Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Design of Multimodal Soft Pneumatic Actuators

Saswath Ghosh, Sitikantha Roy

TL;DR

This work tackles the scarcity of public data for designing soft actuators by proposing a Performance-augmented Generative Design (PaGD) pipeline that uses synthetic Pneu-net data, dimensionality reduction with t-SNE, and Gaussian mixture modeling to generate novel designs. The workflow includes CAD visualization and ABAQUS finite element analysis to assess feasibility and performance, demonstrating bending, twisting, and multimodal actuation in generated actuators. Key contributions are the data-driven generation of multimodal Pneu-net designs, quantitative novelty/diversity metrics, and an automated end-to-end pipeline for design, visualization, and FE validation. The approach has potential to accelerate soft-robot design and enable more capable soft grippers, with future extensions to broader actuator families and public data sharing.

Abstract

The recent advancements in machine learning techniques have steered us towards the data-driven design of products. Motivated by this objective, the present study proposes an automated design methodology that employs data-driven methods to generate new designs of soft actuators. One of the bottlenecks in the data-driven automated design process is having publicly available data to train the model. Due to its unavailability, a synthetic data set of soft pneumatic network (Pneu-net) actuators has been created. The parametric design data set for the training of the generative model is created using data augmentation. Next, the Gaussian mixture model has been applied to generate novel parametric designs of Pneu-net actuators. The distance-based metric defines the novelty and diversity of the generated designs. In addition, it is noteworthy that the model has the potential to generate a multimodal Pneu-net actuator that could perform in-plane bending and out-of-plane twisting. Later, the novel design is passed through finite element analysis to evaluate the quality of the generated design. Moreover, the trajectory of each category of Pneu-net actuators evaluates the performance of the generated Pneu-net actuators and emphasizes the necessity of multimodal actuation. The proposed model could accelerate the design of new soft robots by selecting a soft actuator from the developed novel pool of soft actuators.

Generative Design of Multimodal Soft Pneumatic Actuators

TL;DR

This work tackles the scarcity of public data for designing soft actuators by proposing a Performance-augmented Generative Design (PaGD) pipeline that uses synthetic Pneu-net data, dimensionality reduction with t-SNE, and Gaussian mixture modeling to generate novel designs. The workflow includes CAD visualization and ABAQUS finite element analysis to assess feasibility and performance, demonstrating bending, twisting, and multimodal actuation in generated actuators. Key contributions are the data-driven generation of multimodal Pneu-net designs, quantitative novelty/diversity metrics, and an automated end-to-end pipeline for design, visualization, and FE validation. The approach has potential to accelerate soft-robot design and enable more capable soft grippers, with future extensions to broader actuator families and public data sharing.

Abstract

The recent advancements in machine learning techniques have steered us towards the data-driven design of products. Motivated by this objective, the present study proposes an automated design methodology that employs data-driven methods to generate new designs of soft actuators. One of the bottlenecks in the data-driven automated design process is having publicly available data to train the model. Due to its unavailability, a synthetic data set of soft pneumatic network (Pneu-net) actuators has been created. The parametric design data set for the training of the generative model is created using data augmentation. Next, the Gaussian mixture model has been applied to generate novel parametric designs of Pneu-net actuators. The distance-based metric defines the novelty and diversity of the generated designs. In addition, it is noteworthy that the model has the potential to generate a multimodal Pneu-net actuator that could perform in-plane bending and out-of-plane twisting. Later, the novel design is passed through finite element analysis to evaluate the quality of the generated design. Moreover, the trajectory of each category of Pneu-net actuators evaluates the performance of the generated Pneu-net actuators and emphasizes the necessity of multimodal actuation. The proposed model could accelerate the design of new soft robots by selecting a soft actuator from the developed novel pool of soft actuators.
Paper Structure (10 sections, 7 equations, 7 figures, 2 tables)

This paper contains 10 sections, 7 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Generative Design Methodology
  3. Training dataset preparation and preprocessing
  4. Design generation
  5. Design feasibility
  6. Results and discussions
  7. Evaluating machine learning models
  8. Performance of generated design
  9. Limitations and Future Directions
  10. Conclusion

Figures (7)

  • Figure 1: Flow of Performance-augmented Generative Design (PaGD) methodology for the design of soft actuators.
  • Figure 2: Visualization of parametric data generated in FreeCAD. Four actuator models representing different geometrical features are visualized in FreeCAD.
  • Figure 3: Flow diagram representing the visualization of a newly generated parametric data in FreeCAD open-source software.
  • Figure 4: (a) Visualization of pneu-net actuators using t-Distributed Stochastic Neighbor embedding (t-SNE) of parametric data. Dim_1 and Dim_2 are two embedding dimensions. (b) Gaussian mixture model used to learn the distributions of the data. Random sampling generates a novel design from the distribution.
  • Figure 5: A convex hull of training design data set. Comparison showing the diversity of the generated design data set with respect to training data.
  • ...and 2 more figures