Realizing Target-Directed Throwing With a Real Robot Using Machine Learning Techniques
Malte Wirkus, José de Gea Fernández, Yohannes Kassahun
In ERLARS 2012 - 5th International Workshop on Evolutionary and Reinforcement Learning for Autonomous Robot Systems, (ERLARS-2012), 27.8.-31.8.2012, Montpellier, o.A., pages 37-43, Aug/2012. ISBN: ISSN: 2190-5576.
This paper presents a practical application of machine learning techniques in real-world robotics. Our goal was to make a anthropomorphic robot throw a ball into a bin that is placed at an arbitrary position in front of the robot. We use evolutionary machine learning to optimize a cost function based on a simulation model to aim at the target and generate the necessary motion to throw the ball at the position that is estimated by the simulation model. In order to compensate for the error in the simulation model, we trained an artificial neural network based on data from real-world task executions. We show that a simple simulation model can already result in good throwing performance if machine learning is applied to compensate for the resulting simulation error.