Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

RAPiD distills a frozen diffusion driving planner into a fast deterministic policy that matches the teacher on non-reactive nuPlan splits and tops learning-based planners on interPlan, while cutting per-trajectory inference from 100.91 ms t

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 03:38 UTC pith:MNOVWLZG

load-bearing objection A useful distillation result with a shaky derivation and unpolished evaluation; worth refereeing after the authors reconcile the math and the abstract. the 5 major comments →

arxiv 2602.07339 v2 pith:MNOVWLZG submitted 2026-02-07 cs.AI cs.LGcs.RO

RAPiD: Reward-Guided Consistency Distillation of Diffusion Planners for Real-Time Autonomous Driving

Ruturaj Reddy , Hrishav Bakul Barua , Junn Yong Loo , Thanh Thi Nguyen , Ganesh Krishnasamy This is my paper
classification cs.AI cs.LGcs.RO
keywords autonomous drivingtrajectory planningdiffusion distillationscore-regularized policy optimizationoffline reinforcement learningclosed-loop evaluationnuPlansafety scoring
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.

The paper tries to show that the expressive, multi-modal behavior of a large diffusion-based driving planner can be compressed into a fast deterministic policy without sacrificing closed-loop safety. It does this by using a frozen DiffusionPlanner as a behavior prior: the planner's score function regularizes the student policy's gradient, while a critic trained on PDM-scored expert trajectories supplies safety and comfort guidance. If correct, autonomous driving planners can get diffusion-level behavior understanding at a fraction of the inference cost—12.41 ms per trajectory rather than 100.91 ms—and with few-step, reproducible decisions. On non-reactive nuPlan splits the student matches or beats the teacher, and on interPlan it is the top learning-based planner, with the largest margins in dense traffic and overtaking scenarios.

Core claim

RAPiD's central claim is that a frozen, stochastic diffusion planner need not be sampled at deployment. The paper trains a deterministic transformer policy to maximize a safety critic's value while a frozen DiffusionPlanner's score function—its noise prediction—regularizes the policy toward the manifold of realistic driving behavior. The critic is learned with Implicit Q-Learning on trajectories scored by a Predictive Driver Model (PDM) reward that heavily weights collision avoidance, proximity, and comfort. The result is a one-forward-pass planner that, on nuPlan, scores 90.19 (val14), 89.98 (test14), and 76.09 (test14-hard) in non-reactive closed-loop evaluation—above the teacher's 89.87,

What carries the argument

The load-bearing mechanism is Score Regularized Policy Optimization (SRPO): instead of sampling the diffusion teacher, the student's gradient is computed from the teacher's noise-prediction network, which approximates the score (log-density gradient) of the behavior distribution at the student's action, plus the gradient of an Implicit Q-Learning critic. The critic is trained with expectile regression on PDM-scored trajectories to give dense safety and comfort rewards. Equation (9) fuses these two signals into a surrogate gradient that updates the deterministic policy without iterative denoising.

Load-bearing premise

The gradient update that trains the student assumes the frozen diffusion teacher's score at noised student actions is a valid estimate of the true behavior-distribution score, and that the safety critic, trained on the offline buffer, stays accurate on actions the student actually generates; if either fails, the student is optimized with biased gradients and closed-loop performance may not transfer.

What would settle it

Compute the Stage-3 surrogate gradient (Eq. 9) on a held-out batch and compare it with a Monte Carlo estimate of the true SRPO objective gradient obtained by actually sampling the diffusion teacher; if the two gradients have near-zero cosine similarity, or if Q-values on student actions carry high TD error relative to the offline buffer, the score approximation is not faithful. More directly, run RAPiD in nuPlan reactive test14 with a critic retrained on student rollouts: if the score stays flat, the fixed-critic assumption is not the bottleneck; if it drops, the frozen critic was masking a bi

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

If this is right

  • Per-trajectory inference falls from 100.91 ms to 12.41 ms (about 8x for the policy; 18.41 ms complete pipeline, about 5.5x), making the planner fast enough for real-time closed-loop driving.
  • On non-reactive nuPlan closed-loop splits, the distilled policy exceeds the diffusion teacher: 90.19 vs 89.87 (val14), 89.98 vs 89.87 (test14), and 76.09 vs 75.53 (test14-hard).
  • On interPlan, the aggregate score of 27 is the highest among learning-based planners, with overtaking (26) and high-density lane-change (23 vs 16) showing the largest gains.
  • Safety-critical PDM metrics improve—collision avoidance, time-to-collision, comfort, and proximity—while route progress drops, because the reward weights favor comfort and safety over distance.
  • Training starts from a frozen pretrained diffusion planner, so the speedup does not require retraining the generative model from scratch.

Where Pith is reading between the lines

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

  • The same distillation recipe—a frozen diffusion score as regularizer plus a reward-trained critic—should transfer to other conditional diffusion planners and to other sequential decision domains such as manipulation or navigation where a pretrained generative policy is too slow for closed-loop control.
  • The reported safety and comfort gains come with a deliberate progress penalty; this suggests the PDM weight vector is a tunable aggressiveness dial, and a system designer could re-calibrate it for highway versus urban driving rather than accepting one fixed tradeoff.
  • The critic is only as good as its training distribution; retraining or finetuning the critic on the student's own rollouts could close the small reactive-split gap (81.86 vs 82.93) and is a direct testable extension of the paper's claims.
  • Because the student is deterministic, its multi-modality is funneled through best-of-K selection; evaluating with a small ensemble of student checkpoints or latent-conditioned heads could recover some of the reactive flexibility the teacher had.

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

5 major / 5 minor

Summary. The paper presents RAPiD, a three-stage framework to distill a pretrained DiffusionPlanner into an efficient deterministic policy for closed-loop autonomous driving. Stage 1 encodes raw inputs with the frozen encoder and scores ground-truth trajectories with a PDM-style reward; Stage 2 trains an IQL critic (plus an AWR-initialized actor) on those latent states and rewards; Stage 3 updates the actor with a surrogate gradient containing the Q-derivative and a diffusion-score regularizer. The authors report nuPlan non-reactive scores that are comparable to or slightly above the no-refinement DiffusionPlanner baseline, competitive reactive scores, the best learning-based aggregate on interPlan, and a 12.41 ms inference time versus 100.91 ms for the teacher, i.e., about an 8x speedup.

Significance. If the method worked as claimed, it would be a practically valuable result: it would show that a frozen diffusion planner can be converted into a single-forward-pass policy with a large latency reduction and no large closed-loop performance loss, while also providing a template for safety/reward-guided distillation in driving. Strengths of the manuscript include the breadth of closed-loop evaluation (nuPlan and interPlan), explicit PDM metric breakdowns, qualitative scenario analyses, and a reproducible pseudocode outline. However, the theoretical derivation of the core surrogate gradient is incorrect, the critic is trained and evaluated with the same PDM scorer, and the reported performance margins are small and lack statistical support; these issues undermine the central claims as currently presented.

major comments (5)
  1. [Section 3.3, Eq. (9); Algorithm 1] Eq. (9) is not the gradient of the SRPO objective in Eq. (2). Substituting the score approximation Eq. (4) into Eq. (3) gives a term proportional to -epsilon_psi/sigma_t, not -(epsilon_psi - epsilon). Moreover Eq. (4) is explicitly stated only for t->0, whereas Algorithm 1 samples t ~ U(0,T), and the weighting omega(t) is never defined. Consequently the claim that the Stage 3 update 'distills the diffusion prior' or maximizes Eq. (2) has no derivational support. Either re-derive the surrogate with a correct, justified score estimator, or present Eq. (9) as a heuristic with separate empirical validation.
  2. [Stage 2/3; Eq. (6) and Eq. (9)] The critic in Stage 2 is trained only on ground-truth trajectories (plus PDM rewards), but Stage 3 evaluates nabla_a Q at actions pi_theta(s) that may lie outside the training support. No conservative Q penalty, behavior-cloning constraint, or empirical check of Q accuracy on student actions is provided. Since best-of-K selection and safety claims rely on this Q-function, off-support errors could make the optimized policy and reported scores unreliable.
  3. [Section 4.2; Table 1; Figure 4] The main 'safety' evaluation (no-collision, TTC, comfort, proximity in Figure 4) uses the same PDM scoring function that defines the critic's training reward. Improvements in those metrics are therefore partly an expected consequence of the training objective, not independent confirmation of safer driving. The paper also changes PDM weights relative to the nuPlan scorer and reports a progress-safety trade-off, so the lower reactive nuPlan scores are explained by construction. An evaluation with an independent closed-loop safety/comfort protocol, or an explicit ablation separating reward weighting from the distillation mechanism, is needed.
  4. [Tables 1 and 3, experiment setup] All closed-loop results are single-run scores with no error bars, confidence intervals, or repeated seeds. The claimed improvements over DiffusionPlanner on non-reactive splits are 0.32, 0.11, and 0.56 points, while reactive losses are 0.96, 2.28, and 1.07 points; the interPlan aggregate difference is 2 points. These margins are too small to support 'surpasses' or 'comparable' without variance information, especially because nuPlan closed-loop evaluation is known to be noisy.
  5. [Abstract vs. Section 3; Table 1] The abstract asserts a 2-step student generating K trajectories with best-of-K selection and 'retaining multi-modal generation,' but the body and Algorithm 1 describe a deterministic policy mapping one latent state to one action, and no K-sampling/selection appears in the method or latency measurement (12.41 ms). Additionally, Table 1 lists a stronger 'DiffusionPlanner w/ refine' teacher (e.g., 94.80 vs 89.19 on Test14 non-reactive) that is not used in the speedup/comparison, so 'comparable to the diffusion teacher' is only true for the weaker no-refinement variant. These mismatches need to be resolved or explicitly qualified.
minor comments (5)
  1. [Abstract] The abstract states 18.41 ms and 5.5x speedup, while the main text and Table 2 state 12.41 ms and 8x; these numbers must be reconciled.
  2. [Section 3.5, Eq. (9)] omega(t) is used in Eq. (9) and Algorithm 1 but never defined; if it is a chosen weighting it should be specified, and if it is derived it should be justified.
  3. [Figure 3 caption] The caption labels panels (c-1)/(c-2) as 'Yaw Rate Metrics' and (d-1)/(d-2) as 'Speed Metrics,' but the corresponding scenario descriptions say 'Starting Right Turn' and 'Low Speed Maneuvering'; the mapping between scenario and metric label is confusing.
  4. [Section 4.2] There is a duplicated/truncated sentence at the end of the nuPlan subsection ('progressing distance resulting in lower scores...' repeated), and an earlier occurrence of 'denoisining' in Section 1.
  5. [Title/Abstract] The phrase 'consistency distillation' is not operationalized in the method; no consistency loss or consistency-model objective appears anywhere. Either add the missing definition or replace the term.

Circularity Check

1 steps flagged

Partial circularity: the PDM scorer is both the training reward and the headline safety evaluation metric, so the safety improvements are partly expected; the central distillation claim otherwise has independent content.

specific steps
  1. fitted input called prediction [Section 3.4 / Stage 1 (Algorithm 1) and Section 4.2, Figure 4]
    "we integrate the Predictive Driver Model (PDM) scorer as the reward function... Each trajectory is evaluated using the Predictive Driver Model (PDM) scorer to obtain a reward r based on safety and comfort metrics. ... Both models are evaluated in the same reactive closed-loop setting but scored using the PDM metric rather than the standard nuPlan scorer."

    The reward used to train the IQL critic (Stage 1/2) is the same PDM scorer used to evaluate safety and comfort in Figure 4. Reported gains in TTC, comfort, and proximity are therefore the objective being optimized, not an independent confirmation of safety. The paper even states these improvements 'directly reflect the PDM scorer's penalties on jerk... and vehicle in front distance,' making the safety-validation portion of the central claim an expected consequence of the training signal. The speedup and nuPlan/interPlan rankings remain independent.

full rationale

RAPiD's central contribution—fast deterministic extraction of a policy from a frozen diffusion planner—is not itself circular: the latency comparison, non-reactive nuPlan scores, and interPlan generalization are evaluated against external baselines and do not reduce to the training inputs. The only clear circularity is the safety/comfort validation: the IQL critic is trained with a PDM-style reward (Algorithm 1: r ← S_PDM(...)), and the paper's headline safety evidence (Figure 4: TTC, comfort, proximity) is scored with that same PDM metric. Gains there are expected consequences of the objective, not independent confirmation; the paper even says they 'directly reflect the PDM scorer's penalties.' This partial overlap does not infect the speedup or benchmark rankings. Separately, the derivation has a non-circular flaw: Eq. (9)'s (εψ−ε) term does not follow from substituting Eq. (4) into Eq. (3), and Eq. (4) is only valid at t→0 while Algorithm 1 uses t∼U(0,T). That is a correctness risk, not a self-reduction, so it does not raise the circularity score. No load-bearing self-citations were found; SRPO is cited from external authors.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on borrowed diffusion and offline-RL machinery plus a hand-tuned PDM reward. The paper contributes no new mathematical theorem or new entity. The main free parameters are training hyperparameters and reward weights, none of which are given exact values, and the safety evaluation overlaps with the training reward.

free parameters (6)
  • Expectile tau for IQL = not specified (stated in (0.8,1))
    Eq. (5): controls how strongly low-Q trajectories are down-weighted; central to critic training.
  • SRPO temperature beta = not specified
    Eqs. (2) and (9): balances Q maximization against score regularization.
  • AWR temperature beta_AWR = not specified
    Eq. (7): weights advantage-weighted regression pretraining.
  • Policy update coefficient lambda_pi = not specified
    Eq. (8): step size for the surrogate gradient update.
  • Discount factor gamma = not specified
    Eq. (6): used in the Q-learning TD target.
  • PDM scorer metric weights = comfort 5, proximity 5, progress 2 (relative to nuPlan's 2/0/5)
    Section 3.4 and Section 4.2: hand-changed reward weights directly induce the reported safety-vs-progress tradeoff.
axioms (4)
  • domain assumption The score function of the behavior policy can be approximated by the frozen diffusion denoiser's noise prediction at actions generated by the current policy: grad_a log mu(a|s) ≈ -epsilon_psi(a_t|s,t)/sigma_t (Eqs. 4, 9).
    Algorithm 1 samples t ~ U(0,T) and evaluates at a_t = alpha_t * pi_theta(s) + sigma_t * epsilon, while Eq. (4) states the approximation at t -> 0; the gap is not analyzed.
  • domain assumption The IQL critic trained only on ground-truth trajectories remains accurate for actions produced by the stage-3 policy (Eqs. 5, 6, 9).
    Q is queried at pi_theta(s) during stage 3, but the replay buffer contains only ground-truth actions, so out-of-distribution extrapolation is implicitly assumed benign.
  • domain assumption The PDM scorer is a faithful proxy for closed-loop safety and comfort, so optimizing it improves real-world behavior (Section 3.4).
    PDM defines both the training reward and a main evaluation metric, so its validity as a safety measure is assumed rather than independently established.
  • standard math Standard diffusion/score-matching identities and IQL expectile regression behave as claimed (Eqs. 1, 5).
    Borrowed from Ho et al., Song et al., and Kostrikov et al.; not re-derived, acceptable as background.

pith-pipeline@v1.3.0-alltime-deepseek · 14924 in / 17029 out tokens · 175865 ms · 2026-08-03T03:38:04.428724+00:00 · methodology

0 comments
read the original abstract

Diffusion-based trajectory planners can model multi-modal driving behavior, but their iterative denoising process introduces a latency bottleneck for real-time closed-loop deployment. We present RAPiD, a reward-guided consistency distillation framework that distills a pretrained DiffusionPlanner into a few-step consistency student while retaining multi-modal trajectory generation. The student is trained using deterministic teacher denoising steps from the frozen diffusion planner, together with a low-noise data anchor that keeps generated trajectories grounded in expert demonstrations. To make distillation safety-aware, we train an Implicit Q-Learning critic on a balanced mixture of ground-truth log-replay and DiffusionPlanner rollout trajectories, each scored using a modified PDM-style reward, providing trajectory-level supervision beyond conventional imitation learning. During deployment, the 2-step student generates K trajectories, and the trained critic performs best-of-K trajectory selection conditioned on the latent state. On nuPlan, RAPiD maintains comparable performance to the diffusion teacher on non-reactive closed-loop splits and remains competitive on reactive splits, while reducing complete-pipeline inference latency from 100.91 ms to 18.41 ms, corresponding to a 5.5x speedup. On interPlan, RAPiD achieves the highest aggregate score among learning-based methods, demonstrating competitive generalization in interactive long-tail scenarios. These results show that reward-guided consistency distillation can convert a pretrained diffusion planner into a few-step closed-loop planner that substantially reduces inference cost while preserving safety-oriented trajectory selection. The official website of this work is: https://github.com/ruturajreddy/RAPiD

Figures

Figures reproduced from arXiv: 2602.07339 by Ganesh Krishnasamy, Hrishav Bakul Barua, Junn Yong Loo, Ruturaj Reddy, Thanh Thi Nguyen.

Figure 1
Figure 1. Figure 1: Figures (a-1) and (a-2) illustrate the limitations of the base [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework (Section 3.5). Stage 1 involves offline replay buffer construction, where raw sensor data is processed by a frozen DiffusionPlanner encoder to generate rich latent state embeddings (s). Ground truth trajectories are evaluated by the PDM Scorer to assign rewards (r) based on safety and comfort metrics, creating a scored dataset. Stage 2 focuses on Critic Training via Impli… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results across four scenarios in nuPlan: (a) Fol [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Driving risk emerges from the required two-dimensional joint evasive acceleration

    cs.RO 2026-04 unverdicted novelty 7.0

    Evasive acceleration quantifies driving risk as the minimum 2D constant relative acceleration needed to avoid collision and outperforms time-to-collision on warning timing, discrimination, and information retention ac...

  2. Off the Rails: Hijacking the Scoring Head in Generative End-to-End Driving Planners with Safety-Violating Adversarial Perturbations

    cs.RO 2026-06 unverdicted novelty 6.0

    Derail adversarial perturbations hijack the scoring head in generative E2E driving planners, flipping safe to unsafe trajectory selection with 39-80% score drops and up to 50% collision rates.

Reference graph

Works this paper leans on

29 extracted references · 4 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Is Conditional Generative Modeling all you need for Decision Making? InThe Eleventh International Con- ference on Learning Representations (ICLR),

    [Ajayet al., 2023 ] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua B Tenenbaum, Tommi S Jaakkola, and Pulkit Agrawal. Is Conditional Generative Modeling all you need for Decision Making? InThe Eleventh International Con- ference on Learning Representations (ICLR),

    2023

  2. [5]

    Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704,

    [Chiet al., 2025 ] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704,

    2025

  3. [8]

    Off-policy deep reinforcement learning without exploration

    [Fujimotoet al., 2019 ] Scott Fujimoto, David Meger, and Doina Precup. Off-policy deep reinforcement learning without exploration. InThe Proceedings of the 36th Inter- national Conference on Machine Learning (ICML), pages 2052–2062. PMLR,

    2019

  4. [10]

    Can vehi- cle motion planning generalize to realistic long-tail sce- narios? In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5388–5395

    [Hallgartenet al., 2024 ] Marcel Hallgarten, Julian Zapata, Martin Stoll, Katrin Renz, and Andreas Zell. Can vehi- cle motion planning generalize to realistic long-tail sce- narios? In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5388–5395. IEEE,

    2024

  5. [11]

    IDQL: Implicit Q-learning as an actor- critic method with diffusion policies.arXiv preprint arXiv:2304.10573,

    [Hansen-Estruchet al., 2023 ] Philippe Hansen-Estruch, Ilya Kostrikov, Michael Janner, Jakub Grudzien Kuba, and Sergey Levine. IDQL: Implicit Q-learning as an actor- critic method with diffusion policies.arXiv preprint arXiv:2304.10573,

    Pith/arXiv arXiv 2023

  6. [12]

    [Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems 34: Annual Con- ference on Neural Information Processing Systems 2020 (NeurIPS), 33:6840–6851,

    2020

  7. [14]

    DTPP: Differentiable Joint Conditional Prediction and Cost Evaluation for Tree Policy Planning in Autonomous Driving

    [Huanget al., 2024 ] Zhiyu Huang, Peter Karkus, Boris Ivanovic, Yuxiao Chen, Marco Pavone, and Chen Lv. DTPP: Differentiable Joint Conditional Prediction and Cost Evaluation for Tree Policy Planning in Autonomous Driving. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6806–6812. IEEE,

    2024

  8. [15]

    Towards learning-based planning: The nuplan benchmark for real-world autonomous driving

    [Karnchanachariet al., 2024 ] Napat Karnchanachari, Dim- itris Geromichalos, Kok Seang Tan, Nanxiang Li, Christo- pher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gi- anmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for real-world autonomous driving. In2024 IEEE Inter- national Conference on Rob...

    2024

  9. [16]

    General lane-changing model MOBIL for car-following models.Transportation Research Record, 1999(1):86–94,

    [Kestinget al., 2007 ] Arne Kesting, Martin Treiber, and Dirk Helbing. General lane-changing model MOBIL for car-following models.Transportation Research Record, 1999(1):86–94,

    2007

  10. [19]

    Scalable diffusion models with transformers

    [Peebles and Xie, 2023] William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 4195–4205,

    2023

  11. [20]

    Advantage-weighted regres- sion: Simple and scalable off-policy reinforcement learn- ing.arXiv preprint arXiv:1910.00177,

    [Penget al., 2019 ] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-weighted regres- sion: Simple and scalable off-policy reinforcement learn- ing.arXiv preprint arXiv:1910.00177,

    Pith/arXiv arXiv 2019

  12. [21]

    Deep imitative models for flexible inference, planning, and control.arXiv preprint arXiv:1810.06544,

    [Rhinehartet al., 2018 ] Nicholas Rhinehart, Rowan McAl- lister, and Sergey Levine. Deep imitative models for flexible inference, planning, and control.arXiv preprint arXiv:1810.06544,

    Pith/arXiv arXiv 2018

  13. [22]

    Urban driver: Learning to drive from real-world demonstrations using policy gradients

    [Scheelet al., 2022 ] Oliver Scheel, Luca Bergamini, Maciej Wolczyk, Bła ˙zej Osi ´nski, and Peter Ondruska. Urban driver: Learning to drive from real-world demonstrations using policy gradients. InConference on Robot Learning (CoRL), pages 718–728. PMLR,

    2022

  14. [24]

    Deep unsupervised learning using nonequilibrium thermody- namics

    [Sohl-Dicksteinet al., 2015 ] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermody- namics. InInternational Conference on Machine Learning (ICML), pages 2256–2265. pmlr,

    2015

  15. [26]

    Congested traffic states in empirical obser- vations and microscopic simulations.Physical review E, 62(2):1805,

    [Treiberet al., 2000 ] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical obser- vations and microscopic simulations.Physical review E, 62(2):1805,

    2000

  16. [28]

    Diffusion-Es: Gradient-Free Planning with Diffusion for Autonomous and Instruction- Guided Driving

    [Yanget al., 2024 ] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-Es: Gradient-Free Planning with Diffusion for Autonomous and Instruction- Guided Driving. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15342–15353,

    2024

  17. [29]

    Diffusion-Based Planning for Autonomous Driving with Flexible Guidance

    [Zhenget al., 2025 ] Yinan Zheng, Ruiming Liang, Kexin ZHENG, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Wei- hao Gu, Rui Ai, Shengbo Eben Li, Xianyuan Zhan, et al. Diffusion-Based Planning for Autonomous Driving with Flexible Guidance. InThe Thirteenth International Con- ference on Learning Representations (ICLR), 2025

    2025

  18. [2000]

    Diffusion Policies as an Expressive Pol- icy Class for Offline Reinforcement Learning

    [Wanget al., 2023 ] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion Policies as an Expressive Pol- icy Class for Offline Reinforcement Learning. InThe Eleventh International Conference on Learning Represen- tations (ICLR),

    2023

  19. [2005]

    Score Regularized Pol- icy Optimization through Diffusion Behavior

    [Chenet al., 2024 ] jie chang Chen, Cheng Lu, Zhengyi Wang, Hang Su, and Jun Zhu. Score Regularized Pol- icy Optimization through Diffusion Behavior. InThe Twelfth International Conference on Learning Represen- tations (ICLR),

    2024

  20. [2007]

    Offline Reinforcement Learning with Im- plicit Q-Learning

    [Kostrikovet al., 2021 ] Ilya Kostrikov, Ashvin Nair, and Sergey Levine. Offline Reinforcement Learning with Im- plicit Q-Learning. InDeep RL Workshop NeurIPS 2021,

    2021

  21. [2015]

    Score-Based Generative Modeling through Stochastic Differential Equations

    [Songet al., 2021 ] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. InInterna- tional Conference on Learning Representations (ICLR),

    2021

  22. [2018]

    Parting with Miscon- ceptions about Learning-based vehicle motion planning

    [Dauneret al., 2023 ] Daniel Dauner, Marcel Hallgarten, An- dreas Geiger, and Kashyap Chitta. Parting with Miscon- ceptions about Learning-based vehicle motion planning. InThe 7th Conference on Robot Learning (CoRL 2023), Atlanta, USA, pages 1268–1281. PMLR,

    2023

  23. [2019]

    From prediction to planning with goal conditioned lane graph traversals

    [Hallgartenet al., 2023 ] Marcel Hallgarten, Martin Stoll, and Andreas Zell. From prediction to planning with goal conditioned lane graph traversals. In2023 IEEE 26th In- ternational Conference on Intelligent Transportation Sys- tems (ITSC), pages 951–958. IEEE,

    2023

  24. [2020]

    GameFormer: Game-theoretic modeling and learn- ing of transformer-based interactive prediction and plan- ning for autonomous driving

    [Huanget al., 2023 ] Zhiyu Huang, Haochen Liu, and Chen Lv. GameFormer: Game-theoretic modeling and learn- ing of transformer-based interactive prediction and plan- ning for autonomous driving. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3903–3913,

    2023

  25. [2021]

    [Kumaret al., 2019 ] Aviral Kumar, Justin Fu, Matthew Soh, George Tucker, and Sergey Levine. Stabilizing off-policy q-learning via bootstrapping error reduction.Advances in neural information processing systems: 33rd Conference on Neural Information Processing Systems (NeurIPS), 32,

    2019

  26. [2022]

    Behavior Transformers: Cloning $k$ modes with one stone

    [Shafiullahet al., 2022 ] Nur Muhammad Mahi Shafiullah, Zichen Jeff Cui, Ariuntuya Altanzaya, and Lerrel Pinto. Behavior Transformers: Cloning $k$ modes with one stone. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Infor- mation Processing Systems: 36th Conference on Neural Information Processing Systems (...

    2022

  27. [2023]

    Bekey.Autonomous Robots: From Biological Inspiration to Implementation and Control

    [Bekey, 2005] George A. Bekey.Autonomous Robots: From Biological Inspiration to Implementation and Control. MIT Press, Cambridge, Massachusetts,

    2005

  28. [2024]

    Pluto: Pushing the limit of imitation learning- based planning for autonomous driving.arXiv preprint arXiv:2404.14327,

    [Chenget al., 2024 ] Jie Cheng, Yingbing Chen, and Qifeng Chen. Pluto: Pushing the limit of imitation learning- based planning for autonomous driving.arXiv preprint arXiv:2404.14327,

    Pith/arXiv arXiv 2024

  29. [2025]

    End-to-end Driving via Conditional Imitation Learning

    [Codevillaet al., 2018 ] Felipe Codevilla, Matthias M ¨uller, Antonio L´opez, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end Driving via Conditional Imitation Learning. In 2018 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 4693–4700. IEEE,

    2018