The main form of mobility for paraplegic patients is by wheelchair. However, not moving the legs comes with adverse health effects. Exoskeletons are one solution to get these patients walking again. One of the aims of exoskeleton research is the complete restoration of locomotion for paraplegic patients. The achieved gait must be stable, safe and comfortable for the patients. Most research goes into exoskeleton devices which require the use of balancing aids. In the form of crutches, these aids help the exoskeleton users to maintain stability. One of the goals is to eliminate the reliance on balance aids and let the robot do most work. Until now only two exoskeletons are able to achieve autonomous dynamically stable gait. The gait generation algorithms used in these device are based on inverse dynamics. trajectories are calculated and closely tracked. The main challenges of inverse dynamics control algorithms are slow and static movement, balance recovery issues or computational complexity. In this research the aim is to achieve autonomous walking without balance aids. The Project MARCH exoskeleton is taken as an example in this case study. This device has 4 actuated degrees of freedom per leg. The exoskeleton is modelled in OpenSim. Using predictive forward dynamic simulations, a gait algorithm is implemented and evaluated. The reflex-based control algorithm is based on proportional-derivative controllers. This control algorithm is implemented in SCONE and is optimized using the Covariance Matrix Adaptation - Evolution Strategy method. A second simulation experiment uses the same method to achieve standing balance. After optimization of the control algorithm, dynamically stable gait patterns emerge. The exoskeleton model shows limit cycle behaviour and is able to walk for at least 30 seconds at a speed of 0.7 m/s. The controller can optimized to reject perturbations up to 300 N for 0.1 s. The emerging gait pattern shows two features, which complicate the implementation in the real exoskeleton. The model shows a back-heel rotation during stance phase and hits the joint limits during the liftoff phase. Standing balance is also achieved by a different controller. This research serves as a proof of concept on using SCONE (or more general, predictive forward dynamic simulations) to simulate and test an autonomous exoskeleton. The algorithms are completely feedback controlled require no predefined trajectories. Certain features seen in the emerging gait patterns remain to be resolved. This work demands more research to prevent back-heel rotation, to avoid approaching the joint limits and model the toe-off more adequately in order to reduce the peak torque. Furthermore, interesting research can be done on a randomized perturbation rejection and on how to model the inelastic collision at the joint-ends properly as well as on making a comparison between the gait patterns presented in this research and the patterns currently used in the exoskeleton. Only if these challenges are addressed, the gait algorithm becomes eligible to employ in a real exoskeleton.

When designing musculoskeletal models for forwardsimulation studies, a tradeoff must be made between biologicalfidelity and computational efficiency. A method is proposedthat optimizes the coordinates of via points of muscle paths,used in relatively simple musculoskeletal models, such that themuscle moment arms match those of more complex modelsand experimentally measured moment arms. For this purpose,the via points of the Gait2392 model are optimized to matchmuscle moment arms of the Rajagopal2015 model and otherexperimentally obtained muscle moment arms. The results showthat muscles that can be represented by straight lines, can beoptimized well. More complex muscles that contribute to multi-degree of freedom movements, are still improved, but to aless extent. The main benefit of the optimized model, is thatthe compatution time of human motion simulation is greatlyimproved, making it a viable option to pre-train neuromuscularcontrollers, which can be further fine-tuned with the morecomplex models. To further improve musculoskeletal models ingeneral, more data on muscle moment arms is needed, especiallyfor muscles that contribute to multiple joints and movements inmultiple planes of motion.

The Sit-to-Stand (SiSt) task is one of the most crucial yet mechanically demanding daily tasks. A transfemoral amputee develops high torques on the intact leg to complete the SiSt task. Such high torques are a consequence of limited torques produced by a prosthesis. The main interest of this project is on the sit-to-stand task of a transfemoral amputee fitted with an active knee and passive ankle prosthesis. The objective is to test whether strategies like asymmetric foot placement and reduced weight-bearing asymmetry could reduce the torque produced in the intact knee. Musculoskeletal model of an able-bodied human subject and a transfemoral amputee subject enabled with an active knee prosthesis is developed. Forward dynamic optimisations are performed by defining an appropriate framework required for simulating the sit-to-stand task. The process was verified by implementing the framework on the musculoskeletal model of an able-bodied individual and comparing it with experimental results available in the literature. The comparison showed a good agreement of simulated results with results reported in the literature. Joint torque profiles of the intact limb of the amputee model were then simulated with asymmetric foot placement and reduced weight-bearing asymmetry strategies. Placing the prosthetic leg posterior to the intact leg reduced the peak intact knee torques by 1.5% relative to placing the intact leg adjacent to the prosthetic leg. The peak intact knee joint torques were reduced by 13% in SiSt task simulation with reduced weight-bearing asymmetry. An increased metabolic cost needed to perform the SiSt task also resulted from this strategy.

Musculoskeletal modeling and simulation has become a prominent tool in clinical gait analysis with the ability to provide insight into the underlying mechanisms of human movement. However, generic cadaver-based models have been shown to poorly reflect live subjects, especially those with pathologies such as cerebral palsy (CP). The main purpose of this thesis was to evaluate the effect of model personalization on gait simulation outcomes between models of varying level of personalization. Gait data from 7 children with CP was used for simulations in OpenSim using 3 different model types for each: scaled generic (GS), scaled generic with tibial torsion and femoral anteversion (TTAF), and MRI-based. MRI-based model outcomes saw the greatest differences from GS models in the hip and upper leg, specifically hamstrings and quadriceps, but also experienced moderate differences in the lower leg. Similar results were found when comparing MRI to TTAF models. TTAF models differed from the GS models around the subtalar joint, mainly the tibialis anterior. Larger differences in kinematics, kinetics, and muscle activations were accompanied by changes to the most influential model parameters, in descending order of importance, these were: tendon slack length, moment arm length, and normalized muscle fiber length. Despite the differences between these models, there was no indication that either is more accurate or more suitable for clinical use.