Active Learning Design: Modeling Force Output for Axisymmetric Soft Pneumatic Actuators

TL;DR

This work addresses the challenge of characterizing axisymmetric soft pneumatic actuators under external loading by integrating energy-based theory with data-driven methods. It develops two theoretical membrane models and then employs an active-learning, operator-learning neural network framework to predict force output across a parameterized design space, heights, and pressures. The approach is validated with an automated dataset of 22 EcoFlex membranes (n=22) yielding about 188k data points, achieving a RMSE around 4.0 N and enabling lift trajectories with a single pressure input for open-loop control. The results show that the learned model outperforms theory-based models and baseline curve fits, and can be used to optimize membrane designs for targeted lifting tasks, with potential for parallel actuation and real-time trajectory planning in soft robotics.

Abstract

Soft pneumatic actuators (SPA) made from elastomeric materials can provide large strain and large force. The behavior of locally strain-restricted hyperelastic materials under inflation has been investigated thoroughly for shape reconfiguration, but requires further investigation for trajectories involving external force. In this work we model force-pressure-height relationships for a concentrically strain-limited class of soft pneumatic actuators and demonstrate the use of this model to design SPA response for object lifting. We predict relationships under different loadings by solving energy minimization equations and verify this theory by using an automated test rig to collect rich data for n=22 Ecoflex 00-30 membranes. We collect this data using an active learning pipeline to efficiently model the design space. We show that this learned material model outperforms the theory-based model and naive curve-fitting approaches. We use our model to optimize membrane design for different lift tasks and compare this performance to other designs. These contributions represent a step towards understanding the natural response for this class of actuator and embodying intelligent lifts in a single-pressure input actuator system.

Figures (6)

• Figure 1: Schematic of the membrane (gray) with strain limiter (black) in the A. undeformed state, and B. deformed state while in contact with external force (blue).
• Figure 2: A. Design parameters for example membrane. B. Compressive Testing (top) Testing procedure for each membrane: expansion from a flat plane into a set height load-cell. Procedure is repeated for different heights. (bottom) Example of physical test measuring force for varying pressure at a set height. C. (left) Pressure, height, and six design parameters predict force output for a given membrane. (middle) Model (blue) overlaid with force-pressure training data at varying heights (0-70mm from purple to yellow). (right) 3-D visualization of 6-D design space. Example training set points in black, model from C in blue, model from D in green. D. Optimization for lifting task. (left) Physical testing to verify lift trajectories at a given mass. (right) Model planes for (blue) training parameters and (green) design parameters optimized to hit target trajectories (trajectories in black).
• Figure 3: Model Architecture. (Top) Characterization data, pressure and height, and design inputs, ring radius and width, membrane thickness, and contact radius, are inputs to the actuator model. (Bottom) The model solves for force output relative to pressure.
• Figure 4: Model RMSE [N] of best performing model hyperparameters tracked to the corresponding parameters: output force polynomial degree relative to pressure, multi-layer perceptron depth and width, and ring encoder MLP width.
• Figure 5: Experimental trajectories: (left) Pressure-height-force data from lifts at three different masses for each of five membranes (lines) with target points for each trajectory (circles). (right) RMSE between experimental trajectories and target way-points for each pairing of membrane and target.