Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Beyond Simulation: Benchmarking World Models for Planning and Causality in Autonomous Driving

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

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

Pith's one-line read World models that score well on standard traffic-simulation benchmarks can still be unreliable pseudo-environments for training a closed-loop driving policy, because they mishandle scenes where the ego agent is not under the model's…

desk verdict A useful new evaluation lens for traffic world models, but the causal interpretation of the delta metrics is not yet earned; needs a control condition and error bars. read the letter →

arxiv 2508.01922 v1 pith:EQAS7EHU submitted 2025-08-03 cs.RO

Hunter Schofield , Mohammed Elmahgiubi , Kasra Rezaee , Jinjun Shan This is my paper · ORCID
classification cs.RO
keywords worldmodelsautonomousdrivingtrafficsimulationWOSACmetametriccausalagentsclosed-looppolicytrainingcontroldropoutrobustness
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 argues that the standard way of scoring learned traffic simulators—the WOSAC realism metametric—does not tell you whether a world model can serve as a training environment for a closed-loop driving policy. The authors show that autoregressive world models that predict all traffic agents at once are biased toward scenarios where every object is under the model's control: when the ego vehicle is forced to replay its original trajectory, a large fraction of rollouts degrade, such as the vehicle behind failing to brake when the ego stops at a yellow light. To expose this, they propose delta metrics that compare ordinary rollouts with rollouts conditioned on the ego's ground-truth trajectory, extend evaluation to agents causal to the ego, and introduce control dropout as a training fix. Their measurements on three world models show that a high realism score can coexist with high sensitivity to uncontrollable objects, and that control dropout reduces that sensitivity without hurting the realism score.

What carries the argument

The argument is carried by a factored model of the world model as $q_{\text{world}} = \pi \, q_{\text{sim}}$ (Eq. 1), splitting the ego policy from the traffic simulator, and by a paired-rollout comparison. The authors generate a second set of rollouts in which the ego policy is replaced by the ground-truth trajectory, then compute negative log-likelihoods for the original and replay-conditioned rollouts, restricted to non-ego agents to remove the bias of exact ego states. The delta metrics $\Delta M_i = M_i - \hat{M}_i$ and $\Delta M^{\text{sim}}_i$ formalize the shift, and thresholds $\tau = 0.05$ or $\tau = 0.035$ turn the deltas into two rates: the simulation confusion rate $C_s$ and the policy confusion rate $C_p$. The causal-agent domain is obtained by extending the WOSAC evaluation set to include all agents causally linked to the ego, using causal annotations from the CausalAgents benchmark. Control dropout is a training-time regularizer that, with probability $p_{\text{drop}}$, replaces a randomly selected agent's world-model-predicted future states with ground-truth states, so the model learns to cope with uncontrolled objects during autoregressive training.

What would settle it

One direct test: take an autoregressive world model and condition it on exact ground-truth ego states for a single step while keeping all other agents under model control, then measure whether per-step realism collapses. The paper's controllability-confusion explanation predicts the collapse appears only when the ego is out of the model's control; the distribution-shift alternative predicts the same collapse from exact ego conditioning alone, even with full control. A second test: train versions of the same model with increasing control-dropout rates and check whether $C_s$ decreases monotonically and whether a fully open-loop trained model reaches near-zero $C_s$, which the paper already observes with SMART.

Watch

Extended reading notes

Core claim

The central claim is that autoregressively trained traffic world models are only reliable when they control every object in the scene. When the ego vehicle is replaced by a replay policy that follows the recorded ground-truth trajectory, simulated traffic often fails to react plausibly, and the standard WOSAC metametric hides this because it is averaged over many agents and scenarios. On a 10% sample of the WOMD evaluation set, forcing the ego onto its original trajectory produced a simulation confusion rate $C_s$ of 0.167 for TrafficBots V1.5 at threshold 0.05 on the combined standard-plus-causal domain, and the causal-agent subset was 68% more sensitive than the standard evaluation domain. GUMP scored higher on the standard realism metric yet had more than twice the confusion rate ($C_s = 0.356$), while SMART, trained with an open-loop objective, showed almost no confusion ($C_s = 0.006$). Control dropout with $p_{\text{drop}} = 0.1$ reduced TrafficBots V1.5's confusion rate from 0.167 to 0.072 while leaving the metametric essentially unchanged.

Load-bearing premise

The load-bearing premise is that forcing the ego vehicle to replay its recorded trajectory is a clean intervention that isolates the traffic simulator's confusion about uncontrollable objects, rather than a way of feeding the model out-of-distribution inputs; the paper does not ablate this alternative explanation.

Editorial extensions

If this is right

  • High realism scores will no longer be enough to certify a learned traffic simulator as a closed-loop policy training environment.
  • Policies trained inside sensitive simulators can acquire bad habits, such as refusing to stop at yellow lights because the simulator punishes stopping with a rear-end collision.
  • Evaluation domains that include agents causal to the ego will reveal failures that a standard evaluation misses, so simulation benchmarks should report both.
  • Control dropout offers a fine-tuning recipe that improves robustness to uncontrollable objects without sacrificing the realism score, suggesting a practical patch for existing models.

Reading between the lines

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

  • The same paired-rollout delta logic could be applied to any learned simulator with an explicit or implicit ego policy, such as pedestrian or multi-robot simulators, as a general test of whether the simulator can be trusted when an external controller acts inside it.
  • A direct next test: train a policy inside a control-dropout-tuned world model and measure closed-loop driving metrics against a policy trained in the baseline model; the paper stops at the simulator level, so the promised policy-level benefit is inferred rather than demonstrated.
  • The threshold choice for $C_s$ and $C_p$ is tied to leaderboard spreads at the time of writing; calibrating the threshold to the per-model noise in the metametric could make confusion rates comparable across models with different realism scales.
  • Because control dropout preserves the standard realism score, it could plausibly be used as a general regularizer for any multi-agent autoregressive predictor, not only autonomous driving simulators.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes extending the WOSAC realism metametric with two delta metrics, ΔM and ΔM_sim (Eqs. 4–5), that compare a world model's rollout realism under its own ego policy versus a ground-truth ego replay, and uses these to define simulation/policy confusion rates C_s and C_p (Eqs. 7–8). It also introduces a causal-agent evaluation domain and a 'control dropout' fine-tuning procedure. Experiments on TrafficBots V1.5, GUMP, and SMART over a 10% (4,393-scenario) subset of WOMD report that models with high WOSAC scores can be strongly sensitive to uncontrollable agents, and that control dropout reduces this sensitivity. The authors conclude that WOSAC-style realism scores are insufficient to judge a world model as a pseudo-environment for policy training.

Significance. If the central claim holds, the paper addresses an important practical gap: a traffic simulator that looks realistic under full closed-loop control may still be a poor training environment when a separate policy controls the ego. The causal-agent evaluation domain and the control-dropout training scheme are concrete, potentially useful contributions, and the SMART open-loop comparison is a sensible contrasting case. However, the causal interpretation of the delta metrics and the link to actual policy-training outcomes are not yet validated. The paper is honest about computational constraints, but the lack of confidence intervals and the absence of an end-to-end policy-training experiment mean the significance is currently conditional on additional control experiments.

major comments (4)
  1. [III.A, Eqs. (4)-(6)] The interpretation of ΔM and ΔM_sim as measuring 'simulation confusion' caused by uncontrollable agents rests on the premise that the only difference between q_world and q̂_world is the identity of the ego controller. For a model trained autoregressively with full control, conditioning on the exact ground-truth ego trajectory is an out-of-distribution input: during training the model conditions on its own predictions, not on externally imposed states. A large ΔM can therefore mean 'this model is unstable to any non-model-conditioned input' rather than 'this model fails to model causal reactions to an uncontrollable agent.' The paper does not include a control condition—for example, replaying a non-ego agent's trajectory, replaying a perturbed ego trajectory, or injecting actions through the action channel instead of replaying states—so the reported effects conflate OOD sensitivity with simulator confusion. The causal-agent results and the GUMP/SMART comparison inherit this confound, and the central claim in Section V.A is underdetermined without such an ablation.
  2. [IV and VI (end-to-end validation)] The paper motivates the delta metrics as indicators of a world model's quality as a pseudo-environment for training a policy, and Section VI recommends control dropout for producing 'better training environments.' Yet no policy is ever trained inside these simulators, and no correlation is established between C_s/ΔM and any downstream planning outcome (e.g., collision rate, goal reaching, or sim-to-real transfer). Table III shows that control dropout reduces C_s, but it is left untested whether that reduction actually improves policy training. This gap is load-bearing for the paper's stated purpose: a metric that does not predict training quality cannot yet justify the title's claim about 'planning.'
  3. [Tables I-III and Section IV.B (statistics)] All quantitative comparisons report single point estimates on one 10% (4,393-scenario) subset without confidence intervals, significance tests, or multiple evaluation subsets. Differences that drive the conclusions—e.g., ΔM_sim=0.025 on D_eval vs 0.042 on D_causal, or C_s=0.356 for GUMP vs 0.167 for TrafficBots—may be within sampling noise. The authors should report bootstrap confidence intervals or evaluate on multiple independent subsamples, and should specify how the 10% subset was selected (e.g., random seed) so the reader can judge representativeness.
  4. [IV.C and Table II (model training setup)] The comparison of GUMP, TrafficBots V1.5, and SMART is based on models 'trained for 10 epochs on only the WOMD data,' but it is not stated whether these are official released checkpoints or retrained models. If they are retrained for only 10 epochs, the statement that GUMP 'performs better on the standard WOSAC realism metric' is true only for these retrained checkpoints, not necessarily for the models as used on the WOSAC leaderboard. Additionally, the claim that SMART 'is trained in an open-loop manner' is not supported by the cited reference: SMART is described as a next-token autoregressive simulator for multi-agent trajectories, and the distinction between its training objective and its rollout mode needs clarification.
minor comments (6)
  1. [III.A, Eq. (3)] The denominator in Eq. (3) appears as '1/N M' and is ambiguous; it should likely be '1/(N·M)'.
  2. [III.A, Eqs. (6)-(8) and Table I] The per-scenario delta metrics are defined with a subscript i, but the confusion-rate definitions use ΔM > τ and ΔM_sim > τ without the subscript and without an explicit absolute value; clarify that the per-scenario absolute shift |M_i − M̂_i| is used in these counts.
  3. [Figure 2 and Table II] The axes and table columns labeled 'metametric' would benefit from an explicit statement of the sign convention (e.g., 'lower is better' for the WOSAC realism score) and from a definition of the evaluation domain (D_eval vs D_eval ∪ D_causal) used for each reported value.
  4. [References] The in-text citation 'Roelofs et al. [32]' does not match the reference list entry, which is Sun et al., 'CausalAgents: A robustness benchmark for motion forecasting'.
  5. [Abstract and Section III.A] The claim 'This is the first work to explore the robustness of traffic-simulating world models when only part of a scenario is controllable' is strong; prior partial-replay or intervention-based evaluations should be discussed or the claim should be softened.
  6. [Table III caption] The caption should state the control-dropout probability p_drop=0.1 used in the reported run, since the value is only mentioned in the body text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the proposed delta metrics are defined as differences of the external WOSAC realism score under GT ego replay, not fitted to the target claim, and the central findings are empirical comparisons against external world models.

full rationale

The paper's derivation chain is self-contained. Eqs. 4-5 define Delta M_i and Delta M_sim_i as differences between the WOSAC NLL realism score for the original rollout and the score for a rollout with the ego vehicle forced to replay ground truth; these are new measurement instruments, not quantities equivalent to their inputs by construction. The thresholds tau=0.035 and tau=0.05 are hand-chosen from leaderboard statistics and are not fitted to the confusion rates they label. The conclusion that autoregressive world models are sensitive to uncontrollable objects is an empirical finding on TrafficBots V1.5, GUMP, and SMART against the external WOMD benchmark, and the control dropout result is a training intervention evaluated with the same metrics rather than a prediction that assumes its own conclusion. The paper cites its own prior work [16] once in related work, but that citation is not load-bearing for any claim. The main interpretive risk, that ground-truth ego replay may be an out-of-distribution input and therefore Delta metrics may measure perturbation sensitivity rather than simulator confusion, is a validity concern about what the metric means, not a circularity in how it is derived.

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

The evaluation rests on hand-chosen thresholds, a single control dropout rate, and an unseeded 10% data subset, all of which affect the reported numbers. The main conceptual assumption is the factorization and replay intervention used to separate simulator confusion from policy behavior. No new physical entities are introduced.

free parameters (3)
  • threshold tau = 0.035 and 0.05
    Used in Section III.A to define simulation confusion rate Cs and policy confusion rate Cp. The values are taken from WOSAC leaderboard statistics and the difference between the leading model and TrafficBots V1.5, not derived or swept. The reported failure rates change materially with tau.
  • control dropout rate pdrop = 0.1
    Used in Section IV.D. A single value is tested without an ablation over pdrop, so the robustness gain is demonstrated for only one hyperparameter setting.
  • evaluation subset fraction = 10% (4393 samples) of the WOMD evaluation set
    All quantitative results depend on one unseeded 10% subset, with no repeated subsampling or confidence intervals. This choice affects every reported mean and confusion rate.
assumptions (5)
  • domain assumption The world model factorizes as q_world(o_t | o_<t, a_{t-1}) = pi(a_{t-1}|o_<t) * q_sim(o_t^sim | o_<t), as stated in Eq. 1.
    This underlies the attribution of benchmark-score changes to either the traffic simulator or the ego policy. If the policy and simulator are not separable, the delta metrics conflate their contributions.
  • domain assumption The WOSAC metametric, computed from negative log-likelihoods across nine realism metrics, is a valid and calibrated measure of simulation quality.
    All proposed delta metrics are differences of this metametric, so any bias or noise in the base metric is inherited by the new metrics. The paper does not independently validate the base metric.
  • domain assumption The causal agent labels from Roelofs et al. [32] transfer correctly to the world model evaluation setting.
    Section III.B extends the evaluated object set using D_causal. If the causal labels are wrong or not applicable to world model rollouts, the causal-domain comparisons lose their intended meaning.
  • domain assumption Ground-truth ego replay is a representative intervention for what happens during closed-loop policy training.
    The core experiments compare q_sim and q_hat_sim under forced replay. This assumes exact ground-truth ego states resemble the states a separately trained policy would produce, which is not tested.
  • domain assumption TrafficBots V1.5, GUMP, and SMART, each retrained for 10 epochs on WOMD data, are representative of state-of-the-art world models and of the top-ranking models mentioned in the abstract.
    The paper generalizes about existing simulation models from three retrained models, which may not match official leaderboard checkpoints or training budgets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Simulation: Benchmarking World Models for Planning and Causality in Autonomous Driving." pith.science (2026). https://pith.science/paper/EQAS7EHU

@misc{pith2026250801922,
  author       = {Pith},
  title        = {Pith review of: Beyond Simulation: Benchmarking World Models for Planning and Causality in Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQAS7EHU}},
  note         = {Machine review of arXiv:2508.01922}
}
read the original abstract

World models have become increasingly popular in acting as learned traffic simulators. Recent work has explored replacing traditional traffic simulators with world models for policy training. In this work, we explore the robustness of existing metrics to evaluate world models as traffic simulators to see if the same metrics are suitable for evaluating a world model as a pseudo-environment for policy training. Specifically, we analyze the metametric employed by the Waymo Open Sim-Agents Challenge (WOSAC) and compare world model predictions on standard scenarios where the agents are fully or partially controlled by the world model (partial replay). Furthermore, since we are interested in evaluating the ego action-conditioned world model, we extend the standard WOSAC evaluation domain to include agents that are causal to the ego vehicle. Our evaluations reveal a significant number of scenarios where top-ranking models perform well under no perturbation but fail when the ego agent is forced to replay the original trajectory. To address these cases, we propose new metrics to highlight the sensitivity of world models to uncontrollable objects and evaluate the performance of world models as pseudo-environments for policy training and analyze some state-of-the-art world models under these new metrics.

Figures

Figures reproduced from arXiv: 2508.01922 by the authors.

Figure 1
Figure 1. This birds-eye-view image of a WOMD scene represents a motivating [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We plot the per-scenario forced replay versus original prediction metametric and minADE for the TrafficBots V1.5 world model. A majority of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. TrafficBots V1.5 per-scenario metametric on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Distribution of ∆M over 4393 evaluation scenarios. Most cases have ∆M ≈ 0, but the large tail of the distribution for scenarios with |∆M| > 0.05 indicates potential for high sensitivity in the world model traffic simulation. Following the same procedure for evaluation,…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

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

    Z. Huang, H. Liu, and C. Lv, “Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving,” 2023. [Online]. Available: https://arxiv.org/abs/2303.05760

    arXiv 2023

  2. [2]

    Parting with misconceptions about learning-based vehicle motion planning,

    D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta, “Parting with misconceptions about learning-based vehicle motion planning,” in 7th Annual Conference on Robot Learning , 2023. [Online]. Available: https://openreview.net/forum?id=o82EXEK5hu6

    work page 2023

  3. [3]

    Imitation with Spatial-Temporal Heatmap: 2nd Place Solution for NuPlan Challenge

    Y . Hu, K. Li, P. Liang, J. Qian, Z. Yang, H. Zhang, W. Shao, Z. Ding, W. Xu, and Q. Liu, “Imitation with spatial-temporal heatmap: 2nd place solution for nuplan challenge,” 2023. [Online]. Available: https://arxiv.org/abs/2306.15700

    work page Pith review arXiv 2023

  4. [4]

    Trafficbots: Towards world models for autonomous driving simulation and motion prediction,

    Z. Zhang, A. Liniger, D. Dai, F. Yu, and L. Van Gool, “Trafficbots: Towards world models for autonomous driving simulation and motion prediction,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 1522–1529

    work page 2023

  5. [5]

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

    W. Wu, X. Feng, Z. Gao, and Y . Kan, “Smart: Scalable multi-agent real-time simulation via next-token prediction,” 2024. [Online]. Available: https://arxiv.org/abs/2405.15677

    arXiv 2024

  6. [6]

    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. 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 2021 IEEE/CVF International Conference on Computer Vision (I...

    work page 2021

  7. [7]

    Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles,

    K. T. e. a. H. Caesar, J. Kabzan, “Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles,” in CVPR ADP3 work- shop, 2021

    work page 2021

  8. [8]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An open urban driving simulator,” in Proceedings of the 1st Annual Conference on Robot Learning , 2017, pp. 1–16

    2017

Show all 33 references
  1. [9]

    Smarts: Scalable multi-agent reinforcement learning training school for autonomous driving,

    M. Zhou, J. Luo, J. Villella, Y . Yang, D. Rusu, J. Miao, W. Zhang, M. Alban, I. Fadakar, Z. Chen, A. C. Huang, Y . Wen, K. Hassanzadeh, D. Graves, D. Chen, Z. Zhu, N. Nguyen, M. Elsayed, K. Shao, S. Ahilan, B. Zhang, J. Wu, Z. Fu, K. Rezaee, P. Yadmellat, M. Rohani, N. P. Nie...

    2020 arXiv

  2. [10]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,

    Q. Li, Z. Peng, L. Feng, Q. Zhang, Z. Xue, and B. Zhou, “Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,” IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 2022

    2022

  3. [11]

    Sim-to-real transfer in deep reinforcement learning for robotics: a survey,

    W. Zhao, J. P. Queralta, and T. Westerlund, “Sim-to-real transfer in deep reinforcement learning for robotics: a survey,” in 2020 IEEE Symposium Series on Computational Intelligence (SSCI) , 2020, pp. 737–744

    2020

  4. [12]

    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. [Online]. Available: https://arxiv .org/abs/2407.02797

    2024 arXiv

  5. [13]

    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, ...

    2023

  6. [14]

    Learning interactive driving policies via data-driven sim- ulation,

    T.-H. Wang, A. Amini, W. Schwarting, I. Gilitschenski, S. Karaman, and D. Rus, “Learning interactive driving policies via data-driven sim- ulation,” in 2022 International Conference on Robotics and Automation (ICRA), 2022, pp. 7745–7752

    2022

  7. [15]

    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. [Online]. Available: https://arxiv.org/abs/2407.12940

    2024 arXiv

  8. [16]

    Vectorized representation dreamer (vrd): Dreaming-assisted multi- agent motion forecasting,

    H. Schofield, H. Mirkhani, M. Elmahgiubi, K. Rezaee, and J. Shan, “Vectorized representation dreamer (vrd): Dreaming-assisted multi- agent motion forecasting,” in 2024 IEEE Intelligent Vehicles Sympo- sium (IV), 2024, pp. 2012–2017

    2024

  9. [17]

    Model- based imitation learning for urban driving,

    A. Hu, G. Corrado, N. Griffiths, Z. Murez, C. Gurau, H. Yeo, A. Kendall, R. Cipolla, and J. Shotton, “Model- based imitation learning for urban driving,” in Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, E...

    2022

  10. [18]

    Dream to control: Learning behaviors by latent imagination,

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to control: Learning behaviors by latent imagination,” in International Conference on Learning Representations , 2020. [Online]. Available: https://openreview.net/forum?id=S1lOTC4tDS

    2020

  11. [19]

    Mastering atari with discrete world models,

    D. Hafner, T. P. Lillicrap, M. Norouzi, and J. Ba, “Mastering atari with discrete world models,” in International Conference on Learning Representations , 2021. [Online]. Available: https: //openreview.net/forum?id=0oabwyZbOu

    2021

  12. [20]

    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, C. Li, and D. Anguelov, “TNT: Target- driven Trajectory Prediction,” in Proceedings of the 2020 Conference on Robot Learning , ser. Proceedings of Machine Learning Research, vol....

    2020

  13. [21]

    Simulation-based reinforcement learning for real-world autonomous driving,

    B. Osinski, A. Jakubowski, P. Ziecina, P. Milos, C. Galias, S. Homo- ceanu, and H. Michalewski, “Simulation-based reinforcement learning for real-world autonomous driving,” in 2020 IEEE International Con- ference on Robotics and Automation (ICRA) , 2020, pp. 6411–6418

    2020

  14. [22]

    A deep q-network reinforce- ment learning-based model for autonomous driving,

    M. Ahmed, C. P. Lim, and S. Nahavandi, “A deep q-network reinforce- ment learning-based model for autonomous driving,” in 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , 2021, pp. 739–744

    2021

  15. [23]

    Imitation Is Not Enough: Robustifying Imitation with Reinforcement Learning for Challenging Driving Scenarios,

    Y . Lu, J. Fu, G. Tucker, X. Pan, E. Bronstein, R. Roelofs, B. Sapp, B. White, A. Faust, S. Whiteson, D. Anguelov, and S. Levine, “Imitation Is Not Enough: Robustifying Imitation with Reinforcement Learning for Challenging Driving Scenarios,” in 2023 IEEE/RSJ Inter- national C...

    2023

  16. [24]

    Reinforcement Learning based Control of Imitative Policies for Near-Accident Driving,

    Z. Cao, E. Biyik, W. Wang, A. Raventos, A. Gaidon, G. Rosman, and D. Sadigh, “Reinforcement Learning based Control of Imitative Policies for Near-Accident Driving,” in Proceedings of Robotics: Science and Systems , Corvalis, Oregon, USA, July 2020

    2020

  17. [25]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. J. Gordon, and J. A. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” 2011. [Online]. Available: https://arxiv.org/abs/1011.0686

    2011 arXiv

  18. [26]

    Toward causal representation learning,

    B. Sch ¨olkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y . Bengio, “Toward causal representation learning,” Proceedings of the IEEE , vol. 109, no. 5, pp. 612–634, 2021

    2021

  19. [27]

    Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning,

    V . Ramanishka, Y .-T. Chen, T. Misu, and K. Saenko, “Toward driving scene understanding: A dataset for learning driver behavior and causal reasoning,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7699–7707

    2018

  20. [28]

    Learning causal representations for robust domain adaptation,

    S. Yang, K. Yu, F. Cao, L. Liu, H. Wang, and J. Li, “Learning causal representations for robust domain adaptation,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 3, pp. 2750–2764, 2023

    2023

  21. [29]

    Learning causal semantic representation for out-of-distribution prediction,

    C. Liu, X. Sun, J. Wang, H. Tang, T. Li, T. Qin, W. Chen, and T.-Y . Liu, “Learning causal semantic representation for out-of-distribution prediction,” in Advances in Neural Information Processing Systems , M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, ...

    2021

  22. [30]

    Generative causal representation learning for out-of-distribution motion forecasting,

    S. S. G. Bagi, Z. Gharaee, O. Schulte, and M. Crowley, “Generative causal representation learning for out-of-distribution motion forecasting,” 2023. [Online]. Available: https://arxiv .org/abs/ 2302.08635

    2023 arXiv

  23. [31]

    Sim- to-real causal transfer: A metric learning approach to causally- aware interaction representations,

    Y . Liu, A. Rahimi, P.-C. Luan, F. Raji ˇc, and A. Alahi, “Sim- to-real causal transfer: A metric learning approach to causally- aware interaction representations,” 2023. [Online]. Available: https: //arxiv.org/abs/2312.04540

    2023 arXiv

  24. [32]

    Causalagents: A robustness benchmark for motion forecasting,

    L. Sun, R. Roelofs, B. Caine, K. S. Refaat, B. Sapp, S. Ettinger, and W. Chai, “Causalagents: A robustness benchmark for motion forecasting,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 6820–6827

    2024

  25. [6170]

    Available: https://proceedings .neurips.cc/paper files/ paper/2021/file/310614fca8fb8e5491295336298c340f-Paper.pdf

    [Online]. Available: https://proceedings .neurips.cc/paper files/ paper/2021/file/310614fca8fb8e5491295336298c340f-Paper.pdf

    2021

Pith tools

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