This work investigates the two major mechanisms of human behavior control: habitual and goal‑directed. Existing human‑like agents mainly model goal‑directed actions and largely ignore habitual behavior, which is essential in everyday life. We propose a framework inspired by human control systems, consisting of a Habitual Controller, a Goal‑directed Controller, and an Arbiter. The Habitual Controller retrieves cue‑triggered actions from a personalized habit memory; the Goal‑directed Controller uses a context‑aware world model to predict action consequences and estimate their values; the Arbiter dynamically balances the influence of both systems based on individual differences and momentary internal states. To generate diverse human‑level instructions in 3D environments, we also develop a keyframe‑guided 3D motion generation module. Extensive experiments, human studies, and ablation analyses demonstrate that incorporating habitual behavior and coordinating multiple control systems markedly improves the human‑likeness of embodied agents.
Review