As a worker-centred assembly mode developed in the electronics industry in Japan, seru is receiving growing attention due to improved flexibility and responsiveness. In seru implementation, cross-trained worker assignment is a vital problem. Most previous studies focused on assigning cross-trained workers into pure divisional or rotating seru separately, but overlooked the problem for a hybrid seru production system that includes both seru types. This research fills this gap by minimizing the makespan and balancing the workers’ workload of each seru in a bi-objective mathematical model. For medium-scale instances, the exact solutions are obtained. For large-scale instances, we propose an NSGA-II-based memetic algorithm that uses two-level encoding and incorporates the bat algorithm as a local search and two K-means-based NSGA-II algorithms. The experimental results illustrate that the K-means-based NSGA-II not only outperforms other algorithms with respect to common proximity and diversity metrics but also runs an order of magnitude faster (in seconds versus minutes required by others on the same computer). Some management insights are obtained based on many numerical experiments.