Planning to achieve a goal requires knowledge of objects, actions, preconditions, and consequences. These abstract concepts are at a much higher level than the ''pixel-level'' sensory and motor interfaces between an embodied robot and the continuous world. Our goal is to show how high-level concepts of object and action can be learned autonomously from experience with low-level sensorimotor interaction.
We hypothesize that these concepts are part of a larger package of foundational concepts that can be learned in approximately the following sequence: using motion to discriminate objects from background; detecting tight, reliable control loops to distinguish self from non-self objects; learning preconditions and consequences of actions applied to objects; identifying ''grasp'' actions that temporarily transform a non-self object to a self object; learning actions and effects that are achievable only with such an object (a tool!).
The learning process depends on representing sensorimotor interaction with the world as a stochastic dynamical system. A ''curiosity'' drive rewards improvements in prediction reliability. Evaluation uses a simulated robot child with two arms, stereo vision, and a tray of blocks and other objects. This research will help robots learn their own high-level concepts, and could provide insights into human learning disabilities.
