Soft Robots' Dynamic Posture Perception Using Kirigami-Inspired Flexible Sensors with Porous Structures and Long Short-Term Memory (LSTM) Neural Networks

Jing Shu, Junming Wang, Sanders Cheuk Yin Lau, Yujie Su, Kelvin Ho Lam Heung, Xiangqian Shi, Zheng Li, Raymond Kai-Yu TongSensors 2022 · DOI

TL;DR

This paper introduces a novel approach to soft robot posture perception by combining kirigami-inspired flexible sensors made from conductive sponge with Long Short-Term Memory (LSTM) neural networks. This system achieves high accuracy in predicting the dynamic bending angles of soft actuators while minimizing interference with their movement, paving the way for more sophisticated soft robotic control.

What Problem Does This Paper Address?

Soft robots, inspired by nature, offer significant advantages in human-robot interaction due to their inherent flexibility. However, accurately perceiving their posture and deformation is crucial for precise closed-loop control, especially in applications like rehabilitation. Existing soft sensors often struggle to balance flexibility with sensing capability, and their integration can affect the robot's natural movement. Furthermore, the non-linear and hysteresis-prone nature of soft materials makes accurate calibration a significant challenge. This paper aims to address these issues by developing a soft, flexible sensor that minimally impacts actuator dynamics and a robust calibration method capable of handling material complexities.

Methods

The core of this work lies in the novel sensor design and the machine learning approach for calibration.

  • Kirigami-Inspired Flexible Sensors: The researchers designed flexible sensors using a conductive sponge material. To enhance stretchability and reduce the resistive force during deformation, they incorporated a kirigami-inspired structure. This involves cutting specific patterns into a polyimide film backing, which allows the sponge to bend and deform without being directly stretched. The kirigami patterns were varied (two, three, and four units) to explore their impact on sensor performance. The porous nature of the sponge material also contributes to its low resistance.
  • Long Short-Term Memory (LSTM) Neural Networks: To calibrate the sensors and predict the actuator's posture, an LSTM neural network was employed. LSTMs are a type of recurrent neural network well-suited for handling sequential data and are known to effectively address the non-linearity and hysteresis inherent in soft sensor signals. The network was trained using sensor resistance data and corresponding actuator bending angles. Two calibration modes were explored: an "all-to-one" mode where multiple sensors contribute to predicting the overall actuator angle, and a "one-to-one" mode for predicting individual joint angles.

Datasets

The paper does not explicitly mention the creation of a new benchmark dataset. Instead, it describes the process of generating a dataset for training and testing the LSTM network. This involved:

  • Data Collection: The soft bending actuator was pressurized with varying pressures (0-200 kPa) and at different pressurizing rates (10, 50, and 100 kPa/s) for approximately 15 minutes. This generated a dataset of sensor resistance signals corresponding to specific actuator states.
  • Data Splitting: 70% of the collected data was used for training the LSTM neural network, and the remaining 30% was reserved for testing and evaluating its performance.

Key Insights and Ideas

  • Kirigami for Stretchability and Low Resistance: The combination of kirigami patterns with a porous conductive sponge proved highly effective. The kirigami structure allows for significant deformation without directly stretching the sensing material, while the sponge's porous nature keeps the overall resistance low. This is a clever way to achieve high stretchability without sacrificing sensor performance or significantly impacting actuator dynamics.
  • LSTM for Handling Soft Sensor Complexities: The choice of LSTM for calibration was critical. It successfully learned the complex, non-linear relationships and hysteresis present in the sensor's resistance signals, which are common challenges with soft, deformable materials. This demonstrates the power of deep learning in overcoming inherent material limitations.
  • Minimal Impact on Actuator Motion: A key design goal was to minimize the resistive torque introduced by the sensors. The paper quantifies this, showing a significant reduction in resistive torque compared to sensors made from conductive silicon, highlighting the practical advantage of their approach for soft robot integration.
  • "Three-Unit" Structure Optimization: The study found that the "three-unit" kirigami structure offered the best balance of performance, suggesting an optimal design for this specific application.

Key Results

  • Posture Prediction Accuracy: The kirigami-inspired sponge sensors, when calibrated with LSTM, achieved high accuracy in predicting actuator posture. The "three-unit" structure sensor demonstrated the best performance in the "all-to-one" mode, with a Root Mean Square Error (RMSE) of 3.85° and a Coefficient of Determination (R²) of 0.97.
  • Low Resistive Torque: The sensors generated a minimal resistive torque of 0.96 mNm at maximum bending, significantly lower than conductive silicon sensors, ensuring minimal interference with actuator movement.
  • Improved Performance with Multiple Sensors: The study indicated that using multiple sensors together (in the "all-to-one" mode) generally led to higher prediction accuracy compared to individual sensors.
  • Accurate Joint Angle Prediction: In the "one-to-one" mode, the system accurately predicted individual joint angles (PIP and MCP) of a wooden finger model, with RMSE values as low as 2.51° for the MCP joint.

Limitations and Future Work

  • Training Requirement: A primary limitation is that the sensor requires training with a neural network before it can be used for posture perception. This involves a data collection and training phase, which takes time (15 minutes for data collection and 10 minutes for training on their setup).
  • Future Work: The authors suggest exploring the sensor's performance on actuators with three degrees of freedom (3-DOF) motion. They also propose comparing different materials to further enhance sensor performance. The potential for applying these sensors to various flexible exoskeleton systems is highlighted as a promising avenue for future research.

Summary updated

BibTeX
@article{shu2022soft,
  author = {Jing Shu and Junming Wang and Sanders Cheuk Yin Lau and Yujie Su and Kelvin Ho Lam Heung and Xiangqian Shi and Zheng Li and Raymond Kai-Yu Tong},
  title = {{Soft Robots' Dynamic Posture Perception Using Kirigami-Inspired Flexible Sensors with Porous Structures and Long Short-Term Memory (LSTM) Neural Networks}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22207705},
}
Figure