Periodic Reporting for period 2 - CO-MAN (Safe data-driven control for human-centric systems)
Período documentado: 2022-03-01 hasta 2023-08-31
CO-MAN is set to advance on theoretical foundations for control of human-centric systems by combining model-based control and machine learning algorithms. This framework is to be evaluated as a user-adaptive data-driven control in various human-machine interaction applications with a special focus on translation to clinical domains. The research and development activities of the project centers three main objectives:
Objective 1- Develop the system-theoretic basis for probabilistic non-parametric dynamics models in sparse data regimes,
Objective 2- Develop a novel data-driven control theory for complex uncertain systems explicitly considering model uncertainty and incorporating active exploration,
Objective 3- Develop approaches to integrate individual user preferences /models into control design.
With rehabilitation robotics as a project application case, we had explored the utility of machine learning techniques for modelling abnormal human movements as well as a design of a model-based patient-specific interaction strategy in the control design. We envisage an interaction scenario of a patient using a motorized upper body exoskeleton with functional electrical stimulation, also known as a “hybrid exoskeleton”. A combined use of functional electrical stimulation offers great opportunities as it is known to bring great clinical benefits to neurological patients while the control challenge due to its strongly nonlinear, user-specific nature of the system performance limits active use in clinical practice. Therefore, we have been orchestrating the theoretical understandings of machine learning algorithms and robust control frameworks for neurorehabilitation with a hybrid exoskeleton.
Objective 1- Develop the system-theoretic basis for probabilistic non-parametric dynamics models in sparse data regimes,
Objective 2- Develop a novel data-driven control theory for complex uncertain systems explicitly considering model uncertainty and incorporating active exploration,
Objective 3- Develop approaches to integrate individual user preferences /models into control design.
With rehabilitation robotics as a project application case, we had explored the utility of machine learning techniques for modelling abnormal human movements as well as a design of a model-based patient-specific interaction strategy in the control design. We envisage an interaction scenario of a patient using a motorized upper body exoskeleton with functional electrical stimulation, also known as a “hybrid exoskeleton”. A combined use of functional electrical stimulation offers great opportunities as it is known to bring great clinical benefits to neurological patients while the control challenge due to its strongly nonlinear, user-specific nature of the system performance limits active use in clinical practice. Therefore, we have been orchestrating the theoretical understandings of machine learning algorithms and robust control frameworks for neurorehabilitation with a hybrid exoskeleton.
We addressed the project objectives from two perspectives: development of novel learning-based control schemes and advancement of machine learning techniques for individualized control. The former perspective has strengthened theoretical bounds for learning-based control which are evaluated in terms of stochasticity and associated error bounds, performance and safety guarantees, and computational efficiency. The resulting probabilistic machine learning techniques were incorporated into control loops for evaluations. The latter perspective saw progress on a novel framework for user-individualized data-driven control in human-machine interaction scenarios. With rehabilitation robotics as a project application case, we had explored the utility of reinforcement learning techniques for modelling abnormal human movements as well as a design of a model-based patient-specific interaction strategy in the control design. Furthermore, we had proposed an incorporation of preference-based feedback from users into control loops through machine learning, which we believe will accelerate the modelling progress.