Waymo patents method to create fake road scenarios for self-driving training

1. A system for generating and utilizing synthetic scenes, comprising: a memory device; and a processing device, operatively coupled to the memory device, to: receive, at the system, a set of input data including a roadgraph, the roadgraph representing drivable paths and object configurations associated with a segment of road driven and imaged by a first autonomous vehicle, and the drivable paths comprising an autonomous vehicle driving path of the first autonomous vehicle, wherein the autonomous vehicle driving path comprises a plurality of path connection nodes disposed on the autonomous vehicle driving path; modify the roadgraph by adjusting a trajectory of the autonomous vehicle driving path, wherein adjusting the trajectory of the autonomous vehicle driving path comprises: generating between a first pair of path connection nodes of the plurality of path connection nodes, a first synthetic path segment to address a disabled path interval of a path of the roadgraph, the disabled path interval being bounded by the pair of path connection nodes; and generating between a second pair of path connection nodes of the plurality of path connection nodes, a second synthetic path segment connecting the autonomous vehicle driving path to a neighboring drivable path of the drivable paths; place a set of artifacts along one or more lane boundaries of the modified roadgraph comprising the first synthetic path segment and the second synthetic path segment to generate a synthetic scene, wherein placing the set of artifacts comprises placing candidate artifacts along one or more lane boundaries of the first synthetic path segment, one or more lane boundaries of the second synthetic path segment, and one or more lane boundaries of the disabled path interval; train a machine learning model to navigate a second autonomous vehicle based on the synthetic scene; and cause a steering mechanism of the second autonomous vehicle to perform a steering maneuver.