Pith. sign in

REVIEW 3 major objections 47 references

DRIFT generates executable mixed-autonomy traffic by conditioning diffusion on vehicle type and AV share, then filtering, scoring, and reweighting candidates inside a closed-loop simulator.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 13:45 UTC pith:GFAHQG4R

load-bearing objection Competent closed-loop systems paper: real integration of penetration-aware conditioning, executable diffusion candidates, and long-tail feedback—trade-off claims hold inside Flow/SUMO, not beyond it. the 3 major comments →

arxiv 2606.16589 v2 pith:GFAHQG4R submitted 2026-06-15 cs.DC

DRIFT: Risk-Constrained Diffusion with Imitation Priors for Mixed-Autonomy Traffic Generation

Yaoshen Yu , Minghui Liwang , Wenbo Zhu , Xinlei Yi , Zhang Liu , Yiguang Hong , Xianbin Wang , Seyyedali Hosseinalipour This is my paper
classification cs.DC
keywords mixed-autonomy trafficdiffusion modelsimitation learningclosed-loop evaluationlong-tail safetyexecutable trajectory generationAV penetration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

As human-driven and autonomous vehicles share the road, traffic models must track how behavior changes with AV share, produce controls that other vehicles can actually live with, and still catch rare near-collisions and deadlocks. DRIFT answers with one rolling-horizon pipeline: it encodes each vehicle's history, neighbors, road layout, and the current AV penetration into a condition vector; a diffusion model then samples several candidate control sequences that are rolled out and feasibility-checked; a scorer that mixes imitation realism, efficiency, safety margins, and dynamic cost picks one sequence to execute for a short window; rare high-risk candidates are reweighted so the generator learns from the long tail. In Flow/SUMO Ring, figure-eight, and merge scenarios across 0–100% AV penetration, this yields a competitive safety–efficiency balance and smoother penetration trends than several executable baselines, while ablations show that multi-candidate selection, penetration conditioning, and long-tail feedback each change the closed-loop outcome.

Core claim

The paper claims that mixed-autonomy traffic generation becomes useful for closed-loop evaluation only when heterogeneity-aware conditioning, diffusion of executable control candidates, feasibility filtering, multi-term online selection, and risk-aware long-tail reweighting are run as one receding-horizon loop—and that this combination delivers a strong safety–efficiency trade-off across stylized Flow/SUMO benchmarks at six AV penetration rates.

What carries the argument

DRIFT's three-module closed loop: Module A maps history, neighbors, map, control authority, and AV penetration into a condition vector; Module B diffusively samples K executable control sequences and rolls them out; Module C and the online selector score, filter, execute, and reweight candidates by realism, efficiency, risk, and feasibility.

Load-bearing premise

The method treats AV behavior as scaled synthetic versions of human driving parameters and tests only three stylized simulator networks; if those priors or scenes misrepresent real mixed traffic, the reported trade-off does not transfer.

What would settle it

Re-run the same Ring/F8/Merge closed-loop protocol with AV controllers drawn from real fleet logs (or a held-out naturalistic AV corpus) instead of the Table I scalings; if DRIFT's hard-braking and throughput advantage over the same baselines disappears, the central claim fails for realistic AV behavior.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper proposes DRIFT, a closed-loop mixed-autonomy traffic generation framework that unifies three modules: heterogeneity-aware condition encoding (Module A, including vehicle type and AV penetration), conditional diffusion over executable control sequences with feasibility-aware rollout (Module B), and progressive adversarial imitation with deterministic long-tail reweighting (Module C). Traffic evolution is cast as a receding-horizon observation–generation–selection–execution loop in Flow/SUMO, with a composite candidate score and soft rollout-level realism/tail-risk targets (P0 and subproblems P1–P3, Psel). Experiments on Ring, Figure-eight, and Merge across 0–100% AV penetration compare DRIFT to executable baselines (FollowerStopper, PI, IDM, Flow-RL, Flow-AIL) and ablations, reporting a competitive safety–efficiency–executability trade-off rather than uniform dominance, and attributing gains to multi-candidate generation, feasibility filtering, online selection, and long-tail feedback.

Significance. If the closed-loop results hold under the stated protocol, the paper offers a useful systems-level integration for mixed-autonomy traffic generation: control-space diffusion with explicit executability constraints, penetration-aware conditioning, and risk-aware candidate selection inside a single Flow/SUMO feedback loop. Strengths include a clear problem decomposition (P0 into P1–P3 and Psel), transparent soft treatment of CVaR-style tail risk via finite-sample reweighting rather than a false closed-form guarantee, multi-metric evaluation (return, speed/outflow, THW/TTC, hard braking, min acceleration), and ablations/selector/feasibility/stress diagnostics that make mechanism contributions checkable. The work is primarily empirical systems engineering rather than a new theoretical guarantee; its value is a reproducible closed-loop stack and evaluation protocol for executable mixed-autonomy generation.

major comments (3)
  1. The central claim (Abstract; §VI–VII; Table V) is a strong closed-loop safety–efficiency trade-off under mixed autonomy. That claim is measured only under synthetic AV priors obtained by scaling calibrated HV parameters (Eqs. 58–59, Table I) and three stylized Flow/SUMO networks (Ring, F8, Merge). Appendix C correctly flags this limitation, but the introduction and RQs frame long-term mixed-autonomy ecosystems with takeover, cooperation, and non-stationary composition. The manuscript should either (i) substantially narrow the claim language to “stylized Flow/SUMO benchmarks with synthetic AV priors,” or (ii) add at least one sensitivity study with qualitatively different AV priors (e.g., stronger cooperation, communication-aware CACC-like behavior, or takeover-induced control switches) and report whether the Table V joint pattern and Fig. 6 ablation deltas persist. Without that, transfer
  2. Baseline selection undercuts the “strong trade-off” framing. §VI-B and Table V emphasize Flow-native controllers and an adapted Flow-AIL interface, while excluding ACC/CACC/MPC and recent generation methods (SceneDiffuser, TrafficMCTS, DiffAIL, ControlTraj, DragTraffic) on interface grounds (Appendix D). That choice is defensible for protocol consistency, but then claims should be stated as “relative to the reported executable Flow baselines,” not as a general mixed-autonomy generation advance. Please either (a) add at least one stronger closed-loop learning/control baseline under the same interface, or (b) revise Abstract/§VII wording so that competitiveness is scoped to the reported baseline set, and move broader literature comparison fully into the positioning table rather than the main claim.
  3. P0’s soft tail objective (Eqs. 28–30) is implemented by deterministic candidate reweighting χ^{i,k}_t = 1+β(R^{i,k}_t+D^{i,k}_t) and L_tail (Eqs. 56–57; §V-C), not by optimizing CVaR over rollouts. The paper is careful in places, but Abstract and contribution bullets still speak of “risk-constrained” generation and “long-tail feedback” as if they close the P0 tail term. Please state explicitly in §IV–V and the Abstract that Δ_tail is a soft finite-sample surrogate, report the empirical distribution of R_tail (or hard-brake/collision tails) with vs. without reweighting, and avoid language that could be read as a formal risk constraint or CVaR guarantee.

Circularity Check

1 steps flagged

No load-bearing circular derivation: DRIFT is an empirical closed-loop systems paper whose safety–efficiency claims rest on external Flow/SUMO metrics, not on predictions forced by their own fits.

specific steps
  1. self definitional [§V-C Module C; Eqs. (54)–(55), (27), (60); Appendix F-A]
    "During inference, the trained discriminator is reused as the behavior-similarity scorer in (27). ... Si,k_t = σ(D_ϑ(τi,k_t, si_t)), where D_ϑ(·) denotes the discriminator scoring function"

    The same Module-C discriminator trained to separate expert vs generated trajectories under L_align is reused as the behavioral-realism term Si,k_t inside online selection. That dual use is mildly self-referential inside the pipeline. It is not load-bearing circularity for the paper’s main claim: Table V and Fig. 5–6 report external simulator efficiency/safety diagnostics, not discriminator scores, and the paper does not present Si,k_t as an independent prediction of those outcomes.

full rationale

DRIFT does not present a first-principles derivation whose outputs reduce to its inputs. The core claim (Abstract; §VI–VII; Table V) is an empirical closed-loop trade-off under Ring/F8/Merge and six AV penetrations, measured by simulator-external quantities—return, average speed, merge outflow, collision/deadlock, hard braking, min acceleration, THW/TTC violation rates—against executable baselines (FollowerStopper, PI, IDM, Flow-RL, Flow-AIL) and ablations. Training uses expert/reference trajectories and feasibility/risk losses; evaluation is not a re-report of those losses. Module C’s dual use of the discriminator (alignment loss then Si,k_t = σ(D_ϑ(·)) in Eq. 27/60) and the deterministic long-tail weights χi,k_t = 1+β(Ri,k_t+Di,k_t) (Eq. 56) are standard imitation/reweighting design choices; the paper explicitly treats them as finite-sample surrogates for Δreal/Δtail and CVaR (Eqs. 29–30, 33), not closed-form predictions of the evaluation metrics. Synthetic AV priors (Eqs. 58–59, Table I) and stylized scenarios are modeling assumptions that limit transfer (Appendix C), not circular reductions of a claimed derivation. No uniqueness theorem, self-citation chain, or fitted parameter renamed as prediction carries the central result. Score 1 only for the mild self-referential imitation loop, which is not load-bearing for Table V.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on simulator fidelity, synthetic AV priors, hand-chosen scoring/feasibility weights, and the assumption that candidate-level risk reweighting adequately proxies rollout-level tail risk. These are domain and design choices, not free-floating particles; the method invents a modular architecture rather than new physical entities. Exhaustive free parameters below are those that materially shape reported safety–efficiency numbers.

free parameters (7)
  • Candidate count K
    Default K=5; sensitivity (Table X) shows non-monotonic Merge behavior—headline results depend on this choice.
  • Module-C risk weight λ_risk
    Default 0.65 with sweep {0.35,0.65,0.95}; directly trades efficiency vs safety tail (Table XV).
  • Composite score weights (ω_sim, ω_eff, ω_risk, ω_dyn) and executable penalties
    Eq. (27) and Table VI heuristic/learned coefficients (e.g., w_risk, w_ttc, merge conflict penalty) are hand-set and scenario-guarded.
  • AV prior scaling ratios α_v, α_T, α_s0, α_a, α_b, α_τ, α_σ, ω_coop
    Table I scenario-specific scalings define synthetic AVs from HV stats; not fit to real AV fleets.
  • Safety thresholds δ_THW, δ_TTC, d_safe and hard-brake cutoffs
    Feasible set Ω and diagnostics use fixed thresholds (e.g., THW<1s, TTC<2s; Merge TTC guard 2.8s).
  • Long-tail gain β and adversarial warm-up E_warm
    χi,k_t = 1+β(Ri,k_t+Di,k_t) and λ_adv schedule control how strongly rare events reweight Module B.
  • Planning/execution horizons and history length
    H_plan/H_exec and L_h=6 with stride 5 fix the receding-horizon interface used in all closed-loop metrics.
axioms (5)
  • domain assumption Flow/SUMO closed-loop dynamics F(·) and vehicle transition f(·) are adequate proxies for real mixed-autonomy traffic evolution.
    All claims are evaluated only inside this simulator stack (§VI-A, Eq. 25).
  • domain assumption HV behavior can be summarized by calibrated IDM-like priors from highD/rounD/exiD/inD without frame-by-frame replay.
    §VI-A calibration of z^HV_s; drives both training samples and online HV agents.
  • ad hoc to paper Finite-sample adversarial alignment L_align plus deterministic risk reweighting adequately implements the soft Δ_real and CVaR-style Δ_tail objectives in P0.
    §IV-B and §V-C explicitly state CVaR is not solved in closed form; Module C is a surrogate.
  • domain assumption Local feasibility set Ω^i,feas_t (bounds, topology, clearance, THW/TTC) is a sufficient executability filter for closed-loop safety claims.
    Eq. (23) and Psel; fallback to SUMO fail-safe when empty.
  • domain assumption Standard DDPM forward/reverse process on vectorized controls yields useful multi-modal executable candidates under condition c^i_t.
    Module B, Eqs. (41)–(48), citing Ho et al. / Nichol & Dhariwal.
invented entities (2)
  • DRIFT three-module closed-loop stack (heterogeneity-aware encoder + control-space diffusion + progressive AIL with long-tail χ reweighting) no independent evidence
    purpose: Unify penetration-aware representation, executable multi-candidate generation, and risk-aware selection/feedback for mixed-autonomy traffic.
    Architectural invention of the paper; evaluated only within the authors’ Flow protocol, not independently measured outside this work.
  • Heterogeneous local state s^i_t and penetration embedding e_ρ(ρ_t) as the conditioning object for generation no independent evidence
    purpose: Encode HV/AV authority and global AV share so generation adapts across penetration regimes (RQ1).
    Modeling construct introduced in §III; usefulness shown only via ablations (w/o Pen.), not external theory.

pith-pipeline@v1.1.0-grok45 · 38019 in / 4109 out tokens · 48201 ms · 2026-07-12T13:45:53.331474+00:00 · methodology

0 comments
read the original abstract

Future intelligent transportation systems are envisioned to evolve toward a long-term mixed-autonomy paradigm, where human-driven vehicles (HVs) and autonomous vehicles (AVs) coexist within highly coupled traffic ecosystems. Such coexistence introduces pronounced heterogeneity, amplified uncertainty, and increasingly intricate interaction dynamics. In this context, it remains fundamentally challenging to simultaneously capture the heterogeneous behavioral distribution shifts arising from dynamic AV penetration, generate diverse yet executable trajectories under strong inter-vehicle coupling, and conduct reliable closed-loop safety and stability diagnostics for rare but high-impact events. To this end, we present Diffusion with Risk constraints, Imitation priors, and long-tail Feedback for mixed-autonomy Traffic generation (DRIFT), a mixed-autonomy traffic generation framework that unifies heterogeneity-aware conditional encoding, conditional diffusion-based executable trajectory generation, and progressive adversarial alignment enhanced by risk-aware long-tail feedback, thereby enabling traffic behaviors to be iteratively generated, filtered, selected, and validated within a closed-loop execution pipeline. In addition, a unified evaluation protocol is developed to jointly characterize safety, efficiency, and closed-loop stability across representative traffic scenarios and AV penetration regimes. Experimental results demonstrate that DRIFT achieves a strong safety-efficiency trade-off in closed-loop mixed-autonomy benchmarks, while further revealing the critical influence of candidate executability, online selection, and long-tail feedback on executable traffic evolution.

Figures

Figures reproduced from arXiv: 2606.16589 by Minghui Liwang, Seyyedali Hosseinalipour, Wenbo Zhu, Xianbin Wang, Xinlei Yi, Yaoshen Yu, Yiguang Hong, Zhang Liu.

Figure 1
Figure 1. Figure 1: Illustration of temporal-window definitions in this work. TSs are [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall workflow of DRIFT, separating offline training from online [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative benchmark scenarios. Blue vehicles denote AVs, gray [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop simulation protocol. Real trajectory data calibrate [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Efficiency trends across AV penetration rates. Rows report return, [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation and replacement-control deltas relative to full DRIFT. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hard-braking diagnostics. 0 20 40 60 80 100 AV pen. (%) −50 0 Min. accel. (m/s 2 ) Ring 0 20 40 60 80 100 AV pen. (%) −100 0 F8 0 20 40 60 80 100 AV pen. (%) −100 0 Merge FollowerStopper PI Flow-RL DRIFT [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Minimum realized acceleration diagnostics. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 2 linked inside Pith

  1. [1]

    Flow: A modular learning framework for mixed autonomy traffic,

    C. Wu, A. R. Kreidieh, K. Parvate, E. Vinitsky, and A. M. Bayen, “Flow: A modular learning framework for mixed autonomy traffic,”IEEE Trans. Robot., vol. 38, no. 2, pp. 1270–1286, Apr. 2022

    2022

  2. [2]

    SceneDiffuser: Efficient and controllable driving simulation initialization and rollout,

    C.-M. Jianget al., “SceneDiffuser: Efficient and controllable driving simulation initialization and rollout,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 37, 2024

    2024

  3. [3]

    TrafficMCTS: A closed-loop traffic flow generation framework with group-based monte carlo tree search,

    Z. Fu, L. Wen, P. Cai, D. Fu, S. Mao, and B. Shi, “TrafficMCTS: A closed-loop traffic flow generation framework with group-based monte carlo tree search,”IEEE Trans. Intell. Transp. Syst., vol. 26, no. 10, pp. 15 453–15 470, Oct. 2025

    2025

  4. [4]

    Transferring causal driving patterns for generalizable traffic simulation with diffusion-based distillation,

    Y . Chen, J. Sun, J. Fan, and J. Sun, “Transferring causal driving patterns for generalizable traffic simulation with diffusion-based distillation,” in Proc. AAAI Conf. Artif. Intell., vol. 40, 2026, pp. 110–118

    2026

  5. [5]

    Cooperative driving of connected autonomous vehicles in heterogeneous mixed traffic: A game theoretic approach,

    S. Fang, P. Hang, C. Wei, Y . Xing, and J. Sun, “Cooperative driving of connected autonomous vehicles in heterogeneous mixed traffic: A game theoretic approach,”IEEE Trans. Intell. Veh., 2024, early access

    2024

  6. [6]

    Hybrid system stability analysis of multilane mixed-autonomy traffic,

    S. Li, R. Dong, and C. Wu, “Hybrid system stability analysis of multilane mixed-autonomy traffic,”IEEE Trans. Robot., vol. 40, 2024

    2024

  7. [7]

    En- hancing safety in mixed traffic: Learning-based modeling and efficient control of autonomous and human-driven vehicles,

    J. Wang, Y . V . Pant, L. Zhao, M. Antkiewicz, and K. Czarnecki, “En- hancing safety in mixed traffic: Learning-based modeling and efficient control of autonomous and human-driven vehicles,”IEEE Trans. Intell. Transp. Syst., vol. 25, no. 9, pp. 12 784–12 799, Sep. 2024

    2024

  8. [8]

    Study on traffic flows with connected vehicles and human-driven vehicles,

    B.-L. Cen, Y . Xue, K. Zhang, L. Jia, and H.-D. He, “Study on traffic flows with connected vehicles and human-driven vehicles,”Appl. Math. Comput., vol. 490, p. 129182, 2025

    2025

  9. [9]

    Energy and environmental implications of automated vehicles under mixed autonomy traffic environment,

    Z. Wang, Q. Cheng, P. Liu, W. Yu, J. Wang, and Z. Liu, “Energy and environmental implications of automated vehicles under mixed autonomy traffic environment,”IEEE Trans. Intell. Veh., vol. 10, no. 2, Feb. 2025

    2025

  10. [10]

    Analysis of roadway capacity for heterogeneous traffic flows considering the degree of trust of drivers of HVs in CA Vs,

    Y . Pei, H. Wang, and X. Han, “Analysis of roadway capacity for heterogeneous traffic flows considering the degree of trust of drivers of HVs in CA Vs,”Physica A, vol. 639, p. 129693, 2024

    2024

  11. [11]

    Exploring the impact of conditionally automated driving vehicles transferring control to human drivers on the stability of heterogeneous traffic flow,

    C. Wang, W. Ren, C. Xu, N. Zheng, C. Peng, and H. Tong, “Exploring the impact of conditionally automated driving vehicles transferring control to human drivers on the stability of heterogeneous traffic flow,” IEEE Trans. Intell. Veh., vol. 10, no. 2, pp. 912–928, Feb. 2025

    2025

  12. [12]

    Learning to control and coordi- nate mixed traffic through robot vehicles at complex and unsignalized intersections,

    D. Wang, W. Li, L. Zhu, and J. Pan, “Learning to control and coordi- nate mixed traffic through robot vehicles at complex and unsignalized intersections,”Int. J. Robot. Res., 2025

    2025

  13. [13]

    Modeling and robustH inf tycontrol synthesis of the CA V-HDV heterogeneous traffic system with different car-following modes,

    Y . Gong and W.-X. Zhu, “Modeling and robustH inf tycontrol synthesis of the CA V-HDV heterogeneous traffic system with different car-following modes,”IEEE Trans. Intell. Transp. Syst., vol. 25, no. 10, Oct. 2024

    2024

  14. [14]

    Urban vehicle trajectory generation based on generative adversarial imitation learning,

    M. Wang, J. Cui, Y . W. Wong, Y . Chang, L. Wu, and J. Jin, “Urban vehicle trajectory generation based on generative adversarial imitation learning,”IEEE Trans. Veh. Technol., vol. 73, no. 12, pp. 18 237–18 249, Dec. 2024

    2024

  15. [15]

    Ctx2TrajGen: Traffic context-aware microscale vehicle trajectories us- ing generative adversarial imitation learning,

    J. Jin, Q. Sun, Y . Ding, Y . Chen, Y . Zhang, X. Wang, and M. Wang, “Ctx2TrajGen: Traffic context-aware microscale vehicle trajectories us- ing generative adversarial imitation learning,” 2025, arXiv:2507.17418. [Online]. Available: https://arxiv.org/abs/2507.17418

    Pith/arXiv arXiv 2025

  16. [16]

    DiffAIL: Diffusion adversarial imitation learning,

    B. Wang, G. Wu, T. Pang, Y . Zhang, and Y . Yin, “DiffAIL: Diffusion adversarial imitation learning,” inProc. AAAI Conf. Artif. Intell., vol. 38, no. 14, 2024, pp. 15 447–15 455

    2024

  17. [17]

    A fast and stable framework for generative adversarial imitation learning,

    F. Shahabi-Nejad and M. M. Ebadzadeh, “A fast and stable framework for generative adversarial imitation learning,”Eng. Appl. Artif. Intell., vol. 162, p. 112460, 2025

    2025

  18. [18]

    Gen- eralizable multi-modal adversarial imitation learning for non-stationary dynamics,

    Y .-C. Li, N. Chao, Z. Zhang, F. Zhang, L. Yuan, and Y . Yu, “Gen- eralizable multi-modal adversarial imitation learning for non-stationary dynamics,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 7, pp. 5600–5612, Jul. 2025

    2025

  19. [19]

    ControlTraj: Controllable trajectory generation with topology-constrained diffusion model,

    Y . Zhu, J. J. Q. Yu, X. Zhao, Q. Liu, Y . Ye, W. Chen, Z. Zhang, X. Wei, and Y . Liang, “ControlTraj: Controllable trajectory generation with topology-constrained diffusion model,” inProc. ACM SIGKDD Conf. Knowl. Discov. Data Mining, Barcelona, Spain, Aug. 2024, pp. 4676–4687

    2024

  20. [20]

    Diff-RNTraj: A structure-aware diffusion model for road network- constrained trajectory generation,

    T. Wei, Y . Lin, S. Guo, Y . Lin, Y . Huang, C. Xiang, Y . Bai, and H. Wan, “Diff-RNTraj: A structure-aware diffusion model for road network- constrained trajectory generation,” 2024, arXiv:2402.07369. [Online]. Available: https://arxiv.org/abs/2402.07369

    Pith/arXiv arXiv 2024

  21. [21]

    Diffusion-based planning for autonomous driving with flexible guidance,

    Y . Zheng, R. Liang, K. Zheng, J. Zheng, L. Mao, J. Li, W. Gu, R. Ai, S. E. Li, X. Zhan, and J. Liu, “Diffusion-based planning for autonomous driving with flexible guidance,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025

    2025

  22. [22]

    Context-aware trajectory prediction for autonomous driving in heterogeneous environments,

    Z. Li, Z. Chen, Y . Li, and C. Xu, “Context-aware trajectory prediction for autonomous driving in heterogeneous environments,”Comput.-Aided Civ. Infrastruct. Eng., vol. 39, no. 1, pp. 120–135, 2024

    2024

  23. [23]

    Interaction- aware and driving style-aware trajectory prediction for heterogeneous vehicles in mixed traffic environment,

    Q. Zhang, Y . Xing, J. Wang, Z. Fang, Y . Liu, and G. Yin, “Interaction- aware and driving style-aware trajectory prediction for heterogeneous vehicles in mixed traffic environment,”IEEE Trans. Intell. Transp. Syst., vol. 26, no. 7, Jul. 2025

    2025

  24. [24]

    Post-interactive multimodal trajectory prediction for autonomous driving,

    Z. Huang, Y . Li, D. Li, Y . Mu, H. Qin, and N. Zheng, “Post-interactive multimodal trajectory prediction for autonomous driving,”Transp. Res. Part C, Emerg. Technol., vol. 179, p. 105271, 2025

    2025

  25. [25]

    Multi-agent reinforcement learning with transformer-based spatio-temporal fusion for autonomous driving in mixed traffic,

    R. Li, J. Liu, T. Sun, and T. Xu, “Multi-agent reinforcement learning with transformer-based spatio-temporal fusion for autonomous driving in mixed traffic,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Hangzhou, China, Oct. 2025, pp. 8283–8289

    2025

  26. [26]

    A planning-oriented autonomous driving framework: From image to trajectory with intent-aware prediction,

    H.-Y . Kang, J.-G. Lu, M. Wei, Q.-H. Zhang, Z. Zhu, and Y . Wang, “A planning-oriented autonomous driving framework: From image to trajectory with intent-aware prediction,”IEEE Trans. Intell. Transp. Syst., pp. 1–15, 2026

    2026

  27. [27]

    DragTraffic: Interactive and controllable traffic scene generation for autonomous driving,

    S. Wang, G. Sun, F. Ma, T. Hu, Q. Qin, Y . Song, L. Zhu, and J. Liang, “DragTraffic: Interactive and controllable traffic scene generation for autonomous driving,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Abu Dhabi, UAE, Oct. 2024, pp. 14 241–14 247

    2024

  28. [28]

    LD-Scene: LLM- guided diffusion for controllable generation of adversarial safety-critical driving scenarios,

    M. Peng, Y . Xie, X. Guo, R. Yao, H. Yang, and J. Ma, “LD-Scene: LLM- guided diffusion for controllable generation of adversarial safety-critical driving scenarios,”Transp. Res. Part C, Emerg. Technol., 2026

    2026

  29. [29]

    Modelling two-dimensional driving behaviours at unsignalised intersection using multi-agent imitation learning,

    J. Sun and J. Kim, “Modelling two-dimensional driving behaviours at unsignalised intersection using multi-agent imitation learning,”Transp. Res. Part C, Emerg. Technol., vol. 165, p. 104702, 2024

    2024

  30. [30]

    DiffScene: Diffusion-based safety-critical scenario generation for autonomous vehicles,

    C. Xu, A. Petiushko, D. Zhao, and B. Li, “DiffScene: Diffusion-based safety-critical scenario generation for autonomous vehicles,” inProc. AAAI Conf. Artif. Intell., vol. 39, no. 9, 2025, pp. 8797–8805

    2025

  31. [31]

    SceneControl: Diffusion for controllable traffic scene generation,

    J. Lu, K. Wong, C. Zhang, S. Suo, and R. Urtasun, “SceneControl: Diffusion for controllable traffic scene generation,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2024

    2024

  32. [32]

    Optimizing diffu- sion models for joint trajectory prediction and controllable generation,

    Y . Wang, C. Tang, L. Sun, S. Rossi, Y . Xie, C. Peng, T. Hannagan, S. Sabatini, N. Poerio, M. Tomizuka, and W. Zhan, “Optimizing diffu- sion models for joint trajectory prediction and controllable generation,” inProc. Eur. Conf. Comput. Vis. (ECCV), vol. 15087, 2024, pp. 324– 341

    2024

  33. [33]

    Intention-aware denoising diffusion model for trajectory prediction,

    C. Liu, S. He, H. Liu, and J. Chen, “Intention-aware denoising diffusion model for trajectory prediction,”IEEE Trans. Intell. Transp. Syst., vol. 26, no. 5, pp. 5915–5930, May 2025

    2025

  34. [34]

    Traffic flow impact of mixed heterogeneous platoons on highways: An approach combining driving simulation and microscopic traffic simulation,

    W. Ren, J. Liu, J. Wang, L. Zhao, and Y . Du, “Traffic flow impact of mixed heterogeneous platoons on highways: An approach combining driving simulation and microscopic traffic simulation,”Physica A, vol. 643, p. 129803, 2024

    2024

  35. [35]

    A dynamic test scenario generation method for autonomous vehicles based on conditional generative adversarial imitation learning,

    L. Jia, D. Yang, Y . Ren, C. Qian, Q. Feng, B. Sun, and Z. Wang, “A dynamic test scenario generation method for autonomous vehicles based on conditional generative adversarial imitation learning,”Accid. Anal. Prev., vol. 194, p. 107279, 2024. IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 17

    2024

  36. [36]

    An efficient high-risk lane-changing scenario edge cases generation method for autonomous vehicle safety testing,

    S. Jing, Y . Zhao, X. Zhao, F. Hui, and A. J. Khattak, “An efficient high-risk lane-changing scenario edge cases generation method for autonomous vehicle safety testing,”IEEE Trans. Intell. Veh., vol. 10, no. 1, pp. 459–471, Jan. 2025

    2025

  37. [37]

    Detection and analysis of corner case scenarios at a signalized urban intersection,

    C. Schicktanz and K. Gimm, “Detection and analysis of corner case scenarios at a signalized urban intersection,”Accid. Anal. Prev., vol. 210, p. 107838, 2025

    2025

  38. [38]

    Application of uncertainty to out-of-distribution detection for autonomous driving perception safety,

    K. Wang, Q. Ma, C. Shen, and J. Lu, “Application of uncertainty to out-of-distribution detection for autonomous driving perception safety,” IEEE Trans. Intell. Transp. Syst., vol. 26, no. 8, pp. 11 276–11 293, Aug. 2025

    2025

  39. [39]

    Optimization of conditional value- at-risk,

    R. T. Rockafellar and S. Uryasev, “Optimization of conditional value- at-risk,”J. Risk, vol. 2, no. 3, pp. 21–41, 2000

    2000

  40. [40]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, pp. 6840–6851

    2020

  41. [41]

    Improved denoising diffusion proba- bilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion proba- bilistic models,” inProc. Int. Conf. Mach. Learn. (ICML), 2021, pp. 8162–8171

    2021

  42. [42]

    Generative adversarial imitation learning,

    J. Ho and S. Ermon, “Generative adversarial imitation learning,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 29, 2016, pp. 4565–4573

    2016

  43. [43]

    Dissipation of stop-and-go waves via control of autonomous vehicles: Field experiments,

    R. E. Stern, S. Cui, R. Bhadani, M. L. Delle Monache, M. Bunting, M. Churchill, R. Haulcy, N. Hamilton, B. Piccoli, F. Wu, H. Pohlmann, J. Sprinkle, B. Seibold, and D. B. Work, “Dissipation of stop-and-go waves via control of autonomous vehicles: Field experiments,”Transp. Res. Part C, Emerg. Technol., vol. 89, pp. 205–221, Apr. 2018

    2018

  44. [44]

    The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

    R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” inProc. IEEE Int. Conf. Intell. Transp. Syst. (ITSC), 2018, pp. 2118–2125

    2018

  45. [45]

    The round dataset: A drone dataset of road user trajectories at roundabouts in germany,

    R. Krajewski, T. Moers, J. Bock, L. Vater, and L. Eckstein, “The round dataset: A drone dataset of road user trajectories at roundabouts in germany,” inProc. IEEE Int. Conf. Intell. Transp. Syst. (ITSC), 2020, pp. 1–6

    2020

  46. [46]

    The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,

    T. Moers, L. Vater, R. Krajewski, J. Bock, A. Zlocki, and L. Eckstein, “The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany,” inProc. IEEE Intell. Vehicles Symp. (IV), 2022, pp. 958–964

    2022

  47. [47]

    The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,

    J. Bock, R. Krajewski, T. Moers, S. Runde, L. Vater, and L. Eckstein, “The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections,” inProc. IEEE Intell. Vehicles Symp. (IV), 2020, pp. 1929–1934. IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 18 APPENDIXA OVERALLWORKFLOW OFDRIFT This appendix gives the operatio...

    2020