Arc Raiders' AI uses a hidden 'magic' system

At a GDC panel, Arc Raiders machine learning research lead Martin Singh-Blom explained the systems behind the adaptive locomotion that fuels much of the Arc's behavior. Behind the directives and neural net training, a more fantastical-sounding element stood out: a "magic" system that acts as a quiet stop-gap to help the Arc pursue and ultimately kill players.

The "magic" exists alongside physics in the reward scheme for the Arc machine learning. That reward system encourages actions such as orienting or moving toward players; completing those tasks increases the reward value. Combined with standard video game AI behavior, this machine learning produces more dynamic enemies that can handle hard-to-predict situations, like Leapers navigating player-constructed barricades.

Magic is used when physics isn't enough. Singh-Blom said the Arc have access to otherwise inexplicable movement options—literally called "magic" under the hood—often implemented as "magic torque" that allows motion the game's physics would not normally permit.

arc raiders, magic torque, machine learning, adaptive locomotion, reward system, neural net, gdc panel, martin singh-blom, leapers, player barricades