MBD-Planner: A Real-Time Obstacle Avoidance Framework for UAVs via Feature-Domain Motion Blur Decoupling

Abstract

To address UAV obstacle avoidance under motion blur, we propose Feature-Domain Motion Blur Decoupling Planner (MBD-Planner), a real-time feature-domain motion blur decoupling framework that jointly disentangles blurred visual features and optimizes trajectories in an end-to-end manner. Unlike methods relying on explicit 3D reconstruction or ignoring blur in planning, MBD-Planner introduces a lightweight dewarping module that transforms blurred features into virtual sharp ones via depth back-projection and pose compensation, achieving sub-millisecond latency. A blur-aware optimization objective further couples camera motion and trajectory smoothing, while privileged training guided by ESDF gradients enables robust learning without expert demonstrations. Experimental results demonstrate that MBD-Planner enables reliable, smooth, and real-time navigation in visually degraded environments, effectively bridging the gap between theoretical planning and robust navigation.