Robust real-time monitoring of human task advancement for collaborative robotics applications

A crucial problem in human-robot collaboration is to achieve seamless coordination among the agents. Robots have to adapt to human behaviour, which is highly uncertain. In fact, humans can perform each task in many ways and with different speeds, occasional errors and short pauses. This paper offers a robust method to monitor the advancement of the current human activity in real-time in order to predict its duration. The algorithm learns online templates of new variants of the task and uses them as references for a Dynamic Time Warping-based algorithm. The proposed strategy has been tested within a realistic assembly task. Results show its ability to give accurate predictions also in case of peculiar variants, such as those associated with errors.