Deducing human emotions by robots: Computing basic non-verbal expressions of performed actions during a work task

We have established an emotional model to enhance a virtual worker simulation, which could be also used to support robots in a joined human-robot work-task inside an industrial setting. The robot is able to understand people's individual and specific knowledge as well as capabilities, which are ultimately linked to an emotional consequence. As a result, the emotional model outputs the emotional valence calculated as positive or negative values, respective to reward and punishment. This output is applied as value function for a reinforcement learning agent. There we use an actor critic algorithm extended by eligibility traces and task specific conditions to learn the optimal action sequences. We show the influence of emotional reward leads to differences in the learned action sequences in comparison to a simple task performance evaluation reward. Therefore the robot is able to calculate emotional feelings of a human during a given working task, is able to decide if there is a better, more emotional stable path to doing this working task and moreover the robot is able to decide when the human is needed help or even not.