Photorealistic SITL Pipeline for Spatial AI and Autonomous Navigation
Unreal Engine / AirSim / PX4 SITL simulation infrastructure for autonomous UAV navigation, integrating visual-inertial SLAM (ORB-SLAM3) and sensor-realistic simulation for evaluating perception, localization, and control algorithms in GPS-denied environments.
Built a photorealistic Software-in-the-Loop (SITL) simulation pipeline combining Unreal Engine, AirSim, and PX4 autopilot for rigorous evaluation of autonomous UAV navigation algorithms. The pipeline provides sensor-realistic data streams and supports visual-inertial SLAM for GPS-denied localization and control.
Overview
Testing autonomous navigation algorithms on real UAVs is expensive and risky. This infrastructure enables high-fidelity evaluation of perception, localization, and control pipelines in a photorealistic simulation environment before physical deployment.
Key Components
- Unreal Engine + AirSim: Photorealistic rendering with physically accurate camera, IMU, LiDAR, and depth sensor simulation matching real sensor specifications
- PX4 SITL: Full autopilot stack runs in software against the simulated vehicle model, enabling end-to-end flight software testing
- ORB-SLAM3 Integration: Visual-inertial SLAM provides real-time 3D map building and localization using only onboard camera and IMU — no GPS
- GPS-Denied Scenarios: Evaluated navigation and control algorithms in environments where GPS is unavailable (indoor-like and adversarial outdoor)
- Dataset Generation: Produced large-scale labeled datasets for training and evaluating perception models with ground-truth pose and depth