Robotics and Biology Laboratory

Learning of Drift-Prediction Models to Increase the Accuracy of Air-Mass Controllers



Mohamed Rekik Farouk

Adrian Sieler

Oliver Brock


The soft hands developed at RBO are controlled by changing the enclosed air-mass inside the different air-chambers of the actuators that make up a hand. So far the controller relies on a model learned from data to keep track of the air-mass. This approach is tailored to data from a pressure sensor and binary valves that can only be in two states (open/closed). Since the model is purely data-driven, the controller is subject to drift. Drift in this context means that the difference between the estimated air-mass and the actual air-mass is increasing. In addition, opening and closing the values at high frequency leads to a jittery motion of the pneumatic actuators. Most recently we investigated the usage of proportional valves and air-flow sensors to reduce the drift of the controller as well as to make the motion of the actuators smoother. The current version of the controller developed for this setup still exhibits drift over long control sequences. Therefore, the goal of this project is to learn a drift prediction model for a given air-mass controller. This prediction model should then be used to adjust the current air mass estimate after a control update. The approach developed in this project should be applicable to booth controllers - with binary or proportional valves.

Beschreibung der Arbeit

In this thesis you will be guided by the following work-flow to investiagte different data-driven techniquies to counteract the drift an air-mass controller:

  1. Collect a diverse set of data to be used for learning the drift-prediction model.
  2. Investigate different learning algorithms to obtain an accurate drift-predictor.
  3. Incorporate the learned model into the control flow.
  4. Analysis the effect of the drift-prediction model as compared to the baseline controller without drift handling.