The execution of actions in the real world involves physical operations
on physical objects. In order to perform these operations, an embedded
agent needs to anchor its internal representations of these
objects to physical elements in the environment. We sketch a
preliminary account of the anchoring problem where the object of an
action is represented by a description, in terms of perceptual
properties, of the class of physical objects that are ``the right ones''
for that action. We illustrate our account by showing how it has been
used to help autonomous navigation of a mobile robot.