Control of a Pneumatic Robot Arm by means of Reinforcement Learning

We applied Reinforcement Learning on a real robot arm, in order to control its movements. The arm is actuated by two pneumatic artificial muscles, that expose a highly non-linear behavior. To enable a significant speed-up of the learning process, an empirical simulation is constructed, based on real robot observations. Furthermore, we introduce a learning strategy to facilitate learning. Using the simulation model and the learning strategy, Reinforcement Learning was able to quickly learn how to control the robot arm in simulation. Our experiments show that the obtained controller also performs well on the robot arm. Furthermore, Reinforcement Learning is able to adapt itself to changes in the behavior of the robot arm.