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

One of the main challenges in robotic neurorehabilitation 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.