Low-cost soft robotic gripper – combining textile sensors and a lightweight hybrid CNN–LSTM model

Soft robotic grippers are compliant systems for object manipulation and interaction. However, most existing studies focus on individual components, such as materials or structural design, with limited system-level integration. This work presents a soft gripper system that integrates textile-based sensing, compliant actuation, and embedded machine learning within a unified architecture. The CNT-textile stretch sensor exhibits a gauge factor of 5.7 and stable operation over 5000 repeated loading cycles. The pressure sensor, with a thickness of 0.77 mm, achieves a sensitivity of 0.47 kPa −1 , a response/recovery time of 60/55 ms, and durability exceeding 5000 cycles. The soft fingers are fabricated from 3D-printed TPU with a fin-ray structure, enabling adaptive and conformal grasping. A lightweight convolutional neural network–long short-term memory (CNN–LSTM) model (1.09 MB) is deployed on edge hardware for object classification, achieving an overall accuracy of 95.68%. All sensing, preprocessing, and inference processes are executed locally on the embedded platform. A label-only transmission strategy is adopted, reducing payload size by approximately 99.6% compared with raw data streaming. The proposed system features a simple architecture, low fabrication cost, and high reproducibility. It provides a scalable and efficient framework for intelligent soft robotic manipulation with strong potential for practical deployment.