Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Falsification-driven reinforcement learning for maritime motion planning

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Adding falsification-generated adversarial scenarios to RL training improves COLREGs compliance in maritime encounters, with overtaking compliance up 16%.

desk verdict Plausible extension of falsification-driven RL to maritime navigation, but the printed COLREGs formalization is self-negating—the quantitative results rest on an unstated corrected definition. read the letter →

arxiv 2510.06970 v2 pith:MHN2LJ3M submitted 2025-10-08 eess.SY cs.LGcs.SY

Marlon Müller , Florian Finkeldei , Hanna Krasowski , Murat Arcak , Matthias Althoff This is my paper
classification eess.SYcs.LGcs.SY
keywords COLREGsfalsificationsignaltemporallogicreinforcementlearningmaritimemotionplanningCMA-ESrulecomplianceautonomousvessels
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 proposes a falsification-driven reinforcement learning (RL) loop for autonomous surface vessels. Instead of relying only on randomly sampled training scenarios, it uses a CMA-ES optimizer to search for adversary input sequences that make the current RL policy violate formalized COLREGs traffic rules (expressed in signal temporal logic). These counterexample scenarios are added to the training pool, and the policy is retrained on them. The experiments on two-vessel open-sea encounters report that this approach improves rule compliance in crossing, head-on, and overtaking encounters by about 3%, 2%, and 16% respectively, and reduces the variability across random seeds. The authors argue this yields more relevant training scenarios and more consistent rule-compliant behavior.

What carries the argument

The load-bearing machinery is the falsification loop that couples CMA-ES optimization of the adversary's input sequence with an IA-STL robustness objective: the optimizer minimizes ρ_in (the distance to a persistent encounter) when the specification is vacuously satisfied, and ρ_out (the distance to rule compliance) when it is not. This is wrapped around the formalized COLREGs specification φ_colregs, which includes encounter predicates built from position, orientation, and velocity-robustness measures, and maneuver predicates that require a detectable course change.

What would settle it

Check whether the implementation used for the experiments evaluates persistent_encounter at time step k as an encounter at k, rather than as the negation of an encounter at k. Concretely, instrument the simulator to log ρ_in and ρ_out for a hand-crafted crossing scenario where a persistent encounter is clearly present; if ρ_in is positive (indicating vacuity) at the moment of the encounter, the formalization is not doing what Eq. (19) as printed describes.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that counterexample-guided scenario generation — where the adversary is optimized to falsify an STL specification of COLREGs rather than to collide — leads to measurably better and more stable rule compliance than training on random scenarios. The reported improvements are 3% for crossing, 2% for head-on, and 16% for overtaking encounters, with lower standard deviation across seeds in all three situations. The central mechanism is to separate vacuous from non-vacuous satisfaction of the traffic-rule implication, minimizing input robustness to force persistent encounters and output robustness to expose non-compliance.

Load-bearing premise

The entire result hinges on the STL/IA-STL formalization of COLREGs (especially the definition of persistent_encounter) being correct and consistently implemented; as printed, Eq. (19) defines persistent_encounter as 'not an encounter right now, and an encounter for the next tp time steps,' which would make the antecedent false during the actual encounter and render the specification vacuously satisfied, so the reported non-vacuous compliance numbers require an unstated corre

Editorial extensions

If this is right

  • If correct, the method provides a general recipe to inject formal-specification violations into RL training, not just for COLREGs but for any traffic rule set expressible in STL/IA-STL.
  • The reduced standard deviation across seeds suggests that falsification-driven training stabilizes learning of give-way maneuvers, which could translate to more predictable behavior across deployments.
  • The separation of vacuity and output robustness could be reused for reward shaping in other rule-constrained RL settings, since it tells the agent whether it is failing to engage with the rule or failing to comply once engaged.
  • The reported gains in overtaking compliance (16%) indicate that the benefit is largest in encounters where random sampling rarely produces the relevant scenario, pointing to falsification's value as a rare-event sampler.
  • If combined with provably safe RL methods, as the paper suggests, the improved baseline compliance could reduce the frequency of safety interventions, improving runtime efficiency.

Reading between the lines

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

  • A likely extension is to multi-adversary scenarios; the paper notes that with more vessels there may be no rule-compliant solution, so the falsification objective would need to account for conflicts between rules, potentially using prioritized STL or weighted robustness.
  • The method's effectiveness likely depends on the quality of the robustness measures; the time-domain scaling of predicates (dividing distances by velocities, angles by angular accelerations) is a heuristic that could be tuned per scenario, and the paper does not yet demonstrate sensitivity to these scales.
  • The printed definition of persistent_encounter in Eq. (19) appears logically inconsistent with its intended meaning; if that is a typographical error rather than an implementation detail, the reported compliance numbers rest on code that is not fully specified in the paper.
  • A natural testable extension is to compare the falsification-driven agent against an agent trained on a fixed, hand-crafted set of 'challenging' encounters; the paper's claim is that the adversarial search does the crafting automatically, which could be verified by ablating the CMA-ES search.
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

4 major / 6 minor

Summary. The paper proposes a falsification-driven RL framework for maritime motion planning. A CMA-ES-based falsifier searches for adversary ship input sequences that cause the ego policy to violate STL/IA-STL formalizations of COLREGs (crossing, head-on, overtaking). The resulting counterexamples are added to a FIFO scenario pool for PPO training. Experiments in a two-vessel open-sea simulator report that, compared to random-scenario RL, the falsification-driven arm reaches higher non-vacuous encounter rates and higher rule compliance (about 3%, 2%, and 16% for crossing, head-on, and overtaking, respectively) with lower variance across five random seeds.

Significance. If the formalization and implementation are correct, the approach is a practical and interesting way to shape the RL training distribution toward rule-relevant encounters; the substantial reduction in seed variance for overtaking is a notable empirical result. The experimental design has strengths: both arms share the same reward and evaluation metric, the test set is unseen, and multiple seeds are used. However, the central claim depends heavily on the precise semantics of the STL formalization—especially the persistent_encounter predicate—and on the statistical reliability of the reported improvements. The paper does not release code, which makes it impossible to verify the implementation against the printed formulas.

major comments (4)
  1. [Sec. 5.3, Eq. (19); Sec. 3.3; Table 1] There is a mismatch between the prose and the formula. The text states that a persistent encounter is present if the encounter 'is true for at least a duration of tp', but Eq. (19) defines persistent_encounter(xE,xA,k) ⇔ ¬encounter(xE,xA,k) ∧ G[Δt,tp] encounter(xE,xA,k). As written, the predicate is false whenever the encounter currently holds, so the non-vacuous branch of Eq. (5) is triggered only at the step immediately before the encounter begins. If the implementation follows Eq. (19), the semantics is a rising-edge detector, not the persistence predicate described in the text; if the implementation follows the prose, then the formula in Eq. (19) is not the one used in the experiments. Since the same φ_colregs enters the falsification objective (Eq. (7)), the reward (Eq. (10)), and the evaluation (Table 1), the reported 3%, 2%, and 16% improvements cannot be interpreted without resol
  2. [Sec. 6.2, Table 1] The main quantitative claim lacks statistical support. The overtaking row reports baseline η|=φ = 78% with std 29.1% versus falsification 94% with std 2.9% over five seeds. With n=5, this difference should be accompanied by per-seed values, confidence intervals, or a paired significance test. Without this, the claimed 16-point improvement and the variance reduction are not established beyond anecdote.
  3. [Sec. 6.1, Table A.2, Alg. 1] The comparison conflates falsification with scenario-distribution shift. CMA-ES with n_generation=10, λ=10, and n_samples=6 evaluates only about 100 candidate input sequences per falsification call, and the baseline uses random scenarios from the same initial-state distributions. To support the claim that falsification—rather than simply replacing random scenarios with a different, non-uniform distribution—is responsible for the improvement, add an ablation with random search or uniform sampling using the same computational budget and pool dynamics.
  4. [Sec. 4, Sec. 6.2] The evaluation metric is exactly the formalization used to generate counterexamples and to shape the reward. An independent behavioral indicator (e.g., DCPA/TCPA, minimum separation distance, or expert labeling of rollouts) would help rule out that the improvements on η|=φ are artifacts of the formalization or of reward hacking. At minimum, quantify the qualitative rollouts in Fig. 6 with a geometric safety metric.
minor comments (6)
  1. [Alg. 1, line 7] S ← S ∩ (χ0,uA) should be S ← S ∪ {(χ0,uA)}; as written, the intersection with a tuple is undefined and the FIFO pool would be empty.
  2. [Eq. (19)] The typesetting '⇐ ⇒' should be a single biconditional '⇔'.
  3. [Sec. 5.3] The parameter values for the encounter predicates (the β and γ angular bounds for crossing, head-on, and overtaking) are deferred to a self-cited reference; include them in an appendix for reproducibility.
  4. [Table 1] Please define the denominator of the η|=φ percentages explicitly: are they computed over all scenarios or only over non-vacuous scenarios of the respective encounter type?
  5. [Figs. 4 and 5] The captions label panels only as '(a)' and '(b)'; indicate which panel corresponds to the baseline and which to the falsification-driven agent.
  6. [Sec. 6.1] Clarify whether the 10,000 test scenarios are disjoint from the scenario pool used during training and whether the initial-state distributions for the test set are identical to those used for generating falsification scenarios.

Circularity Check

2 steps flagged · score 4.0 of 10

Self-cited formalization and a self-negating persistent-encounter definition make the compliance metric self-referential, though the relative baseline comparison retains independent content.

  1. self definitional [Sec. 5.3, Eq. (19); deployed in Sec. 3.3 Eq. (5), Sec. 4 Eq. (7), Sec. 5.1 Eq. (10), Sec. 6.2 Table 1]
    "For all give-way rules, an encounter predicate must persist for tp: persistent_⟨encounter⟩(xE,x A, k)⇐ ⇒ ¬⟨encounter⟩(xE,x A, k)∧ G[∆t,tp] ⟨encounter⟩(xE,x A, k)"

    As printed, Eq. (19) makes the persistent-encounter antecedent false at every step where the encounter predicate currently holds, so the implication in φ_colregs is vacuously true during actual encounters. The same φ_colregs is simultaneously the falsification objective (Eq. 7), the RL reward (Eq. 10), and the test metric (Table 1's η|=φ). The reported 'non-vacuous' scenario counts and 'rule compliance' percentages are therefore defined by this self-negating formula rather than by an independent check that the vessel maneuvers during the encounter. The central quantitative claim reduces to the paper's own definition of persistent_encounter.

  2. self citation load bearing [Sec. 3.3 and Sec. 5.3; Refs. (Krasowski and Althoff, 2021, 2024)]
    "we adapt the formalization of maritime traffic rules from Krasowski and Althoff (2024) ... For detailed derivations of the maritime traffic rules, we refer the interested reader to (Krasowski and Althoff, 2021, 2024) ... The predicates overtake and head_on, as well as the values of β and γ, are specified in Krasowski and Althoff (2021)."

    All quantitative claims are measured against φ_colregs, and the only authority offered for that formalization is prior work by the same co-authors; the actual definitions of overtake, head_on, and the angular bounds are not reproduced here. The evaluation therefore certifies performance with respect to a specification whose fidelity to COLREGs is supported only by the authors' own prior papers. This is load-bearing because if that self-cited formalization is wrong — as Eq. (19) suggests — the reported 3%, 2%, and 16% improvements are not evidence about COLREGs compliance.

full rationale

The paper's relative comparison — falsification-selected scenarios versus random scenarios under identical PPO and reward — is a meaningful experiment, and the improvement on the stated metric is not syntactically forced. I do not find a direct algebraic equivalence between the falsification objective and the test statistic. However, two features prevent a higher-confidence innocent verdict. First, the only evaluation of 'rule compliance' is satisfaction of φ_colregs, the same formula used for falsification and reward; no external expert-labeled COLREGs benchmark or independently implemented rule set is used. Second, Eq. (19) as printed defines persistent_encounter so that the antecedent of the φ_colregs implication is false during exactly those steps where the encounter holds, making the non-vacuous semantics of the reported compliance statistics depend on an unstated correction or implementation detail. The formalization itself is also inherited via self-citation from Krasowski and Althoff (2021, 2024). These issues make the central numbers self-referential to the authors' own specification, although the scenario-selection contribution itself has independent algorithmic content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method adds no new physical entities. Its numerical content is dominated by hand-chosen reward weights, schedule hyperparameters, and domain parameters inherited from a self-cited formalization; no ablation or sensitivity analysis is provided. The central comparison could change under different choices, but none of these are fitted to the test set, so this is not a hidden data-fitting loop. The most serious ledger item is the printed persistent_encounter definition, which appears inconsistent with the reported results.

free parameters (3)
  • reward coefficients (c_Δgoal=+0.0005, c_v=-0.25, c_ω=-1, c_goal=+3, c_colregs=+3, c_zone=-3) = +0.0005, -0.25, -1, +3, +3, -3
    Hand-chosen in Sec. 5.1; no sensitivity study. The comparison between baseline and falsification policies could shift under different weights.
  • falsification/training hyperparameters (f_falsification=5000, n_samples=6, n_generation=10, λ=10, n_pool=100, T=5e6) = as listed in Table A.2
    Chosen without ablation; they control how strongly the falsification loop biases the training distribution.
  • traffic-rule/scenario parameters (t_p=50s, t_m=70s, δ=20°, t_h=420s, d_zone=450m) = as listed in Table A.2
    Inherited from the cited COLREGs formalization; scenario difficulty and compliance rates depend on these values, yet no robustness analysis is given.
assumptions (5)
  • standard math STL/IA-STL robustness semantics (Donzé and Maler; Ferrère et al.) are a valid quantitative measure of rule compliance
    Used in Sec. 3.2 and 4 to define ρ_in, ρ_out and the falsification objective.
  • domain assumption The Krasowski/Althoff COLREGs formalization extended in Eqs. (5)-(25) faithfully represents the traffic rules
    The entire reward, falsification target, and test metric are built on this formalization; no independent validation against expert judgments is provided.
  • domain assumption The vessel dynamics in Eq. (1) with LSODA discretization adequately model open-sea ship motion
    Kinematic model with acceleration bounds; no waves, wind, current, or actuator dynamics; results are simulation-only.
  • ad hoc to paper CMA-ES with 10 generations and 10 candidates finds meaningful falsifying scenarios rather than merely sampling
    Algorithm 1 uses these settings; the paper reports no falsification success statistics or comparison to random search.
  • ad hoc to paper The printed persistent_encounter definition in Eq. (19) is not a typo
    If read literally, it makes φ_colregs vacuous during encounters; the reported compliance rates therefore assume an implementation different from the printed formula or an interpretation not explained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Falsification-driven reinforcement learning for maritime motion planning." pith.science (2026). https://pith.science/paper/MHN2LJ3M

@misc{pith2026251006970,
  author       = {Pith},
  title        = {Pith review of: Falsification-driven reinforcement learning for maritime motion planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHN2LJ3M}},
  note         = {Machine review of arXiv:2510.06970}
}
read the original abstract

Compliance with maritime traffic rules is essential for the safe operation of autonomous vessels, yet training reinforcement learning (RL) agents to adhere to them is challenging. The behavior of RL agents is shaped by the training scenarios they encounter, but creating scenarios that capture the complexity of maritime navigation is non-trivial, and real-world data alone is insufficient. To address this, we propose a falsification-driven RL approach that generates adversarial training scenarios in which the vessel under test violates maritime traffic rules, which are expressed as signal temporal logic specifications. Our experiments on open-sea navigation with two vessels demonstrate that the proposed approach provides more relevant training scenarios and achieves more consistent rule compliance.

Figures

Figures reproduced from arXiv: 2510.06970 by the authors.

Figure 1
Figure 1. Vessel encounters and rule-compliant maneuvers. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. Every [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the robustness measures (a) [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Distribution of RL reward [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Distribution of encounter types [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Policy and adversary rollouts where vacuous encounters are gray and green [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. pacSTL: PAC-Bounded Signal Temporal Logic from Data-Driven Reachability Analysis

    cs.LO 2025-11 conditional novelty 6.0 of 10

    pacSTL composes PAC-bounded reachable sets with interval STL to compute spec-level robustness intervals that contain an unseen trajectory's robustness with probability ≥ 1−ε.

  2. Staying on Spec: Real-Time Monitoring under Uncertainty with a Maritime Case Study

    cs.RO 2026-08 conditional novelty 5.0 of 10

    A monitoring framework based on pacSTL and data-driven reachable sets, with a sim-to-real disturbance calibration, enables real-time COLREG compliance monitoring and earlier collision avoidance.

Reference graph

Works this paper leans on

66 extracted references · 2 linked inside Pith · cited by 2 Pith papers

  1. [1]

    , author Fainekos, G

    author Abbas, H. , author Fainekos, G. , author Sankaranarayanan, S. , author Ivan c i\' c , F. , author Gupta, A. , year 2013 . title Probabilistic temporal logic falsification of cyber-physical systems . journal ACM Trans. Embed. Comput. Syst. volume 12 . note Article 95

    2013

  2. [2]

    , author Fainekos, G.E

    author Annapureddy, Y.S.R. , author Fainekos, G.E. , year 2010 . title Ant colonies for temporal logic falsification of hybrid systems , in: booktitle Conference on IEEE Industrial Electronics Society , pp. pages 91--96

    2010

  3. [3]

    , author Glad, I.K

    author Bakdi, A. , author Glad, I.K. , author Vanem, E. , year 2021 . title Testbed scenario design exploiting traffic big data for autonomous ship trials under multiple conflicts with collision/grounding risks and spatio-temporal dependencies . journal IEEE Trans. Intell. Transp. Syst. volume 22 , pages 7914--7930

    2021

  4. [4]

    , author Deshmukh, J

    author Bartocci, E. , author Deshmukh, J. , author Donz \'e , A. , author Fainekos, G. , author Maler, O. , author Ni c kovi \' c , D. , author Sankaranarayanan, S. , year 2018 . title Specification-based monitoring of cyber-physical systems: A survey on theory, tools and applications , in: booktitle Lectures on Runtime Verification: Introductory and Adva...

    2018

  5. [5]

    , author Roh, M.I

    author Chun, D.H. , author Roh, M.I. , author Lee, H.W. , author Ha, J. , author Yu, D. , year 2021 . title Deep reinforcement learning-based collision avoidance for an autonomous ship . journal Ocean Eng. volume 234 . note Article 109216

    2021

  6. [6]

    , author Moss, R

    author Corso, A. , author Moss, R. , author Koren, M. , author Lee, R. , author Kochenderfer, M. , year 2022 . title A survey of algorithms for black-box safety validation of cyber-physical systems . journal J. Artif. Intell. Res. volume 72 , pages 377--428

    2022

  7. [7]

    , author Guan, W

    author Cui, Z. , author Guan, W. , author Luo, W. , author Zhang, X. , year 2023 . title Intelligent navigation method for multiple marine autonomous surface ships based on improved PPO algorithm . journal Ocean Eng. volume 287 . note Article 115783

    2023

  8. [8]

    , author Horvat, M

    author Deshmukh, J. , author Horvat, M. , author Jin, X. , author Majumdar, R. , author Prabhu, V.S. , year 2017 . title Testing cyber-physical systems through Bayesian optimization . journal ACM Trans. Embed. Comput. Syst. volume 16 . note Article 170

    2017

Show all 66 references
  1. [9]

    , author Maler, O

    author Donz \'e , A. , author Maler, O. , year 2010 . title Robust satisfaction of temporal logic over real-valued signals , in: booktitle Formal Modeling and Analysis of Timed Systems , pp. pages 92--106

    2010

  2. [10]

    , author Dang, T

    author Dreossi, T. , author Dang, T. , author Donz \'e , A. , author Kapinski, J. , author Jin, X. , author Deshmukh, J.V. , year 2015 . title Efficient guiding strategies for testing of temporal properties of hybrid systems , in: booktitle NASA Formal Methods , pp. pages 127--142

    2015

  3. [11]

    , author Blanke, M

    author Enevoldsen, T.T. , author Blanke, M. , author Galeazzi, R. , year 2023 . title Autonomy for ferries and harbour buses: a collision avoidance perspective . journal IFAC-PapersOnLine volume 56 , pages 5735--5740

    2023

  4. [12]

    , author Pappas, G.J

    author Fainekos, G.E. , author Pappas, G.J. , year 2006 . title Robustness of temporal logic specifications , in: booktitle Formal Approaches to Software Testing and Runtime Verification , pp. pages 178--192

    2006

  5. [13]

    , author Nickovic, D

    author Ferr\` e re, T. , author Nickovic, D. , author Donz\' e , A. , author Ito, H. , author Kapinski, J. , year 2019 . title Interface-aware signal temporal logic , in: booktitle ACM International Conference on Hybrid Systems: Computation and Control , pp. pages 57--66

    2019

  6. [14]

    , author Shiller, Z

    author Fiorini, P. , author Shiller, Z. , year 1998 . title Motion planning in dynamic environments using velocity obstacles . journal Int. J. Robotics Res. volume 17 , pages 760--772

    1998

  7. [15]

    , author Karg \'e n, U

    author Frey, D. , author Karg \'e n, U. , author Varr \'o , D. , year 2025 . title Assessing scene generation techniques for testing COLREGS -compliance of autonomous surface vehicles . journal Proc. ACM Softw. Eng. volume 2 , pages 985--1008

    2025

  8. [16]

    , author Dasgupta, P

    author Gangopadhyay, B. , author Dasgupta, P. , year 2021 . title Counterexample guided RL policy refinement using Bayesian optimization , in: booktitle Advances in Neural Information Processing Systems , pp. pages 22783--22794

    2021

  9. [17]

    , author Luo, W

    author Guan, W. , author Luo, W. , author Cui, Z. , year 2024 . title Intelligent decision-making system for multiple marine autonomous surface ships based on deep reinforcement learning . journal Robotics Auton. Syst. volume 172 . note Article 104587

    2024

  10. [18]

    , year 2016

    author Hansen, N. , year 2016 . title The CMA evolution strategy: A tutorial . journal arXiv:1604.00772

    2016 arXiv

  11. [19]

    , author Auger, A

    author Hansen, N. , author Auger, A. , author Ros, R. , author Finck, S. , author Posik, P. , year 2010 . title Comparing results of 31 algorithms from the black-box optimization benchmarking BBOB-2009 , in: booktitle ACM Genetic and Evolutionary Computation Conference , pp. p...

    2010

  12. [20]

    , author Ostermeier, A

    author Hansen, N. , author Ostermeier, A. , year 1996 . title Adapting arbitrary normal mutation distributions in evolution strategies: t he covariance matrix adaptation , in: booktitle IEEE International Conference on Evolutionary Computation , pp. pages 312--317

    1996

  13. [21]

    , author Stilwell, D.J

    author He, H.J. , author Stilwell, D.J. , author Farhood, M. , author Muniraj, D. , year 2022 . title Use of falsification to find rare failure modes of a ship collision avoidance algorithm , in: booktitle OCEANS 2022 - Chennai , pp. pages 1--9

    2022

  14. [22]

    , year 1983

    author Hindmarsh, A.C. , year 1983 . title ODEPACK , A systematized collection of ODE solvers . journal IMACS Transactions on Scientific Computation volume 1 , pages 55--64

    1983

  15. [23]

    title COLREGs : Convention on the international regulations for preventing collisions at sea

    author International Maritime Organization , year 1972 . title COLREGs : Convention on the international regulations for preventing collisions at sea

    1972

  16. [24]

    , author Perez, T

    author Johansen, T.A. , author Perez, T. , author Cristofaro, A. , year 2016 . title Ship collision avoidance and COLREGs compliance using simulation-based control behavior selection with predictive hazard assessment . journal IEEE Trans. Intell. Transp. Syst. volume 17 , page...

    2016

  17. [25]

    , author Formica, F

    author Khandait, T. , author Formica, F. , author Arcaini, P. , author Chotaliya, S. , author Fainekos, G. , author Hekal, A. , author Kundu, A. , author Lew, E. , author Loreti, M. , author Menghi, C. , author Nenzi, L. , author Pedrielli, G. , author Peltom\" a ki, J. , auth...

    2024

  18. [26]

    , author Esposito, J.M

    author Kim, J. , author Esposito, J.M. , author Kumar, V. , year 2006 . title Sampling-based algorithm for testing and validating robot controllers . journal Int. J. Rob. Res. volume 25 , pages 1257--1272

    2006

  19. [27]

    , author Pek, C

    author Koschi, M. , author Pek, C. , author Maierhofer, S. , author Althoff, M. , year 2019 . title Computationally efficient safety falsification of adaptive cruise control systems , in: booktitle IEEE Intelligent Transportation Systems Conference , pp. pages 2879--2886

    2019

  20. [28]

    , author Althoff, M

    author Krasowski, H. , author Althoff, M. , year 2021 . title Temporal logic formalization of marine traffic rules , in: booktitle IEEE Intelligent Vehicles Symposium , pp. pages 186--192

    2021

  21. [29]

    , author Althoff, M

    author Krasowski, H. , author Althoff, M. , year 2024 . title Provable traffic rule compliance in safe reinforcement learning on the open sea . journal IEEE Trans. Intell. Veh. volume 9 , pages 7617--7634

    2024

  22. [30]

    , author Wilthil, E

    author Kufoalor, D.K.M. , author Wilthil, E. , author Hagen, I.B. , author Brekke, E.F. , author Johansen, T.A. , year 2019 . title Autonomous COLREGs -compliant decision making using maritime radar tracking and model predictive control , in: booktitle European Control Confere...

    2019

  23. [31]

    , author Teigen, H

    author Larsen, T.N. , author Teigen, H. . , author Laache, T. , author Varagnolo, D. , author Rasheed, A. , year 2021 . title Comparing deep reinforcement learning algorithms' ability to safely navigate challenging waters . journal Front. Robot. AI volume 8 . note Article 738113

    2021

  24. [32]

    , author Mengshoel, O.J

    author Laursen, T.S. , author Mengshoel, O.J. , year 2025 . title Evolutionary adaptive stress testing for collision avoidance in sustainable maritime transportation , in: booktitle Genetic and Evolutionary Computation Conference Companion , pp. pages 2215--2223

    2025

  25. [33]

    , author Nickovic, D

    author Maler, O. , author Nickovic, D. , author Pnueli, A. , year 2007 . title On synthesizing controllers from bounded-response properties , in: booktitle Computer Aided Verification , pp. pages 95--107

    2007

  26. [34]

    , author Heiberg, A

    author Meyer, E. , author Heiberg, A. , author Rasheed, A. , author San, O. , year 2020 . title COLREG -compliant collision avoidance for unmanned surface vehicle using deep reinforcement learning . journal IEEE Access volume 8 , pages 165344--165364

    2020

  27. [35]

    , author Shibata, M

    author Nomura, M. , author Shibata, M. , year 2024 . title cmaes : A simple yet practical python library for CMA-ES . journal arXiv:2402.01373

    2024

  28. [36]

    , author Seita, D

    author Pan, X. , author Seita, D. , author Gao, Y. , author Canny, J. , year 2019 . title Risk averse robust adversarial reinforcement learning , in: booktitle IEEE International Conference on Robotics and Automation , pp. pages 8522--8528

    2019

  29. [37]

    , author Davidson, J

    author Pinto, L. , author Davidson, J. , author Sukthankar, R. , author Gupta, A. , year 2017 . title Robust adversarial reinforcement learning , in: booktitle International Conference on Machine Learning , pp. pages 2817--2826

    2017

  30. [38]

    , author Azimi, S

    author Porres, I. , author Azimi, S. , author Lilius, J. , year 2020 . title Scenario-based testing of a ship collision avoidance system , in: booktitle Euromicro Conference on Software Engineering and Advanced Applications , pp. pages 545--552

    2020

  31. [39]

    , author Hill, A

    author Raffin, A. , author Hill, A. , author Gleave, A. , author Kanervisto, A. , author Ernestus, M. , author Dormann, N. , year 2021 . title Stable-Baselines3 : Reliable reinforcement learning implementations . journal J. Mach. Learn. Res. volume 22 , pages 1--8

    2021

  32. [40]

    , author Sehi'c, K

    author Ramezani, Z. , author Sehi'c, K. , author Nardi, L. , author A kesson, K. , year 2025 . title Falsification of cyber-physical systems using Bayesian optimization . journal ACM Trans. Embed. Comput. Syst. volume 24 . note Article 41

    2025

  33. [41]

    , author Hongwei, X

    author Rongcai, Z. , author Hongwei, X. , author Kexin, Y. , year 2023 . title Autonomous collision avoidance system in a multi-ship environment based on proximal policy optimization method . journal Ocean Eng. volume 272 . note Article 113779

    2023

  34. [42]

    , author Fainekos, G

    author Sankaranarayanan, S. , author Fainekos, G. , year 2012 . title Falsification of temporal properties of hybrid systems using the cross-entropy method , in: booktitle ACM International Conference on Hybrid Systems: Computation and Control , pp. pages 125--134

    2012

  35. [43]

    , author Waga, M

    author Sato, S. , author Waga, M. , author Hasuo, I. , year 2021 . title Constrained optimization for hybrid system falsification and application to conjunctive synthesis . journal IFAC -PapersOnLine volume 54 , pages 217--222

    2021

  36. [44]

    , author Sato, K

    author Sawada, R. , author Sato, K. , author Minami, M. , year 2024 . title Framework of safety evaluation and scenarios for automatic collision avoidance algorithm . journal Ocean Eng. volume 300 . note Article 117506

    2024

  37. [45]

    , author Wolski, F

    author Schulman, J. , author Wolski, F. , author Dhariwal, P. , author Radford, A. , author Klimov, O. , year 2017 . title Proximal policy optimization algorithms . journal arXiv:1707.06347

    2017 arXiv

  38. [46]

    , author Kochenderfer, M.J

    author Shahrooei, Z. , author Kochenderfer, M.J. , author Baheri, A. , year 2023 . title Falsification of learning-based controllers through multi-fidelity Bayesian optimization , in: booktitle European Control Conference , pp. pages 1--6

    2023

  39. [47]

    , author Wang, S

    author Shi, J. , author Wang, S. , author Yuan, Q. , author Liu, X. , author Hsieh, T.H. , year 2023 . title Generation of naturalistic and adversarial sailing environment (NASE) for intelligent test of autonomous ships . journal Ocean Eng. volume 285 . note Article 115438

    2023

  40. [48]

    , author Glomsrud, J.A

    author Torben, T.R. , author Glomsrud, J.A. , author Pedersen, T.A. , author Utne, I.B. , author S rensen, A.J. , year 2023 . title Automatic simulation-based testing of autonomous ships using Gaussian processes and temporal logic . journal Proc. Inst. Mech. Eng. O volume 237 ...

    2023

  41. [49]

    , author Fainekos, G

    author Tuncali, C.E. , author Fainekos, G. , year 2019 . title Rapidly-exploring random trees for testing automated vehicles , in: booktitle IEEE Intelligent Transportation Systems Conference , pp. pages 661--666

    2019

  42. [50]

    , author Eriksen, B.O.H

    author Vatle, J.M. , author Eriksen, B.O.H. , author Mengshoel, O.J. , year 2024 . title Green urban mobility with autonomous electric ferries: Studies of simulated maritime collisions using adaptive stress testing . journal Swedish Artificial Intelligence Society , pages 47--56

    2024

  43. [51]

    , year 2020

    author Waga, M. , year 2020 . title Falsification of cyber-physical systems with robustness-guided black-box checking , in: booktitle International Conference on Hybrid Systems: Computation and Control . note Article 11

    2020

  44. [52]

    , author Gheneti, B

    author Wang, W. , author Gheneti, B. , author Mateos, L.A. , author Duarte, F. , author Ratti, C. , author Rus, D. , year 2019 . title Roboat: An autonomous surface vehicle for urban waterways , in: booktitle IEEE International Conference on Intelligent Robots and Systems , pp...

    2019

  45. [53]

    , author Huang, L

    author Wang, W. , author Huang, L. , author Liu, K. , author Zhou, Y. , author Yuan, Z. , author Xin, X. , author Wu, X. , year 2024 a. title Ship encounter scenario generation for collision avoidance algorithm testing based on AIS data . journal Ocean Eng. volume 291 . note A...

    2024

  46. [54]

    , author Nair, S

    author Wang, X. , author Nair, S. , author Althoff, M. , year 2020 . title Falsification-based robust adversarial reinforcement learning , in: booktitle IEEE International Conference on Machine Learning and Applications , pp. pages 205--212

    2020

  47. [55]

    , author Xu, H

    author Wang, Y. , author Xu, H. , author Feng, H. , author He, J. , author Yang, H. , author Li, F. , author Yang, Z. , year 2024 b. title Deep reinforcement learning based collision avoidance system for autonomous ships . journal Ocean Eng. volume 292 . note Article 116527

    2024

  48. [56]

    , author Gang, L

    author Xie, W. , author Gang, L. , author Zhang, M. , author Liu, T. , author Lan, Z. , year 2024 . title Optimizing multi-vessel collision avoidance decision making for autonomous surface vessels: A COLREGs -compliant deep reinforcement learning approach . journal J. Mar. Sci...

    2024

  49. [57]

    , author Liu, K

    author Xin, X. , author Liu, K. , author Yu, Y. , author Yang, Z. , year 2025 . title Developing robust traffic navigation scenarios for autonomous ship testing: an integrated approach to scenario extraction, characterization, and sampling in complex waters . journal Transp. R...

    2025

  50. [58]

    , author Fainekos, G

    author Yaghoubi, S. , author Fainekos, G. , year 2019 . title Gray-box adversarial testing for control systems with machine learning components , in: booktitle ACM International Conference on Hybrid Systems: Computation and Control , pp. pages 179--184

    2019

  51. [59]

    , author Liu, S

    author Yamagata, Y. , author Liu, S. , author Akazaki, T. , author Duan, Y. , author Hao, J. , year 2021 . title Falsification of cyber-physical systems using deep reinforcement learning . journal IEEE Trans. Software Eng. volume 47 , pages 2823--2840

    2021

  52. [60]

    , author Hoxha, B

    author Yamaguchi, T. , author Hoxha, B. , author Ni c kovi \' c , D. , year 2024 . title RTAMT -- runtime robustness monitors with application to CPS and robotics . journal Int. J. Softw. Tools Technol. Transf. volume 26 , pages 79--99

    2024

  53. [61]

    , author Hasuo, I

    author Zhang, Z. , author Hasuo, I. , author Arcaini, P. , year 2019 . title Multi-armed bandits for Boolean connectives in hybrid system falsification , in: booktitle Computer Aided Verification , pp. pages 401--420

    2019

  54. [62]

    , author Roh, M.I

    author Zhao, L. , author Roh, M.I. , year 2019 . title COLREGs -compliant multiship collision avoidance based on deep reinforcement learning . journal Ocean Eng. volume 191 . note Article 106436

    2019

  55. [63]

    , author Niu, Y

    author Zhu, F. , author Niu, Y. , author Wei, M. , author Du, Y. , author Zhai, P. , year 2025 . title A high-risk test scenario adaptive generation algorithm for ship autonomous collision avoidance decision-making based on reinforcement learning . journal Ocean Eng. volume 32...

    2025

  56. [64]

    , author Zhou, Z

    author Zhu, F. , author Zhou, Z. , author Lu, H. , year 2022 . title Randomly testing an autonomous collision avoidance system with real-world ship encounter scenario from AIS data . journal J. Mar. Sci. Eng. volume 10 . note Article 1588

    2022

  57. [65]

    , author Alexander, R

    author Zou, X. , author Alexander, R. , author McDermid, J. , year 2014 . title Safety validation of sense and avoid algorithms using simulation and evolutionary search , in: booktitle Computer Safety, Reliability, and Security , pp. pages 33--48

    2014

  58. [66]

    , author Gehrt, J

    author Zweigel, R. , author Gehrt, J. , author Liu, S. , author Roy, S. , author B\" u skens, C. , author Kurowski, M. , author Jeinsch, T. , author Schubert, A. , author Gluch, M. , author Simanski, O. , author Pairet-Garcia, E. , author Siemer, F. , author Abel, D. , year 20...

    2019

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.