AirSim360 is the first platform to systematically model the 4D real world under an omnidirectional setting for UAVs. Built on a cutting-edge rendering engine (Unreal Engine 5 series), it addresses the critical lack of large-scale and diverse 360-degree aerial data.
The platform enables closed-loop simulation for omnidirectional aerial systems and offers an integrated toolchain for intelligent data acquisition across diverse flight scenarios.
Please wait for our realease source including python tookit, UE plugin and collected dataset!
AirSim360's design focuses on three key components for high-quality data generation:
- Panoramic Image: Adopts the Equirectangular Projection (ERP) as the omnidirectional representation. It uses a GPU-side texture copying mechanism (based on RHI) to quickly and seamlessly stitch six cube-face views for one-time image stitching.
- Depth Information: Defined as the distance (slant range) from the camera center to the point along the viewing ray, aligning with the omnidirectional perspective.
- Segmentation: Provides pixel-level Semantic Segmentation and Entity Segmentation annotations. Entity segmentation ensures complete coverage of all entities (static mesh actors, skeleton mesh actors, and landscape elements) across the entire image.
- Synchronous Rendering: A custom Event Dispatcher synchronizes all sensors with a unified trigger signal to ensure simultaneous data acquisition.
- Human Behavior Modeling: Simulates movable pedestrians with various actions and realistic interactions.
- Autonomous Interaction: Combines NPC Behavior Trees with State Machines for autonomous interaction, such as switching from walking to chatting when agents meet.
- 3D Human Keypoints: Generates keypoints in real time, ensuring temporally consistent annotations for downstream perception tasks, such as 3D monocular human localization.
- Minimum Snap Planning: Adopts the Minimum Snap trajectory planning method to automatically generate a smooth, realistic trajectory from sparse user-defined waypoints.
-
Dynamic Feasibility: The generated trajectory adheres to realistic dynamic constraints on maximum velocity (
$v_{max}$ ) and acceleration ($a_{max}$ ) of the UAV, making it directly applicable to real quadrotors.
Built on AirSim360, we collected the large-scale Omni360-X dataset (more than 60K non-duplicate frames), structured into three subsets:
| Subset | Focus | Samples/Frames | Key Annotations |
|---|---|---|---|
| Omni360-Scene | Panoramic scene parsing | ~61,000 images | Depth, Semantic/Entity Segmentation |
| Omni360-Human | Pedestrian behavior understanding | ~100,700 frames | 3D Human Keypoints |
| Omni360-WayPoint | UAV Navigation and Control |
|
Physics-consistent trajectories |
- Project Website: https://insta360-research-team.github.io/AirSim360-website/