Learning to control a real micropositioning system in the STM-Q framework

This paper presents a new reinforcement learning algorithm named Short Term Model-based learning (STM-Q) able to learn quickly to control stochastic systems. Our objective is to control the positioning of small mechanical parts with a micromanipulation system for microfactory applications. The micropositioning platform is made of three linear actuators that move points to push the object toward a goal position. At the micrometric scale, it is really difficult to model the behaviour of manipulated objects. Moreover, due to the gap between real world and discrete world perceived by the camera, the motion of parts can be seen as a little stochastic. To overcome these problems and design an automated device, we use reinforcement learning methods and develop a new model-based algorithm taking into account the stochastic behaviour of the positioning task.