Towards a generic workload estimation for human robot interaction in real world applications
Presenter: Mathias Trampler
Authors: Mathias Trampler1, Marc Tabie1, Elsa Andrea Kirchner1,2
1 German Research Center for Artificial Intelligence (DFKI), Bremen, Germany
2 University of Bremen
Humans more and more start to collaborate and work side-byside with a robotic coworker but while the spatial separation disappears, a robotic coworker is still a very isolated system with only very limited interfaces like a keyboard or a touch screen with almost no possibility to communicate subtle changes regarding the workers mental states.
Humans are very sensitive in perceiving a coworkers state and adjusting to it, but working with robots is a rather static experience. Detecting a workers mental state and adjust the robots behavior to e.g. the workload the worker can currently handle or the fatigue would improve the collaborative experience and make it more intuitive and safer.
With dry electrode headsets which don’t rely on extensive preparation sessions this may become applicable in real world work situations. To achieve a reasonable classification performance on the estimation of workload with EEG data often a training session for each user is necessary.
Applying Auto-ML tools like TPOT and genetic algorithms allow finding better preprocessing and learning algorithms for optimizing the generalization across subjects. This showed promising results for the classification of different levels of workload for 6 subjects based on the number of interactions with a robotic simulation.
We used the multitaper PSD for each channel and used the XGBoost classifier with shrinkage and subsampling. Furthermore we developed a scenario to gather annotated training data which induces different levels and different kinds of workload to avoid classifier overfitting to a specific workload kind. With this data we want to find a generic processing and classification pipeline to detect workload and fatigue to enable the adjustment of the working environment according to the user’s needs.
Keywords: machine learning, brain reading, workload