Abstract
Soft robotic systems often struggle to integrate tactile sensing without compromising mechanical flexibility or introducing signal noise from the actuator's own movement. This paper introduces SoftMag, a soft pneumatic actuator that unifies a pneumatic muscle (M-PAM) with a magnetic tactile sensor. By employing a neural-network-based decoupling strategy, SoftMag effectively filters out mechanical parasitic effects, allowing for the real-time, non-destructive evaluation of object firmness. The system demonstrates high precision in tri-axial force sensing and object localization, achieving a strong correlation with ground-truth firmness measurements in agricultural applications.
Introduction
The ability to handle delicate objects while simultaneously assessing their physical properties is a critical requirement for intelligent soft robotics. However, integrating sensors into soft actuators presents two primary challenges: the mechanical parasitic effect, where actuator deformation generates false sensor signals, and hardware redundancy, which leads to bulky and inconsistent designs. This research addresses these hurdles by developing a "material-aware" soft actuator capable of intelligent manipulation and firmness assessment, specifically targeted at delicate produce like apricots.
Design and Fabrication
The SoftMag actuator is designed with a unified architecture that integrates a pneumatic muscle with a magnetic tactile sensing layer. The core of the device utilizes a porous silicone foam (Soma Foama 15) that houses four permanent magnets and a Hall-effect sensor. This shared material framework ensures mechanical coherence between the actuation "muscle" and the sensing "skin."
The fabrication process involves a multi-step molding technique to embed the magnets and sensors within the biocompatible polymer, resulting in a compact, sensorized actuator capable of significant bending and force application.
Modeling and Decoupling Strategy
To manage the complex interactions between the actuator's internal pressure and the magnetic sensor's output, the researchers utilized multiphysics modeling. COMSOL simulations characterized the coupling between solid mechanics and electromagnetics, revealing how internal pressure changes the magnetic flux independently of external contact.
To isolate true contact signals, a Multi-Layer Perceptron (MLP) was trained to predict and subtract the "noise" caused by internal pressure. Following this decoupling, a multi-task neural network—incorporating Long Short-Term Memory (LSTM) layers—was used to infer tri-axial contact forces ($F_x, F_y, F_z$) and contact locations ($P_1$ through $P_4$).
Experimental Results
The SoftMag system was validated through a series of indentation and actuation tests:
- Sensing Precision: The multi-task network achieved a Mean Absolute Error (MAE) of 0.085 N for shear force and 0.138 N for normal force, with a 96% accuracy in identifying contact positions.
- Mechanical Performance: The actuator demonstrated a payload-to-weight ratio of 8.9:1, successfully grasping objects as heavy as a 700g coconut.
- Firmness Evaluation: In real-world trials with apricots, the gripper's firmness estimates showed a Pearson correlation coefficient ($r$) between 0.818 and 0.920 compared to standard puncture tests.
- Durability: The actuators maintained performance over hundreds of cycles, showing no significant sensor drift or mechanical degradation.
Discussion and Conclusion
The SoftMag actuator represents a significant step toward truly integrated soft robotic perception. By treating parasitic effects as a source of proprioceptive data rather than just noise, the system achieves a high degree of environmental awareness. While current limitations include spatial resolution and manual fabrication variability, future work will focus on multi-fingered configurations and the integration of computer vision to enhance grasp planning and object recognition. This technology paves the way for automated, non-destructive quality control in the food and agriculture industries.
Summary updated