Learning collision-free reaching skill from primitives

Abstract

Reaching is a fundamental skill for a robot. The purpose of robot reaching is to bring a robot hand to an object without any obstacle collision. Conventional handcrafted methods were complicated to implement reaching skill. Thus, this paper proposes a method using primitives acquired from human demonstrations to learn collision-free reaching skill. End-effector and joint trajectories of primitives are encoded by Gaussian Mixture Model (GMM). The way to avoid an obstacle for a primitive uses reinforcement learning to adjust the part of its end-effector trajectory near the obstacle by adapting the associative Gaussian parameters. By doing this, the movement pattern of robot reaching is goal-directed, collision-free, and similar to human-like reaching. In this paper, we validated the proposed method on an Staibli TX-40 industrial robot. The results showed that the TX-40 robot was able to perform human-like skillful reaching for an object placed in an untrained location of a table.