A couple of weeks ago I wrote about the need for a good robocar driving simulator. Others have been on the track even earlier and are arranging a pair of robotic driving contests in simulator for some upcoming AI conferences.
The main contest is a conventional car race. It will be done in the TORCS simulator I spoke of, where people have been building robot algorithms to control race cars for some time, though not usually academic AI researchers. In addition, they’re adding a demolition derby which should be a lot of fun, though not exactly the code you want to write for safety.
This is, however, not the simulator contest I wrote about. The robots people write for use in computer racing simulators are given a pre-distilled view of the world. They learn exactly where the vehicle is, where the road edges are and where other cars are, without error. Their only concern is to drive based on the road and the physics of their vehicle and the track, and not hit things — or in the case of the derby, to deliberately hit things.
The TORCS engine is a good one, but is currently wired to do only an oval racetrack, and the maintainers, I am told, are not interested in having it support more complex street patterns.
While simulation in an environment where all the sensing problems are solved is a good start, a true robocar simulation needs simulated sensors — cameras, LIDAR, radar, GPS and the works — and then software that takes that and tries to turn it into a map of where the road is and where the vehicles and other things are. Navigation is also an important thing to work out. I will try to attend the Portland version of this conference to see this contest, however, as it should be good fun and generate interest.