Developing grounded goals through instant replay learning

This paper describes and tests a developmental architecture that enables a robot to explore its world, to find and remember interesting states, to associate these states with grounded goal representations, and to generate action sequences so that it can re-visit these states of interest. The model is composed of feed-forward neural networks that learn to make predictions at two levels through a dual mechanism of motor babbling for discovering the interesting goal states and instant replay learning for developing the grounded goal representations. We compare the performance of the model with grounded goal representations versus random goal representations, and find that it is significantly better at re-visiting the goal states when using grounded goal representations.