The playground experiment demonstrates curiosity-driven learning on an autonomous four-legged AIBO robot platform. The robot is equipped with basic motor primitives (control of its head direction, arm and mouth movements) which are controlled by a generic developmental engine, which we call Intelligent Adaptive Curiosity. With this engine, the robot actively chooses its actions and its learning situations. The engine can measure the effects of the actions taken on the environment through a camera, IR sensors, touch sensors and motion feedback.

The developmental engine is composed of

(1) prediction systems that learn the effects of actions in a particular context; These prediction systems are set of experts specialized in particular areas of the sensori-motor space;

(2) meta-prediction systems that learn to predict the error of prediction systems (1) and its evolution over time. In particular, these meta-prediction systems can compute an expected error reduction corresponding to a given action. This is done by comparing the error rate of the new expected sensory-motor context to the error rate in similar sensory-motor contexts in the past, as opposed to the error rate in the most recent sensory-motor context;

(3) an action selection module that chooses actions with maximum expected error reduction (as computed by (2));

As a consequence, this system produces action sequences that are expected to maximize learning progress.

Such a self-motivated robot focuses on tasks with are neither too predictable nor too difficult to predict. It looks for “progress niches”, sensory-motor situations which are optimal for learning given its embodiment (physical body and learning algorithms), the structure of its environment and its current stage of development. This ensures a continuous development of increasingly complex behaviors.