Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for machine learning for home understanding and notification. In one aspect, a method includes obtaining reference videos from a camera within a premises of a home, determining, from the reference videos, timing of actions in a routine that a particular person performs before leaving the home, determining from a sample video from the camera within the home that the particular person appears to be out of sync in performing a particular action based on the timing of actions in the routine determined from the reference videos, and in response, providing a notification to the particular person.