U2UData-2: Long-Horizon Swarm UAV Flight Dataset

• U2UData-2 is a comprehensive swarm UAV dataset designed for long-horizon flight tasks with 15 UAVs across diverse scenes.
• It provides synchronized multimodal sensor data including LiDAR, RGB, depth, and environmental inputs, enabling robust cooperative tracking.
• The platform offers a configurable simulation environment with closed-loop algorithm verification to support persistent multi-agent autonomy.

Searching arXiv for the primary paper and closely related dataset papers. U2UData-2 is a large-scale swarm UAV autonomous flight dataset for long-horizon (LH) tasks and an associated scalable online data collection and algorithm closed-loop verification platform. It is captured by 15 UAVs in autonomous collaborative flights across 12 scenes, 720 traces, 120 hours, and 600 seconds per trajectory, and it includes 4.32M LiDAR frames, 12.96M RGB frames, 12.96M depth frames, and environmental modalities comprising brightness, temperature, humidity, smoke, and airflow along all flight routes. In the paper’s formulation, LH tasks are not mere concatenations of basic tasks; they require handling long-term dependencies, maintaining persistent states, and adapting to dynamic goal shifts. U2UData-2 is therefore positioned as an extension of swarm-UAV dataset design from short-horizon cooperative perception toward persistent multi-UAV autonomy (Feng et al., 25 Aug 2025).

1. Dataset identity and position within swarm-UAV research

U2UData-2 is explicitly introduced as the first large-scale swarm UAV autonomous flight dataset for LH tasks and the first scalable swarm UAV data online collection and algorithm closed-loop verification platform. Its central novelty lies in combining long-duration trajectories, multi-UAV autonomy, multimodal sensing, configurable simulation, and an evaluation stack for collaborative tracking in LH settings (Feng et al., 25 Aug 2025).

The dataset should be distinguished from the earlier U2UData release, which is a separate 2024 dataset rather than a second version. The 2024 paper names the dataset simply “U2UData,” and that release targets cooperative 3D object detection and cooperative 3D object tracking with three UAVs, short 15 s scenarios, and fixed dataset settings rather than LH tasks (Feng et al., 2024).

Dataset	UAVs and horizon	Scope
U2UData	3 UAVs; ET-Length: 15 s; TLT: 8.75 h	Fixed dataset size/settings; basic tasks only
U2UData-2	15 UAVs; ET-Length: 600 s; TLT: 120 h	Scalable; adds visual control window, online collection, closed-loop algorithm verification, and LH wildlife conservation task

Relative to earlier synthetic datasets summarized in the paper, such as CoPerception-UAVs and CoPerception-UAVs+, U2UData-2 preserves the same broad UAV cooperation focus while increasing the number of UAVs, duration, sensing richness, and configurability. The comparison presented in the paper also frames U2UData-2 as a response to limitations of short, basic-task datasets and to domain gaps associated with AirSim/CARLA-based synthetic collections (Feng et al., 25 Aug 2025).

2. Data scale, modalities, and synchronization

U2UData-2 is organized around 15 multirotors flying in autonomous formation. The paper reports 12 scenes, 720 traces, 120 hours of flight, and trajectories of 600 seconds each. The total storage size is over 3.62 TB. The sensor record comprises 12.96M RGB frames at 1920×1080, 12.96M depth frames, 4.32M LiDAR frames, 25.92M airflow frames, and 12.96M frames each for brightness, temperature, humidity, and smoke (Feng et al., 25 Aug 2025).

Each UAV carries five RGBD cameras mounted front, back, left, right, and bottom, with 1920×1080 resolution, 90° FOV, and 30 Hz sample rate. The LiDAR is a 64-channel top-mounted sensor with throughput of 1 million points/second, range of 200 m, accuracy of ±3 cm, vertical FOV of −30° to 30°, horizontal FOV of −180° to 180°, and 10 Hz sample rate. Environmental sensing includes brightness, temperature, humidity, and smoke sensors mounted on the bottom, together with two airflow sensors mounted on the back and right for latitudinal and longitudinal wind speeds. GPS and IMU provide odometry (Feng et al., 25 Aug 2025).

The synchronization description is an important part of the dataset specification. The paper states that all sensors are sampled at 0.03 seconds and synchronized in real time, while also listing 30 Hz RGBD and 10 Hz LiDAR sample rates. This suggests that the implementation couples modality-specific acquisition settings with a synchronized timestamping regime. The same section also states that relative poses are initialized per frame using IMU positional information from the UAVs (Feng et al., 25 Aug 2025).

The sensing design is broader than standard cooperative perception datasets because it couples geometric modalities with environmental variables. A plausible implication is that U2UData-2 is intended not only for perception-stack benchmarking but also for condition-aware flight control, adaptive coordination, and task re-allocation under weather and terrain changes.

3. Real-world mapping simulator and configurable platform

The data are generated in a real-world mapping simulator built with Unreal Engine 5.2 and based on Yunnan Province. The environment is a 3 km × 3 km scaled map with elevation range [56.6, 3000] m, four terrain types—mountains, hills, plains, and basins—and 58 types of original forest vegetation with more than 15 superposition methods, including epiphytic growth and diagonal staggered growth. Leaves respond dynamically to wind and snow. The simulator also includes 15 animal types, with predator–prey dynamics and resource competition affecting collective motion patterns (Feng et al., 25 Aug 2025).

Meteorology is mapped into the simulator from real weather data from the China Meteorological Center by latitude and longitude. Temperature and humidity are treated as scalars with missing values filled by moving average with interval 5 m. Wind speed and direction are decomposed along longitude and latitude, missing values are filled by sliding averages, and the field is reconstructed via vector synthesis. Weather is deployed in specific non-overlapping areas of the map except for wind, which is present across the entire map; the map itself is divided into 6 areas, and weather and terrain are described as strongly coupled (Feng et al., 25 Aug 2025).

The platform is designed as a configurable system rather than a fixed dataset generator. It supports customization of simulators, UAVs, sensors, flight algorithms, formation modes, and LH tasks, together with one-click online dataset collection in a visual control window and closed-loop simulation for algorithm verification. The UAV and sensor stack is configured through setting.json, which exposes 132 customizable parameters, including sensor type, quantity, position, angle, resolution, and LiDAR Range and Number-Of-Channels. UE5.2 controls include animal quantity and activity radius, weather intensity and spatial range via sliders, and selectable UAV “starting point–weather–task” combinations; F11 toggles visual adjustment in the simulator startup interface (Feng et al., 25 Aug 2025).

The platform description also includes operational commands. On Windows the simulator is launched by double-clicking Landscape3.exe; on Linux it is launched by running .Landscape3.sh; the visual control window is started with python AirDroneClient.py. Keyboard control exposes “Take off,” “Land,” “Up,” “Down,” “Move Forward,” “Move Backward,” “Move Left,” and “Move Right” after clicking keyboardControl, and XBOX controller support is provided, although controller and keyboard control are mutually exclusive (Feng et al., 25 Aug 2025).

4. Long-horizon task design and annotation model

The paper describes four preset LH tasks: wildlife conservation, logistics distribution, patrol security, and disaster rescue. Wildlife conservation is the collected public dataset focus. The task is defined around adaptive animal tracking using real-time behavior prediction across variable terrains and vegetation, with synchronized RGBD and LiDAR, condition-aware flight control using environmental sensors, and multi-UAV coordination to mitigate single-view occlusions. Missions run for 600 seconds per trajectory with continuous state and goal adaptation to weather shifts or animal movement (Feng et al., 25 Aug 2025).

The scene set comprises 12 weather-oriented configurations. The single-weather scenes are Sunny, Rain, Snow, Sandstorm, Thunder, and Fog. The cross-weather scenes are Sunny→Rain, Sunny→Snow, Sunny→Fog, Sunny→Sandstorm, Rain→Thunder, and Rain→Snow. The paper also states that single-weather scenes have 5 trajectories each and cross-weather scenes have 3 trajectories each. Independently, it reports 720 traces overall. This suggests that the paper uses scene-level route design and dataset-scale trace counts at different descriptive levels rather than reducing the corpus to a single per-scene enumeration (Feng et al., 25 Aug 2025).

Annotations are produced with SusTechPoint. The dataset provides 15 object classes and 7-DoF 3D boxes with position $(x, y, z)$ and rotation as quaternions $(w, x, y, z)$, with location referring to the box center. Each UAV’s sensor data is treated independently with its own global coordinate system, which supports single-agent detection per UAV. Relative poses are initialized per frame using IMU positional information from the UAVs. The split is random train/val/test at 0.7/0.15/0.15 (Feng et al., 25 Aug 2025).

The paper gives partial but operationally important information about formats and access. Images are stored as PPM, conversion to PNG is supported through python FileConverter.py, the storage layout is illustrated on the project page, one scenario of 252 GB is uploaded, and the full dataset of 3.6–3.62 TB is available via Baidu Cloud after acceptance by email request. The paper does not specify ROS bag usage or a formal metadata schema beyond the described fields, and licensing is not explicitly stated (Feng et al., 25 Aug 2025).

5. Benchmark protocol and quantitative tracking results

U2UData-2 provides an LH collaborative tracking benchmark with nine cooperative tracking algorithms: No Fusion, Late Fusion, Early Fusion, When2Com, DiscoNet, V2VNet, V2X-ViT, CoBEVT, and Where2com. The benchmark uses AB3Dmot as the base tracker with LiDAR detections and a 3D Kalman filter with birth/death memory, PointPillar as the LiDAR feature backbone, and 32× feature compression/decompression for all models. Training randomly designates one UAV as ego, testing uses a fixed ego UAV, and the tracking input consists of the previous three frames plus the current frame (Feng et al., 25 Aug 2025).

The evaluation metrics are AMOTA, AMOTP, sAMOTA, MOTA, MT, and ML. The paper does not provide their formulas, but it does supply exact benchmark values. The baseline No Fusion model records AMOTA 9.36, AMOTP 25.48, sAMOTA 32.19, MOTA 23.47, MT 18.67, and ML 65.52. Cooperative methods significantly outperform this baseline, with reported improvements of at least 35.97% AMOTA and 32.88% sAMOTA (Feng et al., 25 Aug 2025).

Among the tracked models, CoBEVT attains the highest AMOTA, AMOTP, and MOTA, with AMOTA 24.63, AMOTP 45.76, and MOTA 51.18. V2VNet attains the highest sAMOTA and MT, with sAMOTA 57.82 and MT 35.68. Where2com attains the lowest ML at 26.42. The paper also states that intermediate fusion can improve tracking performance up to 27.48% AMOTA versus Late Fusion and up to 8.33% AMOTA versus Early Fusion (Feng et al., 25 Aug 2025).

The full benchmark progression is consistent with the cooperative tracking literature summarized by the paper: Late Fusion improves substantially over No Fusion; Early Fusion improves further on some metrics; and intermediate fusion methods such as When2Com, DiscoNet, V2VNet, V2X-ViT, CoBEVT, and Where2com define the leading performance band. In this sense, U2UData-2 functions both as a long-duration dataset and as a stress test for persistent multi-UAV coordination under dynamic weather and terrain, where the paper argues that short-horizon settings are insufficient (Feng et al., 25 Aug 2025).

6. Interpretation, limitations, and common confusions

The most common confusion surrounding the name is historical. The 2024 paper “U2UData: A Large-scale Cooperative Perception Dataset for Swarm UAVs Autonomous Flight” introduces U2UData, not “U2UData-2,” and centers on cooperative 3D object detection and cooperative 3D object tracking over 15-second scenarios collected by three UAVs in U2USim (Feng et al., 2024). U2UData-2 is the later 2025 dataset and platform devoted to LH tasks, larger swarms, longer trajectories, and platform-level scalability (Feng et al., 25 Aug 2025).

The paper also states several limitations explicitly. Only one LH task, wildlife conservation, is collected in the public dataset because of size, and users are encouraged to collect other LH tasks through the platform. The paper does not provide MDP or POMDP formalisms, reward functions, or control equations; formal algorithmic objectives are therefore unspecified. It also does not include additional ablations, scene-generalization tests, or long-horizon error accumulation analyses beyond the reported collaborative tracking comparisons (Feng et al., 25 Aug 2025).

Other constraints are infrastructural and methodological. Licensing terms are not stated. Although the simulator maps real environmental data, it remains a simulated environment, described in the paper as a digital twin with Yunnan-specific mapping, so potential domain bias may remain. Ethical considerations are noted only at a high level: wildlife conservation tasks should avoid causing disturbance, and low-altitude operations should ensure privacy and safety, but the paper does not discuss ethical protocols (Feng et al., 25 Aug 2025).

These limitations define the present status of U2UData-2 as a research instrument. It is comprehensive in scale, modality coverage, platform configurability, and benchmark integration, but it remains bounded by simulator realism, partial formalization of decision-making objectives, and a public release centered on a single LH task. Within those bounds, it marks a shift in swarm-UAV dataset design from fixed short-horizon cooperative perception toward configurable, closed-loop, long-duration multi-agent autonomy.