Researchers have built a framework that lets users steer how an AI agent behaves — not just what it achieves — while a task is already running.
The system combines a technique called universal value function approximators with curated training scenarios, learning algorithms, and data augmentation. The result is an AI that can absorb style instructions at run time — think "drive more aggressively" or "fight defensively" — without abandoning its core objective. The team tested this in two commercial AAA titles, Horizon Forbidden West and Gran Turismo, as well as an open-source humanoid locomotion environment. Across car racing, stylized combat, and bipedal walking, agents held to the requested style while still completing the underlying task.
Most reinforcement learning research chases a single near-optimal policy and calls it done. This work treats behavioral style as a first-class variable, which matters for any application where a human needs ongoing influence over an AI's decisions — games, robotics, or assistive systems where one-size-fits-all behavior is a liability rather than an asset. The AAA game validation is notable: these are complex, commercially shipped environments, not toy benchmarks.
The paper arrives as game studios and robotics labs race to make AI feel less like a vending machine and more like a collaborator — though shipping this kind of run-time coaching in a consumer product remains a different challenge than demonstrating it in a lab paper.