Meta Motivo

Meta's simulated-humanoid behavior foundation model that does motion tracking, reaching a target pose, and reward-driven behavior with no further training.

Meta Motivo was released by Meta FAIR in December 2024, with the paper published at ICLR 2025; the company calls it the first behavior foundation model for humanoids. It controls a virtual humanoid character in physics simulation (the HumEnv environment), not a real robot. Its core algorithm, FB-CPR, builds on forward-backward representations, an unsupervised reinforcement-learning method that encodes state, reward, and policy into the same latent space; a discriminator is added on top, constraining the policy to stay close to unlabeled motion-capture data, so the learned motion looks human while still generalizing. Once pretrained, giving it a reference motion, a target pose, or a reward function directly yields the corresponding policy, with no additional training or planning needed; it is also somewhat robust to changes like gravity, wind, and external shoves. Code and several models, ranging from 24.5 million to 288 million parameters, are all open-source.

ExampleFeed the same pretrained model a motion-capture reference clip, a target standing pose, or a “spin in place” reward function, and it can directly control the simulated humanoid to do each one, with no retraining needed per task.

Also called

FB-CPR, Forward-Backward Representations with Conditional-Policy Regularization, Zero-Shot Whole-Body Humanoid Control via Behavioral Foundation Models

Related

Behavior Foundation Model · Unsupervised Skill Discovery · Motion Tracking · Zero-shot · BFM-Zero · Adversarial Motion Priors

Sources

Zero-Shot Whole-Body Humanoid Control via Behavioral Foundation Models (arXiv 2504.11054)
Meta FAIR: Sharing new research, models, and datasets (2024-12-12)
facebookresearch/metamotivo (GitHub)

As of

2025-04