Claude Sammut

Tak Fai's work continues a theme originally developed by Malcolm Ryan, namely, combining symbolic and numerical learning techniques. The task was to have a small humanoid robot learn to walk. Having 23 degrees of freedom meant that naive reinforcement learning could not work. Instead, a planner generated a sequence of qualitative actions. A constraint solver determined a valid range of numerical parameters needed to drive the motors and finally, an optimisation method guided trial-and-error learning to determine motor actions that produced a stable walk.