Data-driven Actuator Selection for Artificial Muscle-Powered Robots

Taylor West Henderson, Yuheng Zhi, Angela Liu, Michael C. YipICRA 2021 · OpenReview

TL;DR

This paper introduces a data-driven approach to help engineers select the right artificial muscle actuator for their robotic applications. By leveraging machine learning on a comprehensive dataset of existing actuators, they've created a tool that simplifies the complex decision-making process and provides a benchmark for new technologies.

What Problem Does This Paper Address?

Artificial muscles, with their inherent compliance, flexibility, and compact nature, are increasingly attractive for building more biomimetic and safer robots. However, the sheer diversity of artificial muscle technologies (like SMAs, DEAs, SCPs, PZTs, and SFAs) and their varied properties (compliance, bandwidth, strain, stress, etc.) make it incredibly challenging for designers to choose the most suitable actuator for a specific application. There's no easy way to compare these actuators or to integrate new ones into the existing landscape. This lack of a standardized selection and benchmarking framework hinders the broader adoption of artificial muscle technology in robotics.

Methods

The core of this work is a data-driven approach using Support Vector Machines (SVMs) to classify and select artificial muscle actuators. Here's a breakdown of their methodology:

  • Data Collection and Normalization: They compiled a dataset of existing artificial muscle actuators by conducting an extensive literature review. This involved extracting quantitative performance metrics such as bandwidth, strain, stress, efficiency, and power density. To handle the wide range of values across these metrics, they applied a two-step normalization process: first, a natural logarithm transformation, followed by shifting to zero mean and scaling to unit variance.
  • Defining a Muscle Performance Space: To create a usable selection tool, they built a multi-class classification model. Since the data isn't always linearly separable and can have missing values, they opted for a strategy of training multiple binary SVM classifiers. Specifically, they trained SVMs on every possible pair of the five key features, resulting in 10 distinct bi-variate classifiers. For each pair of features, they used a "one-vs-one" policy to create multiple binary SVMs to handle the 7 actuator classes.
  • Confidence Score for Selection: To provide a more robust prediction than simple voting, they developed a confidence score that combines the number of votes from the binary classifiers with a measure of their decision certainty. This helps users make more informed choices, especially when confidence levels are close.
  • Hyperparameter Tuning: They performed cross-validation to find the optimal hyperparameters for their SVM models, particularly the regularization term (λ), and determined that a confidence score-based decision method was more effective than probabilistic estimation on their dataset.

Datasets

The paper introduces a novel, comprehensive dataset of artificial muscle actuators. This dataset was compiled from an extensive literature review and includes hundreds of data points (678 in total) across 7 types of actuators and 5 key performance parameters: bandwidth, strain, stress, efficiency, and power density. This dataset serves as the foundation for their machine learning models and is intended to be an open-access resource for the community.

Key Insights and Ideas

  • Bridging the Gap: The most significant insight is the creation of a practical, data-driven framework that demystifies artificial muscle selection. This tool lowers the barrier to entry for engineers and researchers who may not be experts in this specific domain.
  • Actuator Groupings: The bi-variate classifier plots reveal natural groupings of actuators with similar properties. For instance, SMA and SCP actuators tend to cluster together, as do DEA and EAP actuators, highlighting shared performance characteristics.
  • Feature Pair Importance: The results suggest that certain pairs of features are more discriminative than others for classifying specific actuator types. This provides valuable guidance on which parameters are most critical to consider when evaluating actuators.
  • Benchmarking New Technologies: The proposed approach offers a standardized way to benchmark new artificial muscle technologies against existing ones, facilitating their integration and comparison within the broader field.

Key Results

  • The developed SVM-based selection models achieve reasonable accuracy in classifying actuator types. For example, the paper shows top-1 accuracies for individual classes can be quite high (e.g., 100% for PZT with certain feature pairs).
  • The confidence score mechanism proved to be a more effective decision-making tool than probabilistic estimation on their dataset.
  • The models were tested on real-world robotic applications (e.g., artificial bicep, active endoscope, rehabilitative glove), demonstrating the ability to provide relevant actuator predictions with high confidence.
  • An open-access web toolkit has been developed, allowing users to interactively input performance criteria and receive actuator recommendations.

Limitations and Future Work

  • Missing Data: A significant limitation is the presence of missing data points for certain features across different actuators. This forced the researchers to train multiple bi-variate classifiers rather than a single multi-dimensional one, which can lead to more complex decision boundaries and potentially less optimal performance.
  • Dataset Size and Balance: While the dataset is comprehensive, it's acknowledged that improving its size and balance across all actuator classes could further enhance model accuracy.
  • Normalization Impact: The logarithmic normalization, while useful for handling wide value ranges, can cause data points to become clustered, potentially introducing errors in the SVM training.
  • Future Directions: The authors propose several avenues for future work:
  • Developing more sophisticated methods for handling missing data, such as smart imputation techniques (e.g., nuclear norm minimization, expectation maximization).
  • Exploring alternative normalization strategies.
  • Developing a data-driven regression strategy to predict optimal actuator configurations (length, thickness, material composition) based on desired performance metrics, moving beyond just selection.

Summary updated

Abstract

Even though artificial muscles have gained popularity due to their compliant, flexible and compact properties, there currently does not exist an easy way of making informed decisions on the appropriate actuation strategy when designing a muscle-powered robot; thus limiting the transition of such technologies into broader applications. What’s more, when a new muscle actuation technology is developed, it is difficult to compare it against existing robot muscles. To accelerate the development of artificial muscle applications, we propose a data-driven approach for robot muscle actuator selection using Support Vector Machines (SVM). This first-of-its-kind method gives users insight into which actuators fit their specific needs and actuation performance criteria, making it possible for researchers and engineers with little to no prior knowledge of artificial muscles to focus on application design. It also provides a platform to benchmark existing, new or yet-to-be discovered artificial muscle technologies. We test our method on unseen existing robot muscle designs to prove its usability on real-world applications. We provide an open-access, web-searchable interface for easy access to our models that will additionally allow for the continuous contribution of new actuator data from groups around the world to enhance and expand these models.
BibTeX
@inproceedings{henderson2021data,
  author = {Taylor West Henderson and Yuheng Zhi and Angela Liu and Michael C. Yip},
  title = {{Data-driven Actuator Selection for Artificial Muscle-Powered Robots}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  url = {https://openreview.net/forum?id=GAI7oHAhDf},
}
Figure