Learning to Explore Efficiently: Heterogeneous Topological Graphs and Lightweight Global Reasoning for Robotic Exploration

Abstract

Autonomous exploration in large-scale, unknown environments remains a significant challenge in mobile robotics. In this paper, we propose a scalable exploration framework that integrates heterogeneous topological representations, lightweight global-local graph reasoning, and reinforcement learning. Our framework is computationally efficient, comprehensively considers global spatial context, and generalizes effectively across diverse environmental scenarios. We introduce a compact topological abstraction to encode crucial spatial and semantic information, substantially reducing map complexity. A novel hybrid inference module, combining linear global attention with local graph convolutions, effectively integrates long-range exploration with comprehensive local coverage. We further design a minimalist reward function paired with a curriculum learning to ensure stable training and enhanced generalization. A viewpoint-based action masking mechanism further refines the action space, accelerating learning convergence. Extensive simulations demonstrate our method consistently surpasses state-of-the-art baselines, achieving up to 14.5% shorter exploration time and 18.8% reduced path length while maintaining low computational overhead. Real-world experiments further validate the practical effectiveness of our approach for robotic exploration tasks.