Task-Oriented Occupancy Grid Mapping Through Active Data Management

Abstract

Occupancy mapping is essential for robotic systems operating in unknown or dynamic environments with limited computational resources. However, many existing frameworks rely on a fixed design choice prioritizing either local or global mapping, which limits adaptability to the other domain. This paper presents a flexible, hash-based occupancy grid mapping framework that addresses this limitation through active data management. The framework integrates three key features: Sequential Voxel Pruning (SVP) for direct control of voxel counts and memory, Dynamic Boundary Filtering (DBF) for focusing updates on task-relevant regions, and Decoupled Obstacle Inflation (DOI) for maintaining a compact inflation layer and efficient collision checking. Benchmark results demonstrate that the framework outperforms other methods in update rate and memory management across diverse setups, while effectively maintaining memory usage below a user-defined limit. Simulation experiments on global pathfinding and UAV local planning show that it adapts effectively to navigational tasks of different planning scales, where specialized maps often struggle. Real-world experiments confirm stable mapping under the computational load of active robotic platforms.