Although image classification networks such as ResNet have shown remarkable accuracy across classification tasks, they experience significant performance degradation when applied to OOD data such as corrupted images. Furthermore, the large model size inherent to most deep learning architectures prevents their use in applications that place constraints on power availability or compute resources. To address these issues, we propose the use of Neural Cellular Automata (NCA) based feature extraction. NCAs are a recently developed class of lightweight and robust models that iteratively apply an update rule to each cell of an image. We present the Localized Autoencoding Neural Cellular Automata (LANCA) pipeline, a novel approach to training NCAs to perform robust feature extraction by encoding their local environment. We demonstrate that we can successfully employ LANCA on out-of-distribution data such as images of different resolutions. We further demonstrate that LANCA encoders can be leveraged to build a robust classification pipeline that outperforms a ResNet baseline by 4–34% on 7 out of 9 noise and blur corruptions selected from the CIFAR-10-C dataset and shows generally improved robustness on corrupted versions of the Imagenette dataset compared to ResNets of equal model size.