ABSTRACT
This study investigates machine learning techniques used to develop a smart deriving control system of a service robot that is adapted to handgrip strength. Experiments have been conducted throughout 10 trials whereby the data has been collected to evaluate supervised learning regression, supervised learning classification, IRL's policy, and arriving domain for each machine learning algorithm, as well as the probability of crashing. For our results, several machine learning techniques were conducted, all of which demonstrated promising results. In supervised learning regression, the mean absolute percentage error ranged from 4.7% to 6.1%, showing that the handgrip strength levels were predicted accurately. Meanwhile, supervised learning classification maintained high accuracy from 91.2% to 93.5%, and the precision, recall, and F1 score were robust. The machine learning-based algorithms have made the learning of user preferences and accomplishing tasks by adjusting the weights. The reward prediction error ranges from 0.030 to 0.038, and the success rates range from 86.4% to 90.2%. The domain adaptation methods have effectively reduced the domain discrepancy from 0.011 to 0.015 and the transfer accuracy from 91.7% to 93.2%. The Bayesian optimization has effectively offered optimal control parameters, and an improvement in the system performance could be seen with a range of performance from 0.91 to 0.93. It could be concluded that the machine learning-based approaches can improve human-robot interaction, and assistive robotics can be advanced.
