Imitation Learning for Robotic Control based on VR Oracle

Description :

In this thesis, we intend to provide an imitation-learning approach, following the recent progress and advances in this field. In order to train the robot by showing how a human oracle achieves the task through VR interface. Following such approach arguably enhances the system’s efficiency in term of timeliness and performance, and reduces manual handcrafting and programming effort. The key idea is to train the Robotics Control agent from human’s demonstration, based on the action/reward concept: During the learning phase, the Robot tries to imitate the human’s actions following a policy that minimizes the difference between the current robot’s state and what a human would typically do under similar circumstances. Due to these try-and-error interactions, the system will eventually determine the optimal sequence of actions in order to detect and recognize the policy to operate the robot.