Force-based robot learning of pouring skills using parametric hidden Markov models
Demonstration phase: The human provides examples of the pouring task by teleoperating the robotic arm using a haptic device. The force perceptions sensed over the course of the task, along with their initial value (i.e., the task parameter), are used to train a parametric hidden Markov model.
Robot execution stage: The robot carries out the pouring skill using Gaussian mixture regression to retrieve joint-level commands given the force-torque inputs at each time step.
NOTE: The execution carried out by the robot is slow and sometime also shows "jerkiness", this occurs because two aspects, namely, (i) the provided demonstrations showed low velocity profiles which do not vary smoothly, and (ii) the robot controller only provides a position controller, thus velocity control is not possible here in order to get smoother and more realistic reproductions.