Figure 1.

Exemplary sketch of an episodic and compositional memory (ECM). A random walk always starts at a percept clip (e.g., clip 1, clip 2) and ends with a behavior clip (e.g., clip 7, clip 8). A transition c → c′ from clip c to clip c′ is done with the probability $p_{c\to c^{\prime }}$. Intermediate clips (e.g., clip 3, clip 6) denote simulation of previous experiences, i.e., behaviors and percepts. The extensions made in this work enable the robot to create transitions between perceptual states on Layer D (i.e., active learning) and to add create new preparatory behaviors on Layer E.