Pith. sign in

REVIEW 3 major objections 6 minor 68 references

ADReFT: Adaptive Decision Repair for Safe Autonomous Driving via Reinforcement Fine-Tuning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Adaptive repair fixes 85% of driving collision cases

desk verdict A genuinely useful two-stage repair idea that deserves a serious referee, but cannot be accepted as-is until the evaluation adds a clean train/test split and variance estimates. read the letter →

arxiv 2506.23960 v1 pith:AJGRBEQI submitted 2025-06-30 cs.LG cs.AIcs.SE

classification cs.LGcs.AIcs.SE
keywords autonomousdrivingsafetyonlinerepairruntimeenforcementreinforcementfine-tuningstatemonitoringdecisionadaptationdeepQ-networkCARLAsimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that a deployed autonomous driving system can be repaired online, without retraining the whole system, by a learned module that spots risky driving states before collisions happen and chooses a corrective throttle-braking action scaled to the danger. The proposed framework, ADReFT, is trained in two stages: a supervised warm-up on coarse labels (states just before a violation are 'unsafe'), followed by reinforcement fine-tuning that relaxes the warm-up's overly harsh braking to find the mildest repair that still works. On CARLA simulations, the paper reports that ADReFT fixes 85% of collision cases for the end-to-end system Roach and 76% for the modular system Pylot, improving overall safety by an average of 80.5% and beating the best baseline by 32.8% while keeping average intervention intensity to 0.77 and 0.72. A sympathetic reader would care because online repair is what can keep already-shipped vehicles safe after testing finds corner-case failures, without waiting for a full software redesign.

What carries the argument

The load-bearing mechanism is the two-headed ADReFT model on a transformer encoder. A scene tokenizer turns the ego vehicle, nearby participants, and planned trajectory points into object tokens; a BERT-style transformer captures interactions; the [CLS] token becomes a global state representation. From that representation, a State Monitor head outputs a safety-critical score, and a Decision Adapter head outputs a repair action over ten throttle-braking levels. The training machinery is what makes the design work: a Weak Annotation Generator labels frames as unsafe if they are within 1 meter of an obstacle or within 3 seconds of a collision, and the supervised stage learns to associate unsafe states with maximum braking; the reinforcement fine-tuning stage then freezes everything except the Decision Adapter and uses DQN with a safe-explore reward (one minus a weighted intervention-distance term) plus a -10 collision penalty to find less aggressive repairs that still avoid the collision.

What would settle it

Run a held-out evaluation by splitting the 400 scenarios per ADS into training and test subsets before the ReFT stage; if the percentage of fixed collisions on truly unseen scenarios falls to the baseline range of about 55% for Roach and 40% for Pylot instead of the reported 85% and 76%, the gain came from replaying test cases during training rather than from learned generalization.

Watch

Extended reading notes

Core claim

The central discovery claimed is that adaptive repair decisions, not just collision detection, are learnable from existing test suites. ADReFT's State Monitor scores every driving state by jointly encoding traffic interactions and the ADS's own decision, and its Decision Adapter picks one of ten discrete throttle-braking levels; the Decision Merger only overrides the ADS when the monitor's score exceeds a threshold and the proposed action is more conservative than the ADS's. Positioned this way, the paper claims two results hold at once: safety-critical states are detected earlier and more reliably than rule-based or anomaly-based monitors, and the chosen repair minimizes deviation from the original driving behavior. Concretely, it reports an average overall safety improvement of 85.0% for Roach and 76.0% for Pylot, with intervention intensity of 0.77 and 0.72, and robust behavior across three fuzzing-based testing techniques.

Load-bearing premise

The reported repair rates depend on the collision cases used to score the method not being the same cases the reinforcement fine-tuning stage replays, and the paper never states that the benchmark and the replay corpus are disjoint.

Editorial extensions

If this is right

  • If ADReFT's numbers hold, the same two-stage recipe should transfer to other end-to-end and modular ADSs, since the inputs are object attributes rather than model-specific internals.
  • Because the Decision Merger only overrides when the proposed action is more conservative than the ADS's, correct emergency actions by the ADS are preserved; occasional overrides that break previously successful scenarios remain possible, and the paper reports low but non-zero degradation percentages.
  • The low runtime of 3.421 ms per frame, versus 25 ms for Roach and 89 ms for Pylot, means the repair module can run in the loop without blocking the vehicle's normal control cycle.
  • Robustness tests with three fuzzers show the repaired ADS stays at a stable low violation count as fuzzing time grows, meaning the monitor does not need to know the exact failure scenario in advance.

Reading between the lines

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

  • If a strict train/test split confirms the reported numbers, the weak-annotation-plus-RL-fine-tuning recipe is a general template for runtime repair of learned controllers beyond driving; the same two-stage idea could apply to robots or drones where failure traces are easy to generate but fine-grained state labels are not.
  • The paper's perfect-perception setup means the monitor sees ground-truth object attributes; under real perception noise the learned safety scores would likely degrade, so a natural next test is feeding noisy detections and measuring how much of the reported 85% and 76% survives.
  • The action space is limited to longitudinal throttle-braking; the framework's own logic suggests that adding steering-level repair actions would extend it to collisions that braking cannot avoid, though the paper leaves that to future work.
  • The missing train/test split is the main open question; resolving it by treating the 400 scenarios per ADS as a shared pool and cross-validating would tell whether the improvement is generalization or memorization.
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

3 major / 6 minor

Summary. ADReFT is a runtime repair module for autonomous driving systems. It combines a transformer-based state encoder with two heads: a State Monitor that predicts a safety-critical score and a Decision Adapter that selects a longitudinal throttle-braking repair action. Training proceeds in two stages: supervised warm-up on weakly annotated states (states within 1 m of obstacles or within 3 s before a collision are positive), followed by DQN-based reinforcement fine-tuning of the Decision Adapter only, using a step reward weighted by the frozen monitor and a terminal penalty for collisions. The method is evaluated in CARLA on Roach (end-to-end) and Pylot (modular) over five scenario types with 400 scenarios per ADS, comparing against RuleRepair (TTC-based), ModelRepair (OOD-based), and Random baselines. The paper reports average repair rates of 90.5% (Roach) and 78.5% (Pylot), average Delta-E improvements of 85.0 and 76.0, and average intervention intensities of 0.77 and 0.72, and claims a 32.8% improvement over the best baseline and a 25.7% improvement in intensity.

Significance. If the reported results are evaluated on scenarios that were not seen during ReFT, ADReFT would be a useful contribution to online repair: it learns a state-dependent repair policy rather than applying emergency braking uniformly, and it addresses a real gap in the runtime-safety literature. The two-stage training scheme and the use of a learned monitor as a reward component are interesting ideas. The paper also includes ablations (only SL, only RL, wo SR, Rand-R) that help understand the contribution of each stage. However, the evaluation as presented does not establish generalization, because the train/test relationship is not stated and several headline numbers are internally inconsistent. With a clarified held-out evaluation and corrected reporting, the contribution could be solid.

major comments (3)
  1. [4 (Benchmark, Training Details) and Algorithm 1] The manuscript never states a train/test split: the Benchmark paragraph describes 400 scenarios per ADS collected from testing techniques, and Training Details says that during ReFT 'we randomly sample both success and violation scenarios from the corpus for replay'; Algorithm 1 (Line 5) samples S_e from the same Scorpus. If the 400 scenarios used to build Tables 1 and 2 are the same as those replayed in ReFT, the reported %Fix values of 90.5 (Roach) and 78.5 (Pylot) and the Delta-E gains are training-set performance rather than evidence of generalization. The paper must either state a disjoint split explicitly (including how the 400 scenarios per ADS were partitioned before SL/ReFT) or re-run the evaluation on held-out scenarios and report repair rates there. The robustness experiment in Figure 3, which reports only qualitative violation counts over time, cannot substitute for this held-out evaluation.
  2. [Abstract, Introduction (Section 1), Tables 1-2] The headline numbers 'outperforming the best baseline by 32.8%' and 'by 25.7%' are not supported by Tables 1 and 2. For %Intensity, the best baseline for Roach is Random (0.78), not RuleRepair (1.11); ADReFT's 0.77 is only 0.01 lower, so the claimed 25.7% improvement uses a non-best baseline for Roach. For Delta-E, ADReFT's average advantage over RuleRepair is 80.5 - 47.75 = 32.75 percentage points, which is a percentage-point difference, not a relative improvement of 32.8%; the relative improvement would be about 69%. Please re-report these metrics consistently and separately for each ADS, or state clearly that they are percentage-point differences.
  3. [4.2.1 (Table 3) and Training Details] There is an internal inconsistency in the threshold reporting: the Training Details paragraph says lambda_safe is set so that 'more than 95% of positive samples are correctly recognized,' yet Table 3 reports ADReFT recall on the test set of only 80.09 (Roach) and 80.34 (Pylot). If the training and test sets are identical, recall should be close to the calibration value; if they are disjoint, that split is nowhere stated. Please clarify the calibration set and report Table 3 on the same held-out split used for Tables 1 and 2.
minor comments (6)
  1. [Abstract] The word 'Morevoer' is a typo for 'Moreover'.
  2. [Section 5 (Related Work)] The word 'divded' in 'existing online ADS testing studies are mainly divded into' is a typo for 'divided'.
  3. [4 (Benchmark)] The sentence 'we selected 400 verification scenarios, each containing 40 violations and 40 success cases' is ambiguous: clarify whether there are 400 scenarios total per ADS or 400 scenarios per scenario type, and how the violation/success counts are distributed.
  4. [Table 5 and its text] The text says '3.421 ms vs. 0.487 ms for RuleRepair' but Table 5 reports 0.718 ms for RuleRepair; align the table and the accompanying sentence.
  5. [Tables 1-4] The tables report single percentages without variance, confidence intervals, or the number of independent runs; please add this information or state that the results are deterministic by construction.
  6. [Algorithm 1 and Section 3.1.4] It is not completely explicit that the action a_hat passed to simulator.step in Algorithm 1 is the output of the Decision Merger rather than the raw Decision Adapter action; please clarify the data flow in the algorithm.

Circularity Check

1 steps flagged · score 6.0 of 10

Main empirical claim is evaluated on the same scenario corpus used for ReFT training; no split is stated, so the reported repair gains reduce to training-set performance.

  1. fitted input called prediction [Section 4, 'Benchmark' and 'Training Details'; Algorithm 1, Line 5; Tables 1-2]
    "In total, for each ADS, we selected 400 verification scenarios, each containing 40 violations and 40 success cases to assess repair performance. ... In each episode, we randomly sample both success and violation scenarios from the corpus for replay. (Algorithm 1: Se ← Sample(Scorpus))"

    The paper never states a train/test split. The 400-scenario benchmark is collected from testing techniques and is also the scenario corpus (Scorpus) replayed during ReFT; the text says the corpus 'saves testing cases collected from ADS testing techniques.' Algorithm 1 samples each episode from Scorpus, and Table 1 and Table 2 report %Fix and %Intensity over the same benchmark. On the face of the paper, the repair 'predictions' are therefore computed on scenarios already used to train both the supervised stage and the DQN adapter, so the reported 80.5% improvement is training-set performance rather than held-out generalization.

full rationale

ADReFT's method is not circular in its learning mechanism: the supervised stage uses outcome-derived weak labels, and the ReFT terminal reward is the actual simulator collision outcome, providing external grounding. The transformer encoder and DQN updates are standard and independently grounded, and the paper's self-citations (e.g., BehAVExplor) are not load-bearing. However, the central empirical claim is compromised by the evaluation protocol as written: the 400-scenario corpus described in the Benchmark section is the same corpus used as the replay buffer in Algorithm 1, and no train/test split is stated. Tables 1 and 2 are therefore, on the face of the paper, training-set performance, which is a fitted-input-called-prediction pattern. The reported recall-vs-threshold inconsistency (lambda_safe at 95% positive-sample recall vs. 80.09 test-set recall in Table 3) underscores that the split or threshold protocol is unspecified. Since the headline 80.5% improvement and the intensity comparisons rest on this unspecified overlap, the circularity score is elevated; the method itself retains independent components such as the ablation study and fuzzing robustness, which prevent a higher score.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests mainly on the experimental protocol. The free parameters are all chosen heuristically, and several important ones such as gamma, exploration schedule, and transformer size are not reported. The axioms are standard design assumptions of the simulator and RL setup; the weak-annotation rule is the most fragile because it converts outcomes into per-frame labels that the monitor merely learns to reconstruct.

free parameters (8)
  • delta_d = 1 meter
    Distance threshold in the weak annotation generator (Eq. 5), chosen empirically based on references [30,64].
  • delta_t = 3 seconds
    Pre-collision labeling window in Eq. 5, chosen empirically based on references [30,64].
  • lambda_safe = Safety score achieving >95% recall on positive samples
    Intervention threshold in the Decision Merger, calibrated using positive labels from the evaluation corpus.
  • reward intensity coefficient = 0.5
    Scaling factor in the safe-explore reward, Algorithm 1 line 17, chosen by hand.
  • termination penalty = -10.0
    Fixed penalty for collisions in Algorithm 1 line 19, chosen by hand.
  • DQN discount factor gamma = not reported
    Discount factor in Eq. (1) is not specified in the paper, despite controlling the RL tradeoff.
  • epsilon-greedy exploration schedule = not reported
    Exploration schedule for DQN training is mentioned but its values and decay are not given.
  • transformer hidden size and layer count = not reported
    The BERT-style encoder's hidden dimension H and number of layers are not reported, making exact reimplementation impossible.
assumptions (6)
  • domain assumption Collisions in the testing set are caused by decision-making errors, not perception errors.
    Section 2.1 states all ADSs receive perfect noise-free perception, and Section 4.4 acknowledges imperfect perception as future work.
  • domain assumption Synchronized CARLA, Roach, and Pylot modes make scenario replays deterministic enough for ReFT.
    Section 4 Benchmark relies on synchronization to claim discovered violations can be reliably reproduced through re-simulation.
  • domain assumption Throttle-braking longitudinal actions are sufficient to repair most collision violations.
    Section 3.1.3 restricts the action space to ten throttle-brake bins, and Section 4.4 says steering-based maneuvers are out of scope.
  • ad hoc to paper Weak annotations (distance < 1m or within 3s of collision) approximate true safety-critical states.
    Eq. 5 defines safety labels from outcome and distance thresholds without an oracle or calibration evidence.
  • domain assumption Transformer self-attention over object-level tokens captures the interactions relevant to collision prediction.
    Section 3.1.1 adopts object-centric transformer encoding following prior work; no proof is offered for the collision-prediction setting.
  • standard math DQN updates optimize the repair policy in this environment.
    Section 2.2 and Algorithm 1 assume the DQN loss and target network lead to a good policy; convergence is a standard RL assumption, not verified in this setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADReFT: Adaptive Decision Repair for Safe Autonomous Driving via Reinforcement Fine-Tuning." pith.science (2026). https://pith.science/paper/AJGRBEQI

@misc{pith2026250623960,
  author       = {Pith},
  title        = {Pith review of: ADReFT: Adaptive Decision Repair for Safe Autonomous Driving via Reinforcement Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AJGRBEQI}},
  note         = {Machine review of arXiv:2506.23960}
}
read the original abstract

Autonomous Driving Systems (ADSs) continue to face safety-critical risks due to the inherent limitations in their design and performance capabilities. Online repair plays a crucial role in mitigating such limitations, ensuring the runtime safety and reliability of ADSs. Existing online repair solutions enforce ADS compliance by transforming unacceptable trajectories into acceptable ones based on predefined specifications, such as rule-based constraints or training datasets. However, these approaches often lack generalizability, adaptability and tend to be overly conservative, resulting in ineffective repairs that not only fail to mitigate safety risks sufficiently but also degrade the overall driving experience. To address this issue, we propose Adaptive Decision Repair (ADReFT), a novel and effective repair method that identifies safety-critical states through offline learning from failed tests and generates appropriate mitigation actions to improve ADS safety. Specifically, ADReFT incorporates a transformer-based model with two joint heads, State Monitor and Decision Adapter, designed to capture complex driving environment interactions to evaluate state safety severity and generate adaptive repair actions. Given the absence of oracles for state safety identification, we first pretrain ADReFT using supervised learning with coarse annotations, i.e., labeling states preceding violations as positive samples and others as negative samples. It establishes ADReFT's foundational capability to mitigate safety-critical violations, though it may result in somewhat conservative mitigation strategies. Therefore, we subsequently finetune ADReFT using reinforcement learning to improve its initial capability and generate more precise and contextually appropriate repair decisions. Our evaluation results illustrate that ADReFT achieves better repair performance.

Figures

Figures reproduced from arXiv: 2506.23960 by the authors.

Figure 1
Figure 1. Architecture overview of ADReFT. repair action space as a discrete set of actions and adopt Deep Q￾Network (DQN) as the underlying RL algorithm due to its simplicity and theoretical guarantees of convergence to an optimal policy [60]. Deep Q-Network. DQN extends traditional Q-learning by em￾ploying a deep neural network to approximate the Q-function Q𝜃 (𝑠, 𝑎), where 𝜃 denotes the learnable parameters of the network.… view at source ↗
Figure 2
Figure 2. Illustration of training ADReFT. merging process replaces the original ADS decision 𝑎 ads 𝑡 with the predicted repair decision 𝑎ˆ𝑡 = A [𝑦ˆ repair 𝑡 ] if: (𝑦ˆ safe 𝑡 > 𝜆safe) ∧ (𝑎 ads 𝑡 > 𝑎ˆ𝑡) = 𝑇 𝑟𝑢𝑒 The first condition, 𝑦ˆ safe 𝑡 > 𝜆safe, indicates that the driving state requires intervention if the predicted safety-critical score 𝑦ˆ safe 𝑡 ex￾ceeds a predefined threshold 𝜆safe. The second condition, 𝑎 ads 𝑡 > 𝑎ˆ𝑡 … view at source ↗
Figure 3
Figure 3. Robustness Testing for ADReFT. The X-axis represents the execution time of ADS testing, while the Y-axis indicates the number of detected violations [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 25 canonical work pages

  1. [1]

    Matthias Althoff and Sebastian Lutz. 2018. Automatic generation of safety-critical test scenarios for collision avoidance of road vehicles. In 2018 IEEE Intelligent Vehicles Symposium (IV). IEEE, Changshu, Suzhou, China, 1326–1333

  2. [2]

    Baidu. 2019. Apollo: Open Source Autonomous Driving. https://github.com/ ApolloAuto/apollo

  3. [3]

    CARLA. 2020. Behavior Agent in Carla

  4. [4]

    Jianyu Chen, Bodi Yuan, and Masayoshi Tomizuka. 2019. Deep imitation learning for autonomous driving in generic urban scenarios with enhanced safety. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS) . IEEE, 2884–2890

  5. [5]

    Mingfei Cheng, Xiaofei Xie, Yuan Zhou, Junjie Wang, Guozhu Meng, and Kairui Yang. 2025. Decictor: Towards Evaluating the Robustness of Decision-Making in Autonomous Driving Systems. In Proceedings of the 47th IEEE/ACM International Conference on Software Engineering (ICSE) . 651–651

  6. [6]

    Mingfei Cheng, Yuan Zhou, and Xiaofei Xie. 2023. BehAVExplor: Behavior Diversity Guided Testing for Autonomous Driving Systems. In Proceedings of the 32nd ACM SIGSOFT International Symposium on Software Testing and Analysis . 488–500

  7. [7]

    Erik Coelingh, Andreas Eidehall, and Mattias Bengtsson. 2010. Collision warning with full auto brake and pedestrian detection-a practical example of automatic emergency braking. In 13th International IEEE Conference on Intelligent Trans- portation Systems. IEEE, 155–160

  8. [8]

    comma.ai. 2022. OpenPilot: An open source driver assistance system. Retrieved Nov 7, 2022 from https://github.com/commaai/openpilot

Show all 68 references
  1. [9]

    Yao Deng, Xi Zheng, Mengshi Zhang, Guannan Lou, and Tianyi Zhang. 2022. Scenario-based test reduction and prioritization for multi-module autonomous driving systems. In Proceedings of the 30th ACM Joint European Software Engi- neering Conference and Symposium on the Foundation...

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human...

  3. [11]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv prepri...

  4. [12]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. 2017. CARLA: An Open Urban Driving Simulator. InProceedings of the 1st Annual Conference on Robot Learning (Proceedings of Machine Learning Research, Vol. 78), Sergey Levine, Vincent Vanhoucke...

  5. [13]

    Shuo Feng, Haowei Sun, Xintao Yan, Haojie Zhu, Zhengxia Zou, Shengyin Shen, and Henry X Liu. 2023. Dense reinforcement learning for safety validation of autonomous vehicles. Nature 615, 7953 (2023), 620–627

  6. [14]

    Alessio Gambi, Tri Huynh, and Gordon Fraser. 2019. Generating effective test cases for self-driving cars from police reports. In Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering ...

  7. [15]

    Alessio Gambi, Marc Mueller, and Gordon Fraser. 2019. Automatically testing self- driving cars with search-based procedural content generation. In Proceedings of the 28th ACM SIGSOFT International Symposium on Software Testing and Analysis . ACM, Beijing, China, 318–328

  8. [16]

    Joshua Garcia, Yang Feng, Junjie Shen, Sumaya Almanee, Yuan Xia, and Qi Alfred Chen. 2020. A comprehensive study of autonomous vehicle bugs. In Proceedings of the ACM/IEEE 42nd International Conference on Software Engineering . IEEE, Seoul, South Korea, 385–396

  9. [17]

    Ionel Gog, Sukrit Kalra, Peter Schafhalter, Matthew A Wright, Joseph E Gonzalez, and Ion Stoica. 2021. Pylot: A modular platform for exploring latency-accuracy tradeoffs in autonomous vehicles. In2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 8806–8813

  10. [18]

    Jörg Grieser, Meng Zhang, Tim Warnecke, and Andreas Rausch. 2020. Assuring the safety of end-to-end learning-based autonomous driving through runtime monitoring. In 2020 23rd Euromicro Conference on Digital System Design (DSD) . IEEE, 476–483

  11. [19]

    Jia Cheng Han and Zhi Quan Zhou. 2020. Metamorphic fuzz testing of au- tonomous vehicles. In Proceedings of the IEEE/ACM 42nd International Conference on Software Engineering Workshops. 380–385

  12. [20]

    Seunghee Han, Jaeuk Kim, Geon Kim, Jaemin Cho, Jiin Kim, and Shin Yoo. 2021. Preliminary evaluation of path-aware crossover operators for search-based test data generation for autonomous driving. In 2021 IEEE/ACM 14th International Workshop on Search-Based Software Testing (SB...

  13. [21]

    Fitash Ul Haq, Donghwan Shin, and Lionel Briand. 2022. Efficient online testing for DNN-enabled systems using surrogate-assisted and many-objective optimiza- tion. In Proceedings of the 44th International Conference on Software Engineering . IEEE, Pittsburgh Pennsylvania, 811–822

  14. [22]

    Fitash Ul Haq, Donghwan Shin, and Lionel C Briand. 2023. Many-objective rein- forcement learning for online testing of dnn-enabled systems. In 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE) . IEEE, 1814–1826

  15. [23]

    Carl Hildebrandt, Meriel von Stein, and Sebastian Elbaum. 2023. PhysCov: Phys- ical Test Coverage for Autonomous Vehicles. In Proceedings of the 32nd ACM SIGSOFT International Symposium on Software Testing and Analysis . 449–461

  16. [24]

    David Ke Hong, John Kloosterman, Yuqi Jin, Yulong Cao, Qi Alfred Chen, Scott Mahlke, and Z Morley Mao. 2020. AVGuardian: Detecting and mitigating publish- subscribe overprivilege for autonomous vehicle systems. In 2020 IEEE European Symposium on Security and Privacy (EuroS&P) ...

  17. [25]

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. 2023. Planning-oriented Autonomous Driving. In Proceedings of the IEEE/CVF Conference on...

  18. [26]

    Yuqi Huai, Sumaya Almanee, Yuntianyi Chen, Xiafa Wu, Qi Alfred Chen, and Joshua Garcia. 2023. sceno RITA: Generating Diverse, Fully-Mutable, Test Scenar- ios for Autonomous Vehicle Planning. IEEE Transactions on Software Engineering (2023)

  19. [27]

    Yuqi Huai, Yuntianyi Chen, Sumaya Almanee, Tuan Ngo, Xiang Liao, Ziwen Wan, Qi Alfred Chen, and Joshua Garcia. 2023. Doppelgänger test generation for revealing bugs in autonomous driving software. In 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE) . ...

  20. [28]

    Shinpei Kato, Shota Tokunaga, Yuya Maruyama, Seiya Maeda, Manato Hirabayashi, Yuki Kitsukawa, Abraham Monrroy, Tomohito Ando, Yusuke Fujii, and Takuya Azumi. 2018. Autoware on board: Enabling autonomous vehicles with embedded systems. In 2018 ACM/IEEE 9th International Confere...

  21. [29]

    Seulbae Kim, Major Liu, Junghwan" John" Rhee, Yuseok Jeon, Yonghwi Kwon, and Chung Hwan Kim. 2022. DriveFuzz: Discovering Autonomous Driving Bugs through Driving Quality-Guided Fuzzing. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security . ...

  22. [30]

    Jeong Keun Lee and Kang Wook Lee. 2013. Study on Effectiveness of Pre-Crash Active Seatbelt Using Real Time Controlled Simulation. In 23rd International Technical Conference on the Enhanced Safety of Vehicles (ESV) National Highway Traffic Safety Administration

  23. [31]

    Guanpeng Li, Yiran Li, Saurabh Jha, Timothy Tsai, Michael Sullivan, Siva Ku- mar Sastry Hari, Zbigniew Kalbarczyk, and Ravishankar Iyer. 2020. AV-FUZZER: Finding safety violations in autonomous driving systems. In 2020 IEEE 31st Inter- national Symposium on Software Reliabilit...

  24. [32]

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. 2023. Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA). IEEE, 2774–2781

  25. [33]

    Chengjie Lu, Yize Shi, Huihui Zhang, Man Zhang, Tiexin Wang, Tao Yue, and Shaukat Ali. 2022. Learning configurations of operating environment of au- tonomous vehicles to maximize their collisions. IEEE Transactions on Software Engineering 49, 1 (2022), 384–402

  26. [34]

    Malte Mauritz, Falk Howar, and Andreas Rausch. 2016. Assuring the safety of advanced driver assistance systems through a combination of simulation and runtime monitoring. In Leveraging Applications of Formal Methods, Verification and Validation: Discussion, Dissemination, Appl...

  27. [35]

    Martin Monperrus. 2018. Automatic software repair: A bibliography. ACM Computing Surveys (CSUR) 51, 1 (2018), 1–24

  28. [36]

    Wassim G Najm, John D Smith, Mikio Yanagisawa, et al . 2007. Pre-crash sce- nario typology for crash avoidance research . Technical Report. United States. Department of Transportation. National Highway Traffic Safety

  29. [37]

    Wassim G Najm, Samuel Toma, John Brewer, et al. 2013. Depiction of priority light-vehicle pre-crash scenarios for safety applications based on vehicle-to-vehicle communications. Technical Report DOT HS 811 732. National Highway Traffic Safety Administration, U.S. Department of...

  30. [38]

    Philippe Nitsche, Pete Thomas, Rainer Stuetz, and Ruth Welsh. 2017. Pre-crash scenarios at road junctions: A clustering method for car crash data. Accident Analysis & Prevention 107 (2017), 137–151

  31. [39]

    Jan-Pieter Paardekooper, S Montfort, Jeroen Manders, Jorrit Goos, E de Gelder, O Camp, O Bracquemond, and Gildas Thiolon. 2019. Automatic identification of critical scenarios in a public dataset of 6000 km of public-road driving. In 26th International Technical Conference on t...

  32. [40]

    Qi Pang, Yuanyuan Yuan, and Shuai Wang. 2022. MDPFuzz: Testing Models Solving Markov Decision Processes. In Proceedings of the 31st ACM SIGSOFT International Symposium on Software Testing and Analysis (Virtual, South Korea) (ISSTA 2022). Association for Computing Machinery, Ne...

  33. [41]

    Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, A Sophia Koepke, Zeynep Akata, and Andreas Geiger. 2023. PlanT: Explainable Planning Transformers via Object-Level Representations. In Conference on Robot Learning . PMLR, 459–470

  34. [42]

    Christian Roesener, Felix Fahrenkrog, Axel Uhlig, and Lutz Eckstein. 2016. A scenario-based assessment approach for automated driving by using time series classification of human-driving behaviour. In 2016 IEEE 19th international con- ference on intelligent transportation syst...

  35. [43]

    Saumya Shankar, VR Ujwal, Srinivas Pinisetty, and Partha S Roop. 2020. Formal Runtime Monitoring Approaches for Autonomous Vehicles. OVERLAY 20 (2020), 89–94

  36. [44]

    Andrea Stocco, Paulo J Nunes, Marcelo d’Amorim, and Paolo Tonella. 2022. Thirdeye: Attention maps for safe autonomous driving systems. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering . 1–12

  37. [45]

    Andrea Stocco, Michael Weiss, Marco Calzana, and Paolo Tonella. 2020. Mis- behaviour prediction for autonomous driving systems. In Proceedings of the ACM/IEEE 42nd international conference on software engineering . 359–371

  38. [46]

    Yang Sun, Christopher M Poskitt, Jun Sun, Yuqi Chen, and Zijiang Yang. 2022. LawBreaker: An approach for specifying traffic laws and fuzzing autonomous ve- hicles. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering. 1–12

  39. [47]

    Yang Sun, Christopher M Poskitt, Kun Wang, and Jun Sun. 2025. FixDrive: Automatically Repairing Autonomous Vehicle Driving Behaviour for $0.08 per Violation. arXiv preprint arXiv:2502.08260 (2025)

  40. [48]

    Yang Sun, Christopher M Poskitt, Xiaodong Zhang, and Jun Sun. 2024. REDriver: Runtime Enforcement for Autonomous Vehicles. In Proceedings of the IEEE/ACM 46th International Conference on Software Engineering . 1–12

  41. [49]

    Wenbing Tang, Mingfei Cheng, Yuan Zhou, and Yang Liu. 2025. Moral Testing of Autonomous Driving Systems. arXiv preprint arXiv:2505.03683 (2025)

  42. [50]

    Yun Tang, Yuan Zhou, Yang Liu, Jun Sun, and Gang Wang. 2021. Collision avoidance testing for autonomous driving systems on complete maps. In 2021 IEEE Intelligent Vehicles Symposium (IV) . IEEE, Nagoya, Japan, 179–185

  43. [51]

    Yun Tang, Yuan Zhou, Fenghua Wu, Yang Liu, Jun Sun, Wuling Huang, and Gang Wang. 2021. Route coverage testing for autonomous vehicles via map modeling. In 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, Xi’an, China, 11450–11456

  44. [52]

    Yun Tang, Yuan Zhou, Tianwei Zhang, Fenghua Wu, Yang Liu, and Gang Wang

  45. [53]

    Tesla. 2024. Autopilot and Full Self-Driving

  46. [54]

    Eric Thorn, Shawn C Kimmel, Michelle Chaka, Booz Allen Hamilton, et al. 2018. A framework for automated driving system testable cases and scenarios . Technical Report. United States. Department of Transportation. National Highway Traffic Safety

  47. [55]

    Haoxiang Tian, Yan Jiang, Guoquan Wu, Jiren Yan, Jun Wei, Wei Chen, Shuo Li, and Dan Ye. 2022. MOSAT: finding safety violations of autonomous driving sys- tems using multi-objective genetic algorithm. InProceedings of the 30th ACM Joint European Software Engineering Conference...

  48. [56]

    Haoxiang Tian, Guoquan Wu, Jiren Yan, Yan Jiang, Jun Wei, Wei Chen, Shuo Li, and Dan Ye. 2022. Generating critical test scenarios for autonomous driving systems via influential behavior patterns. In Proceedings of the 37th IEEE/ACM International Conference on Automated Softwar...

  49. [57]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

  50. [58]

    Renzhi Wang, Mingfei Cheng, Xiaofei Xie, Yuan Zhou, and Lei Ma. 2025. MoDi- tector: Module-Directed Testing for Autonomous Driving Systems. Proceedings of the ACM on Software Engineering 2, ISSTA (2025), 137–158

  51. [59]

    Kosuke Watanabe, Eunsuk Kang, Chung-Wei Lin, and Shinichi Shiraishi. 2018. Runtime monitoring for safety of intelligent vehicles. In Proceedings of the 55th annual design automation conference . 1–6

  52. [60]

    Jingda Wu, Chao Huang, Hailong Huang, Chen Lv, Yuntong Wang, and Fei-Yue Wang. 2024. Recent advances in reinforcement learning-based autonomous driving behavior planning: A survey. Transportation Research Part C: Emerging Technologies 164 (2024), 104654

  53. [61]

    Chejian Xu, Wenhao Ding, Weijie Lyu, Zuxin Liu, Shuai Wang, Yihan He, Han- jiang Hu, Ding Zhao, and Bo Li. 2022. Safebench: A benchmarking platform for safety evaluation of autonomous vehicles. Advances in Neural Information Processing Systems 35 (2022), 25667–25682

  54. [62]

    Xudong Zhang and Yan Cai. 2023. Building Critical Testing Scenarios for Au- tonomous Driving from Real Accidents. In Proceedings of the 32nd ACM SIGSOFT International Symposium on Software Testing and Analysis . 462–474

  55. [63]

    Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool

  56. [64]

    Zhiguo Zhao, Liangjie Zhou, Qiang Zhu, Yugong Luo, and Keqiang Li. 2017. A review of essential technologies for collision avoidance assistance systems. Advances in Mechanical Engineering 9, 10 (2017), 1687814017725246

  57. [65]

    In Proceedings of the IEEE/CVF international conference on computer vision

    End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision . 15222– 15232

  58. [66]

    Yuan Zhou, Yang Sun, Yun Tang, Yuqi Chen, Jun Sun, Christopher M Poskitt, Yang Liu, and Zijiang Yang. 2023. Specification-based Autonomous Driving System Testing. IEEE Transactions on Software Engineering (2023), 1–19

  59. [67]

    Ziyuan Zhong, Gail Kaiser, and Baishakhi Ray. 2023. Neural network guided evolutionary fuzzing for finding traffic violations of autonomous vehicles. IEEE Transactions on Software Engineering 49, 4 (2023), 1860–1875

  60. [2021]

    In Proceedings of the 36th IEEE/ACM International Conference on Automated Software Engineering (ASE)

    Systematic testing of autonomous driving systems using map topology- based scenario classification. In Proceedings of the 36th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, Melbourne, Australia, 1342–1346

Pith tools

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