In this Master thesis project, we want to develop learning controllers with Bayesian optimization for adaptive locomotion of soft microrobots.
Soft mobile microrobots based on micro-fabricated photoresponsive liquid-crystal elastomers are a novel kind of robot developed at the Max-Planck Institute Stuttgart. Driven by an external light field, the material deforms locally and, using these deformations, the microrobot can swim. The microrobots' body effectively behaves as a selectively-addressable continuum actuator with many degrees of freedom (DoFs). Many-DoFs systems are typically very hard to model and control accurately and, therefore, controlling the microrobots with classical control theory approaches is unfeasible. To actuate the microrobots we proposed learning the light-field directly from data using Gaussian Processes, used to build a probabilistic model of locomotion performance, and Bayesian Optimization, maximizing locomotion performance. So far this research has been focused on optimization in a setting where we restrict the parameter space of the controller, showing the effectiveness of the chosen approach.
To better exploit the many DoFs of the light-controlled microrobots, this project aims to extend the learning algorithm to work in a high-dimensional parameter space and discover new locomotion patterns automatically. The main challenges for the learning algorithm are data-effciency and robustness since the parameters are learned from noisy, experimental data. The project will involve finding a suitable method for high-dimensional locomotion optimization, as well as implementing and testing the developed learning control algorithm on the microrobotic system.
More information can be found in the project description and the research project page. Please do not hesitate to contact us if you have questions.