Gait learning for soft microrobots controlled by light fields

The study focuses on soft microrobots made from photoresponsive materials that can produce different locomotion gaits when controlled by light fields. Because accurate locomotion models are lacking and microrobots vary from sample to sample, traditional analytical control is not feasible. To address this, the authors propose a probabilistic learning approach using Bayesian Optimization with Gaussian Processes. They design the learning scheme by comparing different Gaussian-process priors and Bayesian-optimization settings on a semi-synthetic dataset, aiming for data efficiency and robustness to microrobot variability. The approach is then validated in microrobot experiments. The reported result is a 115% improvement in locomotion performance using an experimental budget of only 20 tests, supporting the idea of self-adaptive microrobotic systems based on light-controlled soft microrobots and probabilistic learning control.