In one embodiment, a method includes accessing global positioning system (GPS) data indicating a raw location associated with the computing system; accessing, based on the raw location, environmental model data including one or more structures; generating GPS correction data by processing the GPS data and the environmental model data using a machine-learning (ML) model that has been trained to compensate inaccurate GPS readings due to environmental interference; and determining an accurate location by correcting the raw location using the GPS correction data.