Tele-Kinesthetic Teaching of Motion Skills to Humanoid Robots through Haptic Feedback

The control of a humanoid robot is a very complex challenge, especially when models are not available of are very imprecise. One solution is to try learning based on human teleoperation. The main idea is let a user teleoperate a robot with force feedback and, once the user has learnt to manage this not-so-easy task, record his control actions along proprioception of the humanoid during some specific tasks, such as balancing under random perturbation. Then, this set of data relating proprioception and user actions can be used on a learning based computational technique, such as Neural Networks or similar. However, data harvesting during user teleoperation of complex real platforms is itself a demanding task. To reduce risks of teleoperation in complex frameworks, such as dual haptic joysticks, a simulation setup comes in hand, even if the model is imprecise. So, this paper deals with issues related to user operation of one or two haptic joysticks against a real and a simulated version of the platform in V-REP during several tasks of balancing in several different contexts. Data gathering has shown to be possible, and all is ready for the step which will be learning and self balance. I. INTRODUCTION Exploring the full potential of humanoid robots requires endowing them with the ability to learn, reproduce and generalize new tasks, as well as adapt a skill to changes in real-world environments. Instead of creating a robot with built-in knowledge of possible states and actions, teaching robots to perform a specific task appears as a crucial demand gaining widespread acceptance. In this context, robot Learning from Demonstration (LfD) is a powerful approach to automate the tedious manual programming in which the robot acquires training examples from human demonstrations [1], [2]. Acquiring teacher demonstrations is a key step when transferring skills from humans to robots through imitation. These demonstration examples can be obtained in many different ways, such as recording state-action pairs whilst