Pith. sign in

REVIEW 4 major objections 4 minor 30 references

A single continuous style value learned from pairwise preference comparisons controls how aggressively an imitation-learning planner accelerates, turns, and overtakes.

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 · deepseek-v4-flash

2026-08-03 05:24 UTC pith:EFQFUP5G

load-bearing objection A plausible style-knob for driving policies, but the human-alignment claim is not actually tested against humans. the 4 major comments →

arxiv 2607.29517 v1 pith:EFQFUP5G submitted 2026-07-31 cs.RO cs.HC

STAGE: STyle-controllable Action GEneration for personalized autonomous driving

Zihao Liu , Xing Liu , Yizhai Zhang , Panfeng Huang This is my paper
classification cs.RO cs.HC
keywords driving stylepreference learningimitation learningautonomous drivingstyle-controllable action generationaction modalityvariational autoencoderhuman-in-the-loop
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.

This paper tries to show that a single continuous number, a 'style value', can capture a driver's aggressiveness preference and steer an imitation-learning policy to generate matching trajectories and control commands. The proposed method, STAGE, learns this style value from pairwise comparisons of driving data, scored by hand-written aggressiveness rules that approximate human preference without costly manual labels. The style value is made monotonic in aggressiveness and is fed into a Transformer-based policy as a conditional input, alongside a variational-autoencoder latent that captures the remaining diversity of human driving. If the claim holds, users could personalize autonomous driving by moving a slider, rather than choosing from coarse discrete styles or tuning entangled latent codes. The paper supports the claim with qualitative scenario demonstrations and quantitative comparisons showing style controllability, continuity, alignment, and safety in simulation.

Core claim

STAGE learns a continuous style value V(x,a) that scores the aggressiveness of taking action a in state x, trained by preference learning on pairwise comparisons generated by an aggressiveness scoring rule. This style value is forced to be monotonic in aggressiveness and is used, together with a variational-autoencoder latent, to condition a Transformer imitation policy. During inference the user supplies the style value, and the model generates a trajectory and control commands whose behavior — speed, lane choice, overtaking — varies accordingly. The paper reports that in typical road scenarios, generated behaviors align with human expectations, and quantitative comparisons show STAGE combi

What carries the argument

The central object is the learned style value V(x,a), a scalar trained by preference learning on pairwise comparisons of (state, action) pairs; it is made monotonic in aggressiveness and used as a conditional control input. The supporting mechanism is the action modality module, which splits human action patterns into this style value and a variational-autoencoder latent z: the style value carries the controllable aggressiveness dimension, while z captures style-independent behavioral diversity. Both are fed into a Transformer encoder-decoder trained with a combined loss of trajectory/control imitation, preference ranking, and KL divergence. This decomposition is what turns a continuous styl

Load-bearing premise

The load-bearing premise is that the hand-written aggressiveness scoring rule based on speed, throttle, proximity within 20 meters, and lane deviation faithfully captures how humans judge aggressiveness, since that same rule both supervises the style value and scores the outputs used to validate it.

What would settle it

Have independent human raters pairwise-compare aggressiveness of held-out (state, action) pairs across several road scenarios and compute Spearman correlation with the model's style values; the monotonic-alignment claim fails if the correlation is weak or if the same style value is rated aggressive in one scenario and cautious in another.

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

If this is right

  • Users can personalize the autonomous driving policy on the fly by moving a slider for the style value, with no retraining, because the style value is a conditional input to the trained policy.
  • Continuous style values avoid the abrupt, short-term style switches that discrete categories produce, leading to smoother and more stable driving.
  • Separating style from style-independent action patterns via the VAE latent improves both controllability and safety, with higher average completion rates than behavioral cloning and generative adversarial imitation learning.
  • Low-cost rule-based preference comparisons can substitute for human annotations in style learning, lowering the barrier to personalization.
  • Because the style value is trend-consistent and monotonic, the same preference-learning mechanism can be extended to multidimensional style vectors once enough paired data are available.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct user study could strengthen the alignment claim: collect pairwise aggressiveness judgments from independent drivers and compare them against the learned style value, rather than relying only on the rule-supervised style module.
  • The same slider interface could be applied to other dimensions of driving preference, such as comfort, fuel efficiency, or rule-following, by supplying different preference signals during training.
  • A scenario-conditioned or normalized style value would address the paper's noted limitation that the same value can feel different across road contexts, making the slider's meaning globally consistent.
  • The action-modality decomposition could transfer to other imitation-learning domains where users want a continuous, interpretable knob over a learned policy, such as teleoperated manipulation.

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

4 major / 4 minor

Summary. The paper proposes STAGE, a method for style-controllable action generation in autonomous driving. A scalar 'style value' V(x,a) is learned via pairwise preference learning, where preference labels are generated by a hand-crafted aggressiveness scoring rule based on speed, throttle, proximity within 20 m, and lane deviation (Section III.A, Eqs. (1)-(2)). This style value is then used as a conditional input in a Transformer-based imitation learning policy that outputs trajectories and control signals, with the VAE latent capturing style-independent action modalities (Section III.B). The authors report experiments in the MetaDrive simulator with human-collected driving data, claiming that the generated actions align with human expectations across typical scenarios (Section IV.A), that the method achieves superior style alignment compared with discrete-style baselines (Section IV.B, Fig. 6), and that the action modality architecture improves safety and comfort (Section IV.C). The paper also openly acknowledges limitations, including cross-scenario inconsistency of the style value and the need for larger data for multidimensional styles.

Significance. If the central claim holds—that a single continuous style value, learned from pairwise comparisons, can condition an imitation policy to generate actions whose aggressiveness monotonically matches human preferences—STAGE would provide a practical and intuitive personalization mechanism for autonomous driving, filling a gap between discrete classification-based styles and entangled latent representations. The paper contributes a clean architecture combining preference learning with a VAE-based action modality decomposition, and the authors state that code is publicly available. They also explicitly acknowledge limitations, which is a strength. However, the evidence supporting the human-alignment claim is currently self-referential: the same V module trained on the authors' own aggressiveness rule is used both as the benchmark and as the scorer of generated trajectories. No independent human-preference data or user study is presented. The significance of the paper therefore depends on an external validation step that is currently missing; the contribution is promising but not yet convincingly established.

major comments (4)
  1. [§III.A and §IV.B, Figure 6] The central claim of human-style alignment is supported by a circular evaluation. Eq. (2) trains V to match the paper's hand-written aggressiveness scoring rule (speed, throttle, proximity within 20 m, lane deviation). In §IV.B, the same V module computes both the human benchmark style sequence and the generated-trajectory style sequence whose Spearman correlation is reported in Figure 6. The reported alignment therefore measures self-consistency with the rule, not agreement with human drivers. An independent validation against actual human preference judgments (e.g., forced-choice comparisons or user ratings) is required before the abstract's assertion 'significantly align with human expectations' can be accepted.
  2. [§IV.B, Figure 6] The text says 'Spearman correlation coefficient (R2)' and the caption of Figure 6 also calls it 'Spearman correlation coefficient R2'. Spearman's rank correlation is conventionally denoted ρ or r_s, not R² (which is the coefficient of determination). Moreover, no numerical values are given—the figure is qualitative. Please report the actual Spearman ρ values with confidence intervals or effect sizes, and correct the notation. Without these, the claim that STAGE 'achieves superior style imitation performance' cannot be quantitatively assessed.
  3. [§V, Conclusion; §I Contributions] The first listed contribution claims STAGE delivers 'continuous and consistent assessments of driving behaviors across various scenarios.' The conclusion, however, explicitly states that the same style value can correspond to 'fully accelerating at the starting point or making a slow turn at an intersection' and that 'drivers may need to frequently adjust the style value across scenarios.' This cross-scenario inconsistency directly undermines the 'trend-consistent numerical representation' claim in §III.A and the monotonicity claim in the abstract. The paper should either soften the contribution to reflect the scenario dependence of the style value or provide evidence that the inconsistency is benign in practice.
  4. [§IV.A, Figure 5] The behavior evaluation in Section IV.A is entirely qualitative: three simulator scenarios with hand-picked visual examples. The text states this 'demonstrates that the STAGE has the capability to generate style trajectories that align with human user understand of aggressiveness,' but no human subjects were involved and no quantitative measure of alignment is provided. Given that the paper's central claim is human alignment, a controlled user study (e.g., participants rating or ranking generated trajectories from different style values) is needed, at least in the simulator setting. The current evidence is insufficient to support the abstract's claim of 'significantly align with human expectations.'
minor comments (4)
  1. [References] Reference [21] is cited for Diffusion Policy but the bibliographic entry is 'The mechanisms of policy diffusion' (a political science paper) rather than the correct diffusion policy paper. Please correct this citation.
  2. [§IV.B, Table I] The table formatting is confusing: the header row is split as 'Methods Metrics' and the columns contain only one metric (Average Completion Rate), while the main text discusses controllability and continuity. Consider restructuring the table and providing standard deviations consistently.
  3. [Throughout] There are several typographical issues, e.g., 'V AE' with a space, 'align with human user understand' (should be 'understanding'), and 'In Figure 3 describes this process' (ungrammatical). A careful proofread is recommended.
  4. [§IV.C] The safety evaluation relies on a single metric (Average Completion Rate) and throttle/brake statistics. Since the reader cannot see the underlying data distributions, please clarify how the standard deviations in Table I are computed and whether the differences between CV AE and STAGE are statistically significant.

Circularity Check

1 steps flagged

The headline 'align with human expectations' is supported only by a self-referential evaluation: the style value V is trained on the authors' own aggressiveness rule, and the same V scores both the human benchmark and the generated trajectories.

specific steps
  1. fitted input called prediction [Section III.A (style learning / aggressiveness rule) and Section IV.B (Style Value Alignment Evaluation)]
    "We developed a generalized aggressiveness scoring rule, where the speed and throttle amplitude of the ego vehicle are proportional to the aggressiveness score. ... Notably, our style value model, trained through preference learning, aligns with these rules. (III.A) Initially, human driving data is collected, and the style value sequence is calculated based on STAGE’s style value module, serving as the style benchmark. ... The style values of the generated trajectories are then evaluated using their respective style value modules as the outcome of driving style imitation. (IV.B)"

    The style value V is fitted through preference labels generated by the paper's own hand-written aggressiveness scoring rule (Eq. 2), so V is, by construction, a function of that rule. In the alignment evaluation, the same V module is used to compute both the human benchmark style values and the generated-trajectory style values, and the resulting Spearman R2 is presented as evidence that the style-controlled actions 'align with human expectations.' This correlation measures self-consistency between the conditional generator and its own conditioning/training signal, not agreement with human aggressiveness judgments: the benchmark and the metric are the same fitted function. No independent human preference labels or user study are used in this evaluation.

full rationale

The paper is not entirely circular: the policy also minimizes MAE on human actions (Eq. 4), and the safety metrics in Section IV.C are independent of the style value V. However, the load-bearing claim in the abstract—that style-controlled action generation results 'significantly align with human expectations'—is not externally validated. The style value is defined and trained to match the authors' own aggressiveness scoring rule, and the same V is then used to score both the human benchmark and the generated trajectories in Section IV.B. The reported Spearman correlation is therefore a self-consistency statistic, not a human-alignment test. The paper's own conclusion acknowledges that the same style value can correspond to 'fully accelerating at the starting point or making a slow turn at an intersection,' which further undermines the monotonic cross-scenario aggressiveness claim. The central novelty—a continuous, human-aligned style knob—reduces, in the presented evidence, to a learned scalar fitted to the authors' rule and evaluated with that same rule.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central contribution rests on the heuristic aggressiveness scoring rule and the assumption that a V learned from it is externally valid for human preferences. No independent human-preference benchmark is used, and key hyperparameters are unreported.

free parameters (4)
  • lambda_1, lambda_2 (loss weights) = not reported
    Weights for preference and VAE losses in Eq. (5); no ablation or sensitivity analysis is given.
  • aggressiveness scoring rule weights = not reported
    The relative coefficients for speed, throttle, inverse distance, and lane deviation are described only qualitatively but they define the training signal for the style value.
  • proximity threshold = 20 m
    Hand-chosen threshold in Section III.A that switches on the distance and lane-deviation aggressiveness terms.
  • style value normalization range = [-1, 1]
    Post-training scaling in Section IV; style value 0 represents regular driving, negative conservative, positive aggressive.
axioms (5)
  • standard math Pairwise preference probabilities follow the Bradley-Terry/Luce model in Eq. (1).
    Used in Section III.A as the foundation of preference learning; standard choice model.
  • domain assumption Driving style is adequately captured by a single scalar aggressiveness value that is monotonically ordered.
    Invoked throughout Sections I and III.A; the paper itself notes multidimensional style would require more data.
  • ad hoc to paper The hand-crafted aggressiveness scoring rule is a valid stand-in for human preference without further validation.
    Section III.A defines the rule and uses it to generate preference pairs; no human validation of the rule is reported.
  • ad hoc to paper Action modalities can be decomposed into an independent style dimension and a residual VAE latent.
    Section III.B assumes this disentanglement; no identifiability or independence proof is provided.
  • domain assumption MetaDrive simulator data collected from instructed drivers is representative of real-world human driving preferences.
    Section IV relies on this to support claims about human expectations and safety.
invented entities (1)
  • Style value V(x,a) no independent evidence
    purpose: Continuous scalar control knob for aggressiveness and the evaluation metric for style alignment.
    An invented latent construct; its external validity is never tested against human ratings, and the paper evaluates it using the same V module that was trained on the hand-crafted rule.

pith-pipeline@v1.3.0-daily-deepseek · 11038 in / 11587 out tokens · 104919 ms · 2026-08-03T05:24:06.595516+00:00 · methodology

0 comments
read the original abstract

Driving style refers to the behavioral preferences that drivers maintain during driving, shaped by their diverse experiences, habits, and needs, and is typically reflected in varying levels of aggressiveness. If humans choose to use autonomous driving systems, they would expect the driving style of the systems to closely resemble their own habit. However, this is challenging for current industrial autonomous driving systems. To address this, we developed a style controllable action generation method, STAGE, for driving tasks. Its training process is based on imitation learning, incorporating both style value and latent value action modality encoding. Preference learning is then used to identify the user's driving style as a continuous, monotonic style value. And to reduce the cost of human involvement in the preference training process, we also developed a set of rules to compare driving style in data pairs. Then, during inference, the user inputs the style value to control the generated action patterns, dynamically meeting the user's expectations. Using the STAGE method, we verified that the style-controlled action generation results in several typical road scenarios significantly align with human expectations. Furthermore, through comparisons between the STAGE method and various other approaches, we reveal the unique functionalities of STAGE, including its style controllability, style continuity, driving style alignment capability and driving safety. The code for this work is available at: https://github.com/CarlDegio/STAGE

Figures

Figures reproduced from arXiv: 2607.29517 by Panfeng Huang, Xing Liu, Yizhai Zhang, Zihao Liu.

Figure 1
Figure 1. Figure 1: Overview of the STAGE method. The solid line arrows represent the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A network architecture diagram of the STAGE method. The action [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Our simulation engine utilizes the fast and convenient MetaDrive, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Composition of the loss function. The overall loss consists of three [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The results of the STAGE model in three typical scenarios are shown. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The style value alignment comparison results are presented. The [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The figure illustrates a comparison of driving comfort derived from [PITH_FULL_IMAGE:figures/full_fig_p007_7.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

30 extracted references · 4 linked inside Pith

  1. [1]

    Baidu apollo em motion planner,

    H. Fan, F. Zhu, C. Liu, L. Zhang, L. Zhuang, D. Li, W. Zhu, J. Hu, H. Li, and Q. Kong, “Baidu apollo em motion planner,”arXiv preprint arXiv:1807.08048, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    Optimal vehicle path planning using quadratic optimization for baidu apollo open platform,

    Y . Zhang, H. Sun, J. Zhou, J. Pan, J. Hu, and J. Miao, “Optimal vehicle path planning using quadratic optimization for baidu apollo open platform,” in2020 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2020, pp. 978–984

    2020

  3. [3]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang,et al., “Planning-oriented autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 853–17 862

    2023

  4. [4]

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

    Z. Xu, Y . Zhang, E. Xie, Z. Zhao, Y . Guo, K.-Y . K. Wong, Z. Li, and H. Zhao, “Drivegpt4: Interpretable end-to-end autonomous driving via large language model,”IEEE Robotics and Automation Letters, 2024

    2024

  5. [5]

    Take-over performance analysis depending on the drivers’ non-driving secondary tasks in automated vehicles,

    J. Kim, H.-S. Kim, W. Kim, and D. Yoon, “Take-over performance analysis depending on the drivers’ non-driving secondary tasks in automated vehicles,” in2018 International Conference on Information and Communication Technology Convergence (ICTC), 2018, pp. 1364– 1366

    2018

  6. [6]

    Comfort and safety in conditional automated driving in dependence on personal driving behavior,

    L. Vasile, N. Dinkha, B. Seitz, C. D ¨asch, and D. Schramm, “Comfort and safety in conditional automated driving in dependence on personal driving behavior,”IEEE Open Journal of Intelligent Transportation Systems, 2023

    2023

  7. [7]

    Driving style classification using long-term accelerometer information,

    V . Vaitkus, P. Lengvenis, and G. ˇZylius, “Driving style classification using long-term accelerometer information,” in2014 19th international conference on methods and models in automation and robotics (MMAR). IEEE, 2014, pp. 641–644

    2014

  8. [8]

    Incorporating driving style recognition into mpc for energy management of plug-in hybrid electric buses,

    X. Tian, Y . Cai, X. Sun, Z. Zhu, Y . Wang, and Y . Xu, “Incorporating driving style recognition into mpc for energy management of plug-in hybrid electric buses,”IEEE Transactions on Transportation Electrifica- tion, vol. 9, no. 1, pp. 169–181, 2023

    2023

  9. [9]

    Generating drawing/grinding trajectories based on hierarchical cvae,

    M. Aita, Y . Nogi, K. Sugawara, S. Sakaino, and T. Tsuji, “Generating drawing/grinding trajectories based on hierarchical cvae,”Advanced Robotics, pp. 1–11, 2025

    2025

  10. [10]

    Newtonianvae: Propor- tional control and goal identification from pixels via physical latent spaces,

    M. Jaques, M. Burke, and T. M. Hospedales, “Newtonianvae: Propor- tional control and goal identification from pixels via physical latent spaces,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 4454–4463

    2021

  11. [11]

    Modeling human driving behavior through generative adversarial imitation learning,

    R. Bhattacharyya, B. Wulfe, D. J. Phillips, A. Kuefler, J. Morton, R. Senanayake, and M. J. Kochenderfer, “Modeling human driving behavior through generative adversarial imitation learning,”IEEE Trans- actions on Intelligent Transportation Systems, vol. 24, no. 3, pp. 2874– 2887, 2022

    2022

  12. [12]

    A framework for behavioural cloning

    M. Bain and C. Sammut, “A framework for behavioural cloning.” in Machine Intelligence 15, 1995, pp. 103–129

    1995

  13. [13]

    Behavioral cloning from observa- tion,

    F. Torabi, G. Warnell, and P. Stone, “Behavioral cloning from observa- tion,”arXiv preprint arXiv:1805.01954, 2018

    Pith/arXiv arXiv 2018

  14. [14]

    Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,

    M. Bansal, A. Krizhevsky, and A. Ogale, “Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,”arXiv preprint arXiv:1812.03079, 2018

    Pith/arXiv arXiv 2018

  15. [15]

    Training robots without robots: Deep imitation learning for master-to-robot policy trans- fer,

    H. Kim, Y . Ohmura, A. Nagakubo, and Y . Kuniyoshi, “Training robots without robots: Deep imitation learning for master-to-robot policy trans- fer,”IEEE Robotics and Automation Letters, vol. 8, no. 5, pp. 2906– 2913, 2023

    2023

  16. [16]

    Algorithms for inverse reinforcement learning

    A. Y . Ng, S. Russell,et al., “Algorithms for inverse reinforcement learning.” inIcml, vol. 1, no. 2, 2000, p. 2

    2000

  17. [17]

    Generative adversarial imitation learning,

    J. Ho and S. Ermon, “Generative adversarial imitation learning,”Ad- vances in neural information processing systems, vol. 29, 2016

    2016

  18. [18]

    Efficient spatiotemporal transformer for robotic reinforcement learning,

    Y . Yang, D. Xing, and B. Xu, “Efficient spatiotemporal transformer for robotic reinforcement learning,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7982–7989, 2022

    2022

  19. [19]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on Robot Learning. PMLR, 2022, pp. 158– 168

    2022

  20. [20]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    The mechanisms of policy diffusion,

    C. R. Shipan and C. V olden, “The mechanisms of policy diffusion,” American journal of political science, vol. 52, no. 4, pp. 840–857, 2008

    2008

  22. [22]

    Human-like car-following modeling based on online driving style recognition,

    L. Ma, S. Qu, L. Song, J. Zhang, and J. Ren, “Human-like car-following modeling based on online driving style recognition,”Electron. Res. Arch, vol. 31, pp. 3264–3290, 2023

    2023

  23. [23]

    Finite mixture of the hidden markov model for driving style analysis,

    L. Ding, T. Zhu, Y . Wang, and Y . Zou, “Finite mixture of the hidden markov model for driving style analysis,”Journal of advanced trans- portation, vol. 2022, no. 1, p. 4989947, 2022

    2022

  24. [24]

    An embedded driving style recognition approach: Leveraging knowledge in learning,

    C. Zhang, W. Wang, Z. Ju, Z. Chen, G. Venture, and J. Xi, “An embedded driving style recognition approach: Leveraging knowledge in learning,” IEEE Transactions on Intelligent Vehicles, 2024

    2024

  25. [25]

    Safety-balanced driving-style aware trajectory planning in intersection scenarios with uncertain environment,

    X. Wang, K. Tang, X. Dai, J. Xu, J. Xi, R. Ai, Y . Wang, W. Gu, and C. Sun, “Safety-balanced driving-style aware trajectory planning in intersection scenarios with uncertain environment,”IEEE Transactions on Intelligent Vehicles, vol. 8, no. 4, pp. 2888–2898, 2023

    2023

  26. [26]

    A learning- based discretionary lane-change decision-making model with driving style awareness,

    Y . Zhang, Q. Xu, J. Wang, K. Wu, Z. Zheng, and K. Lu, “A learning- based discretionary lane-change decision-making model with driving style awareness,”IEEE transactions on intelligent transportation sys- tems, vol. 24, no. 1, pp. 68–78, 2022

    2022

  27. [27]

    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 intelli- gence, vol. 45, no. 3, pp. 3461–3475, 2022

    2022

  28. [28]

    Safe real-world autonomous driving by learning to predict and plan with a mixture of experts,

    S. Pini, C. S. Perone, A. Ahuja, A. S. R. Ferreira, M. Niendorf, and S. Zagoruyko, “Safe real-world autonomous driving by learning to predict and plan with a mixture of experts,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 10 069–10 075

    2023

  29. [29]

    Vehicle trajectory prediction using intention-based conditional variational autoencoder,

    X. Feng, Z. Cen, J. Hu, and Y . Zhang, “Vehicle trajectory prediction using intention-based conditional variational autoencoder,” in2019 IEEE Intelligent Transportation Systems Conference (ITSC). IEEE, 2019, pp. 3514–3519. LIUet al.: STAGE 9

    2019

  30. [30]

    Driving style-based conditional variational autoencoder for prediction of ego vehicle trajectory,

    D. Kim, H. Shon, N. Kweon, S. Choi, C. Yang, and K. Huh, “Driving style-based conditional variational autoencoder for prediction of ego vehicle trajectory,”IEEE Access, vol. 9, pp. 169 348–169 356, 2021

    2021