Design and evaluation of bio-inspired control strategies for transfemoral prosthesis

Despite recent developments in solutions to replace the missing limb of transfemoral amputees and restore natural locomotion, the use of these solutions remains limited. To date, the potential benefit of using powered prostheses is strongly impacted by the challenges that are yet to be addressed regarding their development, both on their design and control methods. An emerging trend for building the control strategies for the prostheses actuators takes advantage of bio-inspiration, i.e. the control laws rely on biological principles that have been highlighted in healthy locomotion. In this dissertation, bio-inspired control strategies for a powered transfemoral prosthesis are presented. The following biological concepts are explored throughout this work: (i) Central Pattern Generators (CPGs), which can be seen as a set of coupled oscillators responsible for providing the rhythmic characteristics of locomotion; (ii) motor primitives, which are considered to be the principal components of muscles stimulations; and (iii) the inverse internal models of the cerebellum, which plays an essential role in motor learning and adaptation. This dissertation presents three bio-inspired controllers for a representative compliant prosthesis equipped with series-elastic knee and ankle joints. The three control architectures rely on compliant position tracking by combining a feed-forward torque component with an impedance-based torque component for both joints of the prosthesis. The first controller simply incorporates reference torque and angle profiles associated to healthy walking that were taken from the literature, and it is mainly used to validate the feasibility of torque-based control strategies for the prosthesis without the need for torque sensing. The second version of the controller includes artificial, Gaussian-like torque and angular primitives which, through proper recombination, generate the reference torque and angle patterns for both joints. It also integrates an artificial CPG implemented by an adaptive oscillator which continuously detects the global locomotion parameters. The last version of the controller incorporates an iterative learning mechanism to compute a feed-forward prediction torque component for both joints. This adaptive process uses the Locally Weighted Projection Regression (LWPR) algorithm to continuously learn the inverse internal model of the prosthesis joints. The dissertation reports three sets of experiments performed to validate the developed controllers. The first experiment involved a transfemoral amputee walking on a treadmill. It aimed at validating the feasibility of implementing torque control strategies on the actuated elastic transfemoral prosthesis and used the first controller. It showed that static modelling of the prosthesis geometrical and elastic relationships is appropriate for converting desired torques for the prosthesis joints into positions for its actuators. The next set of experiments investigated the use of artificial primitives in the second controller and the role of the feed-forward torque component for the knee and ankle joints. Two experiments were conducted with a transfemoral amputee walking on a treadmill, first at self-selected speed and then at different speeds. They emphasized the relevance of the primitives-based, feed-forward torque component for both joints and its capacity to increase the controller's compliance. They also revealed that combining artificial primitives with an adaptive oscillator is appropriate for the control of transfemoral prostheses, although the capacity of the controller to adapt the delivered torque patterns to the walking speed could not be demonstrated. The third set of experiments assessed the performance of the iterative learning mechanism included in the third controller. It was performed on a simulated biped walker with a transfemoral amputation as a first validation of the controller before its implementation on a real prosthesis. The simulations highlighted the capacity of LWPR to act as the inverse internal model of the prosthesis joints and to provide accurate feed-forward torque commands to achieve a desired state, while being robust to speed changes. The contribution of the feed-forward component allowed to significantly decrease the gains of the feedback torque component, increasing the controller's compliance.