Stroke is one of the leading causes of disability worldwide, and the number of people affected by stroke is expected to increase. Patients affected by stroke may be able to regain strength and functionality of their limbs through rehabilitation. However, the time a patient spends with a physiotherapist or occupational therapist is limited due to cost, time, and resource restraints. With the expected increase in the number of people affected by stroke, it may be more difficult to provide each patient with the full support they need for better recovery.

A hybrid assist-as-need exoskeleton has the potential to enhance rehabilitation and meet the healthcare and economic demands due to stroke. Automated assistance means a patient can require less time guided by a physiotherapist or occupational therapist, and the exercise frequency can be increased. Hybrid assistance, utilising both an actuator and functional electrical stimulation, can combine the benefits of an actuator and FES, allow for rehabilitation that suits a patient’s needs, and allow for fatigue to be effectively managed to extend the exercise duration. Assist-as-need control sets the assistance according to the patient’s capabilities to encourage active participation, which promotes neuroplasticity and can lead to improved recovery from stroke.

However, hybrid assistance and assist-as-need controllers are currently limited. Current FES controllers are imprecise, and so accurate independent control and balancing of the assistance cannot be ensured. For an exoskeleton to provide optimal assistance, there needs to be a direct measurement of a subject’s voluntary efforts and capabilities, which current assist-as-need controllers lack. There are methods for estimating a subject’s voluntary efforts, however, the current methods do not account for the non- linear and time-varying behaviour of the muscle, and the methods are usually only analysed for isometric contractions.

The aim of this thesis was to first develop techniques to accurately estimate a subject’s voluntary efforts and capabilities, and to accurately estimate the muscle’s response to FES, while accounting for the non-linear and time-varying behaviour of the muscle. The aim was then to develop control systems that utilised the estimates of voluntary effort and the muscle’s response to FES to provide improved assist-as-need control and accurately, independently controlled and balanced hybrid assistance.

An elbow exoskeleton was designed based on the requirements for a rehabilitation exoskeleton. The elbow joint was considered because it is involved in many activities of daily living, and because the elbow joint is simple, allowing for easier verification of the techniques developed before applying the techniques to more complicated systems. The actuator was a series elastic actuator (SEA), formed by an electric motor and a polyurethane elastic element, with Bowden cables for distal actuation. The exoskeleton was used throughout the research to test the techniques and control systems developed.

It was hypothesised that a particle filter could be used to improve the estimate of a subject’s voluntary efforts and the muscle’s response to FES. The particle filter makes no assumptions about the linearity of a system or the type of noise expected, and so the particle filter is suitable for a system such as the muscle.

A particle filter was developed to estimate the torque a subject voluntarily exerted during dynamic elbow flexion, based on the electromyography (EMG) signal of the biceps brachii. The particle filter estimate of voluntary torque had a mean normalised error of 6.56%, with a 95% confidence interval of [6.52%, 6.61%], across all subjects flexing at different speeds. The particle filter accounted for the non-linear and time-varying behaviour of the muscle, and improved the estimate of voluntary torque from EMG for dynamic elbow flexion. Therefore, the particle filter is suitable for accurately estimating a subject’s voluntary efforts and capabilities.

A sigmoid model, particle filter, and moving average to allow for simple control and accurate modelling of the muscle’s response to FES. The stimulation was applied to the biceps brachii, and the FES-induced torque about the elbow was analysed. The sigmoid model described the FES-torque relationship, and was easily transformable. The particle filter estimated the parameters of the sigmoid model over time to adapt to changes in the muscle’s response to the FES over time. The moving average described the general trend in the estimated parameters, allowing for more stable control of the stimulation for a desired torque. The estimate of the FES-induced torque when using the moving averaged parameters had a mean normalised RMS error of 7.94% − 8.00% across the subjects, with a moving average window length of ten seconds or three seconds. The sigmoid model, particle filter, and moving average improved the estimate of FES-induced torque, and are suitable for determining the stimulation required for a desired assistive torque.

The estimate of a subject’s voluntary efforts and capabilities lead to a novel assist- as-need controller. The assist-as-need controller determined the assistance the subject required to complete the exercise, and the assistance was provided by the SEA. As the subject demonstrated they were capable of performing the subject unassisted, the assistance went to zero. When the subject fatigued, the desired assistive torque increased and, with the additional support from the SEA, the subject’s following attempts at the exercise were successful. The sum of the voluntary torque and assistive torque had a squared Pearson correlation coefficient with the desired total torque of R2 ≥ 0.94 for each trial. Therefore, with an improved and accurate estimate of a subject’s voluntary efforts, the assistance provided to a subject can be improved to further encourage active participation and promote neuroplasticity, which can lead to greater recovery from stroke.

The assist-as-need controller and the techniques developed for simply and accurately estimating the FES-induced torque lead to a novel hybrid assistance controller. The hybrid assistance controller allocated the assistive torques between the SEA and FES according to the estimated capabilities of the FES. If the FES-induced torque decreased, for example due to fatigue, the desired assistance from the FES decreased, and the desired assistance from the SEA increased, to compensate for the reduction in the FES-induced torque and to effectively manage the fatigue. The mean squared Pearson correlation coefficient between the sum of the assistive torques and the desired total assistive torque was R2 = 0.84. Therefore, the hybrid assistance controller is able to accurately, independently control the SEA and FES and balance the assistance provided to suit a patient’s needs.

The aim of this thesis was to develop a hybrid assist-as-need exoskeleton for stroke rehabilitation with accurate estimation of a patient’s voluntary efforts, accurate estimation of the muscle’s response to FES, optimal assistance according to the patient’s capabilities, and independent control and accurate balancing of the actuator and FES assistance. The novel techniques developed accounted for the non-linear and time- varying behaviour of the muscle, improving the estimation of voluntary torque from EMG and FES-induced torque. These techniques allowed for a novel assist-as-need controller that improved the assistance provided to a subject, and a novel hybrid assistance controller that independently controlled and accurately balanced the assistive torques. Since the hybrid assistance controller balanced the assistive torques, there is the potential to provide hybrid assist-as-need control. The techniques and control systems developed can be further verified with a wider range of subjects, and expanded for rehabilitating other joints and muscle groups. Since the hybrid assistance controller balanced the assistive torques, there is the potential to provide hybrid assist-as-need control. The novel techniques and control systems presented in this thesis allow for assistance suited to a patient’s capabilities and needs, improve fatigue management, and longer and more frequent therapy sessions. As a result, an exoskeleton with the techniques and control systems developed has the potential to meet the healthcare and economic demands due to stroke, and to greatly improve a patient’s recovery from stroke.