Adaptive learning of dynamic movement primitives through demonstration

Abstract

Complex robotics task such as biped walking, tennis-like swing, object grasping etc, depend on state prediction, complex motion generation and stable execution of motion command. Predictions of states get more accurate over time, hence the robot behavior need to be updated continuously. Such state updates cannot be incorporated straight forwardly in most trajectory generation solutions. dynamic movement primitives (DMP) provides such a flexible formulation which can adapt to spacial and temporal variations. In this paper, we present a novel algorithm for adaptive learning of DMP that can be used to generate complex trajectories required by the robot to perform a complex task. The proposed technique uses a piece-wise linear canonical system (PLCS) instead of the standard exponential canonical system (ECS) for learning the DMP parameters. We show that the proposed PLCS learns the parameters faster and has a smaller mean squared error (MSE) as compared to ECS. Proposed learning technique coupled with PLCS uniquely learns DMP parameters as compared to other state of art techniques. The proposed technique is used to learn the primitive trajectories via imitation (learning from demonstration) and an unseen primitive is generated by a kernel based mixture model. In this paper, we present results from a real 4 degree-of-freedom (DOF) Barrett WAM robotic arm and show that the proposed system is able to hit a ball randomly thrown at it.