Pith. sign in

REVIEW 3 major objections 5 minor 43 references

HAMF: A Hybrid Attention-Mamba Framework for Joint Scene Context Understanding and Future Motion Representation Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HAMF claims that joint scene-and-future encoding achieves state-of-the-art Argoverse 2 motion forecasting with a 3.0M-parameter model running at 22 ms.

desk verdict A lightweight motion forecasting model that puts learnable future tokens in every encoder layer; the core idea is sound and well ablated, but Table II's minFDE6 is off by a factor of 10 and the SOTA claim is stronger than the evidence. read the letter →

arxiv 2505.15703 v1 pith:VDCCKNFK submitted 2025-05-21 cs.CV cs.AI

Xiaodong Mei , Sheng Wang , Jie Cheng , Yingbing Chen , Dan Xu This is my paper
classification cs.CVcs.AI
keywords motionforecastingArgoverse2learnabletokensself-attentioncross-attentionMambastatespacemodelsautonomousdriving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes HAMF, a motion-forecasting model that claims to reach state-of-the-art accuracy on Argoverse 2 while being far smaller and faster than comparable models. Instead of decoding trajectories only after a scene encoder finishes, HAMF inserts learnable future-motion tokens into the encoder alongside agent and map tokens, so future-motion representations are refined jointly with scene understanding at every layer. A Mamba sequence-modeling block then processes the learned tokens to keep the six predicted maneuvers distinct and coherent. On the Argoverse 2 test set the single 3.0M-parameter model reports minADE6 0.64, minFDE6 1.23, MR6 0.14, with 22 ms inference latency. The intended significance is that joint future-motion and scene encoding can replace heavy decoders and still be accurate enough for real-time driving.

What carries the argument

The load-bearing object is the set of $K_e$ learnable future-motion tokens $F \in \mathbb{R}^{K_e \times C}$, concatenated with scene tokens $S$ and processed by a unified encoder: self-attention over the combined sequence, cross-attention with $F$ as queries and scene tokens as keys and values, and a per-layer sum $F^l = F^l_{sa} + F^l_{ca}$. This is paired with a one-block unidirectional Mamba decoder that models dependencies among tokens to enforce diversity; Mamba is a selective state-space sequence model with linear-time scaling. The per-layer interaction is what lets future-motion representations absorb multi-scale scene context, and the unidirectional decoder is claimed to preserve token consistency better than bidirectional or attention-based decoders.

What would settle it

Train HAMF and its M2/M3 baselines with identical data, multiple seeds, and the same training budget; if the validation minFDE6 and MR6 differences between HAMF and M3 (1.229 versus 1.260, and 0.145 versus 0.153) fall within one standard deviation across seeds, the claimed advantage of the per-layer interaction is not established. A second check is to rerun the test-set comparison with identical input preprocessing and decoding budgets against SmartRefine and QCNet; if minADE6 0.64 no longer leads, the state-of-the-art claim fails.

Watch

Extended reading notes

Core claim

HAMF's central claim is that future motion features should not be produced only at the end of encoding; they should be present from the start as learnable tokens that co-evolve with scene context. Each encoder layer runs multi-head self-attention over the concatenation of scene tokens and future-motion tokens, then multi-head cross-attention that uses the previous layer's future-motion tokens as queries against the current layer's scene tokens; the two updated token sets are added each layer. The decoder then passes the final tokens through one unidirectional Mamba block before MLPs output six trajectories and probabilities. The paper reports that this design beats or matches prior single-model methods on Argoverse 2 (minADE6 0.64, minFDE6 1.23, MR6 0.14 on test; 0.633/1.229/0.145 on validation) with 3.0M parameters and 22 ms latency, and that ablations attribute the gain to cross-attention interaction, the per-layer sum, and the unidirectional Mamba decoder.

Load-bearing premise

The central comparison rests on the assumption that the small validation gaps over the paper's own baselines (MR6 0.145 versus 0.150 and 0.153) are real improvements and not run-to-run luck, since only a single training run is reported and no code or weights are provided.

Editorial extensions

If this is right

  • If the claim holds, accurate six-mode motion forecasting no longer requires large decoders: a 3.0M-parameter model runs in real time (22 ms on a single RTX 3090), making deployment in vehicles and edge devices feasible.
  • Joint future-token and scene encoding becomes a viable design for structured prediction tasks where output modalities are known ahead of time, such as multi-agent prediction and goal estimation.
  • The per-layer additive fusion of self- and cross-attention suggests that simple interactions can outperform more complex query-only decoders, shifting design attention from model size to where prediction tokens are placed.
  • Unidirectional sequence modeling in the decoder is a useful alternative to bidirectional attention for maintaining mode diversity in multi-modal trajectory generation.
  • The reported scaling pattern—more encoder depth or more Mamba blocks does not help monotonically—implies that the architecture operates near a practical efficiency-accuracy frontier at 3.0M parameters.

Reading between the lines

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

  • A direct extension the paper leaves untested is pretraining the future-motion tokens on large trajectory corpora; if token initialization matters as much as architecture, HAMF's gains might be amplified or even achievable with a frozen encoder.
  • Because the ablations show that more encoder depth, more Mamba blocks, and bidirectional modeling all hurt or plateau, the per-layer additive fusion is likely the active ingredient; a learned gating or attention-based fusion between self- and cross-attention outputs could improve it further, addressing the limitation the authors note.
  • If the learned future-motion tokens are interpretable (for example, one token per maneuver type), they could double as a lightweight controller interface or explainability tool for planning modules, a role the paper does not discuss.
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 / 5 minor

Summary. The paper proposes HAMF, a motion forecasting framework for Argoverse 2 that introduces a set of learnable 'future motion tokens' into the scene-encoding transformer, combining self-attention and cross-attention in each encoder layer and using a Mamba-based decoder for multi-modal trajectory prediction. The authors report competitive state-of-the-art results on the test set (minADE6 0.64, minFDE6 1.23, MR6 0.14 with a single model) and a lightweight design (3.0M parameters, 22 ms latency). The central claim is that jointly encoding scene context and future motion representations, rather than decoding from a final scene feature, improves prediction accuracy and diversity.

Significance. If the reported results are reliable, the paper offers a simple and parameter-efficient architecture that may be practically attractive, and the idea of pushing future-motion query tokens into the encoding stage is a reasonable extension of prior query-based approaches. The paper provides a thorough ablation over encoder and decoder components. However, the reported numbers are not currently presented in a verifiable form: there is an order-of-magnitude inconsistency in the efficiency table, no error bars or multiple-seed statistics, and no code or model weights. The strengths are the simplicity of the design and the breadth of ablations, but the empirical support needs to be made internally consistent and statistically grounded.

major comments (3)
  1. [Section IV-B, Table II] The minFDE6 values reported for QCNet, DeMo, and HAMF (0.129, 0.125, 0.123) are a factor of ten smaller than the same metric for HAMF in Table III (1.229) and Table IV (1.229), and also inconsistent with the published values for QCNet (1.29) and DeMo (1.25) shown in Table I. No unit conversion or protocol difference is stated. Since Table II is the direct evidence for the 'lightweight, real-time' claim, the numerical support is internally inconsistent. The table must be corrected or the discrepancy must be explained before this claim can be assessed.
  2. [Section IV-C, Tables IV-VII] All validation and ablation results are reported as single-run numbers without standard deviations or multiple seeds. The claimed improvements are often small (e.g., MR6 0.145 vs 0.150 vs 0.153 in Table IV; minADE6 0.633 vs 0.636 in Table V) and could easily be explained by run-to-run variation. The paper should report mean and standard deviation over at least three independent training runs, or otherwise provide a statistical test. Without this, neither the component ablations nor the 'state-of-the-art' claim relative to previously published results can be considered supported.
  3. [Section IV-B, Table I] The abstract and conclusion claim 'state-of-the-art' performance, but Table I shows the margin over existing methods is small (e.g., minADE6 0.64 vs SmartRefine 0.63; minFDE6 tied at 1.23). These test-set comparisons are not controlled for training conditions, and the reported differences are within the typical run-to-run noise for this benchmark. Please either temper the claim to 'competitive' or provide a controlled comparison on the validation set with the same preprocessing and training protocol for all methods, together with variance estimates.
minor comments (5)
  1. [Section IV-B, Qualitative results] The word 'visulation' should be 'visualization'; please also ensure the qualitative figure captions match the subfigure references in the text.
  2. [References] Reference [35] is a duplicate of [23] (the Mamba paper); please merge them or cite the original source once.
  3. [Section III-C, Eqs. (6)-(8)] The cross-attention in Eq. (7) uses F^{l-1} as the query, while the text describes the blocks as 'sequential'; the figure and equations should be clarified to show whether the cross-attention is intended to use the updated future tokens F^l_sa or the previous-layer tokens F^{l-1}.
  4. [Section IV-B, Table I] The sentence 'our approach achieves the best results in terms of minADE1, minFDE6 and MR6' is not strictly accurate because minADE1 is tied with RealMotion (1.59) and minFDE6 is tied with SmartRefine (1.23); please rephrase.
  5. [Section IV-B, Table II] The header 'minFDE 6' should be formatted consistently with the rest of the paper (e.g., minFDE6), and the latency values should report variance across multiple runs rather than a single measurement.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; self-citations are minor and disclosed, and the SOTA claim rests on external Argoverse 2 evaluation.

full rationale

The paper's central claim is empirical, not derived from first principles, and it is grounded in held-out Argoverse 2 test and validation sets. The 'future motion tokens' are learnable parameters, not re-encoded ground-truth trajectories; the paper explicitly notes that the ground truth future trajectory is 'not used in the input' (Fig. 2 caption). Training uses a winner-take-all regression and classification loss, and the reported results are benchmark numbers, not fitted constants renamed as predictions. Self-citations are present but non-load-bearing: the preprocessing follows Forecast-MAE [7], and the token-in-encoder idea is credited to TaskPrompter [21], both from overlapping author groups, but neither citation is used to justify correctness or to exclude alternatives. The architecture's contribution is supported by internal ablations and external comparisons. The internal inconsistency in Table II (minFDE6 0.123 vs 1.229 in Table III) is a data-verifiability or reporting issue, not a circularity: it does not make the evaluation reduce to its inputs, though it should be corrected or clarified with released code/weights.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The paper is an empirical architecture study, so it does not derive a physical law. The ledger lists the hand-chosen design constants (token count, depth, width, decoder depth) and the domain assumptions (benchmark validity, preprocessing inherited from previous work, WTA loss) that the reported performance rests on. The only invented construct is the set of learnable future motion tokens, whose usefulness is evidenced only by the paper's own ablations.

free parameters (4)
  • Number of future motion tokens Ke = 6
    Chosen to match the K=6 output modes; ablation in Table VII shows 6 works best among tested values.
  • Encoder depth = 5
    Ablation in Table VI; depth 5 chosen as a trade-off between 0.633 minADE6 and efficiency, though depth 6 gives 0.629/1.226 at more parameters.
  • Latent feature dimension = 128
    Set by hand as part of the lightweight design; not ablated.
  • Uni-Mamba decoder depth = 1
    Ablation in Table V; one block outperforms two blocks and Bi-Mamba/Attention variants.
assumptions (5)
  • standard math Multi-head self-attention and cross-attention, as in Transformer, provide effective token mixing for scene and motion tokens.
    Invoked in Section III-C via MHA blocks from [22].
  • domain assumption Mamba's selective state space model is an effective sequence model for motion modality tokens.
    Assumed in Section III-D; the paper provides ablations but no theoretical justification.
  • domain assumption The Argoverse 2 official metrics and test set are a valid measure of motion forecasting quality.
    Used in Section IV-B for the SOTA claim.
  • domain assumption The data preprocessing and normalization of Forecast-MAE [7] is suitable and is not the source of reported gains.
    Section III-A states the same preprocessing is used; the paper does not ablate this choice.
  • domain assumption Winner-take-all loss with trajectory regression and classification loss is sufficient for multi-modal training.
    Section III-E; standard in the field, not analyzed here.
invented entities (1)
  • Learnable future motion tokens F
    purpose: Represent candidate multi-modal future trajectories inside the encoder so that scene and future representations are refined jointly.
    These are random-initialized trainable vectors; their value is supported only by the paper's internal ablations, not by any external measurement or formal guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HAMF: A Hybrid Attention-Mamba Framework for Joint Scene Context Understanding and Future Motion Representation Learning." pith.science (2026). https://pith.science/paper/VDCCKNFK

@misc{pith2026250515703,
  author       = {Pith},
  title        = {Pith review of: HAMF: A Hybrid Attention-Mamba Framework for Joint Scene Context Understanding and Future Motion Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VDCCKNFK}},
  note         = {Machine review of arXiv:2505.15703}
}
read the original abstract

Motion forecasting represents a critical challenge in autonomous driving systems, requiring accurate prediction of surrounding agents' future trajectories. While existing approaches predict future motion states with the extracted scene context feature from historical agent trajectories and road layouts, they suffer from the information degradation during the scene feature encoding. To address the limitation, we propose HAMF, a novel motion forecasting framework that learns future motion representations with the scene context encoding jointly, to coherently combine the scene understanding and future motion state prediction. We first embed the observed agent states and map information into 1D token sequences, together with the target multi-modal future motion features as a set of learnable tokens. Then we design a unified Attention-based encoder, which synergistically combines self-attention and cross-attention mechanisms to model the scene context information and aggregate future motion features jointly. Complementing the encoder, we implement the Mamba module in the decoding stage to further preserve the consistency and correlations among the learned future motion representations, to generate the accurate and diverse final trajectories. Extensive experiments on Argoverse 2 benchmark demonstrate that our hybrid Attention-Mamba model achieves state-of-the-art motion forecasting performance with the simple and lightweight architecture.

Figures

Figures reproduced from arXiv: 2505.15703 by the authors.

Figure 1
Figure 1. Comparison of existing motion forecasting frameworks (a, b) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed HAMF. The left part presents the input embedding module with an intersection driving scenario. The ground truth [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of our proposed method, the query-based method [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 31 canonical work pages

  1. [1]

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

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J. K. Pontes, et al., “Argoverse 2: Next generation datasets for self-driving perception and forecasting,” arXiv preprint arXiv:2301.00493, 2023. 1, 2, 4

    arXiv 2023

  2. [2]

    A survey on trajectory-prediction methods for autonomous driving,

    Y . Huang, J. Du, Z. Yang, Z. Zhou, L. Zhang, and H. Chen, “A survey on trajectory-prediction methods for autonomous driving,” IEEE Transactions on Intelligent Vehicles , 2022. 1

    work page 2022

  3. [3]

    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 Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2020, pp. 11 525–11 533. 1, 2

    work page 2020

  4. [4]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y . Chen, R. Liao, S. Feng, and R. Urta- sun, “Learning lane graph representations for motion forecasting,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 , 2020. 1, 2

    work page 2020

  5. [5]

    Simpl: A simple and efficient multi-agent motion prediction baseline for autonomous driving,

    L. Zhang, P. Li, S. Liu, and S. Shen, “Simpl: A simple and efficient multi-agent motion prediction baseline for autonomous driving,” IEEE Robotics and Automation Letters , 2024. 1, 2, 5

    work page 2024

  6. [6]

    Scene transformer: A unified architecture for predicting multiple agent tra- jectories,

    J. Ngiam, B. Caine, V . Vasudevan, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal, et al., “Scene transformer: A unified architecture for predicting multiple agent tra- jectories,” arXiv preprint arXiv:2106.08417 , 2021. 1, 2

    arXiv 2021

  7. [7]

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

    J. Cheng, X. Mei, and M. Liu, “Forecast-mae: Self-supervised pre- training for motion forecasting with masked autoencoders,” in Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8679–8689. 1, 2, 3, 5

    work page 2023

  8. [8]

    Gorela: Go relative for viewpoint-invariant motion forecasting,

    A. Cui, S. Casas, K. Wong, S. Suo, and R. Urtasun, “Gorela: Go relative for viewpoint-invariant motion forecasting,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 7801–7807. 1, 5

    work page 2023

Show all 43 references
  1. [9]

    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, et al., “Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,” in 2022 International Conference on Robotics ...

    2022

  2. [10]

    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,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 6531– 6543, 2022. 1, 2, 3, 5

    2022

  3. [11]

    Query-centric trajectory prediction,

    Z. Zhou, J. Wang, Y .-H. Li, and Y .-K. Huang, “Query-centric trajectory prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 17 863–17 873. 1, 2, 5

    2023

  4. [12]

    Decoupling motion forecast- ing into directional intentions and dynamic states,

    B. Zhang, N. Song, and L. Zhang, “Decoupling motion forecast- ing into directional intentions and dynamic states,” arXiv preprint arXiv:2410.05982, 2024. 1, 2, 3, 4, 5

    2024 arXiv

  5. [13]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2023, pp. 21 995–22 003. 1, 5

    2023

  6. [14]

    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. 1, 2

    1910 arXiv

  7. [15]

    Multimodal trajectory prediction conditioned on lane-graph traversals,

    N. Deo, E. Wolff, and O. Beijbom, “Multimodal trajectory prediction conditioned on lane-graph traversals,” in Conference on Robot Learn- ing. PMLR, 2022, pp. 203–212. 1

    2022

  8. [16]

    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, et al. , “Tnt: Target-driven trajectory prediction,” in Conference on Robot Learning . PMLR, 2021. 1, 2

    2021

  9. [17]

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

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory pre- diction from dense goal sets,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021. 1, 2

    2021

  10. [18]

    Learning to predict vehicle trajectories with model-based planning,

    H. Song, D. Luan, W. Ding, M. Y . Wang, and Q. Chen, “Learning to predict vehicle trajectories with model-based planning,” in 5th Annual Conference on Robot Learning , 2021. 1, 2

    2021

  11. [19]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018. 2

    2018 arXiv

  12. [20]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    D. Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv: 2010.11929 , 2020. 2

    2010 arXiv

  13. [21]

    Taskprompter: Spatial-channel multi-task prompt- ing for dense scene understanding,

    H. Ye and D. Xu, “Taskprompter: Spatial-channel multi-task prompt- ing for dense scene understanding,” in ICLR, 2023. 2, 7

    2023

  14. [22]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , 2017. 2

    2017

  15. [24]

    Multiple futures prediction,

    C. Tang and R. R. Salakhutdinov, “Multiple futures prediction,” Advances in neural information processing systems , 2019. 2

    2019

  16. [25]

    Hgcn-gjs: Hierar- chical graph convolutional network with groupwise joint sampling for trajectory prediction,

    Y . Chen, C. Liu, X. Mei, B. E. Shi, and M. Liu, “Hgcn-gjs: Hierar- chical graph convolutional network with groupwise joint sampling for trajectory prediction,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022. 2

    2022

  17. [26]

    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 transactions on pattern analysis and machine intelligence, 2023. 2, 5

    2023

  18. [27]

    Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding,

    Z. Zhang, A. Liniger, C. Sakaridis, F. Yu, and L. V . Gool, “Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding,” Advances in Neural Information Processing Systems , vol. 36, 2024. 2

    2024

  19. [28]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024. 2, 5

    2024

  20. [29]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision . Springer, 2020. 2

    2020

  21. [30]

    Rethinking imitation-based planners for autonomous driving,

    J. Cheng, Y . Chen, X. Mei, B. Yang, B. Li, and M. Liu, “Rethinking imitation-based planners for autonomous driving,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 14 123–14 130. 2

    2024

  22. [31]

    Pluto: Pushing the limit of imita- tion learning-based planning for autonomous driving,

    J. Cheng, Y . Chen, and Q. Chen, “Pluto: Pushing the limit of imita- tion learning-based planning for autonomous driving,” arXiv preprint arXiv:2404.14327, 2024. 2

    2024 arXiv

  23. [32]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” Advances in neural information processing systems , 2025. 3

    2025

  24. [33]

    Videomamba: State space model for efficient video understanding,

    K. Li, X. Li, Y . Wang, Y . He, Y . Wang, L. Wang, and Y . Qiao, “Videomamba: State space model for efficient video understanding,” in European Conference on Computer Vision . Springer, 2024. 3, 6

    2024

  25. [34]

    Cobra: Extending mamba to multi-modal large language model for efficient inference,

    H. Zhao, M. Zhang, W. Zhao, P. Ding, S. Huang, and D. Wang, “Cobra: Extending mamba to multi-modal large language model for efficient inference,” arXiv preprint arXiv:2403.14520 , 2024. 3

    2024 arXiv

  26. [35]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023. 3

    2023 arXiv

  27. [36]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016. 3

    2016 arXiv

  28. [37]

    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 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 1609–1615. 5

    2023

  29. [38]

    Motion forecasting in continuous driving,

    N. Song, B. Zhang, X. Zhu, and L. Zhang, “Motion forecasting in continuous driving,” in NeurIPS, 2024. 5

    2024

  30. [39]

    Qml for argoverse 2 motion forecasting challenge,

    T. Su, X. Wang, and X. Yang, “Qml for argoverse 2 motion forecasting challenge,” arXiv preprint arXiv:2207.06553 , 2022. 5

    2022 arXiv

  31. [40]

    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. Shen, “Macformer: Map-agent coupled transformer for real- time and robust trajectory prediction,” IEEE Robotics and Automation Letters, vol. 8, no. 10, pp. 6795–6802, 2023. 5

    2023

  32. [41]

    Banet: Motion forecasting with boundary aware network,

    C. Zhang, H. Sun, C. Chen, and Y . Guo, “Banet: Motion forecasting with boundary aware network,” arXiv preprint arXiv:2206.07934 ,

    arXiv

  33. [42]

    Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolu- tional recurrent networks,

    X. Gao, X. Jia, Y . Li, and H. Xiong, “Dynamic scenario representation learning for motion forecasting with heterogeneous graph convolu- tional recurrent networks,” IEEE Robotics and Automation Letters , vol. 8, no. 5, pp. 2946–2953, 2023. 5

    2023

  34. [43]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017. 4

    2017 arXiv

  35. [44]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” in Forty-first International Conference on Machine Learning. 6

Pith tools

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