The median-of-means is an estimator of the mean of a random variable that has emerged as an efficient and flexible tool to design robust learning algorithms with optimal theoretical guarantees. However, its use for the clustering task suggests dividing the dataset into blocks, which may provoke the disappearance of some clusters in some blocks and lead to bad performances. To overcome this difficulty, a procedure termed “bootstrap median-of-means” is proposed, where the blocks are generated with a replacement in the dataset. Considering the estimation of the mean of a random variable, the bootstrap median-of-means has a better breakdown point than the median-of-means if enough blocks are generated. A clustering algorithm called K-bMOM is designed, by performing Lloyd-type iterations together with the use of the bootstrap median-of-means strategy. Good performances are obtained on simulated and real-world datasets for color quantization and an emphasis is put on the benefits of our robust intialization procedure. On the theoretical side, K-bMOM is also proven to have a non-trivial probabilistic breakdown point in well-clusterizable situations.