Closing the Control Loop with Time-Variant Embedded Soft Sensors and Recurrent Neural Networks

Area of Science:

• Robotics
• Materials Science
• Machine Learning

Background:

• Soft sensors are crucial for soft-bodied robot design and control.
• Real-world applications of soft sensors are limited by challenges in modeling nonlinear, time-variant systems.
• Accurate state estimation for complex soft systems is difficult.

Purpose of the Study:

• To present a learning-based approach for closed-loop force control using embedded soft sensors and recurrent neural networks (RNNs).
• To develop accurate and robust state estimation models for complex soft-bodied dynamical systems.
• To enable real-world applications of soft sensing technologies in robotic control.

Main Methods:

• Utilized learning protocols to train long short-term memory (LSTM) networks, a class of RNNs.
• Developed a state estimation model for complex dynamical systems.
• Implemented a feedback force controller for a soft anthropomorphic finger.

Main Results:

• Achieved accurate and robust state estimation models within a short training period.
• Developed a closed-loop controller with a control frequency of 25 Hz.
• Demonstrated an average control accuracy of 0.17 N experimentally.
• Showcased the controller's effectiveness despite sensor drift and hysteresis.

Conclusions:

• Soft sensing technologies, when combined with machine learning techniques like LSTMs, can be effectively used in real-world robotic applications.
• The proposed training methodology and control architecture are capable of handling complex soft-bodied systems.
• This approach facilitates the integration of soft sensors into practical, closed-loop control systems for robots.