Autoregressive Meta-Actions for Unified Controllable Trajectory Generation

Bin Li; Hangning Zhou; Jianbo Zhao; Lei Liu; Mu Yang; Qibin Zhou; Taiyu Ban; Xiyang Wang; Zhihao Liu

REVIEW 2 major objections 7 minor 14 references

Autoregressive Meta-Actions for Unified Controllable Trajectory Generation

T0 review · 2 major / 7 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 Frame-level meta-actions fix a misalignment in controllable trajectory generation, boosting decision-following mAP from 0.635 to 0.718 on a curated autonomous-driving benchmark.

desk verdict The frame-level meta-action formulation is a genuinely useful reframing, but the headline mAP gain is not yet trustworthy because labels, benchmark, and metric all come from the same hand-tuned rules. read the letter →

arxiv 2505.23612 v1 pith:KLR6USZM submitted 2025-05-29 cs.RO cs.CV

Jianbo Zhao , Taiyu Ban , Xiyang Wang , Qibin Zhou , Hangning Zhou , Zhihao Liu , Mu Yang , Lei Liu

show 1 more author

Bin Li

This is my paper

classification cs.ROcs.CV

keywords autonomousdrivingtrajectorygenerationmeta-actionsautoregressivepredictionframe-levelconditioningcontrollableWaymoMotionDatasettaskunification

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

This paper claims that long-interval meta-action conditioning is fundamentally misaligned: when a meta-action is assigned to a fixed multi-frame interval, some predicted trajectory segments cover a transition between two different meta-actions, so the model learns an inconsistent mapping from one decision to a stretch of trajectory. The proposed fix is to predict a meta-action at every frame and generate the next frame's state conditioned on that frame's meta-action, removing the semantic ambiguity at decision boundaries. If correct, this unified formulation improves decision-following mAP from 0.635 to 0.718 on a curated Waymo subset while keeping overall trajectory quality unchanged.

What carries the argument

The central object is the frame-level meta-action, a discrete driving-decision label (stationary, keep lane, lane change, turn, U-turn) assigned to each trajectory frame rather than to a fixed multi-frame interval. The mechanism carrying the argument is the autoregressive factorization of Equation (5c): at each step $\tau$ predict the next meta-action $\hat{c}_\tau$ from history, then generate the next state $\hat{s}_{\tau+1}$ conditioned on that newly predicted meta-action, with injected meta-action embeddings placed strictly between environment interaction and history aggregation to preserve causality across agents. The frame-level labels are extracted from trajectories by hand-tuned kinematic and geometric rules.

What would settle it

Train the same model with frame-level labels generated by perturbed thresholds near decision boundaries (e.g., shifting $d_{\min}$ from 1.75 m to 1.0 m or 2.5 m, or varying the curvature sign window) and check whether the decision-following mAP gap between long-interval and frame-level conditioning persists; alternatively, on a hand-annotated subset of 100 scenes containing lane-change transitions, compare human labels against the heuristic rule labels at frames within one second of the lane-change boundary and measure whether disagreement predicts model mispredictions.

Watch

Extended reading notes

Core claim

The paper's central claim is that the established practice of conditioning trajectory generation on a single meta-action assigned over a long future horizon breaks temporal semantic consistency, because a horizon longer than one frame can always contain a meta-action transition, so the learned mapping from one meta-action to a trajectory segment has inconsistent meaning across samples (Example 1, Proposition 3.1). The proposed formulation, expressed in Equation (5c), predicts the next meta-action $\hat{c}_\tau$ at every frame and then generates the next state $\hat{s}_{\tau+1}$ conditioned on that meta-action, jointly modeling $P(\hat{c}_\tau \mid \text{history})$ and $P(\hat{s}_{\tau+1} \mid \text{history}, \hat{c}_\tau)$. Proposition 3.2 asserts that this frame-level autoregressive formulation is unified with respect to meta-action prediction and meta-action-conditioned trajectory generation, satisfying both structural and semantic consistency at every time step. The paper further claims this unification improves decision-following (mAP 0.718 vs 0.635 for long-interval autoregressive conditioning and 0.567 for regression-based conditioning) without degrading generation quality, and that a staged pretraining procedure (frozen foundation model plus fine-tuned meta-action prediction and injection modules) delivers comparable controllability at roughly 2.25x lower training time.

Load-bearing premise

The benefit of frame-level meta-actions rests on the quality of the heuristic labels (thresholds like $\epsilon_v = 0.3$ m/s, $\epsilon_s = 0.1$ m, $d_{\min} = 1.75$ m, $\kappa$ thresholds, $\theta_{\min} = 15^\circ$) that assign a meta-action to every frame; if these labels are noisy near decision boundaries, the reported alignment gain is partly an artifact of the labeling rules rather than of the formulation itself, a limitation the paper explicitly acknowledges.

Editorial extensions

If this is right

A controllable trajectory generator built on long-interval meta-actions cannot be repaired by better network capacity or more data, because the task itself maps one decision to trajectory segments containing multiple decisions; the fix is to change the task formulation to frame-level conditioning.
Predicting the meta-action at every frame lets the model autonomously switch decisions at transition boundaries, as shown in rollouts where an injected left-lane-change meta-action is followed by a return to keep-lane.
Because the foundation model stays frozen and only meta-action prediction and injection modules are fine-tuned, adding new decision types to a deployed trajectory generator becomes a modular update that does not require retraining the whole model.
The same objective used for trajectory generation (cross-entropy on kinematic control actions) can be repurposed for meta-action-conditioned generation without degrading the Waymo SimAgents metrics.
Jointly predicting meta-actions as part of the model doubles as an interpretability output: generated trajectories come with per-frame high-level intent explanations.

Reading between the lines

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

The unification argument is not limited to driving: any hierarchical control scheme that conditions low-level generation on a high-level command over a fixed interval faces the same transition-boundary ambiguity, so the frame-level decomposition could transfer to robot manipulation or multi-agent simulation tasks with discrete intentions.
A testable consequence the paper does not state: if the label-generation rules (thresholds such as $d_{\min} = 1.75$ m and curvature sign checks) are perturbed near decision boundaries, the mAP gap between long-interval and frame-level conditioning should shrink or grow systematically, which would reveal how much of the benefit comes from the formulation versus the specific labeling procedure.
The paper's restriction to a predefined discrete meta-action set suggests a natural extension where meta-actions are learned rather than hand-defined; such an extension would test whether the alignment benefit survives when the semantic labels are not derived from the same kinematic cues the model already sees.
Because historical meta-action conditioning shows a 12.5% mAP drop when removed specifically for right lane changes, one could build a lightweight decision-boundary detector that runs on historical meta-action sequences alone, without generating full trajectories, as a cheap safety monitor.

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.

Desk Editor's Note

read the letter

The quick take: the core idea here is clean and worth taking seriously. Replacing fixed-horizon meta-action conditioning with frame-level autoregressive meta-action prediction and injection is a real reframing, and the staged fine-tuning scheme is a pragmatic engineering contribution. The paper is also honest about its limitations. But the headline empirical claim—mAP jumping from 0.635 to 0.718—is not yet convincing, because the labels, the benchmark, and the metric are all generated from the same hand-tuned rules in Appendix C. The stress-test note lands: the Keep-Lane rule is a disjunction that can label a vehicle mid-lane-change as Keep Lane, and the Lane-Change rule requires a curvature sign change that can miss asymmetric maneuvers. Since the evaluation metric checks the same constraints, a model that learns these threshold artifacts is rewarded as correct. This does not invalidate the formulation, but it means the claimed gain may be partly an artifact of the annotation scheme, and the paper does not quantify label noise or test alternative label definitions.

What is actually new: the decomposition of long-interval meta-actions into frame-level ones, Equation (5c), and the accompanying autoregressive prediction/injection architecture. That is not present in the cited CTRL-SIM or CARPlanner. The controlled comparison—same architecture, only conditioning granularity changes—is the right experiment to run, and the open dataset of frame-level labels is a genuine contribution. The limitations section explicitly acknowledges the heuristic label issue, which is to the authors' credit.

Where the soft spots are, in proportion: the evaluation is self-curated (661 scenes, manual verification), self-defined metrics, no error bars: that is a major limitation for the empirical claim. The internal inconsistency in the headline mAP (0.716 in text vs 0.718 in Table 1) is a minor but telling sloppiness. The propositions are more re-statements of definitions than deep theorems, so the theoretical novelty is light. The claim in Section 4.2 that injecting meta-actions before history aggregation is the "only placement" is overconfident without an exhaustive argument. None of these are fatal to the formulation, but together they mean the paper's evidence base is weaker than its framing suggests.

Who this is for: people working on controllable trajectory generation or motion prediction for autonomous driving. They will get value from the task formulation and the dataset. The paper deserves a serious referee, but the review should focus on evaluation validity: independent labels, standard benchmarks, error bars, and a check on whether the mAP gain persists. I would send it to peer review, but the current experiments are not enough to certify the claimed advantage.

Referee Report

2 major / 7 minor

Summary. The paper proposes an autoregressive frame-level meta-action formulation for controllable trajectory generation. It argues that conventional long-interval meta-action conditioning is temporally misaligned when a prediction horizon overlaps a behavior transition, and it reformulates the task as joint autoregressive prediction of a frame-level meta-action and the next state at every step (Eq. 5c). The proposed model freezes a KiGRAS-style kinematic foundation model and fine-tunes two additional modules for meta-action prediction and injection. On a self-curated subset of 661 Waymo validation scenes with the authors' own frame-level labels, the method achieves decision-following mAP of 0.718 (reported elsewhere as 0.716), versus 0.635 for the long-interval autoregressive baseline and 0.567 for a regression-based baseline. The paper also contributes a heuristic frame-level label-generation method, a dataset, and qualitative case studies.

Significance. If the empirical results hold up, the frame-level autoregressive formulation is a clean conceptual contribution that unifies meta-action prediction and meta-action-conditioned trajectory generation, and the modular staged fine-tuning is a practical asset. The controlled comparison (same architecture, only conditioning granularity varied) is a strength, as is the decision to release the labeled dataset and qualitative rollouts. However, the main empirical claim currently rests on labels and metrics built from the same hand-tuned rules and on a single training run, so the evidence for the headline improvement is not yet robust.

major comments (2)

[5.1, Appendix C.1, D.1-D.2] The headline decision-following comparison is confounded by the fact that both the frame-level meta-action labels and the evaluation constraints are generated from the same hand-tuned rules. In Eq. (31), the Keep Lane condition is a disjunction that includes |κ_t|<ε_κ, so a vehicle at the low-curvature midpoint of a lane change can be labeled Keep Lane despite a large lateral offset; Eq. (32) further requires a curvature sign change within the window, so asymmetric lane changes may never receive a Lane Change label. The evaluation approach in Appendix D.2 applies behavioral constraints derived from these same rules, meaning a model can score well by reproducing the labeler's thresholds rather than by genuinely following the imposed decision. The Limitations section acknowledges label noise near decision boundaries, but no quantification, no threshold-sensitivity analysis, and no validation against independent labels (e.g., human annotation) is provided. A robustness study over the thresholds ε_v, ε_s, d_min, κ_min, κ_max, θ_min, α, or an independent-label check on a subset, is needed before the 0.083 mAP gain over the long-interval autoregressive baseline can be attributed to the formulation.
[5.1] All decision-following results are reported without any measure of variance: no multiple seeds, no confidence intervals, and no statistical tests. With a curated set of 661 scenes, the reported mAP gain from 0.635 to 0.718 could be within training or subset-selection noise. Please report mean±std over at least three seeds and state whether the paired differences are statistically significant.

minor comments (7)

[5.1, Table 1] The text reports the proposed method's mAP as 0.716, while Table 1 and the abstract report 0.718; please reconcile this discrepancy.
[4.2] The claim that injecting meta-action embeddings at the T_env level is "the only placement that preserves both semantic consistency and sufficiency" is not supported by ablation experiments, since no alternative injection points are compared.
[3.1] The "Necessity of Frame-Level Meta-Actions" argument should be qualified: a prediction horizon T>1 can contain multiple meta-actions only if a behavior transition occurs within it; for transition-free trajectories, long-interval conditioning is not misaligned.
[Appendix D.1] The manual verification procedure used to select the 661 scenes is not described; please provide the selection criteria or an inter-annotator agreement measure.
[4.2] There is a typo in the subsection heading "Causal Contitioning", which should read "Causal Conditioning".
[Appendix B.3] In the sentence introducing p_Q and p_K, "repetitively" should be "respectively".
[5.1, 5.2] The tables reporting mAP results are not formatted as proper tables in the text; the header and data rows run together, making them hard to read. Please use actual table environments.

Circularity Check

1 steps flagged · score 3.0 of 10

Formal autoregressive formulation is independent, but the decision-following evaluation is closed-loop with the heuristic label-generation thresholds.

other [Appendix D.2 (Metrics Design), with label rules in Appendix C.1 (eqs. 31–32)]
"Specifically, for a given meta-action, we sample N candidate trajectories. If all sampled trajectories satisfy the behavioral constraints associated with the target meta-action, the corresponding test scenario is counted as a true positive (TP). The recall is then computed as the proportion of TP scenarios among all evaluation samples."

The 'behavioral constraints' in the decision-following metric are the same hand-threshold rules used to generate the frame-level meta-action labels: Keep Lane is |κ_t|<0.015 or |d_y,t|<0.3 or Δl_t=0, and Lane Change requires |d_y,t|>1.75 m plus a curvature sign change (Appendix C.1). A trajectory is scored as decision-following exactly when it satisfies the labeler's threshold criteria, so the mAP comparison (0.635→0.718) rewards models that reproduce the annotation rules rather than an externally verified semantic ground truth. The paper concedes in Limitations that 'our meta-action labels are heuristically derived from trajectory data.' This is an evaluator closure, not a circularity in the Eq. (5c) derivation itself.

full rationale

The paper's formal contribution—the frame-level autoregressive factorization in Eq. (5c)—is a standard probability decomposition and is not circular: it does not assume the mAP gain it reports, and Proposition 3.2 merely unpacks the authors' own Definition 1. The incorporation of the KiGRAS control representation and directional RoPE are prior modules from the same group, but they are used as building blocks, not as evidence for the new unification claim. The main circularity concern is evaluative rather than derivational: the frame-level labels used for training (Appendix C.1) and the behavioral constraints used for the decision-following metric (Appendix D.2) are generated from the same hand-tuned kinematic/geometric thresholds. Thus the reported mAP improvement (0.635→0.718) measures how well a model reproduces the labeler's threshold rules, not an independently grounded semantic ground truth. The authors acknowledge the risk in Limitations. Since the central formulation is independent of this evaluation loop, the circularity score is modest (3).

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim of task unification rests on a standard autoregressive probability factorization and on the causal assumption that decisions precede and determine motion. It does not depend on fitted constants for the formulation itself, but the empirical evaluation depends on roughly a dozen hand-chosen thresholds in the label-generation rules and on the authors' own prior components (KiGRAS control representation and directional RoPE).

free parameters (6)

Stationary velocity threshold epsilon_v = 0.3 m/s
Hand-chosen in Appendix C.1 Equation (30) to define stationary behavior; directly affects label quality.
Stationary displacement threshold epsilon_s = 0.1 m
Hand-chosen in Equation (30); another label-generation threshold.
Lane-change minimum lateral offset d_min = 1.75 m
Hand-chosen in Equation (32); triggers lane-change labels in the curated dataset.
Turn curvature bounds kappa_min, kappa_max = 0.015 and 0.25 per meter
Hand-chosen in Equations (33)-(35); define the turn and U-turn label regimes.
Turn heading threshold theta_min = 15 degrees
Hand-chosen in Equation (33) for turn detection.
U-turn heading window alpha = 15 degrees
Hand-chosen in Equation (34) for U-turn detection.

assumptions (4)

standard math The CTRA kinematic model governs state transitions, x_{t+1}=K(x_t, a_t), and is used to derive discrete control labels.
Invoked in Section 4.1 and Appendix B.4; standard kinematic assumption inherited from prior work.
domain assumption High-level decisions determine motion outcomes, justifying the factorization P(s|c)P(c) in Equation (5c).
Stated in Section 3.2; if decisions do not causally precede motion, the ordering of the two factors is not justified.
domain assumption Meta-actions form a small predefined discrete set, and every trajectory frame has exactly one active meta-action.
Used throughout Sections 3 and 4; the paper concedes in Limitations that a fixed discrete vocabulary may miss nuanced intentions.
ad hoc to paper For any horizon T>1 there exists a prediction window spanning at least two meta-actions, making frame-level representation necessary.
Example 1 in Section 3.1 assumes a meta-action transition inside the horizon; constant-behavior windows violate the claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Autoregressive Meta-Actions for Unified Controllable Trajectory Generation." pith.science (2026). https://pith.science/paper/KLR6USZM

@misc{pith2026250523612,
  author       = {Pith},
  title        = {Pith review of: Autoregressive Meta-Actions for Unified Controllable Trajectory Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KLR6USZM}},
  note         = {Machine review of arXiv:2505.23612}
}

read the original abstract

Controllable trajectory generation guided by high-level semantic decisions, termed meta-actions, is crucial for autonomous driving systems. A significant limitation of existing frameworks is their reliance on invariant meta-actions assigned over fixed future time intervals, causing temporal misalignment with the actual behavior trajectories. This misalignment leads to irrelevant associations between the prescribed meta-actions and the resulting trajectories, disrupting task coherence and limiting model performance. To address this challenge, we introduce Autoregressive Meta-Actions, an approach integrated into autoregressive trajectory generation frameworks that provides a unified and precise definition for meta-action-conditioned trajectory prediction. Specifically, We decompose traditional long-interval meta-actions into frame-level meta-actions, enabling a sequential interplay between autoregressive meta-action prediction and meta-action-conditioned trajectory generation. This decomposition ensures strict alignment between each trajectory segment and its corresponding meta-action, achieving a consistent and unified task formulation across the entire trajectory span and significantly reducing complexity. Moreover, we propose a staged pre-training process to decouple the learning of basic motion dynamics from the integration of high-level decision control, which offers flexibility, stability, and modularity. Experimental results validate our framework's effectiveness, demonstrating improved trajectory adaptivity and responsiveness to dynamic decision-making scenarios. We provide the video document and dataset, which are available at https://arma-traj.github.io/.

Figures

Figures reproduced from arXiv: 2505.23612 by the authors.

**Figure 2.** The overall diagram of the proposed model architecture. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Visual comparison of various models for following decision "Left Lane Change". [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Frame-by-frame visualization of predicted meta-actions and resulting ego trajectories. [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]

**Figure 5.** Figure 5: Illustration of the temporal stages of a lane change maneuver. The process is divided into [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]

**Figure 6.** Figure 6: Inference result without manually injected meta-actions. The ego vehicle performs standard [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

**Figure 7.** Figure 7: Inference result with a manually injected Left Lane Change meta-action at the current [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

**Figure 8.** Figure 8: Sampled rollout in a U-turn scenario: the ego vehicle first performs a Left U-turn (LU), [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]

**Figure 9.** Figure 9: Alternative sample from the same U-turn scenario: after completing the Left U-turn (LU), [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages

[1]

Gpt-4 technical report

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

arXiv
[4]

Revisit mixture models for multi-agent simulation: Experimental study within a unified framework

Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework. arXiv preprint arXiv:2501.17015,

arXiv
[8]

Demo: Decoupling motion forecasting into directional intentions and dynamic states

Bozhou Zhang, Nan Song, and Li Zhang. Demo: Decoupling motion forecasting into directional intentions and dynamic states. InNeurIPS, 2024a. Diankun Zhang, Guoan Wang, Runwen Zhu, Jianbo Zhao, Xiwu Chen, Siyu Zhang, Jiahao Gong, Qibin Zhou, Wenyuan Zhang, Ningzi Wang, et al. Sparsead: Sparse query-centric paradigm for efficient end-to-end autonomous drivin...

arXiv
[9]

Drope: Directional rotary position embedding for efficient agent interaction modeling.arXiv preprint arXiv:2503.15029,

Jianbo Zhao, Taiyu Ban, Zhihao Liu, Hangning Zhou, Xiyang Wang, Qibin Zhou, Hailong Qin, Mu Yang, Lei Liu, and Bin Li. Drope: Directional rotary position embedding for efficient agent interaction modeling.arXiv preprint arXiv:2503.15029,

arXiv
[11]

14 A Related Work In this section, we first review the mainstream scene representation approaches used in trajectory generation in Section A.1. Then, in Section A.2, we introduce two predominant modeling paradigms: generating long-horizon trajectories in a single step versus generating them incrementally in an au- toregressive manner. Finally, in Section ...

work page 2023
[12]

This significantly reduces the memory complexity by a factor of N, the number of agents in the scene, thereby accelerating training while maintaining strong performance

extended this direction by introducing rotary position embedding (RoPE) to the domain, proposing a directional variant to capture heading relationships. This significantly reduces the memory complexity by a factor of N, the number of agents in the scene, thereby accelerating training while maintaining strong performance. Motivated by its efficiency, we ad...

work page 2023
[13]

This adjustment ensures that the embedding retains the inherent2π-periodicity of directional angles

The other operations remain the same as RoPE. This adjustment ensures that the embedding retains the inherent2π-periodicity of directional angles. We adopt a head-by-head design to integrate relative spatial coordinates and heading directions in a mixed manner, following the approach introduced by Zhao et al. [2025, Section IV-A]. B.4 Kinematic Control Ac...

work page 2025
[64]

Both the action prediction module and the meta-action prediction module consist of three transformer layers for temporal aggregation

The environment fusion module is repeated three times to sufficiently capture scene-level interactions. Both the action prediction module and the meta-action prediction module consist of three transformer layers for temporal aggregation. We train the model using the Adan optimizer [Xie et al., 2024] with a fixed learning rate of 1e-2 and no learning rate ...

work page 2024

Show all 14 references

[2017]

Omnidrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning.arXiv preprint arXiv:2504.04348,

Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning.arXiv preprint arXiv:2504.04348,

arXiv
[2019]

Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang

Arxiv preprint: 1910.05449. Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024a. Zhili Chen, Maosheng Y...

1910 arXiv
[2020]

Large trajectory models are scalable motion predictors and planners.arXiv preprint arXiv:2310.19620,

Qiao Sun, Shiduo Zhang, Danjiao Ma, Jingzhe Shi, Derun Li, Simian Luo, Yu Wang, Ningyi Xu, Guangzhi Cao, and Hang Zhao. Large trajectory models are scalable motion predictors and planners.arXiv preprint arXiv:2310.19620,

arXiv
[2021]

Ctrl-sim: Reactive and controllable driving agents with offline reinforcement learning.arXiv preprint arXiv:2403.19918,

Luke Rowe, Roger Girgis, Anthony Gosselin, Bruno Carrez, Florian Golemo, Felix Heide, Liam Paull, and Christopher Pal. Ctrl-sim: Reactive and controllable driving agents with offline reinforcement learning.arXiv preprint arXiv:2403.19918,

arXiv
[2024]

Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning.arXiv preprint arXiv:2410.05582, 2024a

Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning.arXiv preprint arXiv:2410.05582, 2024a. Zhiyu H...

arXiv
[2025]

Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang

URLhttps://openreview.net/forum?id=wM2sfVgMDH. Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863–17873, 2023a. Zikang Zhou, Zihao Wen,...

arXiv

Pith tools

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

[1] [1]

Gpt-4 technical report

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

arXiv

[2] [4]

Revisit mixture models for multi-agent simulation: Experimental study within a unified framework

Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework. arXiv preprint arXiv:2501.17015,

arXiv

[3] [8]

Demo: Decoupling motion forecasting into directional intentions and dynamic states

Bozhou Zhang, Nan Song, and Li Zhang. Demo: Decoupling motion forecasting into directional intentions and dynamic states. InNeurIPS, 2024a. Diankun Zhang, Guoan Wang, Runwen Zhu, Jianbo Zhao, Xiwu Chen, Siyu Zhang, Jiahao Gong, Qibin Zhou, Wenyuan Zhang, Ningzi Wang, et al. Sparsead: Sparse query-centric paradigm for efficient end-to-end autonomous drivin...

arXiv

[4] [9]

Drope: Directional rotary position embedding for efficient agent interaction modeling.arXiv preprint arXiv:2503.15029,

Jianbo Zhao, Taiyu Ban, Zhihao Liu, Hangning Zhou, Xiyang Wang, Qibin Zhou, Hailong Qin, Mu Yang, Lei Liu, and Bin Li. Drope: Directional rotary position embedding for efficient agent interaction modeling.arXiv preprint arXiv:2503.15029,

arXiv

[5] [11]

14 A Related Work In this section, we first review the mainstream scene representation approaches used in trajectory generation in Section A.1. Then, in Section A.2, we introduce two predominant modeling paradigms: generating long-horizon trajectories in a single step versus generating them incrementally in an au- toregressive manner. Finally, in Section ...

work page 2023

[6] [12]

This significantly reduces the memory complexity by a factor of N, the number of agents in the scene, thereby accelerating training while maintaining strong performance

extended this direction by introducing rotary position embedding (RoPE) to the domain, proposing a directional variant to capture heading relationships. This significantly reduces the memory complexity by a factor of N, the number of agents in the scene, thereby accelerating training while maintaining strong performance. Motivated by its efficiency, we ad...

work page 2023

[7] [13]

This adjustment ensures that the embedding retains the inherent2π-periodicity of directional angles

The other operations remain the same as RoPE. This adjustment ensures that the embedding retains the inherent2π-periodicity of directional angles. We adopt a head-by-head design to integrate relative spatial coordinates and heading directions in a mixed manner, following the approach introduced by Zhao et al. [2025, Section IV-A]. B.4 Kinematic Control Ac...

work page 2025

[8] [64]

Both the action prediction module and the meta-action prediction module consist of three transformer layers for temporal aggregation

The environment fusion module is repeated three times to sufficiently capture scene-level interactions. Both the action prediction module and the meta-action prediction module consist of three transformer layers for temporal aggregation. We train the model using the Adan optimizer [Xie et al., 2024] with a fixed learning rate of 1e-2 and no learning rate ...

work page 2024