this topic is based on lg-sim, the advantage is plenty Unity3D resources. to drive planning design, simulation has to keep as real-world env as possible. for L3 and below, the high-fidelity is actually most about HD map, since in L3 ADS, sensors only cares about the road parternors on road, and planning module take hd map as input.
so there is a need to build simulation envs based on hd map.
osm
in the Unity engine, all in the env(e.g. traffic light, road lanes, lane mark etc) are GameObjects. to describe a map info, one of most common used map format is OSM (another is opendrive, but no open parser in unity yet), however which is not born to used in hd map, but still is very flexible to extend to descripe all info, a hd map requires. e.g. lane id, lane width, lane mark type e.t.c
the open-source tool OsmImporter is good enough to parse osm meta data into Unity GameObjects. for different map vendors, first to transfer their map format into the extended-osm format. and then can generate all map info related gameObjects in Unity env. that’s the main idea to create real-world env in simualtor.
the end is to make a toolchain from map vendor input to Unity 3D env.
waypoints
in either lg-sim or carla, the npc vehicles in self-driving mode, is actually controlled by following waypoints in the simulator, and waypoints is generated during creating map in step above.
carla has plenty client APIs to manage the waypoints, and then drive npcs.
hd-map tool
by default, both lg-sim and carla has a tool to create hd-map, that’s basically mark waypoints on the road in an existing map, which is not strong. carla later support Vector one to build in map more efficiently.
L3+ virtual env generator
there are plenty teams working on building/rendering virtual envs directly from sensor data, e.g Lidar cloud point or camera image, and plenty image-AI techs here, which of course gives better immersed experince. and the test task is mostly for perception, data-fusion modules, which is heavier in L3+