Simulator for training machine vision racing AIs, created with Unity.
Clone this repo and open it in Unity. Run and enjoy!
The point is to connect your AI to localhost:11000. The simulator will feed you PNG images from the car's front camera. Like this
- two bytes indicating the size of the next image (higher byte first)
- PNG image data
- goto 1
The simulator accepts control commands as a JSON string terminated with \n from the AI:
{ "action": "forward", "value": 0.5}\nwhere value is the throttle value between 0 and 1.{ "action": "turn", "value": -0.1}\nwhere value is the direction for the front wheels. 1.0 is full left and -1.0 is full right.
The RaceParameters.json file can be used to control the operating mode of the Simulator. First of all, mode can be one of development, race or playback for different modes.
In development mode the simulator accepts bot connections and places cars immediately on the track for free practice with lap timing
In race mode the simulator accepts bot connections while in the "race lobby" phase. Then it starts Qualifying and finally the Race. You can move tho the next phase using SPACEBAR.
It also moves automatically forward based on RaceParameters, see self-documenting code in RaceParameters.cs.
In playback mode the simulators plays back a recorded race log. When running in WebGL, the playback mode is always assumed. See below.
You can build a WebGL build with Unity. It results into an index.html file containing the player. You need to server it over HTTP though - file protocol is not supported.
At the moment it always looks for race data at http://localhost:8000/race-capture.json and this should be parameterized soon.
When running a race, the simulator saves race data to race.log. If you serve the resulting file over HTTP to the player, you should see a race.
The simulator serves a full capture of race events on port 11001 as a newline-separated JSON stream. This could be used for watching the race live.