Pith. sign in

REVIEW 3 major objections 5 minor 29 references

On Learning Closed-Loop Probabilistic Multi-Agent Simulator

T0 review · 3 major / 5 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 NIVA is a hierarchical Bayesian multi-agent traffic simulator that learns separate latents for driving style and intention and matches the top reported realism scores on the Waymo Open Motion Dataset.

desk verdict Reasonable incremental simulator, but the controllability claim is not backed by the training procedure or the evaluation. read the letter →

arxiv 2508.00384 v1 pith:VGFFXQFD submitted 2025-08-01 cs.RO

Juanwu Lu , Rohit Gupta , Ahmadreza Moradipari , Kyungtae Han , Ruqi Zhang , Ziran Wang This is my paper
classification cs.RO
keywords multi-agenttrafficsimulationclosed-loophierarchicalBayesianmodelvariationalinferencetrajectorypredictiondrivingstyleintentionmodelingautonomous
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

Closed-loop traffic simulators that condition each agent's next action on the evolving scene are becoming the standard for testing autonomous vehicles, but they tend to be black boxes. NIVA is a probabilistic simulator that explicitly factors each agent's behavior into a continuous driving-style variable and a discrete intention, then autoregressively samples future motion through a two-stage generative process: an open-loop prior conditioned only on history and intention, followed by closed-loop resampling conditioned on map features and other agents. The paper argues that this hierarchical Bayesian formulation unifies earlier sequence-to-sequence trajectory predictors with newer next-token-prediction simulators, and that the learned latents remain semantically meaningful. On the Waymo Open Motion Dataset, NIVA matches the realism metrics of leading baselines while reporting the lowest minimum average displacement error and the best kinematic realism, with a one-million-parameter model.

What carries the argument

The machinery is a three-level hierarchical generative model. Agent count is drawn from a Poisson prior; each agent's style $b_n$ is drawn from a standard Gaussian; each agent's intention $z_n$ is drawn from a categorical prior. Conditioned on history $O_{\le T_h}$ and intention, temporal self-attention blocks parameterize an open-loop Gaussian prior over a high-dimensional latent state $s_t$. Then adaptive-norm Transformer blocks—Transformer blocks whose layer-normalization scale and shift are MLP outputs of the style variable, conditioning attention on behavior—resample $s_t$ by cross-attending to map features and other agents, and a linear Gaussian emission model $p(o_t|s_t)$ yields closed-form marginal and posterior distributions for observations and latents. Training alternates between a variational E-step that updates the posterior of $b$ and $z$ (with hard intention assignment) and an M-step that maximizes the ELBO with respect to generative parameters.

What would settle it

Generate many rollouts from the same history with the same style code but two different intention codes, and measure the distance between the empirical distributions of their endpoints; if the two destination distributions are statistically indistinguishable, or a classifier cannot predict which intention code was used above chance, then the intention latent is inert and the disentanglement claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that a hierarchical Bayesian generative process with two latent levels—a continuous style $b_n$ drawn from a standard Gaussian and a discrete intention $z_n$ drawn from a categorical prior—can drive a decoder-only Transformer whose adaptive layer-normalization parameters are modulated by those latents, producing closed-loop multi-agent trajectories that are both realistic and coachable. The generative process first samples a latent trajectory $\hat{s}_{T_h+t}$ from an open-loop prior conditioned only on history and intention (Eq.~3), then resamples the observation $o_{T_h+t}$ from a closed-loop distribution conditioned on map features, other agents' latent states, and the style variable (Eq.~4). A linear Gaussian emission model makes the marginal predictive distribution over $o_t$ and the posterior over the latent state $s_t$ available in closed form, which is what lets the model sample efficiently (Algorithm~2) and condition on ground-truth observations when they are available. Training maximizes an evidence lower bound with a mean-field variational family, and the E-step assigns each agent the single intention with highest posterior probability, i.e. $k^*=\arg\max_k q(z_k)$. The paper reads this as a Bayesian unification of open-loop sequence-to-sequence prediction and closed-loop next-token-prediction simulation, and its experimental section supports the claim that the disentangled latents control destination and aggressiveness in the generated scenarios.

Load-bearing premise

NIVA's controllability relies on the assumption that fixing each agent to a single intention via the argmax in training preserves the multimodality of real driver intentions; if that hard assignment collapses the latent space, the style and intention controls will not behave as advertised.

Editorial extensions

If this is right

  • Autonomous-vehicle developers could use NIVA to generate large batches of scenarios with chosen intentions and driving styles, enabling targeted testing of rare or adversarial situations such as aggressive cut-ins.
  • The closed-form marginal and posterior for the emission model mean the same trained parameters support both open-loop prediction and closed-loop simulation, so a single model can serve trajectory forecasting and scenario generation.
  • The one-million-parameter size of the model (versus tens of millions for some baselines) suggests the disentangled hierarchical prior is an efficient inductive bias, and the model degrades gracefully when trained on as little as 1 percent of the data.
  • Because the ELBO objective reduces to minimizing the data log-likelihood plus KL penalties on style and intention posteriors, the framework supplies a principled objective for future simulators that want to condition on interpretable factors.
  • If the latent factors are truly disentangled, the model should allow interpolation between aggressive and conservative driving styles, giving a continuous dial for scenario difficulty.

Reading between the lines

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

  • A testable extension of the paper's framework would replace the hard intention assignment $k^*=\arg\max_k q(z_k)$ with a soft mixture-weighted objective or a Gumbel-softmax relaxation; if the hard assignment is collapsing multimodal intentions, the soft version should improve diversity metrics such as the realism score without hurting minADE.
  • The Poisson process for agent count is stated but then fixed to the observed $N$ in experiments, so the framework does not yet simulate variable traffic density; a natural extension is to learn the arrival rate from data and couple it to the style prior, making scenario generation open-ended.
  • The unification claim could be made quantitative by deriving the open-loop objective of a sequence-to-sequence model as a special case of the NIVA ELBO when the closed-loop resampling step is ignored; the paper gestures at this but does not show it formally.
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 NIVA, a hierarchical Bayesian model for closed-loop multi-agent traffic simulation. Each agent is associated with a continuous "behavior style" latent b_n and a discrete "intention" latent z_n, and the observed states are generated through latent motion states s_t with a linear-Gaussian emission model. Training uses a mean-field variational approximation and an EM-style iterative algorithm, while sampling is autoregressive through a decoder-only Transformer with adaptive LayerNorm modulated by the latents. On the Waymo Open Motion Dataset, NIVA reports competitive simulation metrics (Table II), including the best minADE among the listed methods, and Fig. 3 qualitatively shows trajectories that change with different intentions and style values. The paper's central claims are that NIVA unifies open-loop and closed-loop simulation paradigms and provides interpretable, controllable generation of traffic behaviors.

Significance. If the latent variables genuinely and independently steer generated behavior, NIVA would be a valuable contribution: it is parameter-efficient (1.0M parameters), competitive on a standard benchmark, and it connects probabilistic open-loop prediction with closed-loop simulation under one generative model. The paper is also transparent in presenting its probabilistic assumptions and closed-form Gaussian computations. However, the contributions as stated are only partially supported by the current evidence: the control claim rests on a training procedure that may not train all intention modes, and the paper provides no quantitative controllability evaluation. The benchmark results are plausible but their significance is mostly confirmatory with respect to existing simulators.

major comments (3)
  1. [Section III-E, Algorithm 1, Eq. (12)-(13), Algorithm 2] The training objective is inconsistent with the stated ELBO and with the sampling procedure. After computing q(z_n^k) via Eq. (13), Algorithm 1 selects k* = argmax_k q(z_n^k) and computes the loss only for that single intention. This means the generator is trained primarily on the most probable intention per agent rather than on the full categorical distribution q(z_n) used in Eq. (12). At sampling time, Algorithm 2 draws z_n ~ Cat(pi), so it can query intention tokens that may have been trained only rarely or never for a given history. The paper provides no quantitative evidence that all K intentions produce realistic behaviors: Fig. 3 is qualitative, and there are no per-intention minADE, intention classification accuracy, controllability metrics, or trajectory diversity metrics. The abstract's claim of "embellishing control over intentions and driving styles" is therefore not currently supported. Please either train with the full ELBO expectation over q(z_n), use a soft/random assignment during training, or provide explicit validation that all codebook entries are meaningfully learned and controllable.
  2. [Equations (9)-(10) and Algorithm 2] The posterior precision in Eq. (10) is mis-specified: P = (σ²_s_t^{-1} + A^T(ε I)A)^{-1} should be (σ²_s_t^{-1} + A^T ε^{-2} A)^{-1} to be consistent with the marginal covariance A(σ²_s_t I)A^T + ε²I in Eq. (9) and with the posterior mean formula A^T(ε^{-2} I)o_t. As printed, the covariance term has inconsistent dimensions and would lead to incorrect posterior updates. The same issue appears in Algorithm 2, where P^{-1} is written as Σ^{-1}_{s_t} + A^T A, omitting the ε^{-2} factor. Please correct the derivation and the algorithm, or state explicitly that ε is set to 1 in all experiments.
  3. [Section III-A vs. Algorithm 2] The generative process described in Section III-A does not match the autoregressive sampling in Algorithm 2. Equations (3)-(4) define an open-loop prior p(s_t | O≤Th, z_n) and a per-step emission p(o_t | S_t, M, b_n) with no dependence on previously sampled observations o_{<t}. Algorithm 2 instead conditions NIVA on O_{<Th+t} at each time step, and the temporal self-attention in Eq. (7) uses keys o_j for j=1,...,t-1, which for t > Th include simulated observations. This is a genuine closed-loop autoregressive model, but the formal generative model and the variational objective in Eq. (12) are written for a different, non-autoregressive process. Please make the formal generative process and the training objective consistent with the implemented sampler, or clearly explain how Eq. (3)-(4) are intended to represent the autoregressive model actually used.
minor comments (5)
  1. [Section IV-E(b), Table V] The text says "In V, we compare performance..." but should refer to Table V; also the dash for Nblocks=3 should be explained in the caption or text as a training failure due to GPU memory.
  2. [Section III-C, Eq. (8)] The equations for the adaptive-norm block contain a likely typo: "Q^n_{t,3} ← δ1(b_n)⊙" is incomplete, and the subscript in the second cross-attention key/value "m_j + R_n j" is missing a comma. Please clarify the residual structure.
  3. [Algorithm 2] The noise vector ν_n is sampled once per agent before the t loop and then reused for every time step. This is an unusual reparameterization that couples all timesteps through the same noise; the authors should state whether this is intentional and discuss its effect on trajectory diversity.
  4. [Introduction, Contributions] The claimed unification of sequence-to-sequence prediction and next-token-prediction closed-loop models is not formally developed beyond the hierarchical generative process; a short derivation or explicit mapping to NTP would make this contribution clearer.
  5. [Table II] NIVA's minADE is the best among the listed methods, but its realism, interactive, and map-based scores are all slightly below KiGRAS. The word "competitive" is appropriate, but the conclusion's statement that the model "encodes driving styles effectively" is not backed by quantitative results and should be softened or supported.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the ELBO derivation is self-contained and the simulator is evaluated on held-out Waymo data.

full rationale

The paper's core derivation is self-contained. The variational objective (Eq. 12) is the ELBO of the hierarchical generative process in Eqs. (3)-(4), with a mean-field family (Eq. 11) and the analytic optimal q(z) given by Eq. (13); no fitted parameter is later reused as the predicted quantity. Algorithm 2 samples from the model's own generative prior and the closed-form marginal Gaussian in Eq. (9), which is the standard use of a trained generative model. Table II reports minADE and realism scores on the Waymo test set with 32 rollouts, so the central performance claims are external and not fixed by construction. Self-citations [6], [8], [19] are background references on uncertainty quantification and generalization; they are not invoked as an unverified premise, uniqueness theorem, or fitted input, so they are not load-bearing. The abstract's statement that NIVA 'unifies' seq2seq and NTP models is a high-level framing rather than a formal reduction: Eqs. (3)-(4) define a new generative process that contains both open-loop and closed-loop components, but the paper does not claim an equation-level equivalence to prior models. If anything, this is an under-supported expository claim, not a circular derivation. The hard argmax in Algorithm 1 (k* = argmax_k q(z_n^k)) is a training approximation that could compromise the claimed controllability of non-selected intentions, but that is a correctness/validation risk, not a circularity, because the intention latents are learned from data rather than being defined in terms of the trajectory they are asked to predict.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

Everything the central claim rests on is either standard probability assumptions or modeling choices made for tractability. The model introduces two latent variables and an intermediate state without external evidence, so their semantic meaning is only inferred from the qualitative figure. The hard intention assignment and linear-Gaussian emission are the most consequential ad hoc choices.

free parameters (6)
  • intention codebook size K = 6
    Number of discrete intention categories is chosen by the authors without a data-driven criterion.
  • latent feature dimension = 128
    Dimension of agent and map feature vectors, chosen by the authors.
  • emission noise constant ε = unspecified
    Small constant in the linear Gaussian emission model (Eq. 9-10); value not reported.
  • Fourier feature scale σ = unspecified
    Variance of the Gaussian used to sample F in the Fourier feature mapping (Eq. 6); not reported.
  • intention prior π = unspecified, likely uniform
    Categorical prior for z in the generative process; no learned or stated values.
  • number of adaptive-norm blocks = 1
    Ablation shows one block used for reported results; memory limits prevent more.
assumptions (6)
  • domain assumption Agents arrive by a Poisson process; in practice N is set independent of other parameters
    Section III-A; the paper acknowledges this is not the focus and treats N as fixed input.
  • domain assumption Behavior style features b_n are i.i.d. standard Gaussian
    Section III-A, generative process step 2.
  • domain assumption Intentions z_n follow a fixed categorical distribution Cat(π)
    Section III-A, generative process step 3a.
  • ad hoc to paper The emission model p(o_t|s_t) is linear Gaussian with learned weight A and fixed ε
    Section III-D; this choice is made to obtain closed-form marginal and posterior, not derived from data.
  • ad hoc to paper Mean-field variational approximation q(z,b)=∏ q(z_n)q(b_n)
    Section III-E, Eq. 11; assumes posterior independence for tractability.
  • ad hoc to paper Hard assignment of intention k* = argmax q(z_n) during training
    Algorithm 1; the paper describes soft EM but the algorithm selects the most consistent intention, potentially discarding multimodality.
invented entities (3)
  • behavior style latent b_n
    purpose: Continuous latent variable intended to encode driving style for controllable generation
    No external ground truth or quantitative validation; only qualitative visualization in Fig. 3.
  • intention latent z_n
    purpose: Discrete latent variable encoding long-term destination/intention for controllable generation
    No external labels; the posterior is learned purely from trajectory data.
  • latent motion state s_t
    purpose: Intermediate representation to enable closed-form Gaussian marginalization for sampling and posterior updates
    A modeling device with no physical meaning; introduced to obtain closed-form equations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Learning Closed-Loop Probabilistic Multi-Agent Simulator." pith.science (2026). https://pith.science/paper/VGFFXQFD

@misc{pith2026250800384,
  author       = {Pith},
  title        = {Pith review of: On Learning Closed-Loop Probabilistic Multi-Agent Simulator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGFFXQFD}},
  note         = {Machine review of arXiv:2508.00384}
}
read the original abstract

The rapid iteration of autonomous vehicle (AV) deployments leads to increasing needs for building realistic and scalable multi-agent traffic simulators for efficient evaluation. Recent advances in this area focus on closed-loop simulators that enable generating diverse and interactive scenarios. This paper introduces Neural Interactive Agents (NIVA), a probabilistic framework for multi-agent simulation driven by a hierarchical Bayesian model that enables closed-loop, observation-conditioned simulation through autoregressive sampling from a latent, finite mixture of Gaussian distributions. We demonstrate how NIVA unifies preexisting sequence-to-sequence trajectory prediction models and emerging closed-loop simulation models trained on Next-token Prediction (NTP) from a Bayesian inference perspective. Experiments on the Waymo Open Motion Dataset demonstrate that NIVA attains competitive performance compared to the existing method while providing embellishing control over intentions and driving styles.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [1]

    Preparing a nation for autonomous vehicles: opportunities, barriers and policy recommendations,

    D. J. Fagnant and K. Kockelman, “Preparing a nation for autonomous vehicles: opportunities, barriers and policy recommendations,” Trans- portation Research Part A: Policy and Practice, vol. 77, pp. 167–181, 2015

    work page 2015

  2. [2]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

    S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou, Z. Yang, A. Chouard, P. Sun, J. Ngiam, V . Vasudevan, A. McCauley, J. Shlens, and D. Anguelov, “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” in Proceedings of the IEEE/CVF International Conference on Co...

    work page 2021

  3. [3]

    Argoverse: 3d tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan, and J. Hays, “Argoverse: 3d tracking and forecasting with rich maps,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    work page 2019

  4. [4]

    Simnet: Learning reactive self-driving simulations from real-world observations,

    L. Bergamini, Y . Ye, O. Scheel, L. Chen, C. Hu, L. Del Pero, B. Osi´nski, H. Grimmett, and P. Ondruska, “Simnet: Learning reactive self-driving simulations from real-world observations,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 5119–5125

    work page 2021

  5. [5]

    The waymo open sim agents challenge,

    N. Montali, J. Lambert, P. Mougin, A. Kuefler, N. Rhinehart, M. Li, C. Gulino, T. Emrich, Z. Yang, S. Whiteson, B. White, and D. Anguelov, “The waymo open sim agents challenge,” in Advances in Neural Information Processing Systems , A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 59 151–59 171

    work page 2023

  6. [6]

    Generalizability analysis of graph-based trajectory predictor with vectorized representation,

    J. Lu, W. Zhan, M. Tomizuka, and Y . Hu, “Generalizability analysis of graph-based trajectory predictor with vectorized representation,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 13 430–13 437

    work page 2022

  7. [7]

    Trafficsim: Learning to simulate realistic multi-agent behaviors,

    S. Suo, S. Regalado, S. Casas, and R. Urtasun, “Trafficsim: Learning to simulate realistic multi-agent behaviors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 10 400–10 409

    work page 2021

  8. [8]

    Quantifying uncertainty in motion prediction with variational bayesian mixture,

    J. Lu, C. Cui, Y . Ma, A. Bera, and Z. Wang, “Quantifying uncertainty in motion prediction with variational bayesian mixture,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 15 428–15 437

    work page 2024

Show all 29 references
  1. [9]

    Editing driver character: Socially-controllable behavior generation for interactive traffic simulation,

    W.-J. Chang, C. Tang, C. Li, Y . Hu, M. Tomizuka, and W. Zhan, “Editing driver character: Socially-controllable behavior generation for interactive traffic simulation,” IEEE Robotics and Automation Letters , vol. 8, no. 9, pp. 5432–5439, 2023

    2023

  2. [10]

    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 (CVPR) , June 2020

    2020

  3. [11]

    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,” in Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Asso...

    2022

  4. [12]

    Trafficbots v1.5: Traffic simulation via conditional vaes and transformers with relative pose encoding,

    Z. Zhang, C. Sakaridis, and L. Van Gool, “Trafficbots v1.5: Traffic simulation via conditional vaes and transformers with relative pose encoding,” arXiv preprint arXiv:2406.10898 , 2024

    2024 arXiv

  5. [13]

    Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction,

    Z. Zhou, H. Hu, X. Chen, J. Wang, N. Guan, K. Wu, Y .-H. Li, Y .- K. Huang, and C. J. Xue, “Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction,” 2024

    2024

  6. [14]

    Smart: Scalable multi-agent real-time motion generation via next-token prediction,

    W. Wu, X. Feng, Z. Gao, and Y . Kan, “Smart: Scalable multi-agent real-time motion generation via next-token prediction,” in Advances in Neural Information Processing Systems , A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Cu...

    2024

  7. [15]

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model,

    Z. Xu, Y . Zhang, E. Xie, Z. Zhao, Y . Guo, K.-Y . K. Wong, Z. Li, and H. Zhao, “Drivegpt4: Interpretable end-to-end autonomous driving via large language model,” IEEE Robotics and Automation Letters , vol. 9, no. 10, pp. 8186–8193, 2024

    2024

  8. [16]

    Deep learning to represent subgrid processes in climate models,

    S. Rasp, M. S. Pritchard, and P. Gentine, “Deep learning to represent subgrid processes in climate models,” Proceedings of the National Academy of Sciences , vol. 115, no. 39, pp. 9684–9689, 2018

    2018

  9. [17]

    Learning particle physics by example: location-aware generative adversarial networks for physics synthesis,

    L. de Oliveira, M. Paganini, and B. Nachman, “Learning particle physics by example: location-aware generative adversarial networks for physics synthesis,” Computing and Software for Big Science, vol. 1, no. 1, p. 4, 2017

    2017

  10. [18]

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

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention query- ing,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 3955–3971, 2024

    2024

  11. [19]

    Towards generalizable and interpretable motion prediction: A deep variational Bayes approach,

    J. Lu, W. Zhan, M. Tomizuka, and Y . Hu, “Towards generalizable and interpretable motion prediction: A deep variational Bayes approach,” in Proceedings of The 27th International Conference on Artificial Intelligence and Statistics , ser. Proceedings of Machine Learning Researc...

    2024

  12. [20]

    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 (CVPR), June 2023, pp. 17 863–17 873

    2023

  13. [21]

    Fourier fea- tures let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Ragha- van, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier fea- tures let networks learn high frequency functions in low dimensional domains,” in Advances in Neural Information Processing Systems , H. L...

    2020

  14. [22]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023

    2023

  15. [23]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2023, pp. 4195–4205

    2023

  16. [24]

    Multiverse transformer: 1st place solution for waymo open sim agents challenge 2023,

    Y . Wang, T. Zhao, and F. Yi, “Multiverse transformer: 1st place solution for waymo open sim agents challenge 2023,” 2023

    2023

  17. [25]

    Solving motion planning tasks with a scalable generative model,

    Y . Hu, S. Chai, Z. Yang, J. Qian, K. Li, W. Shao, H. Zhang, W. Xu, and Q. Liu, “Solving motion planning tasks with a scalable generative model,” 2024

    2024

  18. [26]

    Kigras: Kinematic-driven generative model for realistic agent simulation,

    J. Zhao, J. Zhuang, Q. Zhou, T. Ban, Z. Xu, H. Zhou, J. Wang, G. Wang, Z. Li, and B. Li, “Kigras: Kinematic-driven generative model for realistic agent simulation,” 2024

    2024

  19. [27]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” 2019

    2019

  20. [28]

    Sgdr: Stochastic gradient descent with warm restarts,

    ——, “Sgdr: Stochastic gradient descent with warm restarts,” 2017

    2017

  21. [29]

    Kigras: Kinematic-driven generative model for realistic agent simulation,

    J. Zhao, J. Zhuang, Q. Zhou, T. Ban, Z. Xu, H. Zhou, J. Wang, G. Wang, Z. Li, and B. Li, “Kigras: Kinematic-driven generative model for realistic agent simulation,” IEEE Robotics and Automation Letters, vol. 10, no. 2, pp. 1082–1089, 2025

    2025

Pith tools

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