Designing Personalised Rehabilitation Controllers using Offline Model-Based Optimisation

The use of robotic assistance in rehabilitation is becoming more popular, yet delivering optimal assistance remains an open challenge. In order to accelerate a patient's recovery, assistance that is personalised to the needs of the patient is required. However, controllers of rehabilitation robots have traditionally been designed and tuned heuristically, through trial and error, with one set of parameters used across several patients. In this paper, we propose an offline model-based optimisation approach, which can be used to create personalised rehabilitation controllers. We formulate the process of designing and tuning a rehabilitation controller as a multi-objective optimisation problem, and we solve this problem using Bayesian optimisation. We evaluate our method with forward dynamics simulations and the results demonstrate that a set of controller parameters can be obtained that are both patient-specific and task-specific. Our approach could be used for the personalisation of controllers designed for rehabilitation, injury prevention and human augmentation.