Novel Frontier Detection and Optimisation for Scalable Autonomous Exploration

Autonomous exploration must be fast, scalable, and complete, yet popular frontier based methods either scan exhaustively (e.g., classical BFS/wavefront) or restrict search locally (e.g., Naıve Active Area, Expanding Wavefront), leading to high computational cost or missed opportunities as maps grow. We propose a hybrid framework that introduces a grid-tree frontier detector that hierarchically focuses search and parallelises evaluation to avoid re-scanning, adds a CNN-based learned detector for rapid frontier inference, and employs Bayesian optimisation to auto- tune goal-selection weights. In simulation across various maps, the grid tree method achieves up to 3.8× lower per-cycle computation than a naive baseline while preserving completeness; the learned detector is faster still with minor completeness loss. By reducing frontier detection overhead and automating parameter tuning, the framework addresses the efficiency–completeness gap in frontier exploration.