Goal-directed occupancy prediction for autonomous driving

An autonomous vehicle, comprising:

• a sensor system configured to generate sensor data of an environment external to the autonomous vehicle;
• one or more processors;
• and
• one or more non-transitory computer-readable media that store:
• a machine-learned model configured to estimate occupancies for candidate paths that a detected object may follow in the environment based at least in part on spatial relationships between lanes of a road network associated with the environment;
• and
• instructions for execution by the one or more processors that cause the one or more processors to perform operations, the operations comprising:
• inputting, to the machine-learned model, state data associated with the object and map data including information associated with spatial relationships between at least a subset of lanes of the road network associated with the environment;
• determining, by the machine-learned model in response to the state data and the map data, a candidate path that the object may follow in the environment, wherein the candidate path includes a discretized sequence of spatial cells along the candidate path;
• determining, by the machine-learned model, a predicted occupancy for each spatial cell of the discretized sequence of spatial cells for the candidate path;
• and
• generating prediction data associated with the object based at least in part on the predicted occupancy for each spatial cell of the discretized sequence of spatial cells for the candidate path.

One or more non-transitory computer-readable media that store instructions for execution by one or more processors that cause the one or more processors to perform operations, the operations comprising:

• inputting, to a machine-learned model, state data associated with an object and map data including information associated with spatial relationships between at least a subset of lanes of a road network associated with an environment external to an autonomous vehicle;
• determining, by the machine-learned model in response to the state data and the map data, a candidate path that the object may follow in the environment, wherein the candidate path includes a discretized sequence of spatial cells along the candidate path;
• determining, by the machine-learned model, a predicted occupancy for the respective spatial cells of the discretized sequence of spatial cells for the candidate path;
• and
• generating prediction data associated with the object based at least in part on the predicted occupancy for each spatial cell of the discretized sequence of spatial cells for the candidate path.