Recognising the user’s locomotive intentions is crucial for the correct functionality of exoskeletons and active orthoses. For gait applications, extrapolating control inputs from the arm swing may be worthwhile, since arm oscillations naturally occur during human locomotion. A similar method would be unaffected by severe impairments of the lower limbs, and there is evidence suggesting enhanced results of gait rehabilitation when arms and legs exercise together. In this thesis, we propose a control algorithm to drive online a lower limb exoskeleton through the arm swing. Contrary to a previous EMG-based approach by La Scaleia et al. (2014), our algorithm exploits shoulder kinematic data to mimic “single swinging”, a natural mode of human interlimb coordination which is characterised by each arm swinging in-phase with the contralateral leg. Our proposed control architecture relies on two major modules: an Arm Observer and a Gait Generator. The Arm Observer consists of an adaptive frequency oscillator which extrapolates the frequency and phase of the arm swing by receiving online measurements of the angular shoulder position in the sagittal plane. This data is used by the Gait Generator to compute lower limb trajectories, based on regression models from a previous study by Koopman et al. (2014). We validated our controller through human-subject experiments, involving three participants walking on a treadmill with and without a lower limb exoskeleton, the Lopes II. When feed by data associated with natural walking, our adaptive frequency oscillator could very precisely replicate the arm swing frequency, stride cadence and timing of shoulder flexion peaks when walking faster than 0.5 m/s. When wearing the exoskeleton, our algorithm allowed the participants to cope with constant and variable treadmill velocities in the range of 0.5-1.25 m/s. As such, the results of this thesis show that our proposed approach can extend the applicability of arm-based control to walking speeds suitable for gait rehabilitation and assistance.

An estimated 40,000 people have trauma related above knee amputations in the United States alone. A transfemoral prosthesis is an artificial limb that replaces the amputated limb. Although trauma related amputations are going down annually, there is still the need for prostheses that are capable of restoring normal biological knee function. Active micro-processor knees are a type of transfemoral prosthesis that can supply energy for activities therefore making many activities of daily living like chair and stair negotiation possible, how-ever these devices are not yet commercially available as it does not yet meet the requirement of robust and unambiguous mode switching. One aspect of active microprocessor knees that needs improvement is the intent recognition system that perceives the intent of the user. A novel hybrid intent recognition algorithm based on machine learning using solely mechanical signals was developed and tested in this thesis. The algorithm is capable of distinguishing be-tween the modes Standing, Sitting, Walking, Ramp Ascent, Ramp Descent, Stair Ascent and Stair Descent. The analysis of the algorithm was done on an open source healthy subject gait data set containing a total of 476 trials. The analysis involved determining recognition error rates and decision times for a novel subject’s data. The proposed algorithm fuses data from Inertial Measurement Units(IMUs) worn on the shank and knee joint encoders to make the decisions. The algorithm can achieve an overall error rate of 14.28%, the error rate reduces to 2.62% when grouping the Ramp Ascent and Ramp Descent together with the Walking mode.Decision times are, on average 9.59ms after a transition for critical transitions between stair modes and walking. For transitions between less critical modes like sitting and standing, decisions are taken with a maximum delay of 610ms. All transitions were successfully detected in 229 out of the 476 trials. The remaining trials had misclassifications due to improper labeling and variations in gait speed among the users. A preliminary analysis into adding ground reaction forces and moments indicates that the error rates can be decreased with it’s use. The research concludes that a hybrid classifier in which ramp walking is treated as level ground walking is a good starting point for implementing on the transfemoral prosthesis.

The restoration of the walking function of paraplegic patients requires long and extensive training. Physicians themselves are not able to provide this training due to the physical labour required. Robots and soft exoskeletons are in development to assist the physicians in this task, however, they are either bulky and restrictive or lack a method to actuate the knee joint. This research explores if a hydraulic system with elastic element is able to actuate a human knee joint during gait training.
The hydraulic system is required to provide a torque of 75Nm with a bandwidth of 4Hz, the large force bandwidth, and a torque of 20Nm over a bandwidth of 12Hz, the small force bandwidth. The system must fit on a knee joint and impose the least amount of restrictions on the patient. Concepts that were identified as possible solutions in a previous literature review included a hydrostatic transmission, a system with a master rotary pump and a rotary slave actuator, a dual-acting hydraulic cylinder, and a master slave hydraulic cylinder system. Preliminary open loop simulation results suggest that the Hydrostatic transmission could give better performance in Torque transmission, size, and complexity. The bandwidth performance for both concepts was equal.
The simulation model of the hydrostatic concept is expanded to include DC motor/driver dynamics, knee dynamics in swing phase, and ground reaction forces in stance phase. A physical test bed is designed to evaluate the simulation results with a real-world counterpart.
Open loop and closed loop with PI controller results showed a significant decrease in the output torque magnitude for the real-world system, up to a factor 13 smaller. The output torque is decreased by the higher friction, slow pressure build-up and leakages in the real-world system. These factors are in part caused by the design choices made for the test bed design, such as the over-the-counter parts and usage of a 3D printed spring. Failure of the spring was observed at high torque and low frequency input signal. This was made worse by the large initial overshoot of the system in the closed loop tests. A more precise control scheme than the PI controller should be designed to prevent breakage. The real-world system is stable and the bandwidth of the system was larger than the 4Hz that is required for the application. For future research a smaller custom hydraulic loop will need to be created to alleviate the friction and leakage issues. Only then will a hydraulic system with series elastic element actuate a human knee to help a paraplegic patient walk again.

Currently, lower-body exoskeletons for
paraplegics are investigated as an alternative to wheelchairs
and as an exercise method with medical benefits. Literature
provides little examples for users to influence the
length or frequency of steps taken by the exoskeleton,
complicating stability and practical usability. This study
proposes a novel control paradigm that allows users to
influence step parameters of the exoskeleton, through a
bi-directional haptic interface that also provides feedback.
The exoskeleton handles the cyclic walking pattern while
the patient is enabled to correct for disturbances. I adapted
an existing lower-body exoskeleton trajectory generator to
allow real-time adaptation of step length and swing time.
To demonstrate a proof-of-concept of the control paradigm,
I implemented the controller for a virtual 2D exoskeleton,
to be controlled by an existing bi-manual haptic control
interface. A human-in-the-loop experiment was performed
with the goal to compare the benefits of user control
over either step length or swing time with a situation
with no human control. In the experiment perturbations
of increasing magnitude were applied to a 2D virtual
exoskeleton, participants could increase or decrease either
step length or swing time during swing to correct for these
disturbances. The number of successful step taken before
the perturbations resulted in a fall were measured. The
swing time group succeeded in making the exoskeleton
walk stably significantly longer then when there was no
human input, proving that the proposed control paradigm
is feasible and beneficial for stability.

Ever since human ancestors picked up rocks and used them as tools, technology has enhanced human capabilities. Over the last decades more and more research has been done in trying to enhance strength and endurance by means of an exoskeleton. Reducing the metabolic cost of human walking using lower extremity exoskeletons is one of the challenges that has been taken on by several research groups. Only recently have some of these groups managed to reduce the metabolic cost of walking with the help of an exoskeleton. Understanding how human walking behavior changes and adapts to these new supportive devices is crucial in the quest to better, more efficient exoskeletons. Effects of two control methods, that were proven to be successful in making human walking more energy efficient, on the biomechanics of human subjects were examined by performing lab experiments and using the collected experimental data to run musculoskeletal simulations to compute muscle mechanics and energetics of six lower extremity muscles. Lab experiments, which were completed by eight out of ten recruited subjects, consisted of treadmill walking with and without the exoskeleton, while EMG, respiratory, kinematic and kinetic data was recorded. Processing of this data, the simulation steps taken and the eventual results are presented in this thesis. It was found that walking with the exoskeleton, while it was not giving any assistance, significantly increased the energy consumption rate of subjects, compared to normal walking. This increase was reduced by one of the two support methods, whereas the other had no significant effect on the total metabolic rate.