Certifiable Transformer-Based Sensor Fusion Architecture for Urban Air Mobility

Urban Air Mobility (UAM) vehicles rely on robust multi-sensor perception for safe navigation in complex environments. This paper presents a novel sensor fusion architecture using a Transformer-based model for integrating LiDAR, Electro-Optical/Infrared (EO/IR) cameras, GNSS, ADS-B, IMU, and radar data. We detail a hardware-software co-design for real-time embedded deployment, emphasizing compliance with DO-178C(software) and DO-254 (hardware) certifiability. A mathematical formulation of the fusion algorithm is provided, leveraging cross-attention to achieve multimodal state estimation. We simulate an urban canyon scenario with multiple UAM vehicles (using ROS2/Gazebo and MATLAB) to evaluate performance. Results demonstrate high accuracy, low latency, and stable confidence intervals, even under sensor degradation or GNSS loss. Comparative analysis shows our Transformer-based fusion outperforms legacy Extended Kalman Filter and earlier deep models (including a BART-based approach) in both accuracy and robustness. We also discuss how the design handles adversarial sensor inputs and degrades gracefully. The proposed architecture, supported by certifiable development practices and a safety monitor subsystem, offers a viable path toward certification in UAM.