Abstract: In this project our main aim is to develop a system wherein an individual can train robots and teach them how to perform a task by showing them how it is done by actually performing the task once. This type of system would reflect on the human methods of teaching wherein a teacher teaches a student how to perform a task by showing them how it is done by actually performing it once himself. As a child observes and learns the teacherís methods and actions and tries to repeat the same when he tries it himself, our system will install the same characteristic into our robot as well. For the purpose of presentation of such a system, we will implement an algorithm or a procedure via which the robot will record and learn the actions when performed by the user or the trainer in the learning phase which is nothing but when the user is performing and teaching the action to the robot for the first time. We will derive a technique or a mechanism via which these actions can be stored, altered and replayed. Also, filters, acceleration, slow motion replays are also be included. We are going to create the dummy model of a vehicular robot to demonstrate the system. The user will guide or train the robot to run on certain track or paths by manually driving it once, after which the robot will record, learn and execute the task itself.
Keywords: Real-time control, Experience replay (ER), reinforcement learning (RL), Q-learning, robotics.