructs graph snapshots from previously predicted agent positions (Eq. 2), performs coarse prediction, updates the snapshot, and refines joint predictions, with K=6 modalities conditioned on learnable scenario embeddings (Eq. 6). Training uses a winner-take-all joint loss plus auxiliary midpoint and marginal losses (Eqs. 7–10). The paper reports 1st place on the INTERACTION multi-agent leaderboard (Table 1) and improved B-minJFDE on Argoverse 2 (1.98 vs 2.23 for the strongest listed baseline, Table 2), supported by ablations on the Argoverse 2 validation set (Tables 4–6).

Significance. The core idea—explicitly evolving the interaction graph over the prediction horizon rather than assuming static future interactions—is a timely and potentially valuable step beyond static-graph joint predictors such as FJMP. The INTERACTION leaderboard ranking is an externally validated strong result, and the ablations (e.g., ID-1 vs ID-6 in Table 4) consistently indicate that dynamic graph modeling and multiscale decoding contribute to performance. The paper is clearly written and the architecture is described in sufficient detail to be reproduced in principle. The significance is currently limited by three issues: the Argoverse 2 SOTA claim rests on an incomplete baseline list; no variance/error bars are reported, making small margins in the ablations difficult to assess; and no experiment directly probes the sensitivity to progressive error accumulation. Addressing these would turn a promising empirical contribution into a fully supported one.

major comments (3)
  1. [Abstract / §4.2, Table 2] The abstract's claim of 'state-of-the-art performance on ... the Argoverse 2 multi-world forecasting benchmark' is not supported by the evaluation. Table 2 compares against only eight baselines and omits several strong joint-prediction methods that report on the same benchmark, such as MotionLM, MTR++, QCNet, JFP, SceneTransformer, and MotionDiffuser. The body text in §4.2 honestly says ProgD 'still achieves competitive performance,' which is a weaker and more accurate phrasing. Since the headline claim includes Argoverse 2, the authors must either add the omitted baselines (or explain their absence under the official protocol) and temper the abstract accordingly.
  2. [Tables 2 and 4] No error bars, confidence intervals, or multiple-seed results are provided anywhere. On Argoverse 2, the reported margin over SGPred (B-minJFDE 1.98 vs 2.23) is within the range of differences among baselines, and several ablation comparisons in Table 4 are very small (e.g., ID-1 vs ID-5: 1.941 vs 1.958; ID-4 vs ID-1: 1.983 vs 1.941). Without variance estimates, the reader cannot judge whether these differences are meaningful. Please report mean and standard deviation over at least three seeds for the main tables, or at minimum for the ablations that support the central dynamic-graph and multiscale-decoding claims.
  3. [§3.3.1–3.3.2, Eq. (2)] The progressive construction conditions future graph topology on coarse predicted positions. The paper acknowledges 'potential error accumulation' in §3.3.2, but no experiment isolates this feedback. Table 4 ID-1 vs ID-5 confounds multiscale decoding with the dynamic graph update, and ID-6 replaces the dynamic graph with a static one but does not vary the quality of the coarse predictions used to build edges. A concrete test would be to inject controlled noise into the predicted positions used in Eq. (2), or to use ground-truth future states as an oracle during construction, and then report B-minJFDE sensitivity. This would establish that the method is robust to its own coarse errors rather than self-confirming.
minor comments (8)
  1. [§4.1] Typo: 'Datesets' should be 'Datasets'.
  2. [Table 2] Caption typo: 'Argovese' should be 'Argoverse'.
  3. [Table 1] The DenseTNT row has values run together ('0.22400.00000.0136') and the Graphformer egoCR value 0.3450 looks suspiciously high compared to all other rows; please verify these numbers and reformat the table.
  4. [§4.2 (Inference time)] 'INTERSECTION dataset' should be 'INTERACTION dataset'.
  5. [§4.1 Implementation details] The text says 'the distance threshold ϵ for agent-agent edges is 15 meters,' but §3.3.1 states that agent-agent edges are fully connected and Eq. (2) uses ϵ only for lane-agent edges. Please clarify which edges use ϵ.
  6. [Table 4] The column headers 'G-Stat, G-Dyn, AA+AL, AA, AL' are not defined in the caption or text; please add a legend explaining the ablation components.
  7. [§3.1] 'Notions' should be 'Notations'.
  8. [Eq. (5)] There is an unmatched parenthesis in the message-passing formula. Also, the notation c_{i,p-1} and c_{l,p} appears without a clear definition of the time index for lane-node coordinates.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the central benchmark claims are externally evaluated and the progressive dynamic graph decoding is a feedback architecture, not a definitional reduction. Only minor self-citation to HeteroGCN, which is a component.

full rationale

The paper's central claims are empirical benchmark results on INTERACTION and Argoverse 2, evaluated with external, pre-defined metrics (minJFDE, minJMR, B-minJFDE, etc.). No parameter is fitted to the headline metric, and the reported scores come from a public leaderboard or a fixed validation protocol; the benchmarks are outside the method. The progressive dynamic graph construction in Sec. 3.3.1 (Eq. 2) uses previously predicted agent states to build the next snapshot, and the resulting graph features are then used to predict subsequent states. This is an autoregressive feedback loop, not a logical circularity: the predicted states are an input to the graph, not the same quantity being derived without independent evidence. The paper explicitly acknowledges 'potential error accumulation in the progressive prediction framework' in Sec. 3.3.2 and proposes multi-scale decoding as a mitigation; that is a robustness concern, not a circularity. The only self-citation is the use of HeteroGCN (Gao et al., 2023) as a static encoder component (Sec. 3.2, Eq. 1). Since the current paper's first author is the same as that of Gao et al., this is a self-citation, but it is a component-level design choice and is not the source of the claimed benchmark improvements. The Argoverse 2 'state-of-the-art' wording is weaker in Sec. 4.2 ('still achieves competitive performance') and the baseline table omits several strong recent methods; these are completeness/robustness concerns external to circularity. No step in the derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The ledger contains no invented physical entities. The model does use learnable scenario-modality embeddings m_k (Eq. 6), but these are standard latent parameters, not postulated entities. The load-bearing assumptions are the stability of progressive graph construction and the validity of benchmark metrics as safety proxies; both are domain assumptions with limited analysis in the paper.

free parameters (3)
  • Lane-agent edge distance threshold epsilon = 15 m
    Used to construct dynamic lane-agent edges in Eq. (2); chosen as a hyperparameter without an ablation study for this value specifically (Sec. 3.3.1).
  • Snapshot time interval tau = 1 s
    Length of each dynamic graph snapshot; default set to 1 s and ablated in Table 5 (Sec. 4.1).
  • Loss weights lambda1, lambda2 = 1.0, 1.0
    Weights for coarse and marginal auxiliary losses in Eq. (10); lambda2 is ablated in Table 6, lambda1 is not ablated (Sec. 3.4).
assumptions (4)
  • domain assumption The benchmark metrics (minJFDE, B-minJFDE, Consis-minJMR, collision rates) are valid and sufficient proxies for prediction quality and safety.
    The paper's central claim is defined by these metrics; no independent validation that they align with real-world planning safety is provided (Sec. 4.1).
  • domain assumption Progressive graph construction from predicted states is stable, and multiscale decoding corrects error accumulation.
    Sec. 3.3.2 motivates multiscale decoding by error accumulation, but no error analysis or empirical breakdown of the feedback loop is provided.
  • standard math The message-passing and attention components from prior literature are taken as working building blocks.
    Used to define the heterogeneous graph convolution operator in Eq. (5), citing Gilmer et al. (2017), Hamilton et al. (2017), and Gao et al. (2023).
  • domain assumption Scenario-centric coordinate normalization and DFS-based lane candidates preserve the information needed for prediction.
    Sec. 3.2 and Sec. 4.1 introduce these preprocessing choices; no analysis of their sensitivity or failure cases is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProgD: Progressive Multi-scale Decoding with Dynamic Graphs for Joint Multi-agent Motion Forecasting." pith.science (2026). https://pith.science/paper/54HVTO4K

@misc{pith2026250909210,
  author       = {Pith},
  title        = {Pith review of: ProgD: Progressive Multi-scale Decoding with Dynamic Graphs for Joint Multi-agent Motion Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/54HVTO4K}},
  note         = {Machine review of arXiv:2509.09210}
}
abstract

Accurate motion prediction of surrounding agents is crucial for the safe planning of autonomous vehicles. Recent advancements have extended prediction techniques from individual agents to joint predictions of multiple interacting agents, with various strategies to address complex interactions within future motions of agents. However, these methods overlook the evolving nature of these interactions. To address this limitation, we propose a novel progressive multi-scale decoding strategy, termed ProgD, with the help of dynamic heterogeneous graph-based scenario modeling. In particular, to explicitly and comprehensively capture the evolving social interactions in future scenarios, given their inherent uncertainty, we design a progressive modeling of scenarios with dynamic heterogeneous graphs. With the unfolding of such dynamic heterogeneous graphs, a factorized architecture is designed to process the spatio-temporal dependencies within future scenarios and progressively eliminate uncertainty in future motions of multiple agents. Furthermore, a multi-scale decoding procedure is incorporated to improve on the future scenario modeling and consistent prediction of agents' future motion. The proposed ProgD achieves state-of-the-art performance on the INTERACTION multi-agent prediction benchmark, ranking $1^{st}$, and the Argoverse 2 multi-world forecasting benchmark.

Figures

Figures reproduced from arXiv: 2509.09210 by the authors.

Figure 1
Figure 1. High-level illustration of joint motion prediction through progressive [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the encoder-decoder framework of ProgD. The encoder encodes the observed trajectories of agents and road networks in a scenario, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of joint prediction results on the testing set of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of joint prediction results on the validation set of Argoverse 2. The pose of each agent at the last observation time is denoted by a gray [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 5 linked inside Pith

  1. [1]

    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...

  2. [2]

    , author Liu, W

    author An, J. , author Liu, W. , author Liu, Q. , author Guo, L. , author Ren, P. , author Li, T. , year 2022 . title Dginet: Dynamic graph and interaction-aware convolutional network for vehicle trajectory prediction . journal Neural Networks volume 151 , pages 336--348

    2022

  3. [3]

    , author Boukerche, A

    author Azadani, M.N. , author Boukerche, A. , year 2025 . title Hierarchical transformers for motion forecasting based on inverse reinforcement learning . journal IEEE Trans. Veh. Technol. volume 74 , pages 3751--3764

    2025

  4. [4]

    , author Krizhevsky, A

    author Bansal, M. , author Krizhevsky, A. , author Ogale, A. , year 2019 . title Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst , in: booktitle Robotics: Science and Systems XV , address Freiburg im Breisgau, Germany

    2019

  5. [5]

    , author Gulino, C

    author Casas, S. , author Gulino, C. , author Suo, S. , author Luo, K. , author Liao, R. , author Urtasun, R. , year 2020 . title Implicit latent variable model for scene-consistent motion forecasting , in: booktitle Computer Vision--ECCV 2020: 16th European Conference , address Glasgow, UK . pp. pages 624--641

    2020

  6. [6]

    , author Luo, W

    author Casas, S. , author Luo, W. , author Urtasun, R. , year 2018 . title Intentnet: Learning to predict intention from raw sensor data , in: booktitle 2nd Annual Conference on Robot Learning , address Z \" u rich, Switzerland

    2018

  7. [7]

    , author Sapp, B

    author Chai, Y. , author Sapp, B. , author Bansal, M. , author Anguelov, D. , year 2019 . title Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction , in: booktitle 3rd Annual Conference on Robot Learning , address Osaka, Japan . pp. pages 86--99

    2019

  8. [8]

    , author Wang, J

    author Chen, H. , author Wang, J. , author Shao, K. , author Liu, F. , author Hao, J. , author Guan, C. , author Chen, G. , author Heng, P. , year 2023 . title Traj-MAE : Masked autoencoders for trajectory prediction , in: booktitle Proceedings of the IEEE/CVF International Conference on Computer Vision , address Paris, France

    2023

Show all 49 references
  1. [9]

    , author Mei, X

    author Cheng, J. , author Mei, X. , author Liu, M. , year 2023 . title Forecast-MAE : Self-supervised pre-training for motion forecasting with masked autoencoders , in: booktitle Proceedings of the IEEE/CVF International Conference on Computer Vision , address Paris, France . ...

    2023

  2. [10]

    , author Radosavljevic, V

    author Cui, H. , author Radosavljevic, V. , author Chou, F.C. , author Lin, T.H. , author Nguyen, T. , author Huang, T.K. , author Schneider, J. , author Djuric, N. , year 2019 . title Multimodal trajectory predictions for autonomous driving using deep convolutional networks ,...

    2019

  3. [11]

    , author Zhou, H

    author Feng, C. , author Zhou, H. , author Lin, H. , author Zhang, Z. , author Xu, Z. , author Zhang, C. , author Zhou, B. , author Shen, S. , year 2023 . title MacFormer : Map-agent coupled transformer for real-time and robust trajectory prediction . journal IEEE Robotics Aut...

    2023

  4. [12]

    , author Sun, C

    author Gao, J. , author Sun, C. , author Zhao, H. , author Shen, Y. , author Anguelov, D. , author Li, C. , author Schmid, C. , year 2020 . title VectorNet : Encoding hd maps and agent dynamics from vectorized representation , in: booktitle Proceedings of the IEEE/CVF Conferen...

    2020

  5. [13]

    , author Jia, X

    author Gao, X. , author Jia, X. , author Li, Y. , author Xiong, H. , year 2023 . title Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolutional recurrent networks . journal IEEE Robotics and Automation Letters volume 8 , pages 2946--2953

    2023

  6. [14]

    , author Sabatini, S

    author Gilles, T. , author Sabatini, S. , author Tsishkou, D. , author Stanciulescu, B. , author Moutarde, F. , year 2022 a. title GOHOME: graph-oriented heatmap output for future motion estimation , in: booktitle International Conference on Robotics and Automation , address P...

    2022

  7. [15]

    , author Sabatini, S

    author Gilles, T. , author Sabatini, S. , author Tsishkou, D. , author Stanciulescu, B. , author Moutarde, F. , year 2022 b. title THOMAS : Trajectory heatmap output with learned multi-agent sampling , in: booktitle The 10th International Conference on Learning Representations...

    2022

  8. [16]

    , author Schoenholz, S.S

    author Gilmer, J. , author Schoenholz, S.S. , author Riley, P.F. , author Vinyals, O. , author Dahl, G.E. , year 2017 . title Neural message passing for quantum chemistry , in: booktitle Proceedings of the 34th International Conference on Machine Learning , address Sydney, NSW...

    2017

  9. [17]

    , author Golemo, F

    author Girgis, R. , author Golemo, F. , author Codevilla, F. , author Weiss, M. , author D'Souza, J.A. , author Kahou, S.E. , author Heide, F. , author Pal, C. , year 2022 . title Latent variable sequential set transformers for joint multi-agent motion prediction , in: booktit...

    2022

  10. [18]

    , author Sun, C

    author Gu, J. , author Sun, C. , author Zhao, H. , year 2021 . title DenseTNT : End-to-end trajectory prediction from dense goal sets , in: booktitle Proceedings of the IEEE/CVF International Conference on Computer Vision , address Montreal, QC, Canada

    2021

  11. [19]

    , author Ying, Z

    author Hamilton, W. , author Ying, Z. , author Leskovec, J. , year 2017 . title Inductive representation learning on large graphs , in: booktitle Advances in Neural Information Processing Systems 30 , address Long Beach, CA, USA

    2017

  12. [20]

    , author Sapp, B

    author Hong, J. , author Sapp, B. , author Philbin, J. , year 2019 . title Rules of the road: Predicting driving behavior with a convolutional model of semantic interactions , in: booktitle Proceedings of the IEEE/CVF Conference on Ccomputer Vision and Pattern Recognition

    2019

  13. [21]

    , author Wu, P

    author Jia, X. , author Wu, P. , author Chen, L. , author Liu, Y. , author Li, H. , author Yan, J. , year 2023 . title HDGT: heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding . journal IEEE Transactions on Pattern Analysis and Mac...

    2023

  14. [22]

    , author Cornman, A

    author Jiang, C. . , author Cornman, A. , author Park, C. , author Sapp, B. , author Zhou, Y. , author Anguelov, D. , year 2023 . title MotionDiffuser : Controllable multi-agent motion prediction using diffusion , in: booktitle Proceedings of the IEEE/CVF Conference on Compute...

    2023

  15. [23]

    , author Wang, S

    author Kang, M. , author Wang, S. , author Zhou, S. , author Ye, K. , author Jiang, J. , author Zheng, N. , year 2024 . title FFINet : Future feedback interaction network for motion forecasting . journal IEEE Transactions on Intelligent Transportation Systems

    2024

  16. [24]

    , author Ba, J

    author Kingma, D.P. , author Ba, J. , year 2015 . title Adam: A method for stochastic optimization , in: booktitle The 3rd International Conference on Learning Representations , address San Diego, CA, USA

    2015

  17. [25]

    , author Zhang, L

    author Li, T. , author Zhang, L. , author Liu, S. , author Shen, S. , year 2024 . title Multi-modal integrated prediction and decision-making with adaptive interaction modality explorations . journal arXiv preprint arXiv:2408.13742

    2024 arXiv

  18. [26]

    , author Yang, B

    author Liang, M. , author Yang, B. , author Hu, R. , author Chen, Y. , author Liao, R. , author Feng, S. , author Urtasun, R. , year 2020 . title Learning lane graph representations for motion forecasting , in: booktitle Computer Vision--ECCV 2020: 16th European Conference, , ...

    2020

  19. [27]

    , author Park, C

    author Luo, W. , author Park, C. , author Cornman, A. , author Sapp, B. , author Anguelov, D. , year 2022 . title JFP : Joint future prediction with interactive multi-agent modeling for autonomous driving , in: booktitle Conference on Robot Learning , address Auckland, New Zea...

    2022

  20. [28]

    , author Huang, Z

    author Mo, X. , author Huang, Z. , author Xing, Y. , author Lv, C. , year 2022 . title Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network . journal IEEE Transactions on Intelligent Transportation Systems volume 23 , pages 9554--9567

    2022

  21. [29]

    , author Xing, Y

    author Mo, X. , author Xing, Y. , author Lv, C. , year 2020 . title ReCoG : A deep learning framework with heterogeneous graph for interaction-aware trajectory prediction . journal arXiv preprint arXiv:2012.05032

    2020 arXiv

  22. [30]

    , author Al-Rfou, R

    author Nayakanti, N. , author Al-Rfou, R. , author Zhou, A. , author Goel, K. , author Refaat, K.S. , author Sapp, B. , year 2023 . title Wayformer: Motion forecasting via simple & efficient attention networks , in: booktitle International Conference on Robotics and Automation...

    2023

  23. [31]

    , author Vasudevan, V

    author Ngiam, J. , author Vasudevan, V. , author Caine, B. , author Zhang, Z. , author Chiang, H.T.L. , author Ling, J. , author Roelofs, R. , author Bewley, A. , author Liu, C. , author Venugopal, A. , et al., year 2022 . title Scene transformer: A unified architecture for pr...

    2022

  24. [32]

    , et al., year 2017

    author Paszke, A. , et al., year 2017 . title Automatic differentiation in PyTorch , in: booktitle Advances in Neural Information Processing Systems, Autodiff Workshop , address Long Beach, CA, USA

    2017

  25. [33]

    , author Ethier, M

    author Rowe, L. , author Ethier, M. , author Dykhne, E.H. , author Czarnecki, K. , year 2023 . title FJMP : Factorized joint multi-agent motion prediction over learned directed acyclic interaction graphs , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision...

    2023

  26. [34]

    , author Cera, B

    author Seff, A. , author Cera, B. , author Chen, D. , author Ng, M. , author Zhou, A. , author Nayakanti, N. , author Refaat, K.S. , author Al-Rfou, R. , author Sapp, B. , year 2023 . title MotionLM : Multi-agent motion forecasting as language modeling , in: booktitle Proceedi...

    2023

  27. [35]

    , author Jiang, L

    author Shi, S. , author Jiang, L. , author Dai, D. , author Schiele, B. , year 2024 . title MTR++ : Multi-agent motion prediction with symmetric scene modeling and guided intention querying . journal IEEE Transactions on Pattern Analysis and Machine Intelligence volume 46 , pa...

    2024

  28. [36]

    , author Huang, X

    author Sun, Q. , author Huang, X. , author Gu, J. , author Williams, B.C. , author Zhao, H. , year 2022 . title M2I : From factored marginal trajectory prediction to interactive prediction , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2022

  29. [37]

    , author Regalado, S

    author Suo, S. , author Regalado, S. , author Casas, S. , author Urtasun, R. , year 2021 . title TrafficSim : Learning to simulate realistic multi-agent behaviors , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , address virtua...

    2021

  30. [38]

    , author Sun, W

    author Tang, X. , author Sun, W. , author Hu, S. , author Sun, Y. , author Guo, Y. , year 2024 . title MS-Net : A multi-path sparse model for motion prediction in multi-scenes . journal IEEE Robotics and Automation Letters volume 9 , pages 891--898

    2024

  31. [39]

    , author Hefny, A

    author Varadarajan, B. , author Hefny, A. , author Srivastava, A. , author Refaat, K.S. , author Nayakanti, N. , author Cornman, A. , author Chen, K. , author Douillard, B. , author Lam, C. , author Anguelov, D. , author Sapp, B. , year 2022 . title Multipath++: Efficient info...

    2022

  32. [40]

    , author Qi, W

    author Wilson, B. , author Qi, W. , author Agarwal, T. , author Lambert, J. , author Singh, J. , author Khandelwal, S. , author Pan, B. , author Kumar, R. , author Hartnett, A. , author Pontes, J.K. , et al., year 2023 . title Argoverse 2: Next generation datasets for self-dri...

    2023 arXiv

  33. [41]

    , author Chu, D

    author Xin, G. , author Chu, D. , author Lu, L. , author Deng, Z. , author Lu, Y. , author Wu, X. , year 2025 . title Multi-agent trajectory prediction with difficulty-guided feature enhancement network . journal IEEE Robotics and Automation Letters volume 10

    2025

  34. [42]

    , author Cao, T

    author Ye, M. , author Cao, T. , author Chen, Q. , year 2021 . title TPCN : Temporal point cloud networks for motion forecasting , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2021

  35. [43]

    , author Xu, J

    author Ye, M. , author Xu, J. , author Xu, X. , author Cao, T. , author Chen, Q. , year 2022 . title DCMS : Motion forecasting with dual consistency and multi-pseudo-target supervision . journal arXiv preprint arXiv:2204.05859

    2022 arXiv

  36. [44]

    , author Liang, M

    author Zeng, W. , author Liang, M. , author Liao, R. , author Urtasun, R. , year 2021 . title LaneRCNN : Distributed representations for graph-centric motion forecasting , in: booktitle Proc. Int. Conf. Intell. Robots Syst , address Prague, Czech Republic

    2021

  37. [45]

    , author Sun, L

    author Zhan, W. , author Sun, L. , author Wang, D. , author Shi, H. , author Clausse, A. , author Naumann, M. , author Kummerle, J. , author Konigshof, H. , author Stiller, C. , author de La Fortelle, A. , et al., year 2019 . title Interaction dataset: An international, advers...

    2019 arXiv

  38. [46]

    , author Fu, M

    author Zhang, T. , author Fu, M. , author Yang, Y. , author Song, W. , author Liu, T. , year 2024 . title Edge-enriched graph transformer for multi-agent trajectory prediction with relative positional semantics . journal IEEE Transactions on Instrumentation and Measurement volume 73

    2024

  39. [47]

    , author Liniger, A

    author Zhang, Z. , author Liniger, A. , author Sakaridis, C. , author Yu, F. , author Gool, L.V. , year 2023 . title Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding . journal Advances in Neural Information Processing Systems 36

    2023

  40. [48]

    , author Wang, J

    author Zhou, Z. , author Wang, J. , author Li, Y.H. , author Huang, Y.K. , year 2023 . title Query-centric trajectory prediction , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 17863--17873

    2023

  41. [49]

    , author Ye, L

    author Zhou, Z. , author Ye, L. , author Wang, J. , author Wu, K. , author Lu, K. , year 2022 . title HiVT : Hierarchical vector transformer for multi-agent motion prediction , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ad...

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.