Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation, combining visual perception, language understanding, and action planning in a unified framework. This work applies mechanistic interpretability techniques using Sparse Autoencoders (SAEs) to analyze hidden layer activations in VLA models, revealing sparse dictionary features that provide interpretable bases for model computation. The research discovers that most SAE features correspond to memorized sequences from training demonstrations, while some features represent interpretable, generalizable motion primitives and semantic properties. This analysis offers insights into VLA model generalizability and provides a framework for steering model behavior through identified interpretable features, advancing the understanding of embodied AI systems for robot manipulation tasks.