Transferring Skills to a Simulated Robot

A skill transfer and improvement framework was designed and implemented to transfer skills to a simulated robot without the help of a robotics expert. A skill expert first learns to perform the desired task by maneuvering the robot and giving actuator commands using the keyboard. The framework is then used to transfer the expert’s skills to the robot by teaching the robot’s controller to perform the skill. CMAC neural networks are responsible for learning skills. Learning is accomplished using imitation of the demonstrated movements that are mapped to a set of primitives. After transferring a skill, the expert can give advice to the controller to improve the robot’s performance. INTRODUCTION The problem tackled in this research is how to transfer a skill from a skill expert to a robot. A skill expert is a person who possesses the skill to perform a task but is not a robotics expert. With today’s technology we need robotics experts to program robots to perform some specific task. Consequently, a skill expert who is not a robotics expert cannot easily transfer skills to a robot. Our first hypothesis is that an adaptive learning controller [1] can be used to implement the skill transfer. Adaptive control will allow adaptation to variations in the environment of the robot. The second hypothesis says that the learning controller can be based on the idea of cloning [1]. Cloning is also called learning by example, or imitation learning by researchers [2-4]. Cloning design allows imitating the skill(s) demonstrated by an expert. The third hypothesis proposes that the expert first needs to learn how to execute the skill by maneuvering the robot (controlling the robot’s actuators) while observing the feedback from the robot’s sensors. Then the expert should provide the imitation controller with pairs of state inputs (functions of sensor readings) and desired actuator outputs. The fourth hypothesis proposes that all robot skills should be described using action primitives, where primitives describe simple robot motions taking into account interactions with the environment. All skills should be described as concatenated primitives, where combined primitives represent skills. The fifth hypothesis deals with improving the performance of the imitation controller. The human expert can let the controller execute a skill and then give an advice to the controller on how to improve the performance of the skill. This advice can be advice about desired outcomes and advice about the state space variables. Our research is constrained to robots with a small number of degrees of freedom interacting with a small number of static objects. Thus we will consider a two dimensional (2D) model of a truck with a trailer, a parking spot, and