A Shared-Codebase Ablation Study of GA, PSO, and ACO for Grid-Based Robot Path Planning

Abstract

Grid-based robot path planning must balance two objectives: a short path and a smooth one. A mobile robot pays a time and energy cost at every heading change, a cost a pure shortest-path solver ignores. This paper compares a genetic algorithm (GA), particle swarm optimization (PSO), and ant colony optimization (ACO) against an A baseline, using one shared cell-sequence representation, fitness function, and A-based repair routine across all three. We benchmark the four methods on nine grid instances (10×10, 20×20, and 50×50, at obstacle rates from 10% to 50%) over 30 seeds each, reporting path length, turn count, and Wilcoxon signed-rank significance against A on turns. All three EAs match A's optimal length in eight of nine instances: turn reduction, not length reduction, is where they earn their keep. GA lowers mean turns significantly against A in eight of nine instances (cuts of 4.4–11.9% where significant); PSO reaches significance in all nine; ACO in only two, tying rather than losing to A on the two densest large-grid conditions. A follow-up ablation on the same shared codebase isolates which component actually drives GA's advantage: top 5% elitism changes too few seeds per instance to test at all, A-seeded initialization reaches significance in only one of nine instances, and a stagnation-triggered mutation boost combined with population reinjection, proposed in the original design but never implemented, is by far the strongest lever, significant in seven of nine instances; splitting that combined mechanism further shows reinjection, not the mutation-rate increase, is the active ingredient. The same reinjection idea, transplanted onto ACO's ant-construction step, has no measurable effect on the two instances where ACO underperforms, and shows a benefit only on the small, sparse grids where ACO was already winning; a route-diversity argument explains why. We also document where the comparison breaks down: ACO occasionally returns a longer-than-optimal path on the densest small grid, and grid connectivity itself collapses under density, leaving only 11 of 30 seeds solvable on the densest 50×50 map. Runtime scales sharply with grid size for GA and PSO, both of which route every offspring through the same A-based repair step, while ACO stays cheapest by a wide margin.