Pith. sign in

REVIEW 5 major objections 8 minor 49 references

Anomalous Decision Discovery using Inverse Reinforcement Learning

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Crash-bound driving can be flagged before the crash, without anomaly labels.

desk verdict Promising IRL-based anomaly detection pipeline with a real novelty, but the evaluation shares a noise mechanism between reward learning and test anomaly generation, so the headline 0.90 AUC needs a held-out noise control before generalization claims hold. read the letter →

arxiv 2507.04464 v1 pith:I7QRZ5V6 submitted 2025-07-06 cs.AI

Ashish Bastola , Mert D. Pesé , Long Cheng , Jonathon Smereka , Abolfazl Razi This is my paper
classification cs.AI
keywords inversereinforcementlearninganomalydetectionautonomousvehiclestrajectoryanalysisworst-caseterminustemporalcreditassignmentLiDARnoiserobustnessvariable-horizonsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that autonomous-driving anomalies can be flagged without labeled anomaly data by training a sequence classifier on a proxy task: predicting whether a trajectory segment ends in a worst-case terminus, defined as a crash or driving off the road. It first learns a reward function by ranking expert rollouts from noisiest to cleanest, then crops each crash-labeled trajectory to random lengths and expands it into many prefixes, each still labeled as crash-bound, and fine-tunes a transformer on state-action-reward embeddings of those segments. On more than 14,000 simulated trajectories the resulting TRAP system reaches 0.90 AUC and an 82.2% F1-score, with recall 39% above supervised baselines, and the authors attribute the gain to implicit temporal credit assignment: short random prefixes force the model to learn early warning signs rather than recognizing the crash itself at the end. A reader should care because the recipe replaces expensive anomaly annotations with a cheap binary 'did it end badly' label, which is much closer to what a real deployment can collect.

What carries the argument

The load-bearing mechanism is trajectory-ranked reward extrapolation combined with end-clipped variable-length sampling. The reward is trained with a softmax contrastive loss over expert rollouts ranked by added noise, so it encodes how far a segment deviates from expert behavior rather than an absolute optimality score; this sidesteps the standard maximum-entropy IRL assumption that experts are optimal. Then every trajectory that reaches a worst-case terminus is cropped from the end to a random length and expanded into many prefixes, each still labeled as crash-bound, and a transformer fine-tuned on these prefixes learns to associate early states with eventual catastrophe. The random-length structure is what implements temporal credit assignment: because short prefixes must predict the far-future crash label, the network is forced to detect behavior deviation early, which the paper identifies as the source of its generalization to unseen anomaly types.

What would settle it

Train TRAP exactly as described, then evaluate it on anomalies generated without any added LiDAR noise, for example by injecting a deliberately reckless driving policy in the simulator. If the F1-score collapses toward the supervised baselines, the learned reward is encoding noise intensity rather than behavioral badness; alternatively, compute the correlation between TRAP's anomaly score and the applied noise amplitude after controlling for rule-based anomaly labels, and check whether it remains significant.

Watch

Extended reading notes

Core claim

TRAP's central discovery is that time-to-consequence can be maximized without being explicitly optimized: a reward trained to rank noisy expert rollouts, combined with worst-case-terminus labels and random-length prefix sampling, makes a classifier learn which early states predict later catastrophe. The classification module receives embedding tokens produced by projecting state, action, and the learned reward into a transformer-compatible space, and training on randomly cropped crash-labeled prefixes means short segments inherit the crash label, so the network must flag deviations well before the collision point. In the paper's evaluation, this design reaches an F1 of 0.822 and an AUC of 0.90 across more than 14,000 simulated trajectories, outperforming similarly trained supervised and unsupervised baselines, and it maintains high F1 when entire anomaly groups are removed from the expert training set. The authors state that the method remains unsupervised with respect to anomaly categories: labels are used only for validation, never for training.

Load-bearing premise

The reward function is trained to rank rollouts by their added noise level, and the test anomalies are produced by adding the same kinds of noise to the same driving policy; if the reward is measuring noise magnitude rather than safety-relevant badness, the reported ability to catch unseen anomaly classes may just be detection of that shared noise pattern.

Editorial extensions

If this is right

  • Anomaly detection for autonomous vehicles could be trained from unlabeled expert trajectories plus a cheap crash/off-road signal, removing the need for large hand-annotated anomaly datasets.
  • The same learned reward and embedding can be reused for new anomaly categories without retraining, as the group-removal experiments indicate.
  • Random-length cropping converts a trajectory-level outcome label into segment-level supervision, so detection happens earlier than end-of-trajectory classification would allow.
  • Sensor noise robustness can be achieved without explicit denoising: F1 stays in the 0.82-0.88 range across eight LiDAR noise types, including composite noise.

Reading between the lines

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

  • If the reward genuinely encodes safety-relevant deviation rather than noise intensity, the same reward could be reused as a dense safety signal during online reinforcement learning of driving policies, a use the paper does not explore.
  • The variable-horizon prefix trick should transfer to any temporal domain with a definable worst-case terminus, such as industrial process monitoring or delayed-fraud detection, and could be tested there directly.
  • An ablation that trains TRAP on full trajectories only, without random-length cropping, would isolate how much of the early-detection gain comes from the variable-horizon mechanism; the paper does not report such an ablation.
  • Until the method is evaluated on real-world trajectory logs, the reported 0.90 AUC should be read as a simulation-bound result, since the benchmark's anomalies and expert data come from the same simulator policy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper proposes TRAP, an inverse-reinforcement-learning-based framework for detecting anomalous driving behavior in autonomous vehicles. Reward learning is performed by ranking expert rollouts perturbed with different noise intensities, after which worst-case-terminus trajectories are cropped at variable lengths, projected through a feedforward network, and used to fine-tune DistilBERT for binary worst-case classification. The authors contribute a 14,000+ trajectory MetaDrive dataset with six rule-based anomaly categories and eight LiDAR noise conditions, and report state-of-the-art results (0.90 AUC, 82.2% F1) against supervised, unsupervised, and IRL baselines, including noise-robustness and anomaly-group-removal experiments. The central claims are that the method is robust to sensor noise and generalizes to unseen anomaly types while enabling early detection through implicit temporal credit assignment.

Significance. If the central claims held, TRAP would be a useful contribution: the combination of reward-based trajectory ranking with variable-horizon pretraining for early anomaly detection is a plausible and interesting idea, and the released-style benchmark of 14,000+ trajectories with rule-based behavioral labels and eight noise groups would be a valuable community resource. The paper also deserves credit for including both supervised and unsupervised baselines and for reproducing a prior IRL method rather than only citing it. However, the evaluation as presented does not yet establish that the method detects behavioral anomalies rather than a shared noise signature, and the formal claims about implicit optimization are not supported by derivations or ablations. The benchmark and code availability are stated but the code is not yet released and no download link is provided, which limits reproducibility. The contribution is significant in scope and ambition, but the evidence is currently insufficient to support the headline generalization and robustness claims.

major comments (5)
  1. [IV-A and V-A] The reward function is trained by ranking expert rollouts according to injected noise intensity (Section IV-A, Eq. 4: 'the label being the hierarchical ordering that we already know based on the added noise to expert policy'), while the test anomalies are generated by applying the same eight LiDAR noise groups to the same expert/IDM policy (Section V-A). This shared noise-injection mechanism means the learned reward may encode 'noisier implies lower reward', and the high reported F1/AUC may reflect detection of a common perturbation pattern rather than discovery of distinct behavioral anomalies. Table IV further shows that noise causes substantial trajectory displacement, and Table II's unusual result that F1 and MCC increase with noise intensity is consistent with the model exploiting noise as a shortcut. To support the 'unseen anomaly types' and 'noise robustness' claims, the authors should report a held-out noise-type split (e.g., train reward ranking on some noise groups and test on others), include noise-matched normal trajectories as controls, and state explicitly which noise types, if any, are shared between reward training and test anomaly generation.
  2. [IV-E, Eqs. (8)-(11)] The 'implicit optimization strategy' is presented as solving three interconnected optimization problems, but Eqs. (8)-(11) are not derived from any implemented objective. Equation (4) contains no KL-divergence term or λ weight, so Eq. (8) does not describe the actual loss being minimized. Equation (9) writes min_φ over an expression involving 1_WCT, but no such adversarial minimization is implemented anywhere in the algorithm. Equation (10) asserts a conditional probability proportional to exp(−β T_WCT/t_d Σ r̂_θ) without derivation, and Eq. (11) introduces weights w_k that are never defined. Since 'implicitly learning temporal credit assignments' and 'maximizing time-to-consequence' are the paper's core claimed innovations, the authors must either derive these expressions rigorously from the actual training procedure (Eq. (4), Eq. (6), and the DistilBERT fine-tuning) or remove them and replace them with a precise description of what the variable-horizon sampling actually optimizes.
  3. [Tables I, II, III] All performance numbers are reported as point estimates without error bars, repeated-seed statistics, or significance tests. Given that trajectories are generated with random seeds and the deep models (DistilBERT, LSTM, SPAE) involve stochastic training, the differences between TRAP (F1 0.822) and Isolation Forest (F1 0.688) in Table I, or the small noise-specific differences in Table II, cannot be assessed for reliability. The authors should run at least 3-5 seeds for the main experiments and report means and standard deviations, and ideally paired statistical comparisons for the headline results. This is load-bearing because the abstract claims 'state-of-the-art performance' and specific percentage improvements over baselines.
  4. [III-B] The rule-based anomaly labels that serve as evaluation ground truth are defined using several thresholds and window sizes, but most are not reported: the curvature threshold for zigzag, the braking threshold θ, the lateral-acceleration threshold τ (one value, 0.8 m/s², is given), the lane-change interval threshold τ, the window size w, the temporal window size 40, and the lane-change convolution threshold 0.5. The text says thresholds are set 'based on human observation' or are 'user-validated', which makes the evaluation labels partly subjective. Since all reported metrics depend on these labels, the paper should list all threshold values and include a sensitivity analysis showing how the anomaly labels and final metrics change with reasonable threshold variations.
  5. [V-C] The 'unseen anomaly type' experiments in Table III remove certain anomaly groups from the expert training set, but the test trajectories are still generated by the same noise-injection mechanism that defines the anomaly classes, and the evaluation groups are still formed by the same rule-based kinematics. Without a demonstration that the removed groups are perceptually distinct from the noise signature used during reward training, this experiment does not show generalization to genuinely unseen behavioral anomalies. A more convincing protocol would train on clean expert demonstrations only, inject a truly novel failure mode (e.g., cyber-attack-induced steering or braking faults) at test time, and evaluate whether the model flags those trajectories without ever seeing that perturbation during reward learning.
minor comments (8)
  1. [Abstract and Table I] The abstract's claim of improving 'by 39% on Recall and 12% on F1-score, respectively' does not match Table I: TRAP recall 0.902 vs. IForest 0.539 is a 67% relative improvement, and F1 0.822 vs. IForest 0.688 is a 19.5% relative improvement; please clarify which baseline and which relative formula are used for the '39%' and '12%' figures.
  2. [Tables I and II] The footnote markers '**' and '*' on 'Ours (Final)**' in Table I and 'Ours*' in Table II are never explained; please either define the footnotes or remove the markers.
  3. [IV-C, Eq. (6)] The symbol D' is reused inconsistently: Eq. (6) defines D' as the expanded set of positive-label trajectories, but the following paragraph redefines 'the final training dataset D′ is given by D′ = D ∪ D′', which conflates the expanded set with the final training set. Please use distinct names for the two sets and clarify the sampling procedure for negative trajectories.
  4. [IV-E, Eq. (8)] Equation (8) uses the symbol τ in τ̂_θ(s_t), which is neither defined nor used elsewhere; this appears to be a typo for the reward function r̂_θ(s_t).
  5. [III-B, item 3] In the sudden-turns subsection, 'across' should be 'arccos' in the description of the argument clamping, and the definition of the heading vectors h_i should state whether they come from absolute vehicle heading or from position differences.
  6. [References] Reference [7] cites a DataCamp tutorial as the source for the claim about unsupervised methods' limitations; a peer-reviewed or archival reference is more appropriate for this claim.
  7. [V-A and Fig. 7] The explanation for the counterintuitive result that accuracy increases with noise (Section V-A) is speculative ('noise can enhance separation') and is not backed by a quantitative measure of distribution shift or a control experiment; even if the overall claim is retained, it should be tested, not just asserted with a t-SNE figure.
  8. [General] The statement 'Our code will be available at: https://github.com/abastola0/TRAP.git' appears only in the abstract and the linked repository is not yet populated or described; since reproducibility is a key strength the authors claim, they should either include a detailed repository alongside the paper or state the current availability status more precisely.

Circularity Check

1 steps flagged · score 6.0 of 10

IRL reward is fit to injected-noise orderings while test anomalies are produced by the same added-noise mechanism, so the headline 'unseen anomaly type' result partially reduces to re-detecting the reward's own training signal.

  1. fitted input called prediction [Section IV-A (Reward Learning, Eq. 4) with Section V-A (Noise Robustness) and Fig. 7 caption]
    "With this approach, we first position the expert policy with varying noise intensity and then rank the expert rollouts τ1, . . . , τm, from worst to best. ... This objective is trained by predicting whether one trajectory is preferable to another, with the label being the hierarchical ordering that we already know based on the added noise to expert policy. ... In our implementation, we thus generate these crashes by different ways of traffic density variation and adding noise to the expert policy. ... Figure shows distribution shift with added noise (type: Composite2 high)"

    The reward r_theta is fit by Eq. (4) to pairwise preference labels that are 'based on the added noise to expert policy,' so low learned return means high injected noise. The test set supporting the 'unseen anomaly type' claim is then generated from the same policy under added noise ('we thus generate these crashes by ... adding noise to the expert policy'; Fig. 7 shows test trajectories with 'added noise (type: Composite2 high)'). A detector built on the fitted reward can therefore rank test trajectories by noise magnitude, the very quantity that produced the rule-labeled kinematic deviations. The paper's own statement of 'increased accuracy with increased noise' is consistent with this shortcut.

full rationale

The central circular step is fitted-input-called-prediction, not self-citation. Section IV-A learns r_theta using preference labels derived from the noise level added to expert rollouts; Eq. (4) is optimized so that reward sums reproduce that noise ordering. The robustness evaluation (Section V-A) and Figure 7 then score the model on test trajectories whose rule-based anomalies arise under the same added-noise perturbations, so a model that has simply learned 'noisier is worse' can achieve high recall and F1 without modeling the kinematic anomaly categories named in Section III-B. The authors' own note of 'increased accuracy with increased noise' corroborates the shortcut. This reduces the 'generalization to unseen anomaly types' claim. However, the paper is not wholly circular: rule-based labels are used only for validation, and the no-noise evaluation (Section V-B) gives some independent signal, so a score of 6 (partial reduction) rather than 8-10 is appropriate. No load-bearing self-citation, imported uniqueness theorem, or ansatz-via-citation pattern was found.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The method rests on several unverified assumptions: that noise intensity is a valid ranking for trajectory quality (Section IV-A), that crash/off-road is an adequate worst-case terminus for all safety-critical anomalies (Section IV-B), that rule-based kinematic labels are valid ground truth (Section III-B), and that prefix sampling of crash trajectories transfers to early detection (Section IV-C). These are engineering assumptions, not standard mathematical axioms, and they are load-bearing for the reported generalization claims.

free parameters (7)
  • Lateral acceleration threshold tau = 0.8 m/s^2
    Used to label sudden-turn anomalies (Section III-B.3); value taken from comfort literature [32].
  • Braking threshold theta = not reported
    Threshold on smoothed acceleration for sudden-braking labels (Section III-B.2).
  • Curvature threshold for zigzag = not reported
    Hand-picked based on human observation (Section III-B.1).
  • Lane-change interval threshold tau = not reported
    Minimum interval length for frequent lane-switch anomaly (Section III-B.4).
  • Prefix sample count eta = not reported
    Number of random prefixes generated per crash trajectory (Section IV-C).
  • Max crop length Lthres = not reported
    Upper bound for random trajectory cropping (Section IV-C).
  • Noise intensities (low/med/high) = not reported
    Magnitudes of the eight LiDAR noise types (Section V-A).
assumptions (6)
  • domain assumption Noise intensity ranks trajectory quality
    Section IV-A ranks expert rollouts by added noise and treats higher noise as worse; if noise is not monotonically linked to unsafe behavior, the reward is mis-calibrated.
  • domain assumption Crash or off-road termination is a sufficient worst-case terminus proxy for safety-critical anomalies
    Section IV-B defines WCT as crash/off-road; anomalies that do not lead to such terminus are considered low interest, so detection is aimed at crash precursors.
  • domain assumption Rule-based kinematic labels are valid ground truth for anomaly evaluation
    Section III-B uses thresholds on heading, acceleration, and lane crossings to label zigzag, braking, turns, lane switching, tailgating; these thresholds are chosen by hand or from literature and are used only for validation.
  • ad hoc to paper Prefix sampling of crash trajectories with positive labels teaches early detection
    Section IV-C assumes that labeling all prefixes of a crash trajectory as positive forces the classifier to recognize early warning signs; no proof is given that early prefixes are informative.
  • ad hoc to paper DistilBERT finetuning on projected reward tokens captures temporal credit assignment
    Section IV-D/E assumes the transformer learns length-invariant representations of danger from variable-length inputs; this is the core mechanism but is only empirically motivated.
  • standard math Trajectory-ranked reward extrapolation loss is valid for this setting
    The paper adopts Brown et al. [34] preference-based IRL and assumes its loss (Eq 4-5) yields a useful reward under noise-ranked demonstrations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anomalous Decision Discovery using Inverse Reinforcement Learning." pith.science (2026). https://pith.science/paper/I7QRZ5V6

@misc{pith2026250704464,
  author       = {Pith},
  title        = {Pith review of: Anomalous Decision Discovery using Inverse Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7QRZ5V6}},
  note         = {Machine review of arXiv:2507.04464}
}
read the original abstract

Anomaly detection plays a critical role in Autonomous Vehicles (AVs) by identifying unusual behaviors through perception systems that could compromise safety and lead to hazardous situations. Current approaches, which often rely on predefined thresholds or supervised learning paradigms, exhibit reduced efficacy when confronted with unseen scenarios, sensor noise, and occlusions, leading to potential safety-critical failures. Moreover, supervised methods require large annotated datasets, limiting their real-world feasibility. To address these gaps, we propose an anomaly detection framework based on Inverse Reinforcement Learning (IRL) to infer latent driving intentions from sequential perception data, thus enabling robust identification. Specifically, we present Trajectory-Reward Guided Adaptive Pre-training (TRAP), a novel IRL framework for anomaly detection, to address two critical limitations of existing methods: noise robustness and generalization to unseen scenarios. Our core innovation is implicitly learning temporal credit assignments via reward and worst-case supervision. We leverage pre-training with variable-horizon sampling to maximize time-to-consequence, resulting in early detection of behavior deviation. Experiments on 14,000+ simulated trajectories demonstrate state-of-the-art performance, achieving 0.90 AUC and 82.2\% F1-score - outperforming similarly trained supervised and unsupervised baselines by 39\% on Recall and 12\% on F1-score, respectively. Similar performance is achieved while exhibiting robustness to various noise types and generalization to unseen anomaly types. Our code will be available at: https://github.com/abastola0/TRAP.git

Figures

Figures reproduced from arXiv: 2507.04464 by the authors.

Figure 1
Figure 1. Anomalies classified using a rule-based approach. We classify common [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Hybrid IRL-based reward guidance in supervised fine-tuning of variable horizon sampled trajectories labeled for worst-case terminus. By making easy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Trajectory-Reward fusion to generate DistilBERT compatible embed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: t-SNE plot of anomalous segments of test trajectories labeled using the rule-based approach described in Section III-B. Figure shows distribution shift [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: ROC curve demonstrating the performance comparison of some widely [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: t-SNE plot of 14000+ full test trajectories labeled using the rule-based approach described in III-B for evaluation. Figure shows distribution shift with [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Figure shows the percentage of anomalies identified by unsupervised [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 37 canonical work pages

  1. [1]

    Systematic literature review: Anomaly detection in connected and autonomous vehicles,

    J. R. V . Solaas, E. Mariconti, and N. Tuptuk, “Systematic literature review: Anomaly detection in connected and autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems , 2024

    work page 2024

  2. [2]

    On evaluating black-box explainable ai methods for enhancing anomaly detection in autonomous driving systems,

    S. Nazat, O. Arreche, and M. Abdallah, “On evaluating black-box explainable ai methods for enhancing anomaly detection in autonomous driving systems,” Sensors, vol. 24, no. 11, p. 3515, 2024. 10

    work page 2024

  3. [3]

    Anomaly detection in autonomous driving: A survey,

    D. Bogdoll, M. Nitsche, and J. M. Z ¨ollner, “Anomaly detection in autonomous driving: A survey,” inProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , 2022, pp. 4488–4499

    2022

  4. [4]

    Outlier-robust inverse reinforcement learning and reward-based detec- tion of anomalous driving behaviors,

    D. Li, M. L. Shehab, Z. Liu, N. Ar ´echiga, J. DeCastro, and N. Ozay, “Outlier-robust inverse reinforcement learning and reward-based detec- tion of anomalous driving behaviors,” in 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2022, pp. 4175–4182

    work page 2022

  5. [5]

    Sensor data based anomaly detection in autonomous vehicles using modified convolutional neural network

    S. Rajendar and V . K. Kaliappan, “Sensor data based anomaly detection in autonomous vehicles using modified convolutional neural network.” Intelligent Automation & Soft Computing , vol. 32, no. 2, 2022

    work page 2022

  6. [6]

    Machine learning- based real-time anomaly detection using data pre-processing in the telemetry of server farms,

    D. L. Vajda, T. V . Do, T. B ´erczes, and K. Farkas, “Machine learning- based real-time anomaly detection using data pre-processing in the telemetry of server farms,” Scientific Reports, vol. 14, no. 1, p. 23288, 2024

    work page 2024

  7. [7]

    Introduction to anomaly detection,

    DataCamp, “Introduction to anomaly detection,” 2025, accessed: 2025-03-09. [Online]. Available: https://www.datacamp.com/tutorial/ introduction-to-anomaly-detection

    work page 2025

  8. [8]

    Anomaly detection in connected and autonomous vehicle trajectories using lstm autoencoder and gaussian mixture model,

    B. Wang, W. Li, and Z. H. Khattak, “Anomaly detection in connected and autonomous vehicle trajectories using lstm autoencoder and gaussian mixture model,” Electronics, vol. 13, no. 7, p. 1251, 2024

    work page 2024

Show all 49 references
  1. [9]

    Sequential anomaly detection using inverse reinforcement learning,

    M.-h. Oh and G. Iyengar, “Sequential anomaly detection using inverse reinforcement learning,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & data mining, 2019, pp. 1480–1490

    2019

  2. [10]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,

    Q. Li, Z. Peng, L. Feng, Q. Zhang, Z. Xue, and B. Zhou, “Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,” IEEE transactions on pattern analysis and machine intelli- gence, vol. 45, no. 3, pp. 3461–3475, 2022

    2022

  3. [11]

    Outlier detection by active learning,

    N. Abe, B. Zadrozny, and J. Langford, “Outlier detection by active learning,” in Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , 2006, pp. 504– 509

    2006

  4. [12]

    Neural batch sampling with reinforcement learning for semi-supervised anomaly detection,

    W.-H. Chu and K. M. Kitani, “Neural batch sampling with reinforcement learning for semi-supervised anomaly detection,” in Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVI 16 . Springer, 2020, pp. 751–766

    2020

  5. [13]

    Dolphin: An efficient algorithm for mining distance-based outliers in very large datasets,

    F. Angiulli and F. Fassetti, “Dolphin: An efficient algorithm for mining distance-based outliers in very large datasets,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 3, no. 1, pp. 1–57, 2009

    2009

  6. [14]

    Distance-based outliers: algorithms and applications,

    E. M. Knorr, R. T. Ng, and V . Tucakov, “Distance-based outliers: algorithms and applications,” The VLDB Journal, vol. 8, no. 3, pp. 237– 253, 2000

    2000

  7. [15]

    Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,

    X. Zhang, M. Xu, and X. Zhou, “Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 16 699–16 708

    2024

  8. [16]

    Lof: identifying density-based local outliers,

    M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof: identifying density-based local outliers,” in Proceedings of the 2000 ACM SIGMOD international conference on Management of data , 2000, pp. 93–104

    2000

  9. [17]

    Anomaly detection using local kernel density estimation and context-based regression,

    W. Hu, J. Gao, B. Li, O. Wu, J. Du, and S. Maybank, “Anomaly detection using local kernel density estimation and context-based regression,” IEEE Transactions on Knowledge and Data Engineering , vol. 32, no. 2, pp. 218–233, 2018

    2018

  10. [18]

    Bootstrap fine-grained vision-language alignment for unified zero-shot anomaly localization,

    H. Deng, Z. Zhang, J. Bao, and X. Li, “Bootstrap fine-grained vision-language alignment for unified zero-shot anomaly localization,” https://arxiv.org/abs/2308.15939, 2023

    2023 arXiv

  11. [19]

    Discovering cluster-based local outliers,

    Z. He, X. Xu, and S. Deng, “Discovering cluster-based local outliers,” Pattern recognition letters, vol. 24, no. 9-10, pp. 1641–1650, 2003

    2003

  12. [20]

    Isolation forest,

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in 2008 eighth ieee international conference on data mining. IEEE, 2008, pp. 413–422

    2008

  13. [21]

    Guided cost learning: Deep inverse optimal control via policy optimization,

    C. Finn, S. Levine, and P. Abbeel, “Guided cost learning: Deep inverse optimal control via policy optimization,” in International conference on machine learning. PMLR, 2016, pp. 49–58

    2016

  14. [22]

    Anomaly detection and correction of optimizing autonomous systems with inverse reinforcement learning,

    B. Lian, Y . Kartal, F. L. Lewis, D. G. Mikulski, G. R. Hudas, Y . Wan, and A. Davoudi, “Anomaly detection and correction of optimizing autonomous systems with inverse reinforcement learning,” IEEE Trans- actions on Cybernetics , vol. 53, no. 7, pp. 4555–4566, 2022

    2022

  15. [23]

    Anovox: A benchmark for multimodal anomaly detection in autonomous driving,

    D. Bogdoll, I. Hamdard, L. N. R ¨oßler, F. Geisler, M. Bayram, F. Wang, J. Imhof, M. de Campos, A. Tabarov, Y . Yang et al. , “Anovox: A benchmark for multimodal anomaly detection in autonomous driving,” arXiv preprint arXiv:2405.07865 , 2024

    2024 arXiv

  16. [24]

    Modernisation of carla - can we define the what, and how and where?

    shott92, “Modernisation of carla - can we define the what, and how and where?” https://github.com/carla-simulator/carla/discussions/6237, March 2023, gitHub Discussion #6237

    2023

  17. [25]

    Detecting anomalies in semantic segmentation with prototypes,

    D. Fontanel, F. Cermelli, M. Mancini, and B. Caputo, “Detecting anomalies in semantic segmentation with prototypes,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 113–121

    2021

  18. [26]

    Augmenting anomaly detection for autonomous vehicles with symbolic rules,

    T. Chen et al., “Augmenting anomaly detection for autonomous vehicles with symbolic rules,” Ph.D. dissertation, Massachusetts Institute of Technology, 2019

    2019

  19. [27]

    Over-the-air: How we remotely compromised the gateway, bcm, and autopilot ecus of tesla cars,

    S. Nie, L. Liu, Y . Du, and W. Zhang, “Over-the-air: How we remotely compromised the gateway, bcm, and autopilot ecus of tesla cars,” Briefing, Black Hat USA , vol. 91, pp. 1–19, 2018

    2018

  20. [28]

    Free-fall: Hacking tesla from wireless to can bus,

    S. Nie, L. Liu, and Y . Du, “Free-fall: Hacking tesla from wireless to can bus,” Briefing, Black Hat USA , vol. 25, no. 1, p. 16, 2017

    2017

  21. [29]

    Orads: One class rule-based anomaly detection system in autonomous vehicles,

    A. Kumar and T. K. Das, “Orads: One class rule-based anomaly detection system in autonomous vehicles,” IEEE Sensors Journal, 2025

    2025

  22. [30]

    An association rules-based approach for anomaly detection on can-bus,

    G. D’Angelo, M. Ficco, and A. Robustelli, “An association rules-based approach for anomaly detection on can-bus,” in International conference on computational science and its applications. Springer, 2023, pp. 174– 190

    2023

  23. [31]

    Sce- narionet: Open-source platform for large-scale traffic scenario simulation and modeling,

    Q. Li, Z. M. Peng, L. Feng, Z. Liu, C. Duan, W. Mo, and B. Zhou, “Sce- narionet: Open-source platform for large-scale traffic scenario simulation and modeling,” Advances in neural information processing systems , vol. 36, pp. 3894–3920, 2023

    2023

  24. [32]

    Standards for passenger comfort in automated vehicles: Acceleration and jerk,

    K. N. De Winkel, T. Irmak, R. Happee, and B. Shyrokau, “Standards for passenger comfort in automated vehicles: Acceleration and jerk,” Applied Ergonomics, vol. 106, p. 103881, 2023

    2023

  25. [33]

    Maximum entropy inverse reinforcement learning

    B. D. Ziebart, A. L. Maas, J. A. Bagnell, A. K. Dey et al., “Maximum entropy inverse reinforcement learning.” in Aaai, vol. 8. Chicago, IL, USA, 2008, pp. 1433–1438

    2008

  26. [34]

    Extrapolating beyond suboptimal demonstrations via inverse reinforcement learning from observations,

    D. Brown, W. Goo, P. Nagarajan, and S. Niekum, “Extrapolating beyond suboptimal demonstrations via inverse reinforcement learning from observations,” in International conference on machine learning . PMLR, 2019, pp. 783–792

    2019

  27. [35]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V . Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,” arXiv preprint arXiv:1910.01108, 2019

    1910 arXiv

  28. [36]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo...

    2019

  29. [37]

    Receding horizon control of nonlinear systems,

    D. Q. Mayne and H. Michalska, “Receding horizon control of nonlinear systems,” in Proceedings of the 27th IEEE Conference on Decision and Control. IEEE, 1988, pp. 464–465

    1988

  30. [38]

    An optimization-based receding horizon trajectory planning algorithm,

    K. Bergman, O. Ljungqvist, T. Glad, and D. Axehill, “An optimization-based receding horizon trajectory planning algorithm,” IFAC-PapersOnLine, vol. 53, no. 2, pp. 15 550–15 557, 2020

    2020

  31. [39]

    Efficient driving with automated and connected vehicles algorithms, microsimulations, and cyber-physical experiments,

    A. Vahidi, “Efficient driving with automated and connected vehicles algorithms, microsimulations, and cyber-physical experiments,” IEEE Transactions on Intelligent Transportation Systems , vol. 25, no. 4, pp. 214–256, 2024

    2024

  32. [40]

    One-class svms for document clas- sification,

    L. M. Manevitz and M. Yousef, “One-class svms for document clas- sification,” Journal of machine Learning research , vol. 2, no. Dec, pp. 139–154, 2001

    2001

  33. [41]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  34. [42]

    Induction of decision trees,

    J. R. Quinlan, “Induction of decision trees,” Machine learning, vol. 1, pp. 81–106, 1986

    1986

  35. [43]

    Supervised autoencoders: Improving generalization performance with unsupervised regularizers,

    L. Le, A. Patterson, and M. White, “Supervised autoencoders: Improving generalization performance with unsupervised regularizers,” Advances in neural information processing systems , vol. 31, 2018

    2018

  36. [44]

    R. Smith. (n.d.) Lidar modelling with weather. Rob Smith is identified as ”AD Simulation Engineer” at Claytex. [Online]. Available: https://www.claytex.com/blog/adas/lidar-modelling-with-weather/

  37. [45]

    A methodology to model the rain and fog effect on the performance of automotive lidar sensors,

    A. Haider, M. Pigniczki, S. Koyama, M. H. K ¨ohler, L. Haas, M. Fink, M. Schardt, K. Nagase, T. Zeh, A. Eryildirim et al., “A methodology to model the rain and fog effect on the performance of automotive lidar sensors,” Sensors, vol. 23, no. 15, p. 6891, 2023

    2023

  38. [46]

    Predicting the influence of rain on lidar in adas,

    C. Goodin, D. Carruth, M. Doude, and C. Hudson, “Predicting the influence of rain on lidar in adas,” Electronics, vol. 8, no. 1, p. 89, 2019

    2019

  39. [47]

    Lidar data noise models and methodology for sim-to-real domain generalization and adaptation in autonomous driving perception,

    J. Espadinha, I. Lebedev, L. Lukic, and A. Bernardino, “Lidar data noise models and methodology for sim-to-real domain generalization and adaptation in autonomous driving perception,” in 2021 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2021, pp. 797–803

    2021

  40. [48]

    Dlp ® dmd technology : Lidar ambient light reduction,

    “Dlp ® dmd technology : Lidar ambient light reduction,” 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:186200652

    2018

  41. [49]

    Learning to drop points for lidar scan synthesis,

    K. Nakashima and R. Kurazume, “Learning to drop points for lidar scan synthesis,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 222–229

    2021

Pith tools

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