EEFOLLM: LLM-guided reward shaping for electric eel foraging optimization in robot path planning

Mobile robot path planning is one of the core challenges in autonomous systems. Existing metaheuristic algorithms commonly adopt fixed weights in multi-objective fitness aggregation, which can neither perceive the structural complexity of the map nor adaptively adjust with the progression of search stages, causing convergence to infeasible paths on complex maps or sacrificing path efficiency on simple ones. To address these limitations, this paper proposes EEFOLLM—a novel robot path planning method that embeds a Large Language Model (LLM) as an offline reward-weight configurator within the Electric Eel Foraging Optimization (EEFO) framework. EEFOLLM first extracts a nine-dimensional structural feature vector from the grid map (including obstacle density, corridor width, and clutter score), then drives a locally deployed Qwen2.5-3B-Instruct model to generate three-stage, four-element fitness weights (path length, collision penalty, smoothness, and turning penalty) tailored to each map. A dual-safeguard pipeline—comprising Python-side clipping and normalisation together with MATLAB-side structural validation—guarantees experimental robustness under extreme conditions. A stage scheduler switches the active weight set according to normalised iteration progress, driving EEFO through a guide–perturb–restart update mechanism to evolve the population. Comparative experiments on five benchmark maps spanning $$40{\times }40$$ to $$70{\times }70$$ scales with obstacle densities from 0.078 to 0.185, against nine state-of-the-art metaheuristic algorithms published between 2019 and 2024 (1,000 independent optimisation runs in total), demonstrate that EEFOLLM ranks first on all five maps, achieving a cross-map average Friedman rank of 1.26, significantly outperforming the runner-up HHO (3.96). A five-variant ablation study further reveals that Gaussian perturbation is the core exploration mechanism of EEFO, and that search-kernel integrity and LLM-generated stage weights act as complementary rather than strictly ordered factors, with their synergy yielding the most robust performance across diverse map structures. Extensive supplementary analyses further quantify the LLM-inference overhead, the post-hoc statistical significance under Holm correction, a controlled weight-strategy comparison, and the sensitivity of the framework to stage-boundary, waypoint-number, and large-scale map settings. This work provides one of the first systematic empirical investigations of the “LLM-enhanced metaheuristic” paradigm in the domain of static binary-grid robot path planning, showcasing the practical value and generalisation potential of lightweight LLMs in offline reward shaping.