In labour-intensive cellular production systems, it is important to train operators efficiently because productivity depends on operator skills. In our previous study, we proposed a ‘Skill Index’ to classify operator skills based on the time required for each task and used this method to allocate operators with a primary focus on training. The ‘Skill Index’ uses a learning curve to estimate the time required. However, in actual workplaces, operators are expected to accumulate fatigue caused by repetitive tasks, which affects their execution time. Therefore, in this study, a fatigue-learning curve was employed. We proposed an operator allocation that includes fatigue and compared it with computer experiments in our previous study. As a result, it is possible to represent the decrease in proficiency due to fatigue from repetitive tasks and to include fatigue in the design phase. To validate the proposed method, we conducted an experiment to assemble LEGO robots with the operator allocation from the computer experiment and analysed the assembly time for each operator. The results showed that the Skill Index Values were the same for more than half of the tasks in the computer experiment and the assembly experiment, indicating that the proposed method is effective. Efficient operator training is critical in labour-intensive cellular production systems. Our previous study proposed a ‘skill index’ to classify operator skills based on the time required for each task and used this method to allocate operators for focused training. However, fatigue resulting from repetitive tasks affects operator execution time. Therefore, we proposed an operator allocation method that considers fatigue using a new fatigue learning model and compared it with computer experiment in a previous study. As a result, we proposed a new proficiency curve to capture how fatigue from repetitive tasks reduces operator performance. We conducted an assembly experiment using LEGO robots with the operator allocation obtained from the computer experiment. The results showed that the skill index values were consistent for more than half of the tasks in both experiments, indicating the effectiveness of our proposed method.