CAIR-MMOEA-DL: Contribution-Aware Inter-Robot Reallocation for Multimodal Multi-objective Multi-Robot Task Allocation

Multi-robot task allocation (MRTA) is a fundamental problem in collaborative robot systems, where task allocation and route planning are tightly coupled. In static, offline MRTA problems, relying on a single assignment scheme can limit the flexibility of task allocation. Therefore, it is necessary to obtain high-quality Pareto front approximations and multiple alternative task allocation schemes. Although MMOEA-DL, a multimodal multi-objective evolutionary algorithm based on deep reinforcement learning (DRL) and large neighborhood search (LNS), addresses this problem effectively, some solutions may still suffer from workload imbalance, with a few robots becoming routing bottlenecks. To address this issue, this paper proposes Contribution-Aware Inter-Robot Reallocation MMOEA-DL (CAIR-MMOEA-DL). The method retains the bilevel structure of MMOEA-DL, where the upper level evolves task allocations and the lower level evaluates routes using a pretrained deep reinforcement learning planner. After the evolutionary search, a contribution-aware inter-robot reallocation strategy with coverage-guided target selection is applied to each solution on the first nondominated front of the final population. It identifies the robot with the longest route, ranks its tasks according to their removal contributions, and attempts to transfer selected tasks to lightly loaded or positive-coverage target robots. Large neighborhood search is then used to refine the task-visiting order. Experiments on 16 simulated MRTA instances with 10 independent runs show that CAIR-MMOEA-DL significantly improves hypervolume (HV) and Pareto-set proximity (PSP) compared with four baseline algorithms. Ablation results further show that the performance gains are accompanied by additional post-search assignment evaluations and longer computation time.