Human limits
The AI sees the game 150 ms late, about a fast player's reaction, and makes about 80 actions a minute. It must win by its decisions, not by speed.
It learns from human replays, then plays against itself. When it makes progress, a new checkpoint goes up here: its wins with each race, against itself and Computer Insane (which will be dropped once the AI wins consistently against it). It is currently learning from an older patch, due to the replay catalog being larger. Once the training architecture is on point, we will add more recent patches, replays, and maps.
The checkpoints, the newest first.
Loading the checkpoints…
The AI sees the game 150 ms late, about a fast player's reaction, and makes about 80 actions a minute. It must win by its decisions, not by speed.
Watch right in your browser — move the camera, change the speed, see each player's view.