Soft microrobots based on stimuli-responsive polymers [ ] are able to generate a wide variety of different locomotion patterns [ ]. Their inherent flexibility stems from their continuous actuation, which enables them to adapt their motion to changing conditions. However, the design of controllers for soft microrobots is a challenging task due to the lack of accurate locomotion models. Common data-driven approaches, on the other hand, require running prohibitive numbers of experiments and lead to very sample-specific results.
In [ ], we propose learning locomotion patterns using Gaussian Processes and Bayesian Optimization. We demonstrate that the proposed learning scheme is data-efficient and robust with respect to starting conditions and differences among microrobots. These qualities are achieved by comparing different settings for the Bayesian Optimizer on a semi-synthetic data set. The developed learning scheme is validated on an experiment with a soft microrobot experiment, resulting in a 115% improved locomotion performance with an experimental budget of only 20 tests.
Future work aims to extend the learning scheme such that the controller can adapt to changing environments by learning spatio-temporal correlations, as well as forgetting old uninformative data.