TL;DR
This paper introduces a data-efficient and robust learning framework for optimizing the locomotion of soft microrobots controlled by light fields. By leveraging Bayesian Optimization with Gaussian Processes, they achieve significant improvements in microrobot speed with a minimal number of experiments, paving the way for adaptive microrobotic systems.
What Problem Does This Paper Address?
Soft microrobots, particularly those made from photoresponsive liquid-crystal elastomers (LCEs), offer exciting possibilities for micro-scale manipulation and sensing. These robots can achieve diverse gaits by deforming in response to structured light fields. However, controlling these robots effectively is challenging due to several factors:
- Lack of Accurate Models: The complex, non-linear dynamics of soft materials at the microscale make it difficult to develop precise analytical models for their locomotion.
- Inter-Microrobot Variability: Manual fabrication processes lead to inherent differences between individual microrobots, meaning a controller optimized for one might not work well for another.
- Material Degradation: Over time, the materials themselves can degrade, altering the robot's response and requiring re-optimization.
Traditional methods of hand-tuning light field parameters are extremely inefficient and time-consuming. Data-driven approaches often require a prohibitive number of experiments and can result in highly specialized controllers. This paper aims to bridge this gap by developing a learning scheme that is both data-efficient and robust to these challenges.
Methods
The core of the paper's contribution lies in its probabilistic learning approach for gait optimization. They employ Bayesian Optimization (BO), a derivative-free global optimization algorithm well-suited for expensive and uncertain function evaluations, coupled with Gaussian Processes (GPs).
Here's a breakdown of their methodology:
- Gaussian Processes (GPs): GPs are used to model the unknown cost function (e.g., locomotion speed) as a distribution over functions. This allows them to incorporate prior knowledge about the system and quantify uncertainty in their predictions. They explore different kernel functions (SE, RQ, Matern, and a composite 2Mat kernel) to capture the underlying cost landscape.
- Bayesian Optimization (BO): BO iteratively selects the next experiment (i.e., the next set of light field parameters) to perform by maximizing an acquisition function. This function balances exploring uncertain regions of the parameter space with exploiting regions predicted to have good performance. They compare different acquisition functions like Probability of Improvement (PI) and Expected Improvement (EI).
- Semi-Synthetic Benchmark: To rigorously test and compare different BO settings (kernels, acquisition functions, hyperparameter learning strategies) without the cost of numerous real-world experiments, they create a "semi-synthetic" benchmark. This involves sampling data from a real microrobot on a coarse grid and then interpolating to create a continuous cost function. This benchmark is augmented to simulate inter-microrobot variability.
- Hyperparameter Optimization: They investigate the impact of learning the GP hyperparameters (e.g., kernel length scales, signal variance) during the BO process, comparing it to using fixed hyperparameters.
Datasets
- Real-world experimental data from [6]: This data was used to estimate prior information for the Gaussian Process hyperparameters, including the mean, signal variance, and noise variance.
- Semi-Synthetic Dataset: This dataset was generated by interpolating and smoothing data collected from real microrobot experiments. It was designed to simulate inter-microrobot variability and was used to benchmark different BO settings.
- Newly Tested Microrobot: A distinct microrobot, not used for the benchmark, was employed for the final experimental validation to demonstrate the robustness of the learned scheme.
Key Insights and Ideas
- Data Efficiency is Paramount: For micro-robots, where experiments are costly and time-consuming, a learning scheme that requires very few samples is crucial. BO with GPs excels here by intelligently selecting the most informative experiments.
- Robustness to Variability: The ability of the learning scheme to generalize and perform well across different microrobot samples is a critical requirement. The semi-synthetic benchmark was specifically designed to test this.
- The Power of Composite Kernels: The 2Mat composite kernel, combining two Matern kernels with different length scales, showed superior performance in the benchmark, suggesting that more expressive kernels can better capture the complex cost landscapes of microrobot locomotion.
- Hyperparameter Learning Matters (Sometimes): While not always leading to better results, learning GP hyperparameters during BO can significantly improve robustness, especially when initial fixed hyperparameters are poorly chosen or when dealing with limited data.
- Pessimistic Priors for Better Exploration: Using a pessimistic prior mean for the GP (assuming most controllers perform badly) was found to yield a better trade-off between exploration and exploitation, leading to more effective learning.
Key Results
- 115% Improvement in Locomotion Speed: In real-world experiments with a new microrobot, the developed learning framework achieved a 115% improvement in average speed compared to the initial controller parameters. This was achieved with an experimental budget of only 20 tests.
- Benchmarking Success: The semi-synthetic benchmark demonstrated that the 2Mat kernel with Expected Improvement (EI) acquisition function and learned hyperparameters was a highly effective configuration, significantly reducing normalized regret.
- Reduced Normalized Regret: The best performing configurations in the benchmark showed a median normalized regret of around 3.5% after 20 iterations, indicating a substantial improvement in finding near-optimal gaits.
- Robustness Demonstrated: The learning scheme successfully optimized the gait of a microrobot that was different from the one used to establish the initial prior information, highlighting its ability to handle inter-microrobot variability.
Limitations and Future Work
- Limitations:
- The current work focuses on learning 1D locomotion.
- The semi-synthetic benchmark, while valuable, is still a simplification of real-world complexities.
- The paper acknowledges that the "transient" behavior of the microrobot at the start of an experiment is not fully modeled.
- Future Work:
- Higher-Dimensional Control: Extending the learning to control more complex gaits and utilize the microrobot's multiple intrinsic degrees of freedom (DOFs) for planar locomotion.
- Problem-Specific Kernels: Developing kernels tailored to the specific physics and material properties of soft microrobots.
- Time-Varying Cost Functions: Enabling microrobots to adapt their gaits to changing environments or degrading material properties by learning time-varying cost functions.
Summary updated