Learning Robot Task Planning Primitives by means of Long Short-Term Memory

This work is an attempt to create a robot task planner by exploiting increasingly popular Deep Neural Networks. The aim is to learn how to achieve a robotic manipulation task by selecting the appropriate action to perform, along with its arguments, by observing the robot workspace. This work proposes a model based on Long Short-Term Memory, that reaches up to 97% of accuracy on action prediction, along with an expert policy that is able to generate an artificial dataset used for training.