A real-time path planning approach for underground mining truck without prior map

Abstract

Autonomous driving in underground mining environments faces challenges such as lack of prior high-precision maps, perception noise, and occlusions, which can cause path discontinuities and increase safety risks. This paper presents a real-time path planning framework that ensures the spatiotemporal continuity of local paths without relying on prior maps. First, a unified representation for spatial coordinate transformation is established, and an adaptive-degree Bézier curve fitting method is analyzed. Then, a multi-layer grid map construction method based on real-time perception data is designed. The bidirectional circular search algorithm (BCS) is introduced, establishing a recursive search objective function that considers dynamic grid weights to efficiently guide the search direction. A physics-informed shallow neural network is employed to integrate current and historical control points to suppress perception-induced fluctuations. Finally, comparative experiments were conducted in a full-scale simulated underground roadway using a real diesel mining truck. Results show that the proposed BCS-t10 reduces temporal curvature by 58.80% compared with single-frame planning. It achieves only 22.31% of the mean temporal curvature of state-of-the-art methods, and maintains high robustness and computational efficiency. These results demonstrate its superior adaptability and spatiotemporal continuity in narrow underground mining roadways.