Colloquia

Summary

    While walking is natural for most people, it is a difficult or even impossible activity for people with a spinal cord injury. These people can benefit greatly from an exoskeleton to balance and walk. The biggest challenge in controlling exoskeletons is their lack of flexibility in adapting to disturbances and unstructured terrain. A non-linear model predictive control(NMPC) policy would enable the exoskeleton to deal with this, while greatly increasing the flexibility and safety. Unfortunately, such an NMPC is computationally too expensive to compute in real time.
   
    This work combines two methods for learning an approximate NMPC via machine learning and evaluates its performance. In order to guarantee that the robustness of the NMPC is sufficiently transferred to the approximate NMPC, a statistical validation is performed according to Hoeffding's inequality. The outcome of the work is a machine-learning based approximate controller that is very fast while having similar performance and stability properties as the NMPC it is approximating. The NMPCs are approximated for a linear and a non-linear dynamic system. While the approximate NMPCs for both systems failed the validation, it is still possible to control and stabilize both systems. Evaluations show that the resulting approximate NMPCs are robust to a limited amount of input noise. The presented method, while usable for linear and relatively simple nonlinear systems, is insufficient to guarantee robust behavior for non-linear systems of a higher order, such as walking with a human in an exoskeleton robot . The is because the quadratic Lyapunov function used is unable prove Lyapunov stability due to the higher order dynamics.