Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Separating perception from symbolic rules yields safer, transparent UAV landings, a neuro-symbolic pipeline that outperforms end-to-end deep baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 08:08 UTC pith:FFL7UQRP

load-bearing objection The headline success count in the abstract is missing from the experiments, and the quality metrics are partly circular, but the neuro-symbolic integration itself is a real contribution worth a careful look. the 4 major comments →

arxiv 2510.22204 v3 pith:FFL7UQRP submitted 2025-10-25 cs.RO cs.AI

Human-Inspired Neuro-Symbolic World Modeling and Logic Reasoning for Interpretable Safe UAV Landing Site Assessment

Weixian Qian , Tianyi Yang , Sebastian Schroder , Yao Deng , Jiaohong Yao , Xiao Cheng , Richard Han , Xi Zheng This is my paper
classification cs.RO cs.AI
keywords UAV safe landingneuro-symbolic reasoningprobabilistic semantic scene graphsemantic segmentationlogic-based safety gateinterpretable AIedge deploymentcovariate shift
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

NeuroSymLand is trying to establish that a UAV landing-site safety system should not let a neural network decide safety end-to-end. Instead, a lightweight segmentation model feeds a probabilistic scene graph of regions, attributes, and spatial relations, and offline-synthesized logic rules executed at runtime act as a white-box safety gate. Across 72 simulated and hardware-in-the-loop scenarios, this design selected safe sites in 61 cases, versus 37 to 57 for four baselines, while producing human-readable justifications and running within edge compute budgets. The payoff if true: mission-critical landing decisions can be trustworthy, auditable, and reconfigurable without retraining the perception stack.

Core claim

On the paper's own terms, the central claim is that decoupling perception from safety reasoning is both practical and superior. The authors argue that a frozen quantized segmentation model can be lifted into a probabilistic semantic scene graph of regions, attributes, and spatial relations, and that probabilistic first-order logic rules — synthesized offline by a language model with human refinement — can be executed over that graph as a white-box gate, producing ranked, explained landing candidates. They report that this combination outperforms perception-only, deterministic-logic, and vision-language baselines in success count and in safety-buffer/centering quality, with symbolic inference

What carries the argument

The load-bearing object is the probabilistic semantic scene graph (PSSG), a graph whose nodes are candidate landing regions with calibrated attributes (area, flatness, class confidence) and whose edges encode spatial relations; evidence is represented as probabilistic facts with calibrated confidences. Over this graph, the framework runs probabilistic first-order logic rules in a deductive engine, using top-k proof search, product t-norm proof weights, and noisy-OR aggregation to produce a calibrated per-region safety score; mission-specific weights then rank surviving candidates. The decoupling — rules never see raw pixels, only graph facts — is what makes the gate editable, auditable, and

Load-bearing premise

The whole safety guarantee rests on the upstream segmentation labeling surfaces correctly; if the backbone mistakes a roof for flat ground, the white-box logic can certify an unsafe touchdown, which the paper's own failure case demonstrates.

What would settle it

Feed the pipeline a scene where a roof, a wet asphalt patch, and a shadowed grassy area are present with ground-truth labels; if a 'safe' verdict ever fires for a roof or water category due to segmentation error, the claim that the symbolic gate provides safety robustness under covariate shift is refuted. More simply, flip a single segmentation label from 'roof' to 'ground' in one scene and show the safety score changes to pass.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A landing decision can carry a replayable proof trace showing exactly which safety rules and scene facts fired, enabling certification-style audit rather than acceptance by aggregate metrics.
  • Mission policy (emergency, rescue, payload-safe landing) can be changed by editing rule weights or thresholds in the symbolic layer, with no retraining of the perception network.
  • Because symbolic inference takes roughly 8 ms versus about 997 ms for scene-graph construction, the reasoning layer is essentially free, leaving headroom for other onboard tasks.
  • New safety constraints can be added as rules without catastrophic forgetting of prior guarantees, because the rule set extends rather than overwrites.
  • Robust behavior under covariate shift, relative to end-to-end baselines, follows from the geometric, non-learned computation of attributes like flatness and adjacency.

Where Pith is reading between the lines

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

  • If perception fidelity is the binding constraint, the paper's claimed robustness may improve by adding multi-view or depth consistency checks to the segmentation stage, an extension the authors themselves flag as future work rather than a demonstrated result.
  • The same scene-graph-plus-rule architecture could transfer to other safety-critical robotics decisions, such as ground-vehicle hazard assessment or manipulator collision checks, as long as a symbolic rule set can be authored and maintained for that domain.
  • A testable prediction: if the segmentation backbone is deliberately fed images with roof/ground ambiguity or wet surfaces, the symbolic gate's performance will track the segmentation's semantic label accuracy; measuring that correlation would quantify the boundary of the paper's robustness claim.
  • One could measure the value of the white-box layer directly by taking the same probabilistic scene graph and comparing the logic engine's rankings against those of a fine-tuned transformer on the same candidates; the paper's 61-versus-37-57 comparison already approximates this, but only over simulated scenes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes NeuroSymLand, a neuro-symbolic framework for UAV safe-landing-site (SLZ) assessment that decouples a frozen quantized segmentation backbone from an LLM-synthesized, human-refined Scallop rule base. The pipeline builds a probabilistic semantic scene graph from segmentation and geometric post-processing, then applies probabilistic first-order logic rules as a white-box safety gate, producing ranked landing candidates with provenance-based explanations. The authors claim that across 72 simulated and hardware-in-the-loop scenarios NeuroSymLand achieves 61 successful assessments versus 37-57 for four baselines, as well as Pareto-better quality (MOD/TCD) on successful SLZs, robustness to covariate shift, and negligible edge overhead. The experimental section reports only three lightweight baselines in the quantitative comparison, gives no success counts, and does not define the success criterion; the quality metrics are the same objectives encoded in the symbolic rules.

Significance. If the claimed results were properly supported, the paper would make a useful contribution to interpretable and verifiable UAV landing safety: decoupling perception from symbolic reasoning, using LLM-plus-human synthesis of Scallop rules, and showing that the symbolic layer costs only ~8 ms on a Jetson edge device are all attractive properties. The provenance-based explanations (Table VI) and the qualitative mission-specific ranking demonstration (Fig. 5) are genuinely interesting. However, the manuscript's headline quantitative claim is not backed by the reported experiments; no success counts, no significance tests, and no quantitative robustness metric are presented. The central 'higher accuracy' claim is therefore currently unverified, and the quality comparison is partly circular because MOD and TCD are the same objectives encoded in the safety rules.

major comments (4)
  1. [Abstract and §IV-B] The abstract's central claim of '61 successful assessments' vs. '37-57 successes' for baselines is not reported anywhere in the experimental section. §IV-B defines only MOD and TCD as metrics and Table III reports mean MOD/TCD for three baselines (SegOpticalFlow, DetFOL, LLM-Explain), not four, with no success counts or success-rate table. Furthermore, the described setup (15 maps × 2 SLZs × 5 weather conditions × 3 intensities = 450 configurations) does not map to the claimed 72 scenarios. The authors must either report the per-model success counts with an explicit definition of a 'successful assessment' and a scenario-count reconciliation, or remove/soften the abstract claim. Without this, the headline result cannot be verified or falsified.
  2. [§IV-B1, §IV-B2 and §III-C/III-D] The quality claim is partly circular. MOD (minimum obstacle distance) and TCD (touchdown-centroid distance) are introduced as the two quality metrics, while the symbolic rules in Table II explicitly penalize proximity to water, trees, pools, obstacles, and buildings, and the mission ranking in §III-E weights distance-to-center and safety margins. Thus NeuroSymLand is evaluated on exactly the objectives it was designed to optimize, and its Pareto superiority over baselines that do not encode these same objectives is unsurprising. To support a claim of better SLZ quality, the paper needs an external or at least independent ground truth—for example, human-annotated safe zones on the test images, closed-loop landing success in simulation, or a defined 'successful assessment' criterion that is not simply the rule output. At minimum, the authors should acknowledge this alignment and temper the
  3. [§IV-B, Table III] Table III reports only single mean values for MOD and TCD, with no standard deviations, confidence intervals, or significance tests. The SIL setup randomizes wind and time of day, and the HIL setup runs 100 trials per configuration, so variance information is available and should be reported. Without it, the +17.4%, +53.6%, and +82.4% MOD differences and the corresponding TCD reductions cannot be distinguished from noise. Please add error bars/quantiles and a statistical test (or at least per-trial distributions) for the reported aggregates.
  4. [§IV-B d and §IV-B b] The claimed robustness to covariate shift is not quantitatively demonstrated. The experiments vary weather, wind, and time of day, but no robustness metric (e.g., success rate per weather condition, or segmentation accuracy under fog/rain) is reported. Moreover, the paper's own failure case (§IV-B d, Fig. 3b) shows that when the segmentation backbone misclassifies roof as flat ground, all downstream selectors fail. This is a load-bearing limitation: the white-box safety gate inherits any perception bias, so 'safe' certifications are only as valid as the segmentation labels. The manuscript should quantify perception fidelity under the tested covariate shifts and report how often this propagates to SLZ decisions, or explicitly scope the safety claims to the case where segmentation is accurate.
minor comments (5)
  1. [Abstract and §I] The abstract says 'four competitive baselines' but §IV-B evaluates only three (SegOpticalFlow, DetFOL, LLM-Explain); VisLanding and SafeUAV appear only qualitatively in RQ3. Please align the count with the actual experimental comparison or clarify that the fourth baseline is a qualitative one.
  2. [§II-A / §III-A] The references to prior markerless landing methods are sparse and do not include recent learning-based SLZ detection work beyond a few examples. The authors should position against a broader set of recent baselines, especially if they claim to outperform 'state-of-the-art' methods.
  3. [§IV-C, Table IV] The stage-wise latency sums to 312 + 350 + 997 + 8 = 1667 ms, while the text reports 1668 ms. Also, the FPS values in Fig. 4 (median 0.600 FPS = ~1667 ms) should be cross-checked against the stage-wise sum for consistency.
  4. [Throughout] There are several typos and formatting issues, e.g., 'UA Vs' in the abstract and introduction, 'SEURO' in the title, and the duplicate use of 'V' for UAV. A careful proofreading pass is needed.
  5. [§IV-D c, Table VI] The provenance table is a nice contribution, but the notation 'r_hazard' is used inconsistently (e.g., rule IDs in Table II are R1-R13, while Table VI uses rule names). Please align the naming.

Circularity Check

0 steps flagged

No significant circularity: the symbolic safety rules, scene graph construction, and evaluation metrics are not mutually defining; the MOD/TCD overlap is an evaluation alignment, not a derivation input.

full rationale

The paper's claimed derivation chain is: frozen INT8 SegFormer segmentation -> probabilistic semantic scene graph with geometric attributes -> LLM-plus-human-in-the-loop synthesized Scallop P-FOL rules -> top-k proof safety scoring -> mission-weighted ranking. None of these stages is defined in terms of the evaluation metrics or the claimed outcome. The offline rule corpus is explicitly held out: 'This corpus is used exclusively for rule synthesis and refinement, with no reuse in evaluation.' The rules are grounded in external landing-safety norms (avoid water, obstacles, humans; prefer large, flat, accessible regions) and are executed over a scene graph produced by a fixed perception backbone. The paper's own failure case ('the backbone segmentation model misclassifies roof surfaces as flat ground') shows the pipeline can fail despite the rules, so the safety-gate output is not forced by the rule definitions alone. The MOD/TCD quality metrics do overlap with the stated safety objectives (hazard stand-off and center proximity), making the RQ1 comparison partly an alignment check rather than a fully independent benchmark; however, this is an evaluation-design choice, not a derivation from, or statistical fit to, those metrics. The abstract's 61/72 success count is not reported in the experimental section, but an unsupported or unverifiable claim is a correctness/evidence problem rather than a circularity problem. No self-citation chain, imported uniqueness theorem, or ansatz-by-citation carries a load-bearing argument. Therefore no significant circularity is identified.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests mainly on domain assumptions about perception reliability and rule completeness, plus several unreported thresholds/weights. No new physical entities are introduced. The most exposed premise is that a monocular, quantized segmentation backbone can supply faithful semantic labels across the tested covariate shifts—an assumption the paper's own rooftop failure case shows to be fragile.

free parameters (7)
  • Safety gate threshold τ_mission
    A region passes the white-box gate iff S_safe(i) >= τ_mission; the value is never reported, yet it determines go/no-go behavior.
  • MFV hazard and jitter thresholds τ_haz, τ_jit
    Used in Pass_T(v) for temporal validation; no values given, so the strictness of multi-frame validation is unspecified.
  • Mission weights ω_m,k
    Define emergency/rescue/safe-landing ranking but numeric weights are not reported; the qualitative mission-specific outputs depend on them.
  • Spatial relation thresholds (near_to, adjacent_to, contain radius)
    Rules R3–R11 rely on these relations; their geometric parameters are not specified, changing which hazards trigger.
  • Top-k proof count k
    Unsafe risk is approximated by noisy-OR over the top-k proofs; k is not reported.
  • Multi-frame window T and temporal smoothing
    The MFV window W_T and smoothing horizon are not quantified despite claims of temporal consistency.
  • Fact probability calibration
    Grounded facts P(f) are said to be calibrated from segmentation posteriors and geometric routines, but the calibration method is not described.
axioms (5)
  • domain assumption The semantic classes in Table I are sufficient to characterize landing safety.
    Rules reference only these classes; hazards outside the taxonomy (wires, ice, wet surfaces, unlisted debris) are invisible to the gate.
  • domain assumption 2D monocular proxies for flatness, slope, and area accurately predict 3D touchdown safety.
    Flatness/slope are computed as geometric proxies from a single RGB segmentation without depth or extrinsic validation.
  • domain assumption SegFormer-B0 fine-tuned on Semantic Drone Dataset transfers to AirSim terrains and weather conditions.
    All quantitative claims are tested in AirSim with rain/snow/fog/dust, but no per-condition perception accuracy is reported; the roof misclassification failure in Fig. 3b shows this assumption can fail.
  • ad hoc to paper The LLM + human review loop produces a complete and consistent set of safety rules.
    The rules in Table II are accepted after human review, but completeness and uniqueness are not proven; a different LLM, prompt set, or reviewer set could yield different rules.
  • standard math Scallop's t-norm/noisy-OR aggregation with top-k proofs yields calibrated safety scores.
    The probabilistic semantics are assumed correct; the final S_safe score is never validated against actual landing outcomes or a calibration check.

pith-pipeline@v1.3.0-alltime-deepseek · 13309 in / 13257 out tokens · 138430 ms · 2026-08-04T08:08:53.063445+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Human-Inspired Neuro-Symbolic World Modeling and Logic Reasoning for Interpretable Safe UAV Landing Site Assessment." pith.science (2026). https://pith.science/paper/FFL7UQRP

@misc{pith2026251022204,
  author       = {Pith},
  title        = {Pith review of: Human-Inspired Neuro-Symbolic World Modeling and Logic Reasoning for Interpretable Safe UAV Landing Site Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FFL7UQRP}},
  note         = {Machine review of arXiv:2510.22204}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reliable assessment of safe landing sites in unstructured environments is essential for deploying Unmanned Aerial Vehicles (UAVs) in real-world applications such as delivery, inspection, and surveillance. Existing learning-based approaches often degrade under covariate shift and offer limited transparency, making their decisions difficult to interpret and validate on resource-constrained platforms. We present NeuroSymLand, a neuro-symbolic framework for marker-free UAV landing site safety assessment that explicitly separates perception-driven world modeling from logic-based safety reasoning. A lightweight segmentation model incrementally constructs a probabilistic semantic scene graph encoding objects, attributes, and spatial relations. Symbolic safety rules, synthesized offline via large language models with human-in-the-loop refinement, are executed directly over this world model at runtime to perform white-box reasoning, producing ranked landing candidates with human-readable explanations of the underlying safety constraints. Across 72 simulated and hardware-in-the-loop landing scenarios, NeuroSymLand achieves 61 successful assessments, outperforming four competitive baselines, which achieve between 37 and 57 successes. Qualitative analysis highlights its superior interpretability and transparent reasoning, while deployment incurs negligible edge overhead. Our results suggest that combining explicit world modeling with symbolic reasoning can support accurate, interpretable, and edge-deployable safety assessment in mobile systems, as demonstrated through UAV landing site assessment.

Figures

Figures reproduced from arXiv: 2510.22204 by Jiaohong Yao, Richard Han, Sebastian Schroder, Tianyi Yang, Weixian Qian, Xiao Cheng, Xi Zheng, Yao Deng.

Figure 1
Figure 1. Figure 1: NEUROSYMLAND pipeline. The left is online inference phase : SegFormer with lightweight post-processing (contours, distances, flatness/flow) extracts candidate regions. Cues are lifted to a probabilistic semantic scene graph; a Scallop-based, task-weighted selector infers hazards and ranks SLZs. Right: offline human-in-the-loop produces frozen rules for deployment. flatness/slope proxies (from local geometr… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the LLM–human-in-the-loop rule learn [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative outcomes in the Airsim. Two representa￾tive cases illustrate how the pipeline behaves under different scene conditions. c) Result Analysis:: The experiment shows NEUROSYM￾LAND is Pareto-superior on both safety buffer (MOD↑) and maneuver economy (TCD↓). Compared to SegOpticalFlow, it yields a +17.4% increase in MOD (112.81 vs. 96.13) and a 21.9% reduction in TCD (126.77 vs. 162.36), indicat￾ing … view at source ↗
Figure 4
Figure 4. Figure 4: Inference speed (FPS; higher is better). Boxes show the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case study for a skateboard park. Green: feasible SLZ. Only the top-k candidates (k = 10) are shown in ours to avoid clutter; inference uses the full candidate set. TABLE V: System metrics on Jetson Orin. Model CPU (%) GPU (%) Mem (GB) Power (W) Temp ( ◦C) NEUROSYMLAND (ours) 76.87 33 3.28 1.23 47.91 DetFOL 77.12 33 3.10 1.196 47.21 SegOpticalFlow 59.82 33 2.09 1.07 45.20 optimization. As shown in Table IV… view at source ↗
Figure 6
Figure 6. Figure 6: Semantic node distribution (Ours). Calibrated seg￾mentation is overlaid on the RGB frame; colors indicate classes (e.g., paved, vegetation, person, obstacle). The white grid is the spatial index used for candidate sampling, with zone IDs shown at centers. heuristics—such as assigning higher priority to proximity￾based objectives like “closer to the center” for emergency touchdowns or “closer to a rescue ta… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Vision-and-Language Navigation for UAVs: Progress, Challenges, and a Research Roadmap

    cs.RO 2026-04 unverdicted novelty 4.0

    A survey of UAV vision-and-language navigation that establishes a methodological taxonomy, reviews resources and challenges, and proposes a forward-looking research roadmap.

Reference graph

Works this paper leans on

46 extracted references · 12 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Free lsd: Prior-free visual landing site detection for autonomous planes,

    T. Hinzmann, T. Stastny, C. Cadena, R. Siegwart, and I. Gilitschen- ski, “Free lsd: Prior-free visual landing site detection for autonomous planes,”IEEE Robotics and Automation Letters, vol. 3, no. 3, pp. 2545– 2552, 2018

    2018

  2. [2]

    Multi-resolution elevation mapping and safe landing site detection with applications to planetary rotorcraft,

    P. Schoppmann, P. F. Proenc ¸aet al., “Multi-resolution elevation mapping and safe landing site detection with applications to planetary rotorcraft,” 2021. [Online]. Available: https://arxiv.org/abs/2111.06271

    Pith/arXiv arXiv 2021

  3. [3]

    Visual environment assessment for safe autonomous quadrotor landing,

    M. Secchiero, N. Bobbili, Y . Zhou, and G. Loianno, “Visual environment assessment for safe autonomous quadrotor landing,” 2024. [Online]. Available: https://arxiv.org/abs/2311.10065

    Pith/arXiv arXiv 2024

  4. [4]

    Vislanding: Monocular 3d perception for uav safe landing via depth-normal synergy,

    Z. Tan, B. He, Y . Ji, and L. Wu, “Vislanding: Monocular 3d perception for uav safe landing via depth-normal synergy,” 2025. [Online]. Available: https://arxiv.org/abs/2506.14525

    Pith/arXiv arXiv 2025

  5. [5]

    Emergency landing spot detection algorithm for unmanned aerial vehicles,

    G. Loureiro, A. Dias, A. Martins, and J. Almeida, “Emergency landing spot detection algorithm for unmanned aerial vehicles,”Remote Sensing, vol. 13, no. 10, p. 1930, 2021

    1930

  6. [6]

    Emergency uav landing on unknown field using depth-enhanced graph structure,

    J. Chen, W. Duet al., “Emergency uav landing on unknown field using depth-enhanced graph structure,”T-ASE, 2024

    2024

  7. [7]

    Uav computing-assisted search and rescue mission framework for disaster and harsh environment mitigation,

    S. H. Alsamhi, A. V . Shvetsovet al., “Uav computing-assisted search and rescue mission framework for disaster and harsh environment mitigation,”Drones, vol. 6, no. 7, p. 154, 2022

    2022

  8. [8]

    Design and analysis of propeller for high-altitude search and rescue unmanned aerial vehicle,

    C. Dahal, H. B. Dura, and L. Poudel, “Design and analysis of propeller for high-altitude search and rescue unmanned aerial vehicle,”Interna- tional Journal of Aerospace Engineering, vol. 2021, no. 1, p. 6629489, 2021

    2021

  9. [9]

    Target search in dynamic environments with multiple solar-powered uavs,

    Y . Lun, H. Wanget al., “Target search in dynamic environments with multiple solar-powered uavs,”TVT, vol. 71, no. 9, pp. 9309–9321, 2022

    2022

  10. [10]

    Uav-based delivery systems: A systematic review, current trends, and research challenges,

    F. Betti Sorbelli, “Uav-based delivery systems: A systematic review, current trends, and research challenges,”Journal on Autonomous Trans- portation Systems, vol. 1, no. 3, pp. 1–40, 2024

    2024

  11. [11]

    Drone- aided delivery methods, challenge, and the future: A methodological review,

    X. Li, J. Tupayachi, A. Sharmin, and M. Martinez Ferguson, “Drone- aided delivery methods, challenge, and the future: A methodological review,”Drones, vol. 7, no. 3, p. 191, 2023

    2023

  12. [12]

    Review of navigation methods for uav-based parcel delivery,

    D. Dissanayaka, T. R. Wanasinghe, O. De Silva, A. Jayasiri, and G. K. Mann, “Review of navigation methods for uav-based parcel delivery,” IEEE Transactions on Automation Science and Engineering, vol. 21, no. 1, pp. 1068–1082, 2023

    2023

  13. [13]

    Computer vision for autonomous uav flight safety: An overview and a vision-based safe landing pipeline example,

    E. Kakaletsis, C. Symeonidiset al., “Computer vision for autonomous uav flight safety: An overview and a vision-based safe landing pipeline example,”Acm Computing Surveys (Csur), vol. 54, no. 9, pp. 1–37, 2021

    2021

  14. [14]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,”Nature Machine Intelligence, vol. 1, no. 5, pp. 206–215, 2019

    2019

  15. [15]

    Chatgpt: O3-mini,

    OpenAI, “Chatgpt: O3-mini,” https://openai.com/chatgpt, 2025, large language model

    2025

  16. [16]

    Reluplex: An efficient SMT solver for verifying deep neural networks,

    G. Katz, C. Barrett, D. L. Dill, K. Julian, and M. J. Kochenderfer, “Reluplex: An efficient SMT solver for verifying deep neural networks,” inComputer Aided Verification (CAV 2017), ser. Lecture Notes in Computer Science, vol. 10426. Springer, 2017, pp. 97–117

    2017

  17. [17]

    The marabou framework for verification and analysis of deep neural networks,

    G. Katz, D. A. Huanget al., “The marabou framework for verification and analysis of deep neural networks,” inComputer Aided Verification (CAV 2019), ser. Lecture Notes in Computer Science, vol. 11561. Springer, 2019, pp. 443–452

    2019

  18. [18]

    Scallop: A language for neurosymbolic programming,

    Z. Li, J. Huang, and M. Naik, “Scallop: A language for neurosymbolic programming,”Proceedings of the ACM on Programming Languages, vol. 7, no. PLDI, Jun 2023

    2023

  19. [19]

    Neurosymbolic ai: The 3rd wave,

    A. d’Avila Garcez and L. C. Lamb, “Neurosymbolic ai: The 3rd wave,”

  20. [20]

    Neuro-symbolic continual learning: Knowledge, reasoning shortcuts and concept rehearsal,

    E. Marconato, G. Bontempo, E. Ficarra, S. Calderara, A. Passerini, and S. Teso, “Neuro-symbolic continual learning: Knowledge, reasoning shortcuts and concept rehearsal,” 2023. [Online]. Available: https: //arxiv.org/abs/2302.01242

    Pith/arXiv arXiv 2023

  21. [21]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural Networks, vol. 113, pp. 54–71, 2019

    2019

  22. [22]

    Apriltag: A robust and flexible visual fiducial system,

    E. Olson, “Apriltag: A robust and flexible visual fiducial system,” in2011 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2011, pp. 3400–3407

    2011

  23. [23]

    Apriltag 2: Efficient and robust fiducial detec- tion,

    J. Wang and E. Olson, “Apriltag 2: Efficient and robust fiducial detec- tion,” in2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Daejeon, South Korea, 2016

    2016

  24. [24]

    Automatic generation and detection of highly reliable fiducial markers under occlusion,

    S. Garrido-Jurado, R. Mu ˜noz-Salinas, F. J. Madrid-Cuevas, and M. J. Mar´ın-Jim´enez, “Automatic generation and detection of highly reliable fiducial markers under occlusion,”Pattern Recognition, vol. 47, no. 6, pp. 2280–2292, 2014

    2014

  25. [25]

    J. W. Lloyd,Foundations of Logic Programming, 2nd ed. Berlin, Heidelberg: Springer, 1987

    1987

  26. [26]

    Problog: A probabilistic prolog and its application in link discovery,

    L. De Raedt, A. Kimmig, and H. Toivonen, “Problog: A probabilistic prolog and its application in link discovery,” inIJCAI, Hyderabad, India, 2007, pp. 2462–2467

    2007

  27. [27]

    Deepproblog: Neural probabilistic logic programming,

    R. Manhaeve, S. Duman ˇci´c, A. Kimmig, T. Demeester, and L. De Raedt, “Deepproblog: Neural probabilistic logic programming,” inAdvances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  28. [28]

    Tensorlog: A differentiable deductive database,

    W. W. Cohen, “Tensorlog: A differentiable deductive database,” arXiv preprint arXiv:1605.06523, 2016. [Online]. Available: https: //arxiv.org/abs/1605.06523

    Pith/arXiv arXiv 2016

  29. [29]

    Scallop: From probabilistic deductive databases to scalable differentiable reasoning,

    J. Huang, Z. Liet al., “Scallop: From probabilistic deductive databases to scalable differentiable reasoning,” inAdvances in Neural Information Processing Systems (NeurIPS 2021), 2021

    2021

  30. [30]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohanet al., “Do as i can, not as i say: Grounding language in robotic affordances,” 2022. [Online]. Available: https://arxiv.org/abs/2204.01691

    Pith/arXiv arXiv 2022

  31. [31]

    Code as policies: Language model programs for embodied control,

    J. Liang, W. Huanget al., “Code as policies: Language model programs for embodied control,” 2023. [Online]. Available: https: //arxiv.org/abs/2209.07753

    Pith/arXiv arXiv 2023

  32. [32]

    Pal: Program-aided language models,

    L. Gao, A. Madaan, S. Zhou, U. Alon, P. Liu, Y . Yang, J. Callan, and G. Neubig, “Pal: Program-aided language models,” 2023. [Online]. Available: https://arxiv.org/abs/2211.10435

    Pith/arXiv arXiv 2023

  33. [33]

    Satlm: Satisfiability- aided language models using declarative prompting,

    X. Ye, Q. Chen, I. Dillig, and G. Durrett, “Satlm: Satisfiability- aided language models using declarative prompting,” 2023. [Online]. Available: https://arxiv.org/abs/2305.09656

    Pith/arXiv arXiv 2023

  34. [34]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” 2021. [Online]. Available: https: //arxiv.org/abs/2105.15203

    Pith/arXiv arXiv 2021

  35. [35]

    Semantic drone dataset,

    I. of Engineering Geodesy and G. U. o. T. Measurement Sys- tems (IGMS), “Semantic drone dataset,” https://www.tugraz.at/index. php?id=22387

  36. [36]

    The OpenCV Library,

    G. Bradski, “The OpenCV Library,”Dr. Dobb’s Journal of Software Tools, 2000

    2000

  37. [37]

    Safeuav: Learning to estimate depth and safe landing areas for uavs from synthetic data,

    A. Marcu, D. Costea, V . Licaret, M. P ˆırvu, E. Slusanschi, and M. Leordeanu, “Safeuav: Learning to estimate depth and safe landing areas for uavs from synthetic data,” inECCV Workshops, 2018, pp. 0–0

    2018

  38. [38]

    Accurate autonomous uav landing using vision-based detection of aruco-marker,

    I. Lebedev, A. Erashov, and A. Shabanova, “Accurate autonomous uav landing using vision-based detection of aruco-marker,” inInterac- tive Collaborative Robotics, ser. Lecture Notes in Computer Science. Springer, 2020, vol. 12336, pp. 179–188

    2020

  39. [39]

    Airsim: High-fidelity visual and physical simulation for autonomous vehicles,

    S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for autonomous vehicles,” inField and Service Robotics: Results of the 11th International Conference. Springer, 2018, pp. 621–635

    2018

  40. [40]

    Vision-based autonomous uav navigation and landing for urban search and rescue,

    M. Mittal, R. Mohan, W. Burgard, and A. Valada, “Vision-based autonomous uav navigation and landing for urban search and rescue,” inThe International Symposium of Robotics Research. Springer, 2019, pp. 575–592

    2019

  41. [41]

    Adaptive gain control strategy for constant optical flow divergence landing,

    H. W. Ho, G. C. H. E. de Croon, E. van Kampen, Q. P. Chu, and M. Mulder, “Adaptive gain control strategy for constant optical flow divergence landing,”IEEE Transactions on Robotics, vol. 34, no. 2, pp. 508–516, 2018

    2018

  42. [42]

    Evolution of robust high- speed optical-flow-based landing for autonomous micro air vehicles,

    K. Y . W. Scheper and G. C. H. E. de Croon, “Evolution of robust high- speed optical-flow-based landing for autonomous micro air vehicles,” Robotics and Autonomous Systems, vol. 124, p. 103380, 2020

    2020

  43. [43]

    Vision-guided landing of an autonomous helicopter in hazardous terrain,

    A. E. Johnson, J. F. Montgomery, and L. H. Matthies, “Vision-guided landing of an autonomous helicopter in hazardous terrain,” inProceed- ings of the IEEE International Conference on Robotics and Automation (ICRA), 2005, pp. 3966–3971

    2005

  44. [44]

    Visually-guided landing of an unmanned aerial vehicle,

    S. Saripalli, J. F. Montgomery, and G. S. Sukhatme, “Visually-guided landing of an unmanned aerial vehicle,”IEEE Transactions on Robotics and Automation, vol. 19, no. 3, pp. 371–380, 2003

    2003

  45. [45]

    Phi-3 technical report: A highly capable language model locally on your phone,

    M. Abdinet al., “Phi-3 technical report: A highly capable language model locally on your phone,” 2024. [Online]. Available: https://arxiv.org/abs/2404.14219

    Pith/arXiv arXiv 2024

  46. [2020]

    Available: https://arxiv.org/abs/2012.05876

    [Online]. Available: https://arxiv.org/abs/2012.05876

    Pith/arXiv arXiv 2012