Work

A Bio-Inspired Geomagnetic Navigation Method Based on Hybrid Reinforcement Learning and Greedy Descent

Year: 2026
Type: article
Abstract: Geomagnetic navigation offers geospatial positioning in GNSS-denied environments, using the global geomagnetic reference field as a continuously available reference. Traditional map-free bio-inspired navigation, however, relies on random search, which limits experience accumulation and search efficiency. To overcome this, this paper proposes a hybrid reinforcement learning and greedy descent (RL-GD) method. A state–action value table (Q-table) is trained offline in a simulation environment with randomly sampled start and target points using the 14th generation International Geomagnetic Reference Field (IGRF-14) model. In the online navigation stage, Q-table-based heading suggestions are fused with local trial candidates to form hybrid decisions that minimize normalized geomagnetic costs, while pure greedy descent is automatically activated for untrained states. In 50 paired Monte Carlo trials, RL-GD reduces the mean number of iterations from 3572 (EAS) to 1949 (a 45.4% reduction), lowers the path ratio from 2.15 to 1.16, and decreases the terminal positioning error from 0.078% to 0.037% of the straight-line distance. Ablation tests showed that removing greedy descent reduced success from 100% to 0% across all 50 tasks, while the Q-table prior reduced iterations by 17.3%. Under measurement noise of 2.5–10 nT, RL-GD maintained a 100% success rate, whereas EAS decreased from 100% (≤5 nT) to 28% (10 nT).
Source: Geomatics
Language: English
Open Access status: gold
Estimated APC paid: $1,207
OpenAlex ID: W7236693750
FWCI:
Cites: 32
Cited by:
Related to: 10