Improving Rao-Blackwellised genetic algorithmic filter SLAM through genetic learning

Abstract

A Rao-Blackwellized particle filter (RBPF) approach is an effective means to estimate the full SLAM posterior. To improve the memory efficiency of RBPF-SLAM, our previous work proposed a SLAM framework based on Rao-Blackwellised particle filters (RBPF) and genetic algorithms (GA) for recovering the full SLAM posterior using raw exteroceptive sensor measurements, i.e. without landmarks. The resultant Rao-Blackwellised genetic algorithmic filter SLAM (RBGAF-SLAM) permits the efficient use of any arbitrary measurement model. However, the drawback is that the GA operators have to be determined heuristically and the parameters have to be tuned through try-and-error. These tuning operations are time consuming and the optimized solution is not guaranteed. In this paper, we provide the detailed description of how to optimize the RBGAF-SLAM operators through genetic learning approach. GA operators and parameters are encoded into chromosomes, while the Euclidian localization error is chosen to build the fitness function. This approach guarantees the consistent SLAM results and ensures the global convergence to the optimal solution of RBGAF-SLAM. Simulations and experimental data are used as training data. A robot with laser range scanner is used to demonstrate the effectiveness in actual implementations.