View article

This paper outlines a project proposal to be submitted to EC H2020 call ICT-29-2018. The purpose of the project is to create a digital Baby (dBaby)-an agent perceiving and interacting with the 3D world and communicating with its Teacher via natural language phrases to achieve the goals set by the Teacher. The novelty of the approach is that neither language nor visual capabilities are hard-coded in dBaby-instead, the Teacher defines a language learning Game grounded in the 3D world, and dBaby learns the language as a byproduct of the reinforcement learning from the raw pixels and character strings while maximizing the rewards in the Game. So far such approach successfully has been demonstrated only in the virtual 3D world with pre-programmed Games where it requires millions of episodes to learn a dozen words. Moving to human Teacher and real 3D environment requires an order-ofmagnitude improvement to data-efficiency of the reinforcement learning. A novel Episodic Control based pre-training is demonstrated as a promising approach for bootstrapping the data-efficient reinforcement learning.