Pith. sign in

REVIEW 4 major objections 8 minor 40 references

Rolling Ahead Diffusion for Traffic Scene Simulation

T0 review · 4 major / 8 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 A rolling diffusion traffic planner predicts the next step while carrying a partially denoised future, which the paper claims keeps scenes reactive at a quarter of the compute.

desk verdict A solid application of rolling diffusion to traffic simulation, but the 'beneficial compromise' claim is not supported by the closed-loop reactivity numbers. read the letter →

arxiv 2502.09587 v1 pith:OQ2KRXXW submitted 2025-02-13 cs.LG cs.RO

Yunpeng Liu , Matthew Niedoba , William Harvey , Adam Scibior , Berend Zwartsenberg , Frank Wood This is my paper · ORCID
classification cs.LGcs.RO
keywords trafficsimulationdiffusionmodelsautoregressivegenerationclosed-looprollingmulti-agentmotionpredictionconditioningaugmentationINTERACTIONdataset
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 driving simulation needs surrounding agents that both drive naturally and react to an ego vehicle that may behave unpredictably. Full scene diffusion models are realistic but must replan from scratch at every step to react, which is slow; autoregressive diffusion models are fast but see only one step ahead. This paper proposes a rolling diffusion planner that fully denoises only the next state for all agents while carrying partially denoised future states in a sliding window, so the model keeps a look-ahead plan at low cost. On the INTERACTION dataset, the resulting planner beats an autoregressive diffusion baseline on scene-level displacement and miss rate, cuts prediction time by roughly half, and achieves a collision rate of 0.019 compared with 0.016 for AR and 0.014 for the slowest replanning baseline. The authors argue this is the useful compromise: reactivity close to full replanning at a fraction of its cost.

What carries the argument

The load-bearing object is the rolling diffusion process, applied to a sliding window of $W$ timesteps for all agents jointly. Within the window each future state is assigned a local diffusion time $\tau_w$ that increases with the window index, so the signal-to-noise ratio satisfies $\mathrm{SNR}(\tau_{w+1}) < \mathrm{SNR}(\tau_w)$: the further ahead a state lies, the noisier it is. The score-estimation network $D_\theta$, a transformer that maps a noised window to clean states, takes the noised window, the map, agent conditioning, and the vector of local diffusion times, and jointly predicts clean states for every element; during the rolling stage only the first future state is fully denoised, and the rest are carried forward as partially denoised continuations. The diffusion process uses a continuous noise schedule and a second-order sampler, and conditioning augmentation, applied by adding Gaussian noise to observed states during training, is essential for stable autoregressive rollouts.

What would settle it

Run a closed-loop benchmark with an ego agent that makes sudden, out-of-distribution maneuvers while sweeping the window size and the number of denoising steps, and record collision rate against prediction time. The central claim predicts a regime where RoAD matches AR's collision rate at about half the compute; if every window size either collides more often than AR or takes nearly as long as full replanning, the claimed compromise is refuted.

Watch

Extended reading notes

Core claim

The central claim of the paper's rolling ahead diffusion model (RoAD) is that a diffusion traffic-scene generator can be made both reactive and efficient by rolling a sliding window forward one step at a time, fully denoising only the immediate next state for every agent and leaving later states partially denoised with temporally correlated noise levels. Because the next prediction starts from a partially denoised continuation rather than from white noise, the model needs far fewer denoising steps than a diffusion autoregressive model, yet it retains a look-ahead representation of the near future. The paper reports that RoAD-15 and RoAD-20 improve minSceneADE, minSceneFDE, and miss rate over the AR baseline on INTERACTION validation, and that with a 2-second look-ahead window the method requires four times fewer function evaluations than a typical autoregressive diffusion model for 4-second scenarios. In an adversarial closed-loop test, RoAD-15 reduces collision rate by more than 2.5x compared with the non-reactive one-shot DJINN baseline, at half the prediction time of AR.

Load-bearing premise

The load-bearing premise is that future states that are only partly denoised still carry enough signal to support reactive, realistic next-step planning; if that retained signal degrades reactivity faster than it saves compute, the claimed compromise disappears.

Editorial extensions

If this is right

  • A 2-second look-ahead window yields a fourfold reduction in function evaluations relative to a typical diffusion autoregressive model for 4-second scenarios, cutting wall-clock prediction time roughly in half.
  • Scene-level realism improves: RoAD-20 reaches minSceneADE 0.654 and minSceneFDE 1.553, versus 0.695 and 1.670 for AR, with miss rate falling from 0.168 to 0.142.
  • Reactivity remains close to a full replanning baseline: RoAD-15 has a collision rate of 0.019 in adversarial scenes, versus 0.016 for AR and 0.014 for DJINN-10 (MPC-1), while using about half the prediction time of either.
  • Window size becomes a direct tradeoff control: RoAD-20 runs faster (0.20 min) than RoAD-15 (0.34 min) but collides more often (0.024 vs 0.019).
  • Conditioning augmentation is necessary for stable autoregressive diffusion rollouts; without it RoAD's minSceneADE rises from 0.663 to 0.930 and minSceneFDE from 1.579 to 2.197.

Reading between the lines

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

  • The paper leaves implicit that the same signal-to-noise-versus-reactivity knob could be adjusted at test time: lowering the noise level of carried future states when a near-collision is detected might recover much of MPC's reactivity without a full regeneration. This is an editorial inference, not a paper claim.
  • Because the partially denoised future acts as a compressed plan prior, test-time classifier guidance for adversarial or goal-directed scenario generation becomes a natural next step; the paper lists test-time conditioning only as future work.
  • The one-clean-step-plus-partially-noised-continuation pattern is not specific to road traffic; closed-loop simulators for pedestrian crowds or warehouse robots face the same replanning-cost problem and could reuse the design, though reactivity limits would need re-measuring.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper proposes RoAD, a rolling diffusion model for multi-agent traffic scene simulation. RoAD applies the Rolling Diffusion Model (RDM) framework to the DJINN transformer architecture, predicting the next scene state in fully denoised form while simultaneously producing partially denoised future states in a sliding window, with the goal of combining the reactivity of model-predictive-control-style replanning with the computational efficiency of autoregressive prediction. The authors evaluate on the INTERACTION dataset, comparing against DJINN one-shot, DJINN-10 with MPC replanning, and an autoregressive diffusion baseline (AR). They report scene-level displacement metrics (minSceneADE, minSceneFDE, miss rate), collision rates under a scenario with one adversarial slowed agent, and prediction times. The central claims are that RoAD produces more realistic scenes than AR while requiring fewer function evaluations, and that it offers a beneficial compromise between reactivity and computational efficiency.

Significance. If the claims are substantiated, RoAD would provide a practical method for closed-loop traffic simulation at lower computational cost than MPC-style diffusion models while retaining some look-ahead capacity. The application of RDM to the multi-agent traffic setting is natural, and the forward-process conditioning augmentation adaptation is clearly motivated. However, the empirical support is currently incomplete: the central compromise is not convincingly demonstrated, the efficiency gain is not quantified at the level of function evaluations, and no uncertainty measures are provided. The paper does not provide code or machine-checked proofs; its contributions are empirical. With the additional analysis requested below, the results could become publishable, but as it stands the evidence for the central claims is insufficient.

major comments (4)
  1. [Experiments (Table 2, 'Reactivity')] The paper's abstract claims RoAD 'mixes the benefits of both methods' and achieves a 'beneficial compromise between reactivity and computational efficiency.' However, Table 2 shows that RoAD-15 and RoAD-20 have higher collision rates (0.019 and 0.024) than the AR baseline (0.016), while DJINN-10 (MPC-1) has the lowest collision rate (0.014). Thus the look-ahead design does not improve reactivity relative to AR; it provides only a modest open-loop realism gain (Table 1). The statement in the Methods section that 'the partially denoised future plan affects the reactivity of our model' effectively concedes that the look-ahead window degrades reactivity. Since reactivity is the stated counterweight to efficiency, the claimed beneficial compromise is not established by the reported results.
  2. [Tables 1 and 2] All reported metrics are point estimates, with no confidence intervals, standard errors, or significance tests. The differences in minSceneADE (0.673/0.654 vs 0.695) and collision rate (0.019 vs 0.016) are small relative to the scale of the evaluation (1,440 scenarios, three samples per scenario). Without uncertainty quantification, the claims that RoAD 'produces more realistic scenes' than AR and that the collision-rate difference from AR is acceptable cannot be assessed. Please report per-seed variability, bootstrap confidence intervals, or paired significance tests.
  3. [Abstract; Methods ('Rolling ahead autoregressive model'); Experiments (Table 2)] The abstract states that with a '2-second look-ahead planning window,' RoAD requires 'four times fewer function evaluations' than a typical autoregressive diffusion model for 4-second scenarios. No NFE counts or denoising schedules (e.g., number of Heun steps in warm-up vs rolling) are reported. The wall-clock prediction times in Table 2 show speedups of 2.0x (RoAD-15) and 3.4x (RoAD-20) over AR, not 4x, suggesting overhead that should be explained. Moreover, the 2-second look-ahead claim is inconsistent with the RoAD-20 configuration (W=20, n=10), which predicts only 10 future steps at 10 Hz (1 second). Please clarify the window definition and provide a derivation or count of function evaluations.
  4. [Experiments ('Reactivity')] The reactivity evaluation uses a single adversarial perturbation (an agent slowing to half its replayed trajectory). This protocol is reasonable as a stress test, but the paper's broader claim about closed-loop reactivity would be supported by varying the perturbation type and magnitude (e.g., cut-ins, accelerations, partial deviation). The current single-scenario design, combined with the small observed differences, limits the generality of the reactivity conclusion.
minor comments (8)
  1. [Abstract] The phrase 'stand along motion planner' should be 'standalone motion planner.'
  2. [Background, Eq. (4)] The window notation x_{t:t+W} contains W+1 elements, but the product runs from w=t to t+W-1, which includes only W elements. Please align the indexing or define the convention explicitly.
  3. [Methods, Eq. (9)] The training objective in Eq. (9) does not show how the task ratio beta (discussed in the RDM background) weights the warm-up and rolling losses; please specify the combined loss function.
  4. [Methods, 'Conditioning Augmentation'] The word 'enssential' should be 'essential.'
  5. [Experiments, 'Reactivity'] The text says 'report the average collision rate in Table 1,' but the collision rates appear in Table 2; Table 1's caption also lists 'collision rate' although the table contains no collision column.
  6. [Supplementary Materials] The main text refers to 'Figure S1' for the RoAD-20 failure case, but the supplementary material shows this as Figure 4; please update the cross-reference.
  7. [Methods, 'Rolling ahead autoregressive model'] The vector tau in Eq. (9) is introduced but not fully described; please state how the per-element local diffusion times are computed and embedded in the network (e.g., as additional feature channels).
  8. [Experiments, 'Implementation Details'] The conditioning augmentation ratios (0.2 for AR/DJINN-10, 0.5 for RoAD) are given without explanation; a brief sensitivity study or justification would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RoAD's central claims rest on empirical comparisons against independently trained baselines, and self-citations are architectural prior work, not load-bearing.

full rationale

The paper's derivation chain is an application of RDM (Ruhe et al. 2024) to traffic simulation, using a DJINN-style transformer backbone, followed by empirical evaluation. The core claims are measured, not derived: displacement metrics in Table 1 and collision rates under an adversarial ego agent in Table 2 compare RoAD against separately trained AR and DJINN-MPC baselines. Nothing in the method defines the prediction target in terms of the fitted model output; the objective in Eq. (9) is the standard denoising loss over the sliding window, and the partially noised future states are produced by the learned denoiser rather than by a pre-fit parameter renamed as a prediction. Self-citations to DJINN and other prior work concern architecture, map representation, and conditioning augmentation; these are design choices, not uniqueness theorems, fitted constraints, or forbidden alternatives. The paper also explicitly acknowledges the reactivity limitation ('the partially denoised future plan affects the reactivity of our model') and reports the SNR/window-size tradeoff experimentally, which is an honest limitation rather than a disguised input. The 'four times fewer function evaluations' claim is an engineering count of denoising iterations relative to an autoregressive diffusion baseline, not an identity forced by construction. Concerns about missing confidence intervals, the small size of the realism improvement, or the abstract overstating the compromise are correctness and evidence issues, not circularity.

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

The central method borrows the entire diffusion and score-matching apparatus from prior work and adds domain-specific free hyperparameters. No new physical entities are introduced.

free parameters (4)
  • window length W = 15 or 20
    Sliding window size; larger window improves displacement metrics but raises collision rate in the reactivity test. Chosen manually.
  • observation length n = 10
    Number of clean observed states used as conditioning; fixed without ablation.
  • task ratio beta = 0.1
    Controls the proportion of warm-up versus rolling training tasks; set by hand, no ablation reported.
  • conditioning augmentation ratio = 0.5 for RoAD, 0.2 for AR/DJINN-10
    Tuned per model after finding higher ratios harm baselines; directly affects all reported quality metrics.
assumptions (4)
  • domain assumption Temporally correlated noise with decreasing SNR across the window is a valid inductive bias for multi-agent traffic trajectories.
    Borrowed from RDM and applied to traffic via Eq. (4) to (6); no ablation on the noise schedule for this domain.
  • domain assumption The joint state representation (2D position plus 1D orientation for all agents) and map polyline conditioning from DJINN are sufficient for realistic scene generation.
    Taken from Niedoba et al. 2024; no ablations over state representation or map encoding.
  • ad hoc to paper Learning to denoise partially noised future states in parallel with a clean next step preserves enough reactivity for closed-loop simulation.
    Core design premise; the paper acknowledges reactivity depends on SNR and degrades with larger windows, and the collision rates support that degradation.
  • domain assumption Scene-level displacement metrics and collision rate against a single slowed adversarial agent measure simulation realism and reactivity.
    Used as the main evaluation; no perceptual or downstream-task metrics are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rolling Ahead Diffusion for Traffic Scene Simulation." pith.science (2026). https://pith.science/paper/OQ2KRXXW

@misc{pith2026250209587,
  author       = {Pith},
  title        = {Pith review of: Rolling Ahead Diffusion for Traffic Scene Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQ2KRXXW}},
  note         = {Machine review of arXiv:2502.09587}
}
read the original abstract

Realistic driving simulation requires that NPCs not only mimic natural driving behaviors but also react to the behavior of other simulated agents. Recent developments in diffusion-based scenario generation focus on creating diverse and realistic traffic scenarios by jointly modelling the motion of all the agents in the scene. However, these traffic scenarios do not react when the motion of agents deviates from their modelled trajectories. For example, the ego-agent can be controlled by a stand along motion planner. To produce reactive scenarios with joint scenario models, the model must regenerate the scenario at each timestep based on new observations in a Model Predictive Control (MPC) fashion. Although reactive, this method is time-consuming, as one complete possible future for all NPCs is generated per simulation step. Alternatively, one can utilize an autoregressive model (AR) to predict only the immediate next-step future for all NPCs. Although faster, this method lacks the capability for advanced planning. We present a rolling diffusion based traffic scene generation model which mixes the benefits of both methods by predicting the next step future and simultaneously predicting partially noised further future steps at the same time. We show that such model is efficient compared to diffusion model based AR, achieving a beneficial compromise between reactivity and computational efficiency.

Figures

Figures reproduced from arXiv: 2502.09587 by the authors.

Figure 1
Figure 1. From top to bottom row: DJINN (Niedoba et al. 2024), autoregressive (AR), and RoAD (Ours). The adversarial [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Rolling Diffusion Model. Columns represent sequence timesteps and rows represent diffusion timesteps. Circles are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. From Top to Bottom row, AR, RoAD-20. By looking ahead of the subsequent step, the pedestrian marked with a red [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: From top to bottom row: RoAD-20, RoAD-15. The adversarial agent, marked with a red dot, follows its replay log [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

  1. [1]

    D.; Ho, J.; Tarlow, D.; and Van Den Berg, R

    Austin, J.; Johnson, D. D.; Ho, J.; Tarlow, D.; and Van Den Berg, R. 2021. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34: 17981--17993

    work page 2021

  2. [2]

    Chai, Y.; Sapp, B.; Bansal, M.; and Anguelov, D. 2019. MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction. In Kaelbling, L. P.; Kragic, D.; and Sugiura, K., eds., 3rd Annual Conference on Robot Learning, CoRL 2019, Osaka, Japan, October 30 - November 1, 2019, Proceedings, volume 100 of Proceedings of Machine Learning Re...

    work page 2019

  3. [3]

    Chang, W.-J.; Pittaluga, F.; Tomizuka, M.; Zhan, W.; and Chandraker, M. 2023. Controllable Safety-Critical Closed-loop Traffic Simulation via Guided Diffusion. arXiv preprint arXiv:2401.00391

    arXiv 2023

  4. [4]

    Cui, H.; Radosavljevic, V.; Chou, F.; Lin, T.; Nguyen, T.; Huang, T.; Schneider, J.; and Djuric, N. 2019. Multimodal Trajectory Predictions for Autonomous Driving using Deep Convolutional Networks. In International Conference on Robotics and Automation, ICRA 2019, Montreal, QC, Canada, May 20-24, 2019 , 2090--2096. IEEE

    work page 2019

  5. [5]

    Gulino, C.; Fu, J.; Luo, W.; Tucker, G.; Bronstein, E.; Lu, Y.; Harb, J.; Pan, X.; Wang, Y.; Chen, X.; et al. 2024. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems, 36

    work page 2024

  6. [6]

    Guo, Z.; Gao, X.; Zhou, J.; Cai, X.; and Shi, B. 2023. SceneDM: Scene-level multi-agent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736

    arXiv 2023

  7. [7]

    Han, B.; Peng, H.; Dong, M.; Ren, Y.; Shen, Y.; and Xu, C. 2024. AMD: Autoregressive Motion Diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2022--2030

    work page 2024

  8. [8]

    Harvey, W.; Naderiparizi, S.; Masrani, V.; Weilbach, C.; and Wood, F. 2022. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35: 27953--27965

    work page 2022

Show all 40 references
  1. [9]

    P.; Poole, B.; Norouzi, M.; Fleet, D

    Ho, J.; Chan, W.; Saharia, C.; Whang, J.; Gao, R.; Gritsenko, A.; Kingma, D. P.; Poole, B.; Norouzi, M.; Fleet, D. J.; et al. 2022 a . Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303

    2022 arXiv

  2. [10]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

    2020

  3. [11]

    J.; Norouzi, M.; and Salimans, T

    Ho, J.; Saharia, C.; Chan, W.; Fleet, D. J.; Norouzi, M.; and Salimans, T. 2022 b . Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47): 1--33

    2022

  4. [12]

    A.; Bastings, J.; Poole, B.; Berg, R

    Hoogeboom, E.; Gritsenko, A. A.; Bastings, J.; Poole, B.; Berg, R. v. d.; and Salimans, T. 2021. Autoregressive diffusion models. arXiv preprint arXiv:2110.02037

    2021 arXiv

  5. [13]

    B.; and Levine, S

    Janner, M.; Du, Y.; Tenenbaum, J. B.; and Levine, S. 2022. Planning with Diffusion for Flexible Behavior Synthesis. In Chaudhuri, K.; Jegelka, S.; Song, L.; Szepesv \' a ri, C.; Niu, G.; and Sabato, S., eds., International Conference on Machine Learning, ICML 2022, 17-23 July ...

    2022

  6. [14]

    Jiang, C.; Cornman, A.; Park, C.; Sapp, B.; Zhou, Y.; Anguelov, D.; et al. 2023. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9644--9653

    2023

  7. [15]

    Karras, T.; Aittala, M.; Aila, T.; and Laine, S. 2022. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35: 26565--26577

    2022

  8. [16]

    Kingma, D.; Salimans, T.; Poole, B.; and Ho, J. 2021. Variational diffusion models. Advances in neural information processing systems, 34: 21696--21707

    2021

  9. [17]

    W.; Niedoba, M.; Sefas, J.; Dabiri, S.; Green, D.; Liang, X.; Zwartsenberg, B.; \'S cibior, A.; et al

    Liu, Y.; Lioutas, V.; Lavington, J. W.; Niedoba, M.; Sefas, J.; Dabiri, S.; Green, D.; Liang, X.; Zwartsenberg, B.; \'S cibior, A.; et al. 2023. Video Killed the HD-Map: Predicting Multi-Agent Behavior Directly From Aerial Images. In 2023 IEEE 26th International Conference on ...

    2023

  10. [18]

    S.; and Sapp, B

    Nayakanti, N.; Al - Rfou, R.; Zhou, A.; Goel, K.; Refaat, K. S.; and Sapp, B. 2023. Wayformer: Motion Forecasting via Simple & Efficient Attention Networks. In IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023 , 2980--2987. IEEE

    2023

  11. [19]

    L.; Ling, J.; Roelofs, R.; Bewley, A.; Liu, C.; Venugopal, A.; et al

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

    2021 arXiv

  12. [20]

    Niedoba, M.; Lavington, J.; Liu, Y.; Lioutas, V.; Sefas, J.; Liang, X.; Green, D.; Dabiri, S.; Zwartsenberg, B.; Scibior, A.; et al. 2024. A Diffusion-Model of Joint Interactive Navigation. Advances in Neural Information Processing Systems, 36

    2024

  13. [21]

    R.; Hendy, N.; Wang, Z.; Morales, A.; Wang, K.; and Roy, N

    Pronovost, E.; Ganesina, M. R.; Hendy, N.; Wang, Z.; Morales, A.; Wang, K.; and Roy, N. 2023. Scenario Diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems, 36: 68873--68894

    2023

  14. [22]

    J.; Fidler, S.; and Litany, O

    Rempe, D.; Philion, J.; Guibas, L. J.; Fidler, S.; and Litany, O. 2022. Generating Useful Accident-Prone Driving Scenarios via a Learned Traffic Prior. In Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  15. [23]

    Ruhe, D.; Heek, J.; Salimans, T.; and Hoogeboom, E. 2024. Rolling Diffusion Models. arXiv preprint arXiv:2402.09470

    2024 arXiv

  16. [24]

    \'S cibior, A.; Lioutas, V.; Reda, D.; Bateni, P.; and Wood, F. 2021. Imagining the road ahead: Multi-agent trajectory prediction via differentiable simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), 720--725. IEEE

    2021

  17. [25]

    Sohl-Dickstein, J.; Weiss, E.; Maheswaranathan, N.; and Ganguli, S. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, 2256--2265. PMLR

    2015

  18. [26]

    P.; Kumar, A.; Ermon, S.; and Poole, B

    Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2020. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456

    2020 arXiv

  19. [27]

    Sun, P.; Kretzschmar, H.; Dotiwalla, X.; Chouard, A.; Patnaik, V.; Tsui, P.; Guo, J.; Zhou, Y.; Chai, Y.; Caine, B.; et al. 2020. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern reco...

    2020

  20. [28]

    Suo, S.; Regalado, S.; Casas, S.; and Urtasun, R. 2021. Trafficsim: Learning to simulate realistic multi-agent behaviors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10400--10409

    2021

  21. [29]

    Tashiro, Y.; Song, J.; Song, Y.; and Ermon, S. 2021. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems, 34: 24804--24816

    2021

  22. [30]

    Treiber, M.; Hennecke, A.; and Helbing, D. 2000. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2): 1805

    2000

  23. [31]

    Uria, B.; Murray, I.; and Larochelle, H. 2014. A deep and tractable density estimator. In International Conference on Machine Learning, 467--475. PMLR

    2014

  24. [32]

    Wu, T.; Fan, Z.; Liu, X.; Zheng, H.-T.; Gong, Y.; Jiao, J.; Li, J.; Guo, J.; Duan, N.; Chen, W.; et al. 2024. Ar-diffusion: Auto-regressive diffusion model for text generation. Advances in Neural Information Processing Systems, 36

    2024

  25. [33]

    Xu, D.; Chen, Y.; Ivanovic, B.; and Pavone, M. 2023. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2929--2936. IEEE

    2023

  26. [34]

    o m, H.; and Bj \

    Yin, W.; Tu, R.; Yin, H.; Kragic, D.; Kjellstr \"o m, H.; and Bj \"o rkman, M. 2023. Controllable Motion Synthesis and Reconstruction with Autoregressive Diffusion Models. In 2023 32nd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), 1102--1...

    2023

  27. [35]

    ummerle, J.; K\

    Zhan, W.; Sun, L.; Wang, D.; Shi, H.; Clausse, A.; Naumann, M.; K\"ummerle, J.; K\"onigshof, H.; Stiller, C.; de La Fortelle, A.; and Tomizuka, M. 2019. INTERACTION Dataset : An INTERnational , Adversarial and Cooperative moTION Dataset in Interactive Driving Scenarios with Se...

    2019 arXiv

  28. [36]

    Zhang, Z.; Liu, R.; Aberman, K.; and Hanocka, R. 2023. TEDi: Temporally-entangled diffusion for long-term motion synthesis. arXiv preprint arXiv:2307.15042

    2023 arXiv

  29. [37]

    Zhao, T.; Xu, Y.; Monfort, M.; Choi, W.; Baker, C.; Zhao, Y.; Wang, Y.; and Wu, Y. N. 2019. Multi-agent tensor fusion for contextual trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12126--12134

    2019

  30. [38]

    Zhong, Z.; Rempe, D.; Xu, D.; Chen, Y.; Veer, S.; Che, T.; Ray, B.; and Pavone, M. 2023. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 3560--3566. IEEE

    2023

  31. [39]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  32. [40]

    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 gl...

Pith tools

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