← Blog

Paper F Paper Group Planning v1: UAV safety critical scenario generation, coverage and emergency application

Multiple paper routes are planned for UAV safety-critical scene generation, scene coverage, city-local scene correlation, and high-speed emergency rescue resource allocation directions.

Paper FUAVscene generationscene coverageSafety-Criticalaccelerated testingemergency rescueTR-CT-ITS

Paper F Paper group planning v1: UAV safety critical scenario generation, coverage and emergency application

Overall judgment: In addition to Paper B’s hundred-shelf-level system scheduling, Paper C’s 3DGS/FIM active sensing, and Paper E’s LLM/LTL language planning, you can also open a separate UAV safety-critical scenario engineering paper line.
The core of this line is not to create another obstacle avoidance algorithm, but to answer: **How ​​to systematically generate, cover, filter and reuse key UAV safety scenarios so that subsequent training, testing, emergency dispatch and paper experiments have a credible scenario base. **

1. Overall judgment: What other directions can be written?

There are currently several paper lines focusing on different issues:

Thesis lineCore objectsAlready coveredShould not be repeated
Paper BHundreds of UAV fleetThree-layer hierarchical scheduling, queuing theory, Lyapunov, multi-modal transportationNo longer writing large-scale fleet scheduling separately
Paper CUAV active sensing3DGS, Fisher information, next-best-viewNo longer focused on mapping/perspective selection
Paper ELanguage to planningLLM, TaskIR, LTL/STL, formal verificationNo longer focused on natural language task planning
Paper FScene EngineeringScene generation, coverage, dangerous scenes, emergency applicationsNew directions

The value of Paper F is that it can become the experimental infrastructure for the previous papers:

2. Paper group hierarchical design

Paper F is recommended to be planned as 4 progressive papers:

LevelPaperOne sentence positioningPriority
F1CovUAV-BenchUAV safety critical scene coverage benchmarkHighest
F2Coverage-Guided Accelerated TestingCoverage-guided dangerous scene acceleration generation algorithmHighest
F3City2Local-UAVCreate hierarchical scene generation from overall city ODD to local obstacle combinationMedium to High
F4Scenario-Aware Emergency ResponseDo Shandong high-speed UAV-ground resource collaborative emergency deploymentMedium to high

It is recommended to do F1 + F2 first. F1 provides the dataset, metrics, and problem definition, and F2 provides the algorithmic contributions. F3 and F4 can be used as extensions: F3 turns the benchmark into a city-level system, and F4 turns the scene engineering into real traffic emergency business.

3. Common Background: Why scene coverage is the foundation for UAV safety research

A common weakness in UAV security research is: the algorithm is beautifully done, but the experimental scenario is too arbitrary. Just because an obstacle avoidance algorithm is successful in 20 manual scenarios does not mean that it covers the long-tail risks in urban low-altitude operations.A clear consensus has been reached in the field of autonomous driving: crashes/near-crash on real roads are rare events, and relying directly on natural testing would be extremely inefficient. Therefore, Shuo Feng et al. proposed a naturalistic and adversarial environment, using natural distribution to maintain authenticity, and using adversarial distribution to increase the probability of dangerous events, thereby accelerating intelligent driving testing [1]. They further proposed testing scenario library generation, defining the test scenario library under ODD as a set of representative and critical scenarios, and using criticality to take into account exposure frequency and maneuver challenge [2] [3]. Ding et al.’s review of safety-critical scenario generation also classified the field into three types of methods: data-driven, adversarial and knowledge-based, and pointed out that fidelity, efficiency, diversity, transferability and controllability are core challenges [13].

The UAV scenario requires this set of ideas even more for four reasons:

  1. **Three-dimensional space is a higher dimension. ** UAVs are not just about flat lanes, but also about altitude, obstacle volume, wind field, electrical charge, sensor field of view and flight dynamics.

  2. **Hazardous events are more difficult to collect. ** There are very few samples of real collisions with buildings, line collisions, entering no-fly zones, crossing bridges or high-speed accident scenes, and training cannot rely on real accident data.

  3. **Ordinary random generation wastes computing power. ** A large number of random scenarios are either too simple, physically infeasible, or dangerous but inevitable, making them inefficient for training and evaluation.

  4. **There is no unified measurement of scene coverage. ** Existing UAV papers often report success rate/collision rate, but rarely indicate which obstacle combinations, local geometries, task difficulties, and ODD boundaries are covered by the test set.

Therefore, the common scientific questions for Paper F are:

How to build a UAV scenario generation and evaluation system that is real, controllable, and reproducible, and can effectively cover key safety long-tail risks?

4. Paper F1: UAV safety critical scene coverage Benchmark

4.1 Thesis titleCovUAV-Bench: A Coverage-Oriented Benchmark for Safety-Critical UAV Navigation Scenarios

4.2 Background

SafeBench already provides a unified safety-critical benchmark in autonomous driving, integrating multiple types of scene templates, scene generation algorithms and evaluation indicators [5]. Scenic also proved that using probabilistic programs to express scene distribution, hard constraints and soft constraints is a feasible route [4]. There has been preliminary work on UAV simulation environment generation. For example, Nakama et al. proposed an automated UAV simulation environment generator [10]. FADS also showed that temporal-logic safety specification can enter the autonomous drone safety pipeline [11]. However, in the UAV field, there is still a lack of a coverage-oriented benchmark for 3D obstacle avoidance, low-altitude corridors, urban local spaces, and emergency tasks.

The goal of F1 is not to propose the strongest planner, but to define how the UAV scene space is covered by the system.

4.3 Method

Construct a basic test space of 50m x 50m x 50m, starting with local scenes and then expanding to urban blocks:- Scene Objects: Building blocks, towers, wires, trees, bridges, temporary obstacles, dynamic UAVs, ground vehicles, personnel areas.

Scene coverage is defined as:

Where is the discretized scene space of the target ODD, is the bin covered by the sample set on the th class attribute dimension, and is the dimension weight.

The existing 76 million explorations can be written as “Existing exploration log assets” for statistics:

Note: 76 million explorations are only written as “available experimental basis” and cannot be written as verified conclusions.

4.4 Baselines| Baseline | Purpose |

|----------|------| | Random scenario sampling | The most basic coverage baseline | | Grid sampling | Uniform discretization of parameter space | | Latin hypercube sampling | More efficient parameter coverage | | Scenic-style constrained sampling | Constrained scene generation baseline [4] | | SafeBench-style template suite | Templated security scenario baseline [5] |

4.5 Innovation points

  1. Propose UAV scene coverage taxonomy: ODD, obstacle combination, dynamic disturbance, task type, risk level.
  2. Give a coverage-oriented benchmark instead of just a few manual maps.
  3. Convert exploration log into coverage holes and critical scenario seeds.
  4. Provide a unified scene interface for subsequent Paper B/C/E.

4.6 How to evaluate

IndicatorMeaning
Parameter coverageParameter bin coverage ratio
Pairwise / t-wise coverageMulti-attribute combination coverage
Critical scenario densityNumber of near-miss / collisions discovered per unit test budget
Invalid scenario rateThe proportion of physically infeasible or mission meaningless scenarios
Planner ranking stabilityIs the algorithm ranking stable under different random seeds
Replay reproducibilityWhether the same result can be reproduced with the same seed

5. Paper F2: Accelerate the generation of dangerous scenes guided by coverage### 5.1 Thesis title

Coverage-Guided Accelerated Testing for Safety-Critical UAV Obstacle Avoidance

5.2 Background

The core of accelerated testing for autonomous driving is not to “create inevitable crash scenarios”, but to improve the sampling efficiency of safety-critical events while maintaining scene authenticity and actionability [1] [2] [3]. If the generated scenario is not feasible for any planner, then it cannot help differentiate the algorithm’s capabilities; if the generated scenario is too safe, it cannot expose system weaknesses.

UAV obstacle avoidance training also has the same problem:

5.3 Method

Proposed CGAT-UAV: Coverage-Guided Accelerated Testing for UAVs.

The algorithm consists of four modules:

  1. Scenario encoder Encode the scene into structured vectors: number of obstacles, minimum channel width, target direction, dynamic obstacle speed, wind intensity, sensor noise, battery margin, etc.

  2. Coverage memory Maintain coverage bins, failure types, and planner performance for explored scenes.

  3. Criticality score Refer to Feng’s criticality idea and combine the degree of risk with the frequency of exposure [2]:

    Among them, is used to punish inevitable collisions and physically unreasonable scenarios.4. Adaptive generator Generate new scenes in coverage holes and high-criticality regions using Bayesian optimization, CMA-ES, RL editing, or cross-entropy methods.

5.4 Baselines

BaselineComparison purpose
Random generationTest acceleration rate
Grid / Latin hypercube samplingCoverage efficiency
Bayesian optimizationBlack box dangerous search
CMA-ESContinuous Parametric Hazard Search
RL adversarial scenario generationLearning hazard generation
Scenic constrained generationRules and constraints generation [4]
FREA-style feasibility-guided generationCompare the idea of “reasonable antagonism” [12]

5.5 Innovation points

  1. Migrate accelerated testing from autonomous driving to UAV 3D obstacle avoidance.
  2. Optimize coverage, criticality, and feasibility at the same time to avoid pursuing only the collision rate.
  3. Propose a coverage-guided curriculum to train planners with dangerous but solvable scenarios.
  4. The test acceleration rate is given: the number of simulations required to reach the same confidence interval is significantly reduced.

5.6 How to evaluate| Indicator | Meaning |

|------|------| | Acceleration factor | The multiple reduction in the number of tests required to achieve the same failure discovery rate compared to random testing | | Failure discovery rate | The ratio of collision / near-miss / timeout discovered per unit budget | | Feasible criticality | Proportion of danger and feasible obstacle avoidance strategies | | Naturalness score | Whether the scene conforms to ODD prior | | Coverage gain per 1k tests | New coverage every 1000 tests | | Training efficiency | After training with generated scenarios, planner’s improvement in held-out test |

6. Paper F3: Hierarchical generation of urban overall scenes to local obstacle combinations

6.1 Thesis title

City2Local-UAV: Hierarchical Scenario Generation from Urban ODDs to Local Obstacle Compositions

6.2 Background

F1 and F2 address a local 3D test space, but real urban low-altitude flights are not isolated boxes. Why a local scene appears depends on the overall structure of the city: road grades, building density, functional areas, bridges, service areas, interchanges, hospitals, schools, no-fly zones and emergency points.

ASAM OpenODD/OpenSCENARIO provides a standardized idea from ODD, current operating domain to executable scenario description [6] [7]. The UAV field can learn from this level of abstraction, but will need to incorporate three-dimensional obstacles, airspace constraints, and low-altitude mission semantics.

6.3 Method

Propose a three-layer generation pipeline from city to local:

City-level ODD
  -> district / road / highway segment selection
  -> local 50m x 50m x 50m UAV test cell
  -> concrete obstacle composition
  -> simulator executable scenario

Specific modules:- City ODD parser: Extract city/highway semantics from OSM, road grades, building outlines, POIs, service areas, bridges and highway entrances.

6.4 Baselines

BaselineComparison purpose
Pure random local generationDoes not consider urban context
OSM-to-map direct conversionOnly converts the map, does not control scene coverage
CARLA / OSM digital twin generationGround autonomous driving digital twin baseline [14]
Manual scenario templatesManual rule templates
CityEngine / procedural city generationProcedural city generation baseline

6.5 Innovation points

  1. Associate the urban ODD with the UAV local safety test cell.
  2. Propose hierarchical scene generation of “global city semantics -> local obstacle composition”.
  3. Extend scene coverage from local parameters to urban functional area coverage.
  4. Support real city case studies, such as Jinan, Qingdao, and Shandong key highway hubs.

6.6 How to evaluate| Indicator | Meaning |

|------|------| | ODD coverage | Urban functional areas, road grades, building density coverage | | Local composition diversity | Local obstacle combination diversity | | Realism score | Consistency with OSM/POI/Building Statistics | | Transferability | Is the policy generated from one city still valid when moved to another city | | Criticality preservation | Whether urban context generation preserves high-risk local scenes |

7. Paper F4: UAV-ground resource collaborative deployment for Shandong highway emergency rescue

7.1 Thesis title

Scenario-Aware UAV-Ground Resource Allocation for Highway Emergency Response

7.2 Background

Shandong Expressway already has a business foundation for low-altitude inspection and emergency response. Public information from Shandong Hi-Speed ​​Group shows that its comprehensive inspection flight service system has deployed unattended platforms and industrial drones in key areas for road condition inspections, road inspections, emergency response and data analysis [15]. This shows that high-speed scenarios are not pure imagination, but have application entrances.

Research on highway emergency resource allocation pointed out that there are still several problems in the existing work: insufficient site selection of roadside small/micro emergency facilities during the operation phase, complete information is often assumed in the early stage of the accident but is not actually true, traffic status after the accident is uncertain and time-varying, and the integrated optimization of facility site selection, resource allocation and dispatch is still insufficient [16]. There have been studies on space-time network UAV routing in traffic incident monitoring [17], and there have been studies on UAV real-time deployment and resource allocation in disaster emergency communications [18], but they have not yet formed a unified closed loop with high-speed emergency accident scene coverage, on-site reconnaissance information value, and ground rescue resource allocation.

This is suitable for the introduction of UAV: ​​the drone first arrives at the accident scene to obtain the situation, and then the ground clearance, firefighting, rescue and control resources are dynamically dispatched.

7.3 MethodProposed Scenario-Aware UAV-Ground Emergency Dispatch:

7.4 Problem formulation

Let the highway section set be , the accident set be , the UAV set be , the ground rescue resource set be , and the service station/unattended platform set be .

Decision variables include:

Objective function:

Among them, represents the uncertainty of accident information, which can be reduced by UAV reconnaissance.

7.5 Baselines| Baseline | Comparison purpose |

|----------|----------| | Ground-only dispatch | No drone reconnaissance | | Nearest-resource dispatch | Nearest resources first | | Static facility allocation | Fixed facility allocation | | Two-stage stochastic optimization | Estimate the accident before dispatching | | UAV-first heuristic | UAV reconnaissance first, then ground dispatch | | Scenario-aware rolling optimization | Main method |

7.6 Innovation points

  1. Connect scene coverage and high-speed emergency dispatch, instead of just resource allocation.
  2. Model UAV reconnaissance as a decision-making action that reduces uncertainty in incident information.
  3. Support the real business context of Shandong Expressway: unattended platform, road condition inspection, emergency response and work order circulation.
  4. Unified optimization of response time, clearance time, secondary accident risk and dispatch cost.

7.7 How to evaluate

IndicatorMeaning
First-view timeThe time when the UAV first acquired the accident footage
Response timeArrival time of the first batch of rescue resources
Clearance timeAccident clearance completion time
Wrong dispatch rateThe proportion of wrong dispatch, missed dispatch or insufficient resources
Secondary accident riskSecondary accident risk proxy
Congestion delayTotal delay caused by accident
UAV information valueReconnaissance with UAV reduces uncertainty compared to without UAV

8. Unify experimental platform, data sources and evaluation indicators

8.1 Experimental platform| Hierarchy | Recommended implementation | Purpose |

|------|----------|------| | Lightweight simulation | Python / PyBullet / custom 3D grid | 76 million levels of rapid exploration | | UAV simulation | AirSim, Flightmare | Vision, dynamics, sensor verification [8] [9] | | Scenario language | Scenic-like DSL, JSON schema | Reproducible scene generation [4] | | City data | OpenStreetMap, POI, road grades, building outlines | City to local scene generation | | High-speed emergency | Shandong expressway open cases, accident statistics, synthetic accident flow | Emergency resource allocation experiment |

The main experiment of F1/F2 should prioritize lightweight simulation to ensure large-scale exploration. AirSim/Flightmare is used for small-scale high-fidelity verification and is not relied upon for all experiments.

8.2 Data source

8.3 Unified indicators| Indicator Group | Indicator |

|--------|------| | Coverage | parameter coverage, t-wise coverage, ODD coverage, coverage gain | | Safety | collision rate, near-miss rate, minimum distance, constraint violation | | Danger generation | criticality, failure discovery rate, acceleration factor, feasible criticality | | Training value | sample efficiency, held-out success rate, robustness under ODD shift | | Emergency value | first-view time, response time, clearance time, wrong dispatch rate |

9.1 The first stage: do F1 + F2 first

In the first phase, it is recommended to write two articles directly around “UAV safety critical scene coverage + accelerated testing”:

  1. F1 benchmark paper More stable, suitable as an experimental base for all subsequent UAV papers. Even if the algorithm is not particularly strong, it can still be established based on taxonomy, coverage metric, data set and reproducible experiments.

  2. F2 method paper Methodological contributions to AAAI/ICRA/IROS. The highlight is the migration from Shuo Feng’s accelerated testing of autonomous driving to UAV 3D scenes and the addition of coverage-guided feasible criticality.

9.2 Phase 2: Do F3 + F4 again

F3 and F4 are more suitable for advancement after F1/F2 has a tool foundation:

PaperHow to support Paper F
Paper BProvides peak / shock / highway emergency demand scenarios
Paper CProvides local 3D occlusion, perspective coverage and reconstruction difficult scenes
Paper EProvides natural language tasks, map entities and safety constraint scenarios

Paper F is best suited to be the “scenario infrastructure paper” for the entire UAV research line.

10. References

[1] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X. Liu. “Intelligent Driving Intelligence Test for Autonomous Vehicles with Naturalistic and Adversarial Environment.” Nature Communications, 12:748, 2021. DOI: 10.1038/s41467-021-21007-8. URL: https://doi.org/10.1038/s41467-021-21007-8

[2] Shuo Feng, Yiheng Feng, Chunhui Yu, Yi Zhang, and Henry X. Liu. “Testing Scenario Library Generation for Connected and Automated Vehicles, Part I: Methodology.” IEEE Transactions on Intelligent Transportation Systems, 22(3):1573-1582, 2021. DOI: 10.1109/TITS.2020.2972211. URL: https://doi.org/10.1109/TITS.2020.2972211[3] Shuo Feng, Yiheng Feng, Haowei Sun, Shan Bao, Yi Zhang, and Henry X. Liu. “Testing Scenario Library Generation for Connected and Automated Vehicles, Part II: Case Studies.” IEEE Transactions on Intelligent Transportation Systems, 22(9):5635-5647, 2021. DOI: 10.1109/TITS.2020.2988309. URL: https://doi.org/10.1109/TITS.2020.2988309

[4] Daniel J. Fremont, Tommaso Dreossi, Shromona Ghosh, Xiangyu Yue, Alberto L. Sangiovanni-Vincentelli, and Sanjit A. Seshia. “Scenic: A Language for Scenario Specification and Scene Generation.” Proceedings of the 40th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), 2019. DOI: 10.1145/3314221.3314633. URL: https://people.eecs.berkeley.edu/~sseshia/pubs/b2hd-fremont-pldi19.html[5] Chejian Xu, Wenhao Ding, Weijie Lyu, Zuxin Liu, Shuai Wang, Yihan He, Hanjiang Hu, Ding Zhao, and Bo Li. “SafeBench: A Benchmarking Platform for Safety Evaluation of Autonomous Vehicles.” Advances in Neural Information Processing Systems 35 (NeurIPS 2022) Datasets and Benchmarks Track, 2022. URL: https://proceedings.neurips.cc/paper_files/paper/2022/hash/a48ad12d588c597f4725a8b84af647b5-Abstract-Datasets_and_Benchmarks.html

[6] ASAM. “ASAM OpenSCENARIO DSL: Key Terminology and Conceptual Overview.” URL: https://publications.pages.asam.net/standards/ASAM_OpenSCENARIO/ASAM_OpenSCENARIO_DSL/latest/conceptual-overview/key_terms.html

[7] ASAM. “ASAM OpenODD: Model to ASAM OpenSCENARIO DSL Mapping Reference.” URL: https://publications.pages.asam.net/standards/ASAM_OpenODD/ASAM_OpenODD/latest/specification/09_openscenario_dsl/09_01_overview.html[8] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. “AirSim: High-Fidelity Visual and Physical Simulation for Autonomous Vehicles.” Field and Service Robotics, Springer Proceedings in Advanced Robotics, 2017; arXiv:1705.05065. URL: https://arxiv.org/abs/1705.05065

[9] Yunlong Song, Selim Naji, Elia Kaufmann, Antonio Loquercio, and Davide Scaramuzza. “Flightmare: A Flexible Quadrotor Simulator.” Proceedings of the 4th Conference on Robot Learning (CoRL), PMLR 155, 2021. URL: https://proceedings.mlr.press/v155/song21a.html

[10] Justin Nakama, Ricky Parada, Joao P. Matos-Carvalho, Fabio Azevedo, Dario Pedro, and Luis Campos. “Autonomous Environment Generator for UAV-Based Simulation.” Applied Sciences, 11(5):2185, 2021. DOI: 10.3390/app11052185. URL: https://doi.org/10.3390/app11052185[11] Yash Vardhan Pant, Max Z. Li, Alena Rodionova, Rhudii A. Quaye, Houssam Abbas, Megan S. Ryerson, and Rahul Mangharam. “FADS: A Framework for Autonomous Drone Safety Using Temporal Logic-Based Trajectory Planning.” Transportation Research Part C: Emerging Technologies, 130:103275, 2021. DOI: 10.1016/j.trc.2021.103275. URL: https://doi.org/10.1016/j.trc.2021.103275

[12] Keyu Chen, Yuheng Lei, Hao Cheng, Haoran Wu, Wenchao Sun, and Sifa Zheng. “FREA: Feasibility-Guided Generation of Safety-Critical Scenarios with Reasonable Adversariality.” arXiv:2406.02983, 2024. URL: https://arxiv.org/abs/2406.02983

[13] Wenhao Ding, Chejian Xu, Mansur Arief, Haohong Lin, Bo Li, and Ding Zhao. “A Survey on Safety-Critical Driving Scenario Generation: A Methodological Perspective.” arXiv:2202.02215, 2022. URL: https://arxiv.org/abs/2202.02215[14] CARLA Team. “Digital Twin Tool: Procedural Generation from OpenStreetMap.” CARLA Simulator Documentation. URL: https://carla.readthedocs.io/en/0.9.16/adv_digital_twin/

[15] Shandong Expressway Group Co., Ltd. “‘Shandong Expressway Comprehensive Inspection Flight Service System’ goes online.” 2025. URL: https://www.sdhsg.com/article/72553

[16] Zhao Xiangmo, Zhao Yifei, Lu Nengchao, et al. “A review of research on key resource allocation for highway traffic accident emergency.” Transactions of Transportation Engineering, 2024. DOI: 10.19818/j.cnki.1671-1637.2024.06.001. URL: https://transport.chd.edu.cn/cn/article/doi/10.19818/j.cnki.1671-1637.2024.06.001

[17] Jisheng Zhang, Limin Jia, Shuyun Niu, Fan Zhang, Lu Tong, and Xuesong Zhou. “A Space-Time Network-Based Modeling Framework for Dynamic Unmanned Aerial Vehicle Routing in Traffic Incident Monitoring Applications.” Sensors, 15(6):13874-13898, 2015. DOI: 10.3390/s150613874. URL: https://doi.org/10.3390/s150613874[18] Tan Do-Duy, Long D. Nguyen, Trung Q. Duong, Saeed Khosravirad, and Holger Claussen. “Joint Optimization of Real-Time Deployment and Resource Allocation for UAV-Aided Disaster Emergency Communications.” IEEE Journal on Selected Areas in Communications, 39(11):3411-3424, 2021. DOI: 10.1109/JSAC.2021.3088662. URL: https://doi.org/10.1109/JSAC.2021.3088662

Appendix: This execution plan

Step 1: Freeze Paper F Total Positioning

Step 2: Do F1 benchmark first

Step 3: Advance the F2 Accelerated Testing Algorithm- Implement random/grid/LHS/BO/CMA-ES/RL adversarial baselines.

Step 4: Extend F3 city to local scene

Step 5: Expand F4 High-Speed Emergency Application

Step 6: Submission rhythm

Step 7: Tasks for the Recent Week-Write a formal experimental assignment for F1.