Convolutional neural networks (CNNs) are commonly used to explore the neurocomputational bases of object recognition. While they provide the best current models of the human visual system, CNNs often fail where humans do not. New research suggests this may be because of the way CNNs are trained. Historically, CNNs have been trained on entirely clear images despite the fact that much of what we see is blurry or processed at low spatial resolution. The key to their improvement may therefore lie in the incorporation of blurry training inputs. To test this hypothesis, researchers compared the performance of standard and blur-trained CNNs for a diverse set of neural, visual, and behavioral benchmarks. The blur-trained CNNs outperformed standard CNNs at predicting neural responses to objects across various viewing conditions and showed increased sensitivity to shape information and greater robustness to visual noise. These results suggest that blur-trained CNNs are better aligned with human visual systems and provide multifaceted neurocomputational evidence that blurry visual experiences confer robustness in object recognition.