timcsy / funaikart Goto Github PK

Currently, an episode does not terminate when the agent completes a race. The training process is just stuck in the result screen. (e.g. terminal_steps is always empty)

There isn't any observable signals that could determine that the agent has finished a lap, either, so the agent can't stop on its own.

We would appreciate it if this issue could be addressed soon, so that our agents can actually optimize finishing time rather than just learning how to drive safely (ending an episode when it gets too close to a wall, etc.)

Hello, I'm from the NTHU team, trying to train RL models over the karting game.

Exporting a headless build would make training a lot easier for us, using Google Colab. We would highly appreciate it if you could update the build after every new change made to the project.

Basically, we're asking for Cloud Build to be set up here.

This will reduce the need for us to tamper with the state of the project (e.g. changing Player/Build settings), and will make iteration faster on our end.

Hi Tim,

Some issues happened when I executed the command with "python ml.py",
the game seems not to be functioning on the display window

I've wondered if I am installing it incorrectly or what, and how do I fix this problem?

Thanks,
Jashen