Real-time nonlinear model predictive footstep optimization for biped robots

A well known strategy in biped locomotion to prevent falling in the presence of large disturbances is to modify next footstep positions of the robot. Solving this complex control problem for the overall model of the robot is a challenging task. Published methods employ either linear models or heuristics to determine those positions. This paper introduces a new optimization method using a nonlinear and more accurate model of the robot. The resulting optimization problem to calculate the necessary footstep modification is solved by a direct shooting method. Using a problem formulation in an unconstrained way enables an optimization that performs in real-time rates. Further we present our overall framework that uses sensor feedback in trajectory generation. Experimental results of our biped robot LOLA show the effectiveness of the method under real world conditions.