Meta Motivo: An Advanced Model for Humanoid Control | Facebook | WhatsApp Meta AI | Meta AI chatbot | Turtles AI

Meta Motivo represents a breakthrough in humanoid control, enabling pre-trained models to execute complex movements in simulated environments, leveraging unlabeled data. This innovative approach combines unsupervised reinforcement learning and imitation to achieve zero-shot performance on unexplored tasks.

Key Points:

• A new algorithm integrates forward-backward representations with imitation-based conditional policy regularization.
• Meta Motivo controls a virtual humanoid for complex tasks without retraining.
• Evaluations show competitive performance compared to state-of-the-art task-specific methods.
• The released code provides a novel benchmark and access to detailed specifications for further research.

Meta Motivo is at the forefront of unsupervised reinforcement learning research, offering a new avenue for controlling complex humanoid agents. The system is designed to address typical challenges in this field, such as the need for pre-training on a wide range of downstream tasks. Conventional methods often rely on curated datasets or pre-trained policies that are poorly correlated with the final tasks. The proposed algorithm, called Forward-Backward Representations with Conditional-Policy Regularization (FB-CPR), addresses these limitations by integrating unlabeled trajectories into the learning process.

The key technique is the alignment of forward-backward representations with a shared latent space encoding states, rewards, and policies. This approach allows building policies that cover relevant states, maximizing zero-shot generalization on reward- and imitation-based tasks. During pre-training, the model uses unlabeled datasets, leveraging an embedding network to represent complex states and a policy network to translate them into actions. The framework is optimized through direct access to the simulated environment, where 30 million samples and AMASS motion capture data are used for training.

An important contribution of this research is the introduction of a dedicated benchmark to evaluate the model on tasks such as motion tracking, pose attainment, and reward optimization. Meta Motivo stands out for its ability to outperform unsupervised reinforcement learning baselines and model-based approaches, achieving competitive results without any retraining. Performance ranges from 61% to 88% compared to the first-line methods, with particular excellence in reward- and imitation-based tasks.

Experimental results show how the model evolves towards increasingly human-like behaviors during the training process, despite not having been explicitly programmed for specific motions or poses. This confirms the robustness of the FB-CPR method, which allows learning highly adaptive policies using unsupervised data. Applications demonstrate that Meta Motivo excels in contexts requiring flexibility and generalization, while remaining a reference system for further development.

The combination of algorithmic innovation, high performance, and public release of code and benchmarks positions Meta Motivo as a pillar in behavioral foundation model research. This work opens up new possibilities for zero-shot learning and advanced control of humanoid agents.