Hybrid optimization of multimodal SLAM for underground robots with seismic vibration geomagnetic and tactile sensing

Abstract

Reliable underground SLAM remains difficult when GPS, visual features, LiDAR, or external anchors are unavailable or degraded, while non-visual sensors exhibit noise, drift, and environment-dependent uncertainty. We propose a hybrid-optimized seismic/vibration-geomagnetic-tactile SLAM framework that combines complementary sensor factors in a robust factor graph and uses staged COA-PSO-GSA optimization for automated parameter calibration. Evaluation covered 23 complete trajectories (10.2 km; approximately 7.5 h) from simulation, controlled underground corridors, and field tunnel/mine environments, with non-overlapping trajectory-level optimization, validation, and testing subsets. On the test trajectories, mean absolute trajectory error was 0.53 m, convergence occurred in approximately 37 iterations, and error growth remained between 1.81% and 4.09% under the evaluated magnetic, vibration, and tactile degradation conditions. Loop-detection AUC was 0.76, but its difference from the strongest GWO-SLAM baseline (0.74) was not statistically significant after Holm correction (adjusted p = 0.184). The measured foreground update latency was 57 ± 10 ms (approximately 17.5 Hz), indicating near-real-time online operation on the tested platform rather than hard real-time performance. These results support technical feasibility under the tested conditions, while broader claims require larger multi-site validation.