Gait Learning for Soft Microrobots Controlled by Light Fields

Abstract

Soft microrobots based on photoresponsive materials and controlled by light fields can generate a variety of different gaits. This inherent flexibility can be exploited to maximize their locomotion performance in a given environment and used to adapt them to changing conditions. Albeit, because of the lack of accurate locomotion models, and given the intrinsic variability among microrobots, analytical control design is not possible. Common data-driven approaches, on the other hand, require running prohibitive numbers of experiments and lead to very sample-specific results. Here we propose a probabilistic learning approach for light-controlled soft microrobots based on Bayesian Optimization (BO) and Gaussian Processes (GPs). The proposed approach results in a learning scheme that is dataefficient, enabling gait optimization with a limited experimental budget, and robust against differences among microrobot samples. These features are obtained by designing the learning scheme through the comparison of different GP priors and BO settings on a semi-synthetic data set. The developed learning scheme is validated in microrobot experiments, resulting in a 115% improvement in a microrobot's locomotion performance with an experimental budget of only 20 tests. These encouraging results lead the way toward self-adaptive microrobotic systems based on light-controlled soft microrobots and probabilistic learning control.

Task space adaptation via the learning of gait controllers of magnetic soft millirobots

“…However, the physical data has inherent uncertainty owing to the noise in the measurements and the variations during the experiments. To include these uncertainties in the model, overcome the sparsity in the data, and make probabilistic predictions at unobserved locations, we represent the reward function S(u) using GPs following the study in von Rohr et al (2018):…”

“…In this study, we choose the expected improvement (EI) as the acquisition function a acq (u) due to its better performance compared with its alternatives as demonstrated in von Rohr et al (2018). EI seeks the parameter set for the next step where the EI in reward function is the highest compared with the previously collected data and is defined in Jones et al (1998) as…”

“…There are emerging examples that demonstrate the application of this approach to optimize the locomotion performance of robots on different length scales (Calandra et al, 2016;Liao et al, 2019;Marco et al, 2020;Yang et al, 2018). Despite its potential, there are only a few examples that apply this method to address the controller challenge for untethered soft robots, such as in the gait exploration of a tensegrity system (Rieffel and Mouret, 2018), and the optimization of an undulating motion of a microrobot (von Rohr et al, 2018).…”

“…In our recent work in Culha et al (2020), we demonstrated the controller learning of walking soft millirobots using BO and GPs and showed the improvement of learning efficiency in means of transferring prior mean information between robots as a TL application example. We followed the initial example by von Rohr et al (2018), which designed a learning scheme by comparing different GP priors and BO settings on generating a semi-synthetic dataset that represents the estimated gait controller space and used this estimation to optimize the one-dimensional crawling gait of a light-driven soft microrobot. In our work in Culha et al (2020), we adopted the magnetic soft millirobots from Hu et al (2018) that lacked sufficient predictive kinematic models and was therefore controlled with an open-loop system whose multi-dimensional parameters were tuned manually.…”

Excursion Search for Constrained Bayesian Optimization under a Limited Budget of Failures

“…Therein, the aim is to efficiently exploit the observed data in combination with prior probabilistic models to estimate the global optimum from a few trials. In the context of robot learning, BO has been used to mitigate the effort of manual controller tuning, see, e.g., (Calandra et al, 2016;von Rohr et al, 2018;Rai et al, 2018).…”