Rapid Flight Trajectory Planning for Autonomous Terrain Avoidance via Generative Learning

Abstract

Ensuring aircraft safety against terrain collisions in complex and dynamic environments remains a critical challenge in aviation. To address this, a parallel autonomy system is proposed that can take control from a human pilot to prevent a controlled flight into terrain collision. The proposed system operates in the background, continuously maintaining a forward-looking motion plan that can be executed immediately if a terrain collision is projected to happen, absent its timely intervention. Terrain avoidance motion plans are rapidly generated based on the aircraft’s current state vector and a Digital Elevation Model of the surrounding terrain. The planning process involves two main steps: first, a sampling-based motion planner leverages prior knowledge acquired through generative adversarial learning to bias the search toward escape paths within the most favorable regions of Cartesian space. Second; differential flatness of the aircraft model is utilized to ensure the dynamic feasibility of an associated Cartesian space escape trajectory and flattening it into a state-control trajectory. This converts the output tracking problem in Cartesian space into a ready-to-invoke state-feedback control.