A High-Fidelity Digital Twin Framework for Multi-Modal Sim-to-Real Transfer for Robust UAV Autonomy
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Autonomous Unmanned Aerial Vehicle (UAV) navigation remains an important challenge under GNSS-denied and communication-jammed conditions. Visual-Inertial Odometry (VIO) systems experience compounding biases, while traditional simulators cannot properly model real-world visual and wireless phenomena with sufficient fidelity, leading to "the reality gap." In this paper, we introduce a new dual-fidelity DT paradigm that jointly leverages Unreal Engine 5 (UE5) for real-time visual and dynamic simulations and the high-fidelity robotics simulator, Colosseum-AirSim. We describe how this DT infrastructure can enable high-quality S2R learning by creating synthetic images with DR and using efficient AI models like ResNet-18 for real-time semantic image segmentation, achieving mIoU of 87.0% or higher. This significantly surpasses state-of-the-art performance reported with pure simulation learning or real-image learning alone. These high-fidelity DT-infrastructure capabilities enable high-quality communication-aware, context-aware, autonomous functioning with localization under HITL control for real-world applications such as smart urban networks or disaster response operations.
Fields of Study & Subjects
Reinforcement Learning in Robotics
Aerospace Engineering and Control Systems
Robotics and Sensor-Based Localization
Publication details
- Journal / source
- International Conference on Computing, Communication, Control and Cyber-Physical Systems (I5CPS)
- Pages
- 1–5
- Publication date
- 26 February 2026
- Language
- EN