Versatile graceful degradation framework for bio-inspired proprioception with redundant soft sensors

Area of Science:

• Robotics
• Sensor Technology
• Artificial Intelligence

Background:

• Soft sensors are vital for intelligent soft robots but are prone to damage.
• Maintaining accurate proprioception during sensor degradation is a significant challenge.
• Redundant sensor configurations offer partial solutions but lack consistent reliability.

Purpose of the Study:

• To develop a novel framework for graceful degradation in redundant soft sensor systems.
• To enhance the reliability and robustness of soft robots in real-world applications.
• To improve proprioception accuracy despite sensor damage.

Main Methods:

• Proposed a framework integrating a stochastic Long Short-Term Memory (LSTM) and a Time-Delay Feedforward Neural Network (TDFNN).
• LSTM estimates healthy sensor readings for comparison and identifies abnormal data.
• Abnormal sensor readings are zeroed out, and TDFNN performs proprioception with processed data.

Main Results:

• Simulations with a 40-sensor musculoskeletal leg demonstrated the framework's effectiveness.
• Knee angle proprioception accuracy was maintained across four degradation scenarios.
• Mean proprioception error increased by less than 1.91° (1.36%) even with significant sensor degradation.

Conclusions:

• The proposed framework significantly enhances the reliability of soft sensor proprioception.
• This leads to improved robustness of soft robots operating in unpredictable environments.
• The approach offers a viable solution for maintaining soft robot functionality under sensor damage.