AI Makes Some Progress In Quake III
The DeepMind AI project beat the world's top Go players a few years ago, which was quite an impressive result compared to past performance in the board game. Now they're taking first steps into Quake. Specifically they are playing Quake III CTF, and doing quite well.
They use randomly generated maps of two types. One are flat maze maps and others are outdoor maps with elevation changes. Neither have true verticality, but they still vary the number of players and team composition between matches. So, apparently these AIs can learn to play well in human teams too.
The inputs the AIs receive are the raw RGB values for each display frame, and some general info about the current state of the game, such as the score. They are essentially "looking at the screen," as much as that makes sense when talking about AI.
It's quite an impressive result. It would be interesting to see them play against pro teams on a wider variety of maps, but I don't know if that's likely to happen.
Here's a popular article and a paper about the research.