Closed-Loop Online Safety-Oriented Planning for 3-D UAV Navigation in Unknown Cluttered Dynamic Environments

Abstract

Navigation in unknown dynamic environments must be based on partial and noisy onboard observations rather than a complete prior map. Safety-oriented flight therefore depends on how perception, planning, path checking, and execution interact online. We present a framework for safety-oriented three-dimensional UAV navigation using simulated 16-beam LiDAR observations for environmental sensing and the vehicle state for planning and control. Consecutive scans update a short-term voxel map. Inter-frame differencing and clustering provide moving-object observations, and nearest-neighbor association with Kalman filtering estimates position, velocity, and uncertainty. Predicted occupancy adjusts a variable-weight A* search and the spacing of position references. Predicted covariance enlarges the safety constraints around dynamic obstacles. Dense segment sampling rejects transitions when an interior sample violates an obstacle constraint, while a near-field LiDAR guard limits references in the immediate safety region. We evaluate the framework on three disjoint 100-case panels, seven paired ablations, sensing and obstacle-speed perturbations, and comparisons with FAPP, Dyn-Small-Obs, and PX4 3DVFH+ under a shared local-information protocol. Collision-free completion ranged from 95% to 97% across the three panels. Under the common protocol, the proposed framework achieved 96% strict success, compared with 40%–75% for the external systems. Removing inter-frame differencing or dense segment checking reduced collision-free completion to 80% and 73%, respectively. Under degraded LiDAR sensing or a 25% increase in obstacle speed, the rate remained 89%–92%. Among the paired ablations, inter-frame foreground extraction and dense segment validation had the largest measurable effects on collision-free completion.