Parallel robot learning through an ensemble of predictors able to forecast the time interval before a robot failure

This article describes a proposal to achieve fast robot learning from its interaction with the environment. Our proposal will use an ensemble of elements that will combine dynamic programming and reinforcement to predict when a robot will make a mistake. This information will be used to change the control policy trying to maximize the time interval before a failure. Finally, our proposal will be able to learn simultaneously perception and action. The robot will build a representation of the environment that will dynamically increase, during the learning process, to include new situations that have not been seen before.

keywords: