Hacker News new | comments | show | ask | jobs | submit login

If GTA's physics model is realistic enough, you could train a model in that world "safely", then use the model with transfer learning (perhaps with another dataset from the "real world" - either homegrown or pre-existing) to "fine tune" to the real world. At least, that's one possibility.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact