Long Short-Term Memory for Affordances Learning

This paper addresses the problem of sensorimotor learning from the perspective of affordances learning of simple objects. We are developing a scenario where a robotic arm interacts with a polyflap, a simple 3-dimensional geometrical object. We perform experiments with a simulated arm using a physics simulator, but we plan to use also a real arm. The robot interacts with the object by pushing it in different ways. We use Recurrent Neural Networks to predict the arm and object poses during this interaction, given a discrete set of random actions that the robot can produce.