Overview / Usage

This project intends to develop and apply algorithms of visual reinforcement learning in a simulator, so that a robot can learn and reach a certain objective from images coming from a camera in the automaton.

Methodology / Approach

This project will develop a visual reinforcement learning algorithm that learns from camera images in simulation circumstances, using algorithms like Deep Q-Learning, Double Deep Q-Learning and other deep reinforcement learning algorithms.

The project faces many challenges, such as the need of a great amount of samples to make the learning process possible. In order to try to reduce the time and the computational effort of this task, we will use transfer learning techniques, so that we can transfer the acquired learning to other simulations with different scenarios.