Policy search reinforcement learning for automatic wet clutch engagement

In most existing motion control algorithms, a reference trajectory is tracked, based on a continuous measurement of the system´s response. In many industrial applications, however, it is either not possible or too expensive to install sensors which measure the system´s output over the complete stroke: instead, the motion can only be detected at certain discrete positions. The control objective in these systems is often not to track a complete trajectory accurately, but rather to achieve a given state at the sensor locations (e.g. to pass by the sensor at a given time, or with a given speed). Model-based control strategies are not suited for the control of these systems, due to the lack of sensor data.We are currently investigating the potential of a non-model-based learning strategy, Reinforcement Learning (RL), in dealing with this kind of discrete sensor information. Here, we describe ongoing experiments with a wet clutch, which has to be engaged smoothly yet quickly, without any feedback on piston position.