Source-free domain adaptive object detection (SFOD) focuses on adapting the detector trained on source data to the unlabeled target domain without access to the source data. In this work, we propose a multi-expert progressive learning framework that incrementally acquires target domain knowledge to reduce the influence of noisy pseudo labels and thus alleviate performance degradation commonly encountered in the later stages of training. Specifically, we construct an expert module composed of multiple expert units. We further introduce a progressive learning strategy where each expert is guided by the predictions of its predecessor, leveraging carefully chosen pseudo label thresholds. Additionally, we explore the impact of different non-maximum suppression (NMS) threshold settings during pseudo label generation on the effectiveness of our progressive learning scheme. By employing this multi-expert progressive learning approach, our detector achieves improved performance while mitigating noise-induced degradation throughout training. Extensive experiments conducted across three distinct scenarios validate the effectiveness of our method.