Model Predictive Control to Promote Motor Variability During Robotic Assistance Enhance Learning

Robots provide many advantages for motor learning and neurorehabilitation, such as high haptic rendering accuracy of complex virtual training environments and alternative ways of providing assistance during rehabilitation therapy. Despite the advantages, benefits of haptic guidance -i.e., physically guiding the participants’ limbs through a desired trajectory- are limited. Firstly, the assisting forces from haptic guidance might prompt participants to rely on the assistance, which prevents the perception of the implemented dynamics of the training environment. Secondly, haptic guidance limits the movement variability of the patients by enforcing their movements to predefined/fixed trajectories, which might hamper learning.

We propose that Model Predictive Controllers (MPC) can address these limitations by minimizing the assisting forces and by promoting movement variability with flexible movement trajectories. We recruited 41 healthy participants in a between-subject study to test the effectiveness of using MPCs as assistance strategies during learning a complex dynamic task. The task consisted of swinging a pendulum to hit the incoming targets, and the environment was haptically rendered on a Delta.3 robot (Force Dimension, Switzerland). Two different MPCs were implemented. The first MPC -end-effector MPC- applied its forces directly on the end-effector of the robot. The second MPC -ball MPC- applied its forces on the pendulum ball to further reduce the necessary assisting forces. However, this indirect assistance might negatively affect how the task dynamics are perceived by the participants and their sense of agency. The performance of the participants during training and learning at short/long term retention were compared to a group that trained with conventional haptic guidance, and to a control group that trained without assistance.  We hypothesized that the end-effector MPC would increase the movement variability and decrease the assisting forces, therefore, enhance learning. On the other side, we hypothesized that the ball MPC would limit the sense of agency and learning despite enhancing performance and motivation during training.

Our findings show that the MPCs reduce the assisting forces compared to haptic guidance. Training with the end-effector MPC increases the movement variability and does not hamper the pendulum swing variability, resulting in better learning of the task dynamics, compared to the other groups. Furthermore, when training with the end-effector MPC, the increase in the sense of agency is associated with learning. Overall, using MPCs as robotic assistance strategies during training enhances motor learning of complex dynamic tasks, and is a promising robotic strategy to improve training outcomes in brain-injured patients.