Deathmatch needs another person. In the late 1990s, even finding an opponent did not guarantee an enjoyable game: the connection could be too poor. At home, Eraser Bot, CRBot, Gladiator, ACE Bot and numerous derivatives came to the rescue. They helped players learn maps, memorise weapon locations and practise their aim. By the time OpenTDM-X appeared, computer opponents in Quake II had a long history of their own.
Seeing a weapon ahead is not enough for a bot: it must find a way there and decide whether the trip is worthwhile. Eraser used route data. For Gladiator, BSPC built an AAS navigation system from the map's geometry. Its documentation described personalities, weapon and item selection, observing bots and chat commands; by 1999, additional game modes were supported too. The downloadable game module concealed a substantial collection of decisions, developed further from release to release.
ACE Bot's authors also wanted behaviour closer to a human's. They described client emulation, movement in three dimensions and learning routes travelled by players. A door, a staircase or a jump was supposed to remain a real obstacle for the bot. Learning from people was already an established idea; years later, new projects would have more recordings, computing power and ways to analyse those observations.
Knowing the way does not win a duel, however. A bot can reach a gun first and die immediately by stopping beneath an upper platform. It can aim well yet waste ammunition at the wrong range. Or it can dutifully follow a player around a loop, giving itself away with every footstep while the player escapes and collects armour. OpenTDM-X would eventually face precisely these problems. Correcting one action was relatively quick; coordinating them so that a bot could play a whole match well proved much harder.
That was why real matches and separate movement recordings were collected for the new bot. They showed what people did after spawning, how they escaped an opponent's control and when they committed to finishing a kill. Jumps had to use the mod's ordinary physics; weapon-switching time and rocket flight could not be shortened for the program's convenience. On the server, code and prepared data made the decisions. A local neural model would enter this story later, as a tool for studying recordings outside a match.
That work was still months away. In March 2026, the future Q2PRO-X first needed to restore the response players expected: press, turn, shoot. That was where it began.
