Trial-and-error learning of repulsors for humanoid QP-based whole-body control

Whole body controllers based on quadratic programming allow humanoid robots to achieve complex motions. However, they rely on the assumption that the model perfectly captures the dynamics of the robot and its environment, whereas even the most accurate models are never perfect. In this paper, we introduce a trial-and-error learning algorithm that allows whole-body controllers to operate in spite of inaccurate models, without needing to update these models. The main idea is to encourage the controller to perform the task differently after each trial by introducing repulsors in the quadratic program cost function. We demonstrate our algorithm on (1) a simple 2D case and (2) a simulated iCub robot for which the model used by the controller and the one used in simulation do not match.