Roadmap for Physical Intelligence Aims for “Embodied Brain”

Researchers propose a roadmap for achieving physical intelligence in artificial general intelligence, focusing on an “embodied brain” model. They identify fragmentation across action models, vision-language-action policies, and world models as hindering progress. These models currently utilize incompatible action spaces, prediction targets, datasets, and tasks, and lack standardized evaluation methods.

The proposed roadmap centers on building an “embodied brain” capable of integrating multimodal context and comparing potential actions before issuing commands. World Action Models (WAMs) are seen as prototypes for the predictive functions of this brain.

A crucial element is a “physical harness” which would ground model outputs through tools and controllers, allowing for verification and data logging. Shared contracts between heterogeneous models, data, tasks, and robotic embodiments aim to create a modular and adaptive system capable of self-improvement through verified interactions.