Virtual Trainer

Robot-assisted training is a growing field of application in neuromuscular rehabilition and sports. Thereby, robot-assisted training relieves human trainers or therapists from substantial physical labor. In addition, robot-assisted training can be augmented by virtual reality, providing objective and stimulating feedback as well as motivating game environments. Additionally, kinematic data or interaction forces can be recorded during robot-assisted training to comprehensively monitor performance metrics.

However, selecting and configuring robot-assisted training is challenging. A training session is characterized by the rendered training task (e.g. rowing), the difficulty setting, the robotic support and the chosen feedback. Human trainers are usually not aware of all the possibilities to adjust these characteristics.

However, research in motor learning suggests that selecting robot-assisted exercises carefully is crucial for successful training sessions. Appropriate feedback strategies, the amount of robotic support, skill level dependent task difficulty, and subjects’ motivation have a large impact on the outcome. Therefore, we believe that an appropriate selection and configuration of robot-assisted training has a large impact on training efficiency.

This project applies structured approaches to investigate the effect of robot-assisted trainings on changes in kinematic performance metrics. When robot-assisted trainings and kinematic performance assessments are alternating, the effects of the different trainings on the performance metrics may become observable.

In a first virtual trainer study, we combined autonomous, real-time performance evaluation with the selection of the most effective feedback strategies focusing from previous motor learning studies. Each of these selected feedback strategies focused on special aspects of the rowing task to be learnt. Based on the autonomous real-time performance evaluation, the most promising feedback strategies were selected individually for each rower. In this way, we were the first to realize a human-in-the-loop process for automated human training involving multimodal feedback selection and adaption to the trainee’s needs.

In a second virtual trainer study, we have generalized the concept to a prediction based training selection. The feedback selection was not based on special aspects of the rowing task itself. Instead, the virtual trainer was implemented as a statistically learning agent that directly predicts improvements in spatial and temporal outcome measures dependent on the training configuration and the subjects’ current performance level. From the experiences with this concept validation, we are planning to start transfers of the concept to other fields of robot-assisted training, namely robot-assisted upper arm rehabilitation.