Hierarchical Lightweight UAV Mapping-Planning System Using Deterministic Voxel and Probabilistic Sampling

Autonomous navigation of micro-unmanned aerial vehicles (UAVs) in cluttered and dynamic environments presents significant challenges. These challenges stem from several factors, including strict weight constraints on onboard computing, the necessity for more detailed geometric representations or denser sampling of obstacles, and the demands of continuous map reconstruction and frequent path replanning in response to dynamic obstacles. This article introduces a lightweight mapping and planning system using deterministic voxel and probabilistic sampling, termed DVPS-MPS, which utilizes a hierarchical structure that combines deterministic voxel maps with probabilistic sampling. This system constructs a three-level map hierarchy. The first level is a Euclidean distance field using a block-linked voxel data structure to encode fine-grained obstacle distances around the UAV. The second level features a hybrid voxel-sample fusion graph that combines probability-oriented sampled nodes with downsampled voxel centers, enhancing sampling in high-value areas. The third level consists of a reusable global voxel graph that retains previously explored regions at a coarse resolution. The higher level maps are abstracted from lower level ones, forming a unified representation that captures both local details and global structure. A top-down planning strategy is used, initially generating a coarse path on high-level maps for long-range optimality, followed by refinement at lower levels to enhance feasibility and ensure dynamic safety. The path quality is further enhanced through iterative optimization using adaptive sampling and gradient-based smoothing, followed by B-spline generation that adheres to kinematic constraints. Thus, DVPS-MPS can initialize in both fully unknown and partially known environments, dynamically updating the map by fusing sensor observations to effectively manage moving obstacles and changes in static structures. The proposed mapping-planning system is installed on a 2.5-in UAV platform with a Raspberry Pi and has been verified in various real-world scenarios. This system can achieve real-time operation (≥10 Hz), produce high-quality trajectories, and maintain stable performance in narrow and dynamic environments, even with limited computing resources.