Advisable Deep Driving


At a glance

We are developing advisable deep models. Such models are capable of acting with full autonomy, but can also modify their behavior in response to textual user input. We prefer the term “advice” to “commands” since the deep model still bears ultimate responsibility for driving safety, and may or may not act according to the advice. We consider advice at different levels of temporal granularity:
1. Instantaneous advice. Instantaneous advice typically comprises turn-by-turn directions from a mapping/planning service such as Google’s Directions API. The vehicle should follow this advice closely, but must also take care to change lanes or make turns safely.
2. Ride-specific advice. Advice such as “drive gently (passenger susceptible to car-sickness”, “please hurry”, “avoid bad neighborhoods” would normally be given by a passenger at the start of a trip.
3. Global advice. Global advice consists of soft condition-action rules such as “slow down when pedestrians are near an intersection”, and would normally be given by a driving instructor during development of the vehicle’s controller.

principal investigatorsresearchersthemes
John Canny