Hey everyone! Wanted to show off a little side project I've been working on. A lot of people have been using AI to one-shot games, but they're often quite buggy - glitchy graphics, collisions aren't setup, or regressions in earlier levels. With an AI playtester in the loop, I've been able to one-shot games like Candy Crush or Crossy Road that would previously have bugs like cut-off UIs or stuck retries.
The way it works basically goes like this:
1/ The playtester starts by reading the code to understand what its parent AI is asking it to play, and to decide what breakpoints, hooks and logging to add 2/ The playtester sets up the scene with the needed environment 3/ The playtester sends a sequence of inputs, then s the game 4/ The playtester then uses a vision model to understand the scene, queries the hooks it initially set up, and reads the logs to decide what to do next
After doing a few input loops, it then determines if the game is working, the game is broken, or if the test failed.
It is a paid offering so I'm not expecting anyone to try it out, but I wanted to share it nonetheless ^^. Lmk if you have other ideas on how you think it can be improved
Comments URL: [https://news.ycombinator.com/item?id=49475316](https://news.ycombinator.com/item?id=49475316)
Points: 1