To efficiently interact with our environment, we need to observe, understand and act upon actions performed by other people. According to embodied accounts of cognition, understanding the meaning of actions relies on similar representations that are recruited when we perform the same actions ourselves. In this talk I will present a series of functional magnetic resonance imaging (fMRI) experiments investigating (a) to which degree representations of actions are shared across modalities, and (b) whether shared representations for observed and executed actions are recruited during action understanding. I will discuss the theoretical implications of the results in the framework of embodied cognition as well as future directions.