Reinforcement Learning-based Learning from Demonstrations for Collaborative Robots

Learning from Demonstrations (LfD) can support a human operator to control a collaborative robot (cobot) in an intuitive means. Gaussian Mixture Model and Gaussian Mixture Regression (GMM and GMR) are useful tools for implementing such a LfD approach. However, well-performed GMM/GMR require a series of demonstrations without trembling and jerky features, which is challenging to achieve in practical applications. To address this issue, in this paper, an improved Reinforcement Learning (RL)-based approach for GMM/GMR is devised to carry out a variety of complex tasks. The innovations of the research are twofold: firstly, a Gaussian noise strategy is designed to scatter demonstrations with trembling and jerky features to better support the optimization of GMM/GMR; Secondly, an improved RL-based optimization algorithm is developed to eliminate potential under-lover-fitting GMM/GMR. A cases study was conducted to verify the approach. Experimental results and comparative analyses showed that this developed approach exhibited good performances in computational efficiency and solution quality.