Falling-down avoidance control for acrobat robot by Q-Learning with Function Approximation

In this paper we consider a landing control of an acrobat robot by taking steps to avoid falling-down by applying Q-Learning with function approximation We use Q-leaning since it, is difficult to design controller systematically for an acrobat robot, a typical variable constraint and hybrid system. Since Q-learning requires large computation time on line, we use a function approximation to Q-table, and propose to update the Q-table and the approximate function alternatively. The validity of the proposed method will be demonstrated by numerical simulation and experiment.