Behavior Foundation Model

A large-pretrained humanoid whole-body control model that can adapt to many motion tasks zero-shot or with little tuning.

Behavior foundation model is a term that has emerged over the past two years in humanoid whole-body control (low-level motor control that coordinates the legs, torso, and arms together). The traditional approach trains a separate reinforcement-learning controller for each task; a BFM is instead pretrained first on large amounts of human motion data and a wide variety of tasks, learning reusable base skills and behavioral priors, and then adapts to a new task zero-shot or with minimal tuning through a 'prompt' — a reference motion clip, a target pose, or a reward function — playing a role similar to a foundation model in language. Representative work includes Meta researchers' Meta Motivo in April 2025 (based on unsupervised reinforcement learning with forward-backward representations) and BFM-Zero in November 2025 (deployed on a real Unitree G1, where the same policy can be used for motion tracking, reaching a target pose, and reward optimization). In a hierarchical architecture, it can serve as the lower-level motor controller that takes commands from a higher-level layer.

ExampleBFM-Zero runs a single pretrained policy on the Unitree G1, with no retraining, and can switch by prompt between tracking a human motion clip or walking to a specified pose.

Also called

BFM, Behavioral Foundation Model

Related

Whole-Body Control · Motion Tracking · BFM-Zero · Meta Motivo · Foundation Model · Unsupervised Skill Discovery

Sources

A Survey of Behavior Foundation Model: Next-Generation Whole-Body Control System of Humanoid Robots (arXiv 2506.20487)
BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised RL (arXiv 2511.04131)
arXiv 检索：behavior foundation model humanoid (Chinese)

As of

2025-11