Information guided Levy flight for robot search in unknown environments

Abstract

Efficient robotic search in unknown and noisy environments remains a fundamental challenge due to sparse targets and limited sensing resources. This paper proposes an Information-Guided Lévy Flight (IGL) framework that unifies sparse-grid environmental modelling, probabilistic prediction via Variational Bayesian Gaussian Mixture Models (VBGMM), an entropy and mutual information adaptive switching mechanism, and a connected-region guidance strategy. These modules collectively balance global exploration and local exploitation, enabling real-time, uncertainty-aware decision making. Comprehensive ablation and parameter studies demonstrate that probabilistic prediction and connected-region guidance significantly enhance reliability and convergence speed, while clarifying the trade-off between predictive accuracy and computational efficiency. In extensive 100 $$\times $$ 100 unknown environment simulations, the IGL achieves a 75% success rate, the highest overall performance for static targets, and an 82% success rate for dynamic targets, surpassing six representative baselines in terms of path length and iteration count. Despite smaller coverage ratios, IGL maintains targeted and stable exploration, reflecting strong adaptability under uncertainty. The proposed framework establishes a generalisable information-driven paradigm for single-robot search and provides a foundation for scalable multi-robot cooperation and high-dimensional autonomous exploration.