End-to-End Vision-Based UAV Navigation in Unseen Complex Obstacle Environments

Abstract

Autonomous quadrotor navigation in unseen complex environments and under dynamic-obstacle interference requires safe and efficient decisions using only onboard sensing. This paper proposes an end-to-end vision-based navigation method built upon candidate motion primitive prediction. The proposed framework takes consecutive depth observations, the current flight state, and the local goal as inputs to generate executable local motion primitives. To improve navigation adaptability in environments with varying obstacle complexity, a Complexity-Aware Trajectory Scoring (CATS) module adaptively balances safety clearance and goal-directed efficiency during candidate evaluation. In addition, a Dynamic Collision Risk Prediction (DCRP) module estimates the future collision risk of each candidate primitive from short-term depth variations, without requiring explicit mapping or dynamic-object tracking. The selected primitive is executed in a receding-horizon manner for closed-loop navigation. Experiments in static and dynamic simulation environments, together with real-world quadrotor flight tests, demonstrate the effectiveness and practical deployability of the proposed method.