Embodied Reasoning

Embodied reasoning is a robot's ability to look at what posture its body is currently in and what's in the space in front of it, and decide on its own what order of actions to take next.

Think of the relationship between a navigation app and a driver. The navigation app is the one that decides "turn left next," while the driver is the one who actually turns the wheel and presses the pedals. Embodied reasoning plays the navigation app's role. It talks with people, looks around a room, and plans multi-step tasks that take several minutes. Actually moving the body to pick things up and walk is handled by a separate model.

This ability becomes especially important when things don't go according to plan. If one step fails during a long task that requires hundreds of decisions, instead of starting over from scratch, the robot goes back to that point and fixes it on its own before continuing. When multiple robots split up a single job, like cleaning, deciding the order of tasks and coordinating so they don't collide with each other is also part of this domain.

In articles, you'll see something like "Gemini Robotics-ER 2 is an embodied reasoning model that handles a robot's higher-level judgment." A common misunderstanding is thinking this model directly moves the robot's arms and legs, but the actual motion is handled by a separate vision-language-action model, while the embodied reasoning model only handles planning and judgment.