Learning a repertoire of actions with deep neural networks

We address the problem of endowing a robot with the capability to learn a repertoire of actions using as little prior knowledge as possible. Taking a handwriting task as an example, we apply the deep learning paradigm to build a network which uses a high-level representation of digits to generate sequences of commands, directly fed to a low-level control loop. Discrete variables are used to discriminate different digits, while continuous variables parametrize each digit. We show that the proposed network is able to generalize learned actions to new contexts. The network is tested on trajectories recorded on the iCub humanoid robot.