Learning from Multiple Demonstrations with Different Modes of Operations

Abstract

In this paper, teaching multiple types of complex trajectories at once to a robot in a robust, easy to train model using Learning from Demonstration is studied where the robot is expected to gain the capacity to differentiate between different types of demonstrated trajectories and be able to reproduce these trajectories correctly. Demonstrated trajectories are used to train a Hidden Markov Model (HMM) and a modified version of Gaussian Mixture Regression (GMR) -which utilizes state transition probabilities between states of the HMM, the most probable state the end effector of the robot belongs to in the current reproduction of the trajectory, and previous points in the current reproduction of the trajectory- is used to estimate the trajectory iteratively. A Proportional Derivative (PD) controller is employed for the reproduction. Starting points that are intended to correspond to different types of trajectories which the robot is expected to differentiate between are tested on numerical and simulation experiments. Multiple numerical experiments and simulation experiments showed that our modified algorithm produced comparable results to previous work, and in certain complex trajectories our algorithm was successful where previous work has failed to produce expected results.