At a GDC panel attended by GamesRadar+, Arc Raiders machine learning research lead Martin Singh-Blom explained the systems behind the adaptive locomotion that fuels much of the Arc's behavior. Among the machine rewards and references, behind the directives and neural net training, a more fantastical-sounding element caught my attention: a "magic" system that serves as a quiet stop-gap to help the Arc do what they need to do, namely pursuing and ultimately killing players.
The "magic" of Arc Raiders exists alongside physics in the reward scheme for the Arc machine learning. This reward system essentially compels AI to do certain things, like orienting or moving toward players, in their locomotion. Completing such tasks increases the reward value, which is what the whole system is set up for. Combined with standard video game AI behavior patterns, this machine learning leads to more dynamic enemies capable of tackling a wide range of hard-to-predict situations, like Leapers navigating barricades players have set up.
Magic comes in when physics ain't enough. As Singh-Blom explained, the Arc have access to otherwise inexplicable movement options which are very literally called "magic" under the hood. This is often magic torque – a means to move in a way that the physics of the game normally would not allow.
The clearest and perhaps most common example of magic in action comes from the Arc destruction system. Machine learning is really only used for the Arc with multiple legs in the first place, and when you sever multiple legs from the likes of Leapers or Bastions, they find it understandably hard to move. After all, if you severed one of my legs with an Anvil, I would also probably find it to be an inconvenience.
Missing a limb and struggling to get around as they normally would, the Arc can, like wizards resorting to dark sorcery in times of war, turn to magic. With a little bit of magic torque – but, within the rules of the system, not so much that it looks obviously unnatural – a de-limbed Arc can still move relatively normally. In his presentation, Singh-Blom says this was something of a breakthrough for Embark. The team realized they're making a video game, so as long as it looks correct, who cares how it happens? Bring on the magic.
That said, within the same machine learning reward system, there are actually penalties for using too much magic. Since the Arc always want to increase the overall reward value, they generally only turn on the magic when there's no other option. It's one of those systems that, like an excellent user interface, works best when you don't really notice it, and it's a fascinating edge case showing the invisible hand of Embark at work.
