Machine Learning Control for Assistive Humanoid Robots Using Blackbox Optimization of PID Loops Through Digital Twins

With the latest technological advancements in Machine Learning, the focus shifted significantly from classical data analysis to controlling various robotics systems. Such an approach, compared to classical control methodologies, provides a simultaneous design, validation and robustness analysis that is performed in an autonomous manner. Although the dynamic performance is not formally guaranteed, the more learning iterations are performed, the more the confidence in the designed solution increases. In this work, we address the problem of accelerating Machine Learning Control (MLC) algorithms by parallelizing the learning with the aid of a simulated test environment containing a Digital Twin of a NAO robot. The increase of the modeling robustness and of the generality of the control algorithm is ensured by performing random positioning tasks with each learning episode. The solutions are further leveraged through Transfer Learning using the real robot and the results are validated and compared. Our main goal is to provide a design framework for assistive robots, which bring significant societal benefits, although require high reliability and safe operation. In this respect, a thorough statistical study concerning the comparison of two typical MLC algorithms, namely the Genetic Algorithm and the Bayesian Optimization, is included.