Hacker News Reader

Comment by captainarab | Hacker News Reader

This is incredible! Huge fan of GTA V / Rockstar as well as Open.AI

Will be super interesting to see how all of the randomness in GTA V driving influences autonomous driving AIs.

I could see them being a bit more "cautious" in order to deal with all of the crazy drivers (and events) sharing the roads of Los Santos.

View on HN

Thanks! Yes, the game is highly scriptable[1], so you can test the types of scenarios you'd want. We've disabled violence against the driver, but could enable[1] it on-demand if it's useful for training.

[1] http://www.dev-c.com/nativedb/ [2] http://gtaforums.com/topic/822314-guide-driving-styles/

What happens when a car hits a pedestrian at speed? Is it still a bloody mess as in the normal game?

I'd be super interested in seeing how this reacts to having a single star wanted level. At that level police cars will come to a stop almost perpendicular in front of your car. Will the AI be able to avoid an aggressive obstacle.

Also it'd be interesting to see how it reacts to getting shot at

I'm missing (and the article doesn't exactly state) what the fitness function is here.

It might just as well don't care about being shot at.

We currently don't look at being shot at (although it's an interesting suggestion!) - the current reward function is basically this:

  distance_reward - off_road_penalty - speeding_penalty - slow_penalty - discomfort_penalty;

Plus if you collide or drive against traffic, the episode terminates.

https://github.com/openai/universe-windows-envs/blob/f5aad96...

Well that part was mostly a bit of a joke - It could add in a few interesting variables such as deflated tyres causing the car to drift more than the AI is expecting.

Reply on news.ycombinator.com