Session: SYMP S-5: Mechano-Intelligent Robots

Paper Number: 147900

147900 - Information Perception in a Fabric-Based Soft Robotic Arm via Physical Reservoir Computing 

In recent years, soft robotics has garnered significant attention due to its potential for safe interaction with humans and adaptability in complex environments. However, controlling soft robots accurately presents a considerable challenge due to their inherent nonlinearity and uncertainty. Achieving high-fidelity sensing is crucial for enhancing control performance, particularly in tasks involving posture and contact estimation. Traditional sensing methods, such as camera-based systems and load cells, suffer from limitations in portability, accuracy, and high costs, while embedded sensors may compromise mechanical compliance and flexibility. This study explores an alternative approach by leveraging the nonlinear dynamics of soft robots—typically viewed as a hindrance to control—to extract complex information. Instead of relying on specialized sensors, we focus on collecting distributed pressure data (which is readily available) within a pneumatic soft robotic arm and apply physical reservoir computing techniques to simultaneously estimate posture (i.e., bending angle) and payload status (i.e., payload weight). The underlying concept is to exploit the rich nonlinear dynamic responses of the soft body, as represented by pressure distribution, for information perception. In the big picture, we are considering the soft robotic body itself as part of the sensory system.

 

Our findings validate the potential of the soft arm reservoir, which can accurately predict posture and payload mass with careful selection of training data. We thoroughly examined the minimum training data required for kinematic estimation and payload prediction while maximizing prediction accuracy with an average error within 15%. Correlation matrix of the sensor data collected from different input pressure and payload setup indicates the intrinsic nature of nonlinearity within this pneumatic dynamic behavior, which is crucial for physical reservoir computing. Furthermore, we explore the performance of reduced training data quantity and sensor configurations, revealing that payload estimation needs more extensive training data while both prediction error increases with the decrease of number of sensors. We also demonstrate the feasibility of implementing multi-layer perception for both payload and posture estimation using the same sensor data, requiring only minimal additional training, while there exists trade-off between the prediction range and the prediction accuracy with the increase of the training data. This capability is valuable in addressing the challenges of complex and realistic working conditions. In conclusion, this study not only showcases the effectiveness of employing the framework of physical reservoir computing to achieve “information perception through body mechanics” in a fabric-based pneumatic arm with limited sensor data but also paves the way for future advancements in soft robotics.

Presenting Author: Jun Wang Virginia Tech

Presenting Author Biography: Graduate research assistant

Authors:

Jun Wang
Zhi Qiao
Wenlong Zhang
Suyi Li