Publication Details

SELECT * FROM publications WHERE Record_Number=11192
Reference TypeThesis
Author(s)Lamprecht, M.
TitleBenchmarking Robust Control against Reinforcement Learning Methods on a Robotic Balancing Problem
Journal/Conference/Book TitleMaster Thesis
AbstractAs simulators represent an inexpensive, fast and save test environment, it is a common practice to evaluate and optimize controllers of robotic systems in physics simulators, before applying them to the real robot. Identifying precise physics parameters that are required by the simulators is difficult. Thus one can observe a drop in performance when testing the designed controller on the real system. One approach to bridge this reality gap is to design robust controllers that are able to stabilize the system for parameter uncertainties. Within this thesis a Multi-Model Pole Placement (MMPP) and a H2 fixed-structure robust controller is designed. The robustness of both is examined on a seven DoF Schunk arm, that balances a ball on a plate. The static controllers that were designed with respect to a precise model are able to stabilize the Ball-on-Plate system for different radii and rolling friction coefficients of the ball. The balancing behaviour is simulated in the robot control system environment which is developed by the Honda Research Institute Europa. These controllers are compared against a robust controller that is designed using Proximal Policy Optimization (PPO). In the simulation the neuronal network trained with PPO revealed a faster balancing behaviour compared to the MMPP and H2 fixed-structure controller. On the real robot a Linear Quadratic Regulator (LQR) controller with an additional integrative part was able to balance balls with different radii and for different rolling friction coefficients. Using this controller the reality gap could be crossed.
Link to PDF


zum Seitenanfang