AIOZ AI Research: AeroScene Advances 3D Scene Synthesis for Aerial Robotics

We are delighted to announce that the AIOZ AI Research paper, "AeroScene: Transforming Aerial Robotics with Progressive 3D Scene Synthesis," has been accepted for publication at the IEEE International Conference on Robotics and Automation (ICRA 2026).

As drones expand into delivery, inspection, and search-and-rescue missions, they require realistic 3D environments for training and evaluation. To address this need, the AIOZ AI Research team introduces AeroScene, a hierarchical diffusion framework for generating diverse, physics-aware 3D scenes tailored to aerial robotics.

The Challenge in Aerial Robotics Simulation

Real-world autonomous drone missions require environments with navigable corridors, obstacles, landing zones, and physically plausible object placement.

Existing drone simulators often rely on handcrafted scenes that are costly to create and difficult to scale. The challenge is to generate diverse environments that are both visually realistic and physically suitable for robotics simulation.

AeroScene: Progressive 3D Scene Synthesis

AeroScene addresses this challenge through a hierarchical diffusion architecture that generates environments across multiple levels of detail.

Key components include:

• Hierarchy-aware tokenization and feature extraction to separate large-scale structure from fine-grained geometry
• Cross-scale progressive attention to connect global layouts with local object placement
• Task-aware guidance objectives to improve collision avoidance, structural consistency, and semantic relationships

This structure allows AeroScene to move beyond flat scene layouts and generate environments where large-scale structure and task-specific details work together for aerial robotics evaluation.

How AeroScene Works

AeroScene uses a multi-scale generation pipeline designed around the physical needs of drone navigation.

High-level scene representations capture terrain, buildings, and navigable space, while finer representations model local details such as obstacles, landing areas, and object geometry.

The framework then uses cross-scale progressive attention to connect those levels. Top-down context helps guide local placement, while bottom-up detail keeps the full scene semantically consistent.

During denoising, AeroScene applies three physics-aware guidance mechanisms:

• Collision Avoidance: reducing object overlap
• Coarse-to-Fine Guidance: aligning detailed objects with the broader scene plan
• Semantic Constraint Guidance: preserving realistic relationships between scene elements

Results and Performance

AeroScene was evaluated on the 3D-FRONT benchmark and a newly collected drone-oriented dataset.

It outperformed competing methods, including DiffuScene, PhyScene, ATISS, and Diffusion-SDF, across the reported metrics. The lower collision rate is especially important for drone simulation, where generated scenes must support safe and realistic navigation.

From Synthesis to Drone Simulation

Beyond scene synthesis, AeroScene was designed for practical robotics evaluation.

The research generated more than 1,000 high-fidelity 3D scenes and integrated them into NVIDIA Isaac Sim for downstream testing.

In simulation trials, drones successfully completed navigation and rooftop landing tasks with an overall success rate of 91%, demonstrating the usefulness of the generated environments for autonomous flight evaluation.

Why It Matters

AeroScene supports a broader shift toward scalable simulation for aerial autonomy. Potential applications include:

• Autonomous delivery
• Infrastructure inspection
• Search-and-rescue operations
• Robotics research and benchmarking

For AIOZ AI Research, this work highlights a key direction in robotics intelligence: making high-fidelity simulation more scalable, diverse, and task-aware.

Conclusion

AeroScene advances 3D scene synthesis for aerial robotics by combining hierarchical diffusion, cross-scale attention, and physics-aware guidance.

By automating the creation of realistic, physics-ready environments, AeroScene provides a practical step toward more scalable development and validation of autonomous drone systems.

For more technical details:

Paper: https://arxiv.org/abs/2603.23224

Project site: https://aioz-ai.github.io/AeroScene/