Evaluation of Data-Driven Models in Human-Robot Load-Sharing

Human-robot load-sharing is a potential application for human-robot collaborative systems in production environments. However, knowledge of the appropriate data-driven models for this application type is limited due to a lack of physical real-world data and validation metrics. This paper describes and demonstrates a load-sharing testbed for evaluating data-driven models in a human-robot load-sharing application. Specifically, the testbed consists of a single operator and single robot relocating a payload to a desired destination. In this work, the operator initially communicates to the robot using audio feedback to initiate and alter robotic motion commands. During the payload relocation, human, payload, and robot state data are recorded. The measurements are then used to train three data-driven models (neural network, naïve Bayes, and random forest). The data-driven models are then used to transmit movement commands to the robot during human-robot load-sharing without the use of audio feedback, thus improving robustness and eliminating audio signal processing time. Evaluation of the three data-driven models shows that the random forest model was demonstrated to be the most accurate model followed by naïve Bayes and then the neural network. Hence, the results of this study provide novel insight into the types of data-driven models that can be used in load-sharing applications in addition to development of a real-world testbed.