Ghost’s entire approach is based on an axiom that the human driver is fundamentally correct. It begins by collecting mass amounts of video data from kits that are installed on the cars of high-mileage drivers. Ghost then uses models to figure out what’s going on in the scene and combines that with other data, including how the person is driving by measuring the actions they take.
It doesn’t take long or much data to model ordinary driving, actions like staying in a lane, braking and changing lanes on a highway. But that doesn’t “solve” self-driving on highways because the hard part is how to build a driver that can handle the odd occurrences, such as swerving, or correct for those bad behaviors.
Ghost’s system uses machine learning to find more interesting scenarios in the reams of data it collects and builds training models based on them.
Another article that says mostly the same things but also a few different things: