Paper deep dive
AerialDojo-200K: A Large-Scale Benchmark Suite for Open-World Aerial Object-Goal Search
Tongtong Feng, Xin Wang, Haoran Hou, Ren Wang, Weiran Wang, Shaokai Zhu, Ziqi Jia, Hao Wang, Yu-Wei Zhan, Zongyuan Wu, Jinghao Cui, Wenwu Zhu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/30/2026, 7:50:52 AM
Summary
The paper introduces AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search (AerialOGS). It addresses limitations of existing small, heterogeneous benchmarks by providing 42 simulation scenes across four families (urban, natural, infrastructure, disaster) and over 200,000 task instances. The dataset includes high-quality manual annotations, collision-free reference trajectories, and a unified evaluation framework. Evaluations of multimodal large language models (MLLMs) reveal significant gaps in current capabilities for general-purpose aerial agents.
Entities (10)
Relation Signals (8)
AerialDojo-200K → hastask → Aerial Object-Goal Search
confidence 95% · AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search
Multimodal Large Language Models → evaluatedon → AerialDojo-200K
confidence 90% · our evaluation of five open-source and four closed-source multimodal large language models
AerialDojo-200K → outperforms → CityAVOS
confidence 90% · with 3 times as many scenes and 18.7 times as many task instances as the largest existing benchmark
AerialDojo-200K → outperforms → UAV-ON
confidence 90% · with 3 times as many scenes and 18.7 times as many task instances as the largest existing benchmark
AerialDojo-200K → usessimulator → ProjectAirSim
confidence 90% · and ProjectAirSim (Microsoft, 2023)
AerialDojo-200K → usessimulator → Unreal Engine
confidence 90% · We build the AerialDojo-200K benchmark using Unreal Engine
GPT-5.6 → ispartof → Multimodal Large Language Models
confidence 85% · evaluation of five open-source and four closed-source multimodal large language models... GPT-5.6
Qwen3-VL-4B → ispartof → Multimodal Large Language Models
confidence 85% · evaluation of five open-source and four closed-source multimodal large language models... Qwen3-VL-4B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Open-world aerial object-goal search is a foundational yet challenging task, requiring aerial agents to autonomously explore large-scale, unstructured three-dimensional environments and reach target objects specified by semantic descriptions or reference images, rather than following route-specific instructions. However, research in this task remains at a nascent stage and relies on small, environment-specific benchmarks with heterogeneous action spaces and data formats. These limitations hinder large-scale training and cross-benchmark evaluation, constraining the scalability and generalizability of aerial agents. To address this problem, we propose AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search, with 3 times as many scenes and 18.7 times as many task instances as the largest existing benchmark for this task. Specifically, we construct 42 simulation scenes spanning four scene families and 21 scene types, including 18 urban, 12 natural, six infrastructure, and six disaster scenes. To ensure data quality, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. We further construct 205,732 task instances, comprising over 100K semantic-goal and over 100K image-goal instances across Base, Standard, and Long-Horizon settings. Each task instance includes a collision-free reference trajectory and corresponding multi-view video recordings. We also develop a unified evaluation framework with a scene partition comprising 21 in-distribution scenes and 21 out-of-distribution scenes. Finally, our evaluation of five open-source and four closed-source multimodal large language models reveals that there is still a long way to go toward achieving general-purpose aerial agents. All can be found at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.36066v1
- Canonical: https://arxiv.org/abs/2609.36066v1
Trouble viewing inline? Open PDF directly →
Full Text
46,792 characters extracted from source content.
Expand or collapse full text
AerialDojo-200K AERIALDOJO-200K: A LARGE-SCALE BENCHMARK SUITE FOR OPEN-WORLD AERIAL OBJECT-GOAL SEARCH Tongtong Feng 1 , Xin Wang 1∗ , Haoran Hou 1 , Ren Wang 1 , Weiran Wang 2 , Shaokai Zhu 3 , Ziqi Jia 1 , Hao Wang 1 , Yu-Wei Zhan 4 , Zongyuan Wu 1 , Jinghao Cui 1 , Wenwu Zhu 1∗ 1 Department of Computer Science and Technology, BNRist, Tsinghua University 2 School of Electrical & Electronic Engineering, University College Dublin 3 School of Electronics Engineering and Computer Science, Peking University 4 Department of Computer Science and Technology, Zhejiang University of Technology ABSTRACT Open-world aerial object-goal search is a foundational yet challenging task, re- quiring aerial agents to autonomously explore large-scale, unstructured three- dimensional environments and reach target objects specified by semantic de- scriptions or reference images, rather than following route-specific instructions. However, research in this task remains at a nascent stage and relies on small, environment-specific benchmarks with heterogeneous action spaces and data for- mats. These limitations hinder large-scale training and cross-benchmark evalu- ation, constraining the scalability and generalizability of aerial agents. To ad- dress this problem, we propose AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search, with 3× (times) as many scenes and 18.7× as many task instances as the largest existing benchmark for this task. Specifically, we construct 42 simulation scenes spanning four scene families and 21 scene types, including 18 urban, 12 natural, six infrastructure, and six dis- aster scenes. To ensure data quality, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. We further construct 205,732 task instances, comprising over 100K semantic-goal and over 100K image-goal instances across Base, Standard, and Long-Horizon settings. Each task instance includes a collision-free reference tra- jectory and corresponding multi-view video recordings. We also develop a unified evaluation framework with a scene partition comprising 21 in-distribution scenes and 21 out-of-distribution scenes. Finally, our evaluation of five open-source and four closed-source multimodal large language models reveals that there is still a long way to go toward achieving general-purpose aerial agents. All can be found in https://fengtt42.github.io/AerialDojo/. 1INTRODUCTION Open-world Aerial Object-Goal Search (AerialOGS) (Xiao et al., 2025) requires aerial agents to autonomously search for target objects through goal-driven, long-horizon exploration in large-scale, unstructured three-dimensional environments, and to actively stop when onboard observations match the target object specified by a semantic description or reference image. Unlike aerial Vision-and- Language Navigation (VLN), which follows detailed, step-by-step route instructions that constrain how the aerial agent should travel (Zhao et al., 2026), AerialOGS provides only a high-level, open- ended object-goal and must jointly resolve where to explore, how to navigate, and when the observed object satisfies the goal. AerialOGS has broad potential applications in search and rescue, infrastruc- ture inspection, and autonomous delivery, where operators may specify the object without knowing its exact location or providing a complete flight route. ∗ Corresponding Authors. 1 arXiv:2609.36066v1 [cs.CV] 28 Sep 2026 AerialDojo-200K EnvironmentTask Dataset LeaderboardUnified Benchmark 4 Scene Family · 21 Scene Types · 42 Scenes 2099 Anchors 100K SemanticOGS · 100K ImageOGS 80.0K Base 82.4K Standard43.3K Long-Horizon Goal Start High-Quality · Large-Scale · Human Alignment ID: 20%Testing Comprehensive · Large-Scale · Manual Annotation 109 Landmarks2099 Objects OOD: 80%Testing A Large-Scale Benchmark For Open-world Aerial Object-Goal Search Four-view RGBD Hybrid A* Trajectory 4115KM Four Degree of Freedom Action All Trainset Tasks Recorded GPT-5.6 Sol · Claude Opus 5 · Gemini 3.8 Flash Qwen3-VL-4B · grok-4.5 · InternVL3.5-8B · Ministral 3 8B Phi-4-multimodal-instruct · MiniCPM-V 4.6 Data Format · Action Space · Framework · Metric 5 Open-Source MLLMs 4 Closed-Source MLLMs SR:Success Rate OSR: Oracle Success Rate DTS: Distance to Success SPL: Success weighted Path Length CR: Collision Rate A long way to go Open-Source HuggingFaceGithub Simulators Tasks Records User-Friendly Codes Leaderboard Trajectories Webpage Figure 1: The overview of AerialDojo-200K. Research on AerialOGS tasks remains in a nascent stage. On the one hand, existing aerial agents utilize spatio-semantic representations (Zhang et al., 2026) to organize memories and plan over explored regions, frontiers, and object candidates. On the other hand, recent studies demonstrate that multimodal large models can leverage high-level semantic reasoning to interpret semantic goals (Xiao et al., 2025) and guide search decisions (Chen et al., 2026). Nevertheless, the primary performance bottleneck remains the lack of large-scale benchmarks for training and evaluation. As shown in Table 1, 1) UAV-ON (Xiao et al., 2025) and CityAVOS (Ji et al., 2026) are currently the only AerialOGS benchmarks, and the scale of those is approximately one- tenth as many tasks as OpenFly (Gao et al., 2026); 2) Heterogeneous action spaces and data formats across AerialOGS and VLN benchmarks make it difficult for existing VLN benchmarks to support AerialOGS tasks. Consequently, existing research relies on small, environment-specific benchmarks with heterogeneous action spaces and data formats, failing to support large-scale training and cross- benchmark evaluation, and limiting the scalability and generalizability of aerial agents. To address this problem, we propose AerialDojo-200K, a large-scale benchmark suite for open- world aerial object-goal search, with 3× (times) as many scenes and 18.7× as many task instances as the largest existing benchmark for this task. Specifically, as shown in Figure 1, we construct 42 simulation scenes spanning four scene families and 21 scene types, including 18 urban, 12 nat- ural, six infrastructure, and six disaster scenes. To ensure data quality, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. We further construct 205,732 task instances, comprising over 100K semantic-goal and over 100K image-goal instances across Base, Standard, and Long-Horizon settings. Each task in- stance includes a collision-free reference trajectory and corresponding multi-view video recordings, with the trajectories totaling 4115km across the dataset. We also develop a unified evaluation frame- work with a scene partition comprising 21 in-distribution scenes and 21 out-of-distribution scenes. Finally, our evaluation of five open-source and four closed-source multimodal large language mod- els reveals that there is still a long way to go toward achieving general-purpose aerial agents. Our main contributions are summarized as follows: • We propose AerialDojo-200K, a large-scale benchmark suite for open-world aerial object- goal search, with 3× as many scenes and 18.7× as many task instances as the largest existing benchmark for this task. 2 AerialDojo-200K Table 1: Comparison of aerial benchmarks. N task : the number of task instances; F scene : Scene families; N scene : the number of scenes. VLN: vision-and-language navigation; ObjNav: object-goal navigation; OGS: object-goal search. DoF: degrees of fredom. U: urban; N: natural; I: infrastructure; D: disaster. BenchmarkYear TaskActionGoal Type Goal SpecificationN task F scene N scene AerialVLN2023 VLN4-DoFLocationMovement instruction25.3KU25 CityNav2025 VLN4-DoFLocationMovement instruction32.6KU34 OpenUAV2025 VLN6-DoFLocationMovement instruction12.1KU,N22 OpenFly2026 VLN4-DoFLocationMovement instruction103KU18 UAV-ON2025 ObjNav4-DoFInstanceSemantic11KU,N14 CityAVOS2026 OGS4-DoFInstanceSemantic and Image2.4KU6 AerialDojo-200K2026OGS4-DoFInstanceSemantic or image200KU,N,I,D42 • AerialDojo-200K provides a high-quality dataset with extensive manual annotations. Twelve annotators spent two months manually annotating 109 landmarks, 2,099 objects, and 2,099 object anchors. Each task instance includes a collision-free reference trajectory and corresponding multi-view video recordings. • We develop a unified evaluation framework for this task and evaluate 5 open-source and 4 closed-source MLLMs. The results reveal that there is still a long way to go toward achieving general-purpose aerial agents. 2RELATED WORKS We review related work on aerial object-goal search and aerial navigation benchmarks. We first dis- cuss how existing methods combine spatial representations and semantic reasoning for autonomous search, then examine benchmark task formulations, dataset scale, and environmental coverage to contextualize AerialDojo-200K. 2.1AERIAL OBJECT-GOAL SEARCH Aerial object-goal search requires an agent to reach a target object specified by a semantic descrip- tion or reference image based solely on egocentric visual observations. Existing research can be grouped into two categories. On the one hand, existing aerial agents utilize spatio-semantic repre- sentations, such as semantic maps (Zhou et al., 2026), scene graphs (Gai et al., 2026), and value maps (Zhang et al., 2026), to organize memories and plan over explored regions, frontiers, and ob- ject candidates. On the other hand, recent studies demonstrate that multimodal large models can leverage high-level semantic reasoning to interpret semantic goals (Ji et al., 2026), assess object candidates (Pardyl et al., 2025), and guide search decisions (Wang et al., 2025; Chen et al., 2026). AerialOGS has broad potential applications (Kim et al., 2025) in search and rescue (Feng et al., 2026), infrastructure inspection (Feng et al., 2024a), and autonomous delivery (Feng et al., 2024b), where operators may specify the object without knowing its exact location or providing a com- plete flight route. However, limited benchmark scale and heterogeneous evaluation protocols hinder systematic assessment of these methods across diverse environments. AerialDojo-200K addresses these limitations through large-scale, carefully annotated search tasks and a unified framework for evaluating performance and generalization. 2.2AERIAL BENCHMARK Aerial tasks primarily include vision-and-language navigation (VLN) and object-goal navigation (ObjNav), with aerial object-goal search (OGS) representing a more challenging setting within Ob- jNav that emphasizes autonomous exploration and target identification. As shown in Table 1, most existing aerial benchmarks focus on VLN tasks, where agents follow language instructions to reach specified locations. AerialVLN (Liu et al., 2023) and CityNav (Lee et al., 2025) provide 25.3K and 3 AerialDojo-200K 32.6K instruction-guided navigation tasks across 25 and 34 urban scenes, respectively. OpenUAV (Wang et al., 2025) extends this setting to a 6-DoF action space, providing 12.1K tasks across 22 urban and natural scenes, while OpenFly (Gao et al., 2026) expands the task count to 103K across 18 scenes. Beyond instruction-guided navigation, UAV-ON (Xiao et al., 2025) supports instance-level ObjNav tasks through semantic goal specifications, with 11K tasks across 14 scenes. CityAVOS (Ji et al., 2026) further investigates OGS tasks using semantic and visual goal information, providing 2.4K tasks across six urban scenes. Despite these advances, existing benchmarks remain limited in both task scale and environmental coverage. Existing research relies on small, environment-specific benchmarks with heterogeneous action spaces and data formats, hindering large-scale training and cross-benchmark evaluation and limiting the scalability and generalizability of aerial agents. 3AERIALDOJO-200K AerialDojo-200K, the first large-scale benchmark suite for open-world Aerial Object-Goal Search (AerialOGS), comprises 42 simulated scenes, 205K task instances, 4115 km of flight trajectories, and a unified evaluation framework. The construction process is shown in Figure 2. 1. Scene collection2. Landmark annotation3. Object annotation4. Anchor annotation 5. Trajectory collection6. Task partitioning7. Video collection8. Unified benchmark UrbanNatural Infrastructure Disaster 4 Families | 21 Types | 42 Scenes 1 2 3 2 3 1 Manual annotation | 109 LandmarksManual annotation | 2099 Objects Selected anchor 1m Manual annotation | 2099 Anchors Hybrid A* | Totaling 4115km Base 80.0K | Standard 82.4K | LH 43.3K FrontLeftRightDown Only for training and evaluation 5 Open-source | 4 Closed-source Figure 2: Standard operating procedure for constructing AerialDojo-200K. 3.1SIMULATOR We build the AerialDojo-200K benchmark using Unreal Engine (Epic Games, 2008) and Projec- tAirSim (Microsoft, 2023), enabling AerialOGS across diverse, high-fidelity simulated scenes. Large-scale and comprehensive environment collection. AerialDojo-200K contains 42 scenes spanning four scene families and 21 scene types, with two instances per type. As shown in Ta- ble 1, it provides three times as many scenes as UAV-ON (Xiao et al., 2025), the largest existing AerialOGS benchmark. As shown in Figure 3, the collection includes both open landscapes and constrained built spaces, with diverse layouts and obstacle configurations. Specifically, the urban family comprises malls, stadiums, amusement parks, parking lots, communities, neighborhoods, alleys, factories, and construction sites; the natural family covers deserts, forests, mountains, snow- fields, islands, and coasts; the infrastructure family includes bridges, rail corridors, and harbors; and the disaster family depicts earthquake, flood, and explosion scenes. By incorporating infrastructure and disaster scenarios alongside urban and natural environments, AerialDojo-200K covers a broad range of settings relevant to real-world aerial deployment and applications. Together, these scenes form a comprehensive library of simulation environments for AerialOGS. Sensors. AerialDojo-200K can collect synchronized RGB-D streams from four onboard cameras facing forward, left, right, and downward. Both RGB and depth images have a resolution of 640 × 640 pixels, and each camera has a 90 ◦ field of view (FOV). All four views are sampled synchronously at 15 frames per second (FPS). At acquisition time t, the visual observation is repre- sented as o t = (I v t ,D v t ) | v ∈ V and V = front, left, right, down, where I v t and D v t denote the RGB image and depth map from view v, respectively. The observation interface excludes GPS 4 AerialDojo-200K Table 2: Dataset statistics. LM: landmarks; Base, Standard, and LH denote task-instance counts for the Base, Standard, and Long-Horizon setting, respectively, including both semantic-goal and image-goal instances; ALL is their sum. Path reports the cumulative planned path length in kilome- ters, counted once per underlying navigation task. FamilyTypes Scenes LM Objects AnchorsBase StandardLHALL Path (km) Urban918571,0301,030 42,23255,840 26,992 125,064 2,607.253 Natural61221569569 21,66815,188 12,97649,832 1,008.083 Infrastructure36202802808,9347,8003,08019,814357.067 Disaster36112202207,1403,59628611,022142.910 Total21421092,0992,09979,97482,42443,334205,7324,115.313 42 scenes 21 types Disaster: 6 Infra.: 6 Natural: 12 Urban: 18 (a) Simulator statistics(b) Object-name word cloud Long-Horizon 21.1% Standard 40.1% Base 38.9% task instances (c) Task statistics Figure 3: Simulator and dataset statistics of AerialDojo-200K. measurements, external localization signals, and global maps. Simulator ground truth, including object poses, semantic maps, and scene geometry, is also withheld from the agent. Consequently, the agent must interpret the goal, infer the surrounding spatial structure, and identify the target by integrating egocentric visual evidence over time. Actions. AerialDojo-200K provides eight action types with discrete motion magnitudes, supporting four-DoF control of the aerial agent’s position and yaw. Specifically, MoveForward, MoveLeft, and MoveRight allow horizontal translations of 1, 3, or 5 m, while MoveUp and MoveDown allow vertical translations of 1 or 2 m. The rotational actions, TurnLeft and TurnRight, change the yaw angle by 15 ◦ in the corresponding direction. The Stop action declares target discovery and terminates the episode. Together, these eight action types yield 16 discrete action choices, allowing the aerial agent to select movement magnitudes according to local geometry and target proximity. Backward translation is excluded to encourage forward-looking exploration and reduce redundant oscillatory movements. The aerial agent can reposition by combining yaw rotations with forward or lateral translations, enabling flexible exploration while maintaining a compact action interface. Aerial embodiment. The aerial agent has a default wheelbase of 250 m and a collision radius of 250 m. These settings are intended to represent compact commercial aerial platforms with wheelbases below 500 m. Both parameters are configurable. Manual annotation. To ensure simulator quality, as shown in Table 2, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. AerialDojo-200K selects landmarks and target objects according to their roles in Aeri- alOGS tasks. Landmarks serve as visually salient, readily identifiable spatial references, reflecting a human search strategy in unfamiliar environments: first locating a relevant landmark and then re- fining the target’s position. Landmark selection is inherently scene-dependent; shopping malls or historic monuments may serve as urban landmarks, whereas distinctive mountain peaks or rock for- mations may provide natural landmarks. This variation motivates manual selection based on scene- level salience and contextual relevance. Target objects are selected for their practical search value and semantic distinguishability. Their descriptions must allow human annotators to unambiguously identify the intended instance, reducing ambiguous supervision during training. AerialDojo-200K uses object anchors for task evaluation because a centroid-based success region may lie entirely inside a large object, making successful arrival physically infeasible. An anchor is a collision-free viewpoint located 1 m outside the object’s boundary and oriented toward its centroid. A vision- language model shortlists three candidate anchors based on visual informativeness and image–name correspondence. Human annotators then select one anchor per object as the spatial reference for 5 AerialDojo-200K task construction and success evaluation. We also have developed and open-sourced a lightweight annotation tool that standardizes the entire manual annotation workflow. 3.2TASK AND DATASET Task definition. AerialDojo-200K defines open-world AerialOGS tasks as T = ⟨S,G⟩, where S is the start pose and G is the instance-level goal. S = [x,y,z,ψ], where (x,y,z) ∈R 3 denotes its position and ψ denotes its yaw angle. Unlike UAV-ON (Xiao et al., 2025), which randomly initializes the agent’s start pose, we initialize each task instance at the anchor position of a selected object, with the yaw rotated by 180 ◦ relative to the anchor orientation. This design can substantially reduce the risk of task failure due to collisions at initialization while better reflecting practical aerial search scenarios, in which an agent begins a new search from the location of a previously reached target. The goal is specified as G = [name, landmark, direction, representation], comprising the target object’s name, the name of its nearest landmark, its direction relative to that landmark, and a target representation provided as either a semantic description or a reference image. The direction takes one of eight compass orientations: north, south, east, west, northeast, northwest, southeast, and southwest. This goal reflects a common human search strategy in unfamiliar environments: using a recognizable landmark and a coarse direction to orient the search, then relying on visual evidence to identify the target object. AerialDojo-200K defines the task as SemanticOGS when the representation is a semantic description and ImageOGS when it is a reference image. Task execution. An episode is specified by X =⟨T ,E,B,O,A,C⟩,(1) where E is the scene, B is the aerial embodiment, O and A are the observation and action spaces described above, andC specifies the termination conditions. At each decision step, the agent selects a t ∈A based on G and its observation-action history. An episode terminates upon Stop, collision, or exhaustion of its action budget. Success requires an explicit Stop within 3 m of the target anchor, without any collision and within the assigned budget. Collection trajectory. Candidate tasks pair distinct annotated objects within the same scene. We adapt Hybrid A* (Dolgov et al., 2008) to the aerial agent’s four-DoF action space to plan a collision- free reference trajectory from the start pose to the target anchor. Collision checking accounts for the agent’s configured collision radius, and candidate pairs are discarded if no feasible trajectory is found. Let τ ref = (q 0 ,...,q K ), where q k = (p k ,ψ k ), q 0 = S,p 0 is the start position, and p K =p g is the target anchor position. The translational trajectory length is L(τ ref ) = K−1 X k=0 ∥p k+1 −p k ∥ 2 .(2) Pure yaw rotations contribute no translational length. Reference trajectories reach the target anchor itself, whereas evaluation accepts an explicit Stop within 3 m of the anchor. Target coordinates and reference trajectories are withheld from the agent during evaluation. Task partitioning. We categorize tasks by reference trajectory length and use a detour ratio to filter out near-straight-line routes. Letp 0 andp g denote the start and target anchor positions, respectively. The Euclidean separation and detour ratio are defined as d E =∥p 0 −p g ∥ 2 , κ = L(τ ref ) d E .(3) Table 3: Task settings. SettingL(τ ref ) Action budget Base[5, 30)90 Standard[30, 60)180 Long-Horizon[60, 100)300 We retain candidates satisfying d E > 3 m, 5 m ≤ L(τ ref ) < 100 m, and κ ≥ 1.05. The first condition excludes starts already within the success region. The trajectory-length range balances meaningful search diffi- culty with a manageable evaluation horizon. The detour- ratio threshold filters out near-straight-line routes, aiming to reduce trivial tasks on which models may succeed uni- formly and thereby improve the ability of success rate to distinguish model performance. Retained tasks are par- titioned into Base, Standard, and Long-Horizon settings according to their reference trajectory lengths, with corresponding action budgets listed in Table 3. 6 AerialDojo-200K Collection video. We acquire synchronized multi-view RGB-D observations along reference trajec- tories using the sensor configuration described above. Quality checks flag missing camera outputs and abnormal image frames. Recordings are indexed by navigation task, camera view, and reference trajectory steps. The goal reference image used by ImageOGS is stored separately. Dataset statistics. As shown in Figure 3, AerialDojo-200K contains 205,732 task instances, com- prising 102,866 SemanticOGS and 102,866 ImageOGS instances across 42 scenes. The two in- stances of each navigation task share the same scene, start pose, target object, and reference tra- jectory, differing only in goal representation. The dataset covers 2,099 target objects across 254 semantic categories and reports a total planning distance of 4115.313 km. 3.3BENCHMARK Benchmark protocol.To evaluate AerialOGS capabilities in open-world environments, AerialDojo-200K divides its 42 scenes into 21 in-distribution (ID) scenes and 21 out-of-distribution (OOD) scenes, with one instance of each scene type in each split. Within ID scenes, 80% of the tasks are used for training and the remaining 20% for in-distribution performance evaluation. Within OOD scenes, 20% of the tasks form an adaptation subset, while the remaining 80% are reserved for test- ing. Models trained on ID scenes are transferred to OOD scenes to evaluate generalization across distinct map layouts, obstacle configurations, and target instances within the same scene types. In- spired by adaptation with limited target-domain data, we allow models to use the designated 20% OOD subset for scene-specific adaptation before evaluation on the held-out OOD test tasks. Evalu- ation covers SemanticOGS and ImageOGS across the Base, Standard, and Long-Horizon settings, using the action budgets in Table 3. RGB-D observation Semantic OR image goal Interaction history Termination condition Unifiedprompt Multimodal Large Language Model Action Simulator New observation Evaluator only Target anchor + Reference path Episode scoring SR | OSR | SPL | NE | CR Success: Stop within 3m; Collision-free; Within budget Figure 4: The unified benchmark framework of AerialDojo-200K. Benchmark framework. Figure 4 illustrates the closed-loopevaluation framework. Each episode resets the simulator to the task’s scene and start pose. At each step, a model- specific adapter packages the current RGB-D obser- vations, goal specification, termination conditions, and interaction history using the fixed prompt provided in Appendix A. The returns one action, which maps to the simulator’s action interface using a common motion-magnitude configuration. Execution produces the next observa- tion and updates the episode log. Episodes terminate upon Stop, collision, or budget exhaustion; infrastructure failures are logged and rerun separately. Target coordinates and reference trajectories remain accessible only to the evaluator. Benchmark metrics. We evaluate navigation performance using five complementary metrics. Suc- cess rate (SR) measures the fraction of episodes in which the agent explicitly stops within 3 m of the target anchor without collision and within the action budget. Oracle success rate (OSR) measures whether the agent ever enters this region, regardless of its final stopping decision or subsequent fail- ure. Distance to success (DTS) is the mean final Euclidean distance to the target anchor, measured in meters without subtracting the success radius. Success weighted by path length (SPL) evaluates navigation efficiency: SPL = 1 N N X i=1 σ i L i max(L i ,ℓ i ) ,(4) where N is the number of evaluated episodes, σ i indicates episode success, L i = L(τ ref i ) is the reference trajectory length and ℓ i is the executed trajectory length. This formulation uses the planned reference for normalization. Collision rate (CR) is the fraction of episodes containing a collision. All metrics are computed over both successful and failed episodes. 7 AerialDojo-200K 4EXPERIMENTS We conduct experiments on AerialDojo-200K to evaluate existing MLLMs on AerialOGS. We first introduce the baseline models and experimental setup, followed by quantitative comparisons across scene families and task setting. We further analyze how the success threshold affects model perfor- mance and rankings. 4.1BASELINES AND SETUP Baselines. We evaluate five open-source MLLMs, Qwen3-VL-4B, InternVL3.5-8B, Ministral 3 8B, Phi-4-multimodal, and MiniCPM-V 4.6, and four closed-source MLLMs, GPT-5.6 Sol, Grok- 4.5, Gemini 3.8 Flash, and Claude Opus 5. We test these models under the Base, Standard, and Long-Horizon settings. For brevity, the main tables report model-setting combinations with nonzero overall success rates, covering all nine MLLMs on Base tasks, GPT-5.6 Sol and Claude Opus 5 on Standard tasks. Evaluation setup. Base and Standard tasks have reference lengths of [5, 30) m and [30, 60) m and action budgets of 90 and 180, respectively. We average the all map-level scores equally within each family, then average the four family means equally for overall results. Rates are percentages, and DTS is in meters. The prompt specifies stopping within 3 m, our primary success radius; 5 m is a supplementary evaluation threshold. Table 4: Base-task performance across four scene families. SR, OSR, SPL, and CR are percentages; DTS is in meters. Arrows indicate the preferred direction; bold denotes the best available value. Model3 m (primary)5 m SR↑OSR↑SPL↑DTS↓CR↓SR↑OSR↑ Open-source models Qwen3-VL-4B0.71±0.811.50±1.330.51±0.6016.24±1.85 22.17±14.323.85±2.927.99±3.93 InternVL3.5-8B0.16±0.303.41±2.360.11±0.2225.42±11.6362.06±16.891.54±1.7412.04±4.47 Ministral 3 8B0.17±0.311.28±1.480.09±0.2116.23±3.4351.63±24.411.64±1.1811.37±5.20 Phi-4-multimodal0.13±0.350.60±1.010.09±0.2523.52±7.4737.96±16.671.65±1.799.22±6.17 MiniCPM-V 4.60.19±0.531.35±2.410.15±0.4320.26±6.4442.47±11.071.47±1.989.91±3.84 Closed-source models GPT-5.6 Sol2.81±2.423.06±2.442.20±1.8717.90±4.0132.91±14.458.16±3.7811.22±4.23 Grok-4.50.25±0.690.49±1.390.25±0.6915.46±1.8122.18±12.442.46±2.283.69±2.65 Gemini 3.8 Flash2.40±2.233.68±4.322.37±2.2121.61±7.0733.95±12.21 11.16±8.2717.90±8.25 Claude Opus 53.20±2.047.84±4.602.83±2.0419.69±5.9758.96±15.947.38±3.7318.48±5.64 Table 5: Standard-task performance across four scene families. SR, OSR, SPL, and CR are percent- ages; DTS is in meters. Arrows is the preferred direction; bold denotes the best available value. Model3 m (primary)5 m SR↑OSR↑SPL↑DTS↓CR↓SR↑OSR↑ GPT-5.6 Sol0.26±0.720.26±0.720.20±0.5737.66±5.4736.57±16.792.30±2.022.30±2.02 Claude Opus 5 0.49±0.911.32±2.330.40±0.7336.49±8.8956.30±21.311.29±1.912.61±3.84 4.2EVALUATION RESULTS Performance on base and standard tasks. Tables 4–5 show substantial performance differences among the evaluated models, alongside consistently low task completion rates. At the primary 3 m threshold, Claude Opus 5 achieves the highest Base SR of 3.20%, followed by GPT-5.6 Sol (2.81%) and Gemini 3.8 Flash (2.40%). Qwen3-VL-4B leads the open-source models at 0.71%, exceeding Grok-4.5 at 0.25%, indicating that the advantage of closed-source models is not uniform. Performance deteriorates markedly on Standard tasks: GPT-5.6 Sol and Claude Opus 5 achieve SRs 8 AerialDojo-200K SR ↑ OSR ↑ SPL ↑DTS ↓ CR ↓ 0.25 0.50 0.75 (a) Urban SR ↑ OSR ↑ SPL ↑DTS ↓ CR ↓ 0.25 0.50 0.75 (b) Natural SR ↑ OSR ↑ SPL ↑DTS ↓ CR ↓ 0.25 0.50 0.75 (c) Infrastructure SR ↑ OSR ↑ SPL ↑DTS ↓ CR ↓ 0.25 0.50 0.75 (d) Disaster Open-source Closed-source Qwen3-VL-4BInternVL3.5-8BMinistral 3 8BPhi-4-multimodalMiniCPM-V 4.6 GPT-5.6 SolGrok-4.5Gemini 3.8 FlashClaude Opus 5 Figure 5: Base-task performance across four scene families at the 3 m success threshold. Each curve represents one model. From the center to the outer edge, SR and SPL range from 0 to 5%, OSR from 0 to 15%, DTS from 40 to 0 m, and CR from 100 to 0%; outward therefore indicates better performance. Ring labels show normalized radial values. of only 0.26% and 0.49%, respectively. Their OSRs also decrease to 0.26% and 1.32%, showing that even reaching the target region is uncommon. Together, these results highlight the difficulty of maintaining successful search beyond the Base setting. Success-radius sensitivity. Relaxing the success radius from 3 m to 5 m substantially changes both success rates and model rankings. On Base tasks, Gemini 3.8 Flash increases from 2.40% to 11.16% SR, an improvement of 8.76 percentage points, overtaking GPT-5.6 Sol (8.16%) and Claude Opus 5 (7.38%). A ranking reversal also occurs on Standard tasks: Claude Opus 5 leads at 3 m, whereas GPT-5.6 Sol leads at 5 m with 2.30% SR versus 1.29%. These changes demonstrate that model comparisons depend strongly on the spatial tolerance used to define success. The higher scores at 5 m reflect a more permissive evaluation criterion and do not establish improved navigation behavior. We therefore report both thresholds while retaining the prompt-aligned 3 m criterion as the primary measure of precise task completion. Performance differences across scene families. Figure 5 reveals scene-dependent model strengths that are obscured by overall averages. At 3 m, GPT-5.6 Sol achieves the highest SR on Natural and Infrastructure scenes, at 4.08% and 3.06%, respectively. Claude Opus 5 leads on Disaster scenes with 4.85% SR, but achieves only 0.98% on Infrastructure scenes. On Urban scenes, Claude Opus 5 and GPT-5.6 Sol obtain nearly identical SRs of 3.07% and 3.06%. The radar profiles further show that strengths in successful completion do not consistently coincide with strengths in collision avoidance or final target distance. Thus, no evaluated model dominates all scene families and met- rics, emphasizing the need for family-wise evaluation to expose differences in capability across the evaluated environments. A long way to generalist aerial agents. Collectively, these results reveal a substantial gap be- tween current MLLMs under the evaluated framework and reliable generalist aerial agents. The best reported SR remains only 3.20% on Base tasks and 0.49% on Standard tasks at the primary thresh- old. Moreover, reaching the target region does not ensure successful completion: Claude Opus 5 achieves a Base OSR of 7.84% but an SR of only 3.20%, alongside a collision rate of 58.96%. This gap captures failures to convert target-region visits into valid completion, although the aggregate metrics cannot identify their individual causes. Low completion rates, frequent collisions, and un- even performance across scene families collectively indicate that integrating effective exploration, precise target localization, appropriate stopping, and collision avoidance remains a major challenge on the path toward generalist aerial agents. 5CONCLUSION We present AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search. It comprises 205,732 semantic-goal and image-goal task instances across 42 simulated scenes from four scene families, with three task settings and a unified observation, action, and evalu- ation interface. Our evaluation of existing MLLMs reveals substantial limitations: the best reported 9 AerialDojo-200K success rates at the primary 3 m threshold reach only 3.20% on Base tasks and 0.49% on Standard tasks. Together with frequent collisions and gaps between target-region visits and successful com- pletion, these findings show that current MLLMs under the evaluated framework remain far from reliable generalist aerial agents. AerialDojo-200K provides a common testbed for developing and evaluating methods that integrate exploration, instance-level target grounding, precise stopping, and collision avoidance. REFERENCES Xuecheng Chen, Zongzhuo Liu, Jianfa Ma, Bang Du, Tiantian Zhang, Xueqian Wang, and Boyu Zhou. AirHunt: Bridging VLM Semantics and Continuous Planning for Efficient Aerial Object Navigation. arXiv preprint arXiv:2601.12742, 2026. Dmitri Dolgov, Sebastian Thrun, Michael Montemerlo, and James Diebel. Practical Search Tech- niques in Path Planning for Autonomous Driving. In Proceedings of the First International Sym- posium on Search Techniques in Artificial Intelligence and Robotics (STAIR-08). AAAI, 2008. Epic Games. Unreal Engine. Software, 2008. Tongtong Feng, Qing Li, Xin Wang, Mingzi Wang, Guangyao Li, and Wenwu Zhu. Multi-weather cross-view geo-localization using denoising diffusion models. In Proceedings of the 2nd Work- shop on UAVs in Multimedia: Capturing the World from a New Perspective, p. 35–39, 2024a. Tongtong Feng, Xin Wang, Feilin Han, Leping Zhang, and Wenwu Zhu. U2udata: A large-scale cooperative perception dataset for swarm uavs autonomous flight. In Proceedings of the 32nd ACM International Conference on Multimedia, p. 7600–7608, 2024b. Tongtong Feng, Xin Wang, Feilin Han, Leping Zhang, and Wenwu Zhu. U2udata+: A scalable swarm uavs autonomous flight dataset for embodied long-horizon tasks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, p. 1792–1800, 2026. Weiqi Gai, Yuman Gao, Yuan Zhou, Yufan Xie, Zhiyang Liu, Yuze Wu, Xin Zhou, Fei Gao, and Zhijun Meng. USS-Nav: Unified Spatio-Semantic Scene Graph for Lightweight UAV Zero-Shot Object Navigation. arXiv preprint arXiv:2602.00708, 2026. doi: 10.48550/arXiv.2602.00708. Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, Yiwen Tang, Yuhang Tang, Shuai Liang, Songyi Zhu, Ziqin Xiong, Yifei Su, Xinyi Ye, Jianan Li, Yan Ding, Dong Wang, Xuelong Li, Zhigang Wang, and Bin Zhao. OpenFly: A Comprehensive Platform for Aerial Vision-Language Navigation. In The Fourteenth International Conference on Learning Representations, 2026. Yatai Ji, Zhengqiu Zhu, Yong Zhao, Beidan Liu, Chen Gao, Yihao Zhao, Sihang Qiu, Yue Hu, and Quanjun Yin. Towards Autonomous UAV Visual Object Search in City Space: Benchmark and Agentic Methodology. Proceedings of the AAAI Conference on Artificial Intelligence, 40(22): 18342–18350, 2026. doi: 10.1609/aaai.v40i22.38898. Seungchan Kim, Omar Alama, Dmytro Kurdydyk, John Keller, Nikhil Keetha, Wenshan Wang, Yonatan Bisk, and Sebastian Scherer. RAVEN: Resilient Aerial Navigation via Open-Set Se- mantic Memory and Behavior Adaptation.arXiv preprint arXiv:2509.23563, 2025.doi: 10.48550/arXiv.2509.23563. Jungdae Lee, Taiki Miyanishi, Shuhei Kurita, Koya Sakamoto, Daichi Azuma, Yutaka Matsuo, and Nakamasa Inoue. CityNav: A Large-Scale Dataset for Real-World Aerial Navigation. In Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 5912–5922, 2025. Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yanning Zhang, and Qi Wu. AerialVLN: Vision-and-Language Navigation for UAVs. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), p. 15384–15394, 2023. Microsoft. Project AirSim. Open-source simulation software, 2023. 10 AerialDojo-200K Adam Pardyl, Dominik Matuszek, Mateusz Przebieracz, Marek Cygan, Bartosz Zieli ́ nski, and Ma- ciej Wolczyk. FlySearch: Exploring how vision-language models explore. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-5594. Xiangyu Wang, Donglin Yang, Ziqin Wang, Hohin Kwan, Jinyu Chen, Wenjun Wu, Hongsheng Li, Yue Liao, and Si Liu. Towards Realistic UAV Vision-Language Navigation: Platform, Bench- mark, and Methodology. In International Conference on Learning Representations, p. 7292– 7310, 2025. Jianqiang Xiao, Yuexuan Sun, Yixin Shao, Boxi Gan, Rongqiang Liu, Yanjin Wu, Weili Guan, and Xiang Deng. UAV-ON: A Benchmark for Open-World Object Goal Navigation with Aerial Agents. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 13023– 13029, 2025. doi: 10.1145/3746027.3758251. Daoxuan Zhang, Ping Chen, Xiaobo Xia, Xiu Su, Ruichen Zhen, Jianqiang Xiao, and Shuo Yang. APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal Naviga- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 15232–15242, 2026. Baining Zhao, Jiacheng Xu, Weicheng Feng, Xin Zhang, Zhaolu Wang, Haoyang Wang, Shilong Ji, Ziyou Wang, Jianjie Fang, Zhiheng Zheng, Weichen Zhang, Yu Shang, Wei Wu, Chen Gao, Xinlei Chen, and Yong Li. WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation. arXiv preprint arXiv:2605.15964, 2026. doi: 10.48550/arXiv.2605.15964. Jiacong Zhou, Jiaxu Miao, Yourun Lin, Xianyun Wang, Jun Xiao, and Jun Yu. Memory-Augmented Scene Understanding and Exploration for Open-World Aerial Object-Goal Navigation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 21616–21626, 2026. 11 AerialDojo-200K APROMPT Figure 6 presents the evaluation prompt for SemanticOGS in AerialDojo-200K. The prompt spec- ifies the front-facing RGB-D inputs, depth encoding, eight available actions, aerial embodiment configurations, termination Conditions, and the required single-action JSON response. It defines successful completion as an explicit Stop within 3 m of the target anchor, without collisions and within the assigned action budget. Prompt for AerialDojo-200K ## Task Description You control a aerial agent using its current front-facing RGB image and depth image. Choose exactly one action that best moves toward the described target while avoiding visible obstacles. ## Visual Observations The first image is the front-facing RGB image. The second image is the aligned front-facing depth image. ## Depth Encoding Each depth pixel is an unsigned 8-bit value from 0 to 255, with distance_m = pixel_value / 255 * 100; 0 is the nearest encoded distance and 255 is 100 metres or farther. ## Action Space The action must be one of: moveforward, moveleft, moveright, moveup, movedown, turnleft, turnright, stop. ## Output Requirements Return exactly one valid JSON object in the form "action":"moveforward". Do not include Markdown, explanations, or any other keys. ## Success Criterion Success requires an explicit Stop within 3m of the target anchor, without any collision and within the assigned budget. Figure 6: Prompt for MLLM evaluation in AerialDojo-200K. 12 AerialDojo-200K BTASK COMPOSITION AND REFERENCE LENGTHS AerialDojo-200K contains 79,974 Base, 82,424 Standard, and 43,334 Long-Horizon task instances, representing 38.87%, 40.06%, and 21.06% of the release, respectively. Figure 7 separates task volume, Train/Test allocation, and reference length so that a large task set is not confused with a task set containing longer routes. Urban scenes contribute 60.79% of all instances, followed by Natural (24.22%), Infrastructure (9.63%), and Disaster (5.36%). Mall scenes alone contribute 28.82%. Scene-level counts range from 190 in N Mountain1 to 40,214 in UMall0. These unequal counts motivate reporting the aggregation rule explicitly when comparing methods; a task-weighted score and an equally weighted scene-family score represent different evaluation populations. UrbanNatural Infra. Disaster 0 20 40 60 Task instances (thousands) n=125,064 n=49,832 n=19,814 n=11,022 BaseStandardLH IDOOD 0 40 80 120 Task instances (thousands) 113.6 12.7 28.5 50.9 TrainTest UrbanNatural Infra. Disaster 0 10 20 30 40 50 Mean reference length (m) 41.69 40.46 36.04 25.93 (a) Task counts by difficulty(b) Train/Test allocation (c) Mean reference length Figure 7: Task composition and reference lengths. (a) Goal-conditioned instance counts by family and setting. (b) Train/Test counts in the ID and OOD scene groups. Counts in (a,b) include both goals. (c) Mean reference length, weighted by navigation-task count within each family. CSCENE PARTITIONS AND TARGET COVERAGE The scene partition assigns one instance of each type to ID and the other to OOD, giving 21 scenes per group. Equal scene counts do not imply equal task counts: ID scenes contain 142,094 instances (69.07%), whereas OOD scenes contain 63,638 (30.93%). Table 6 reports the actual integer alloca- tion. ID scenes contain 113,642 training and 28,452 test instances; the corresponding OOD counts are 12,712 and 50,926. Their training proportions are 79.98% and 19.98%, respectively. These realized proportions are close to the nominal 80/20 and 20/80 allocations. Table 6: Task-instance counts by scene group, local Train/Test subset, and difficulty. Both goal modalities are included. ID/OOD identifies the scene group; Train/Test identifies the task subset within that group. ScenesTask subsetBaseStandardLHALL IDTrain40,64048,39824,604113,642 IDTest10,24012,0606,15228,452 OODTrain5,7644,4522,49612,712 OODTest23,33017,51410,08250,926 TotalAll79,97482,42443,334205,732 13 AerialDojo-200K DINVENTORY As summarized in Table 7, AerialDojo-200K contains 205,732 search task instances across 42 simu- lation scenes, covering 21 scene types in four scene families: urban, natural, infrastructure, and dis- aster. Each scene type includes one in-distribution (ID) and one out-of-distribution (OOD) instance. Together, these scenes contain 109 landmarks, 2,099 annotated objects, and 2,099 anchors. The dataset comprises 79,974 Base, 82,424 Standard, and 43,334 Long-Horizon task instances, with both goal modalities included in these counts. The reference trajectories span a total of 4,115.313 km, with each trajectory counted once. Table 7: Scene inventory and search task statistics in AerialDojo-200K. For each scene, LM reports the number of landmarks, and Obj./Anc. reports the numbers of annotated objects and anchors, re- spectively. Base, Standard, and LH report the numbers of search task instances in the Base, Standard, and Long-Horizon settings; ALL is their sum. Task instance counts include both goal modalities. Path (km) reports the total length of reference trajectories, with each trajectory counted once. Scene suffixes 0 and1 indicate in-distribution (ID) and out-of-distribution (OOD) scenes, respectively. SceneLMObj./Anc.BaseStandardLHALLPath (km) Urban Mall06123/12312,41219,6528,15040,214839.245 Mall15102/1025,3907,1446,53619,070459.569 Stadium0564/644,3448,03214612,522213.410 Stadium1141/411,2342,0101,1564,40097.323 Park 0214/143469064425.077 Park1246/461,4521,6166003,66870.006 ParkingLot0778/781,5525,7883,14010,480265.216 ParkingLot 1126/2667619808748.540 Community0465/653927401,1282,26065.735 Community 1331/3130038410478814.582 Neighborhood 0132/321,738996262,76036.245 Neighborhood 1263/634,486680325,19851.460 Alley0327/271,1887141002,00226.514 Alley1267/672,5061,5863064,39865.324 Factory0780/801,0521,6421,0483,74286.301 Factory 1250/5051232412496016.148 ConstructionSite0359/591,7383,4623,6408,840232.868 ConstructionSite1162/629147827502,44653.690 Natural Desert0364/643,0105,0086,29414,312382.722 Desert 1129/297002362896411.950 Forest 0154/547,22041207,63264.708 Forest 1128/281,012626841,72224.887 Mountain0253/536483564001,40426.958 Mountain1132/321325441902.262 Snowfield0121/2147821426948.153 Snowfield1144/441,7068121,0363,55472.382 Island 0579/792,9423,3621,7468,050161.095 Island1260/601,0989228142,83460.992 Coast0150/501,5902,1102,0825,782141.467 Coast 1255/551,1321,0764862,69450.507 Infrastructure Bridge 0328/2838233420892417.401 Bridge1320/20240152204125.796 RailCorridor 0350/502,3262,9866665,978110.070 RailCorridor1554/541,592674362,30227.301 Harbor04110/1104,3363,4661,9749,776184.783 Harbor 1218/185818817642211.716 Disaster Earthquake0250/501,4904001,53010.934 Earthquake1349/491,1326682122,01232.151 Flood0325/2594461001,55420.214 Continued on the next page 14 AerialDojo-200K Table 7 continued SceneLMObj./Anc.BaseStandardLHALLPath (km) Flood1113/1323019064266.205 Explosion 0119/1975244401,19615.626 Explosion1164/642,5921,644684,30457.780 Total1092,099/2,09979,97482,42443,334205,7324,115.313 15