We address the problem of continuously observing and forecasting long-term semantic activities of a first-person camera wearer: what the person will do, where they will go, and what goal they are seeking. In contrast to prior work in trajectory forecasting and short-term activity forec