Research on MAPF has primarily focused on high-level planners and on the comparisons of alternative algorithms in heterogeneous settings, while the role of low-level planners has received little attention. These components are essential for generating agent paths under high-level constraints, and their proper selection can significantly affect performance. This work addresses this gap by comparing four A*-based variants (Uniform Cost Search (UCS), standard A*, Weighted A*, and Adaptive Weighted A*) within the CBS framework. The comparison is carried out through a multi-scenario analysis with varying agent densities and obstacle configurations in grid-based environments. The evaluation includes computational efficiency, conflict resolution capability, and path quality. Results indicate that while weighted planners reduce runtime and node expansions, they generate more conflicts in complex scenarios. Adaptive Weighted A* proves effective overall performance in the empty grid environment, combining high success rates, low computational effort, and moderate conflict levels, even as the number of agents grows. In contrast, in the case of random obstacles, the best performance is achieved by 1.5wA*, although overall planner performances become more comparable, except for UCS, which performs significantly worse. These findings provide actionable insights into the trade-offs involved in low-level planner selection for CBS and pave the way to its application in operational time-based logistic environments.
Ricci et al. (Thu,) studied this question.