Multi-purpose Robotic Training Strategies for Neurorehabilitation with Model Predictive Controllers

Abstract

One of the main challenges in robotic neuroreha-bilitation is to understand how robots should physically interact with trainees to optimize motor leaning. There is evidence that motor exploration (i.e., the active exploration of new motor tasks) is crucial to boost motor learning. Furthermore, effectiveness of a robotic training strategy depends on several factors, such as task type and trainee's skill level. We propose that Model Predictive Controllers (MPC) can satisfy many training/trainee's needs simultaneously, while providing a safe environment without restricting trainees to a fixed trajectory. We designed two nonlinear MPCs to support training of a rich dynamic task (a pendulum task) with a delta robot. These MPCs differ from each other in terms of the application point of the intervention force: (i) to the virtual pendulum mass, and (ii) the virtual rod holding point, which corresponds to the robot end-effector. The effect of the MPCs on task performance, physical effort, motivation and sense of agency was evaluated in fourteen healthy participants. We found that the location of the applied controller force affects the task performance -i.e., the MPC that actuates on the pendulum mass significantly reduced performance errors and sense of agency during training, while the other MPC did not, probably due to low force saturation limits and slow optimization speed of the solver. Participants applied significantly more forces when training with the MPC that actuates on the pendulum holding point, probably because they reacted against the robotic assistance. Although MPCs look very promising for neurorehabilitation, further steps have to be taken to improve their technical limitations. Moreover, the effects of MPCs on motor learning should be evaluated.