Jonas Koenemann

Thursday, November 13, 2014, 16:30 - 17:30

Humanoid Robots Lab Kitchen/Room 009, Georges-Koehler-Allee 074, Freiburg 79110, Germany

Abstract: In this work, we use Differential Dynamic Programming (DDP) to control a humanoid robot. DDP is a direct numerical optimization method which formulates objectives as simple cost functions. We use an implementation of DDP from the MuJoCo framework, a physics engine for fast dynamics computation of dynamic systems with contact. MuJoCo was specially designed for the purpose of control. We integrate the DDP algorithm in various control schemes for model predictive control (MPC). We use concepts of state prediction and feedback control. In the experiments we analyse and compare different control schemes reffering to the applicability of the resulting control inputs on a humanoid robot. In a concluding experiment, we show that we are able to control the humanoid robot HRP-2 with model predictive control and DDP.