AirSim360 addresses the lack of large-scale, diverse equirectangular (ERP) drone data by supporting closed-loop flight with render-aligned multimodal exports. This repository provides release notes, user documentation, one synchronized demo sample, and media for the open release.

📢 Release Notice: The public upload of Omni360-Scene and Omni360-WayPoint may be delayed to around April 20, 2026 due to the large number of small files and network-related constraints. The software release is unaffected: both AirSim360 Air and AirSim360 Pro were updated on Hugging Face on April 17, 2026.

This section will be updated continuously so users can track what has been released and what is coming next.

To handle the stunning realism and simultaneous rendering of panoramic RGB, depth, and semantics, you will need a solid rig.

🐧 Linux user? Support is landing in Q2 2026. 🪟 Windows user? Windows 10 & 11 are supported!

• AirSim360 Air — Download on Hugging Face. Designed for direct use with zero environment setup: download the package, unzip it on Windows, start the remote-control program first, and then launch the main simulator.
• AirSim360 Pro — Download on Hugging Face. Shipped as a compiled package for professional developer control: download it, unzip it, run the simulator, and connect via Python or RPC to access the custom API for programmatic control.
• Detailed usage guides are available in software/AirSim360_Air_User_Guide_EN.md and software/AirSim360_Pro_User_Guide_EN.md.

• Current public dataset tracks include Omni360-Scene and Omni360-WayPoint.
• Omni360-Scene — Download on Hugging Face
• Omni360-WayPoint — Download on Hugging Face
• Note: the final public upload of both tracks may be delayed to around April 20, 2026.
• Omni360-Scene follows the same structure shown in the demo sample: every frame is aligned across modalities by a shared file stem.
• The core released modalities are:
  • Raw (panorama): Equirectangular RGB images.
  • Depth: True Euclidean depth in meters, stored as HDF5 datasets.
  • Semantic (seg_panorama): Semantic labels with class IDs carried in the Alpha channel.
  • Instance (instance_panorama): Instance labels where (Alpha, R, G, B) uniquely identifies an object, with Alpha storing the semantic class ID.
• See data/demo_sample/PANORAMIC_DATA_FORMAT.md for the full format specification, and data/demo_sample/depth/read_depth_h5.py for a minimal depth reader.
• A dedicated Omni360-WayPoint section with dataset breakdown and demo scripts will be added in a future update.

Guides: Air User Guide · Pro User Guide

A short demo video of AirSim360 Air, showing its direct-use workflow and panoramic data collection interface.

The following scene packages are currently listed in the public software release. More scenes will be added over time, with monthly updates rolling out on the 15th of each month starting May 2026.

📥 Download Access: All available scene packages and environment assets can be downloaded directly from the AirSim360 Hugging Face Repository.

Omni360-Scene is the current panoramic scene-understanding release in the Omni360-X collection. The sample under data/demo_sample shows the core organization used by this dataset: the same frame stem is shared across RGB, depth, semantic, and instance labels, which makes multimodal alignment straightforward.

The current public specification focuses on four aligned outputs:

• Raw (panorama): Equirectangular RGB images.
• Depth: True Euclidean depth in meters, saved as an HDF5 /depth dataset with float32 values directly in meters — no rescaling or de-quantization is needed when reading. Maximum supported distance is 1000 m. For backward compatibility, the reader (data/demo_sample/depth/read_depth_h5.py) also accepts the legacy uint16 format with a depth_range_m attribute, recovered via depth_uint16 / 65535 * depth_range_m.
• Semantic (seg_panorama): Semantic labels aligned pixel-by-pixel with the panorama, where the Alpha channel stores the semantic class ID. The mapping of class names to IDs is recorded in the semantic_list text file.
• Instance (instance_panorama): Instance labels where the (Alpha, R, G, B) tuple identifies a specific object instance, with Alpha storing the semantic class ID.

The current public release covers three main scenes with full support for all four modalities (RGB, depth, semantic, and instance). The corresponding semantic label IDs are shipped alongside each dataset.

Notes:

1. Additional premium/paid scenes will be released and updated on a monthly basis.
2. Appropriate randomization has been applied to all 6 degrees of freedom (6-DoF) for the panoramic camera across the provided datasets. This ensures a diverse variety of viewing positions and directions, rather than fixing a single orientation.

• Starting May 2026, we will release 5 new AirSim360 scenes each month, shared across the Air and Pro editions.
• Starting May 2026, we will also release 2 new dataset scenes each month, each containing 2,000 panoramic images. Data types will be adjusted based on developer feedback, and community suggestions are welcome.
• Before the end of Q2 2026, we will provide additional dynamic human assets inside the AirSim360 software packages, along with the corresponding calling methods and automatic ground attachment for human placement — so developers no longer need to know ground coordinates in advance.
• We welcome feedback on software usage. Adopted suggestions will be acknowledged in a new section at the end of this page.
• The panoramic plugin and other useful features will be published separately on FAB, and will be free for individual developers.

We gratefully acknowledge the following open-source projects:

• AirSim
• Fly360
• Unreal Engine

If you find our work useful in your research, please consider citing:

@article{ge2025airsim360,
  title={Airsim360: A panoramic simulation platform within drone view},
  author={Ge, Xian and Pan, Yuling and Zhang, Yuhang and Li, Xiang and Zhang, Weijun and Zhang, Dizhe and Wan, Zhaoliang and Lin, Xin and Zhang, Xiangkai and Liang, Juntao and others},
  journal={arXiv preprint arXiv:2512.02009},
  year={2025}
}