Project name: Meta Motivo
Project function: Human motion generation model
Project Description: A first-of-its-kind behavioral basic model designed to control humanoid agents based on virtual physics and capable of performing a variety of whole-body tasks.
Ability to handle unseen tasks such as motion tracking, posture reaching and reward optimization while testing without additional learning or fine-tuning.
Meta Motivo is a behavioral basic model developed by Meta’s FAIR (Facebook AI Research) team that aims to control virtual physics-based humanoid agents so that they can perform a variety of whole-body tasks.
Main functions:
Pre-trained models: Six pre-trained FB-CPR models are provided to control the humanoid models defined in HumEnv.
Evaluation Script: Provides a fully reproducible script for evaluating model performance in HumEnv.
Training Code: Provides complete code for training FB-CPR in HumEnv, as well as FB training code for faster experiments in DMC.
Installation method:
The project can be installed via pip and requires Python version 3.10 and above.
pip install "metamotivo[huggingface,humenv] @ git+https://github.com/facebookresearch/metamotivo.git"Optional dependencies include humenv[“bench”] and huggingface_hub for testing, training, and loading models from HuggingFace.
Pre-trained model:
To ensure the reproducibility of the results, five models (metativo-S-X) for the paper’s results are provided, each model being trained using a different random seed. In addition, a best-performing model (metamotivo-M-1) is provided that can be used for interactive testing.
For more details, please refer to the project’s README.md file.
GitHub：https://github.com/facebookresearch/metamotivo
Oil tubing: