Enabling to learn for force sensing: A coupling-decoupling under-actuated gripper with multiple-DoFs

Under-actuated robotic grippers have commonly appeared in industrial and logistics applications. However, existing grippers exhibit several main issues, such as low payload, insufficient force sensing and weak grasping stability, hindering their broader adoptions. To address these gaps, we propose a novel under-actuated gripper featuring two 3-joint fingers driven by a single actuator, achieving sensor-free force feedback through a hierarchical, dual-mode architecture that combines systematic mechanism modeling with a long short-term memory (LSTM) network. First, each finger employs an artfully six-linkage mechanism and two stacked four-bar linkages, which adaptively switch between parallel and enveloping grasp modes. Second, we devise theoretical models of inverse kinematics and power transmission based on the proposed gripper, accurately obtaining positions and contact forces. Third, we present a hierarchical force-sensing framework for the proposed robotic gripper, enabling precise force-feedback control across a broad force range (0.1 N–350N) via two complementary sensing principles. The work represents the first solution for high-load enveloping grasps, namely an LSTM model integrated with mathematical statistics and the constructed mechanism model, facilitating force-sensing solely through the actuator’s current loop. We implement position-based force control by the inverse kinematics model for delicate, low-force tasks, effectively overcoming inherent current fluctuations to achieve sub-Newton precision. Finally, a series of experiments are implemented to measure defined quantitative indicators, such as the payload, grasping force, force sensing, grasping stability and the dimension ranges of objects to be grasped, to validate the excellent performance of the proposed gripper. The hierarchical force-sensing framework redefines under-actuated gripper capabilities, resolving the long-standing trade-off between force range and precision in universal object manipulation. To our knowledge, this is the first deep-learning approach to estimate contact forces by analyzing actuator current loops, eliminating the need for dedicated force or torque sensors on the fingers. The introduced AI–Mechanism Synergy brings about a new paradigm for future robotic construction, which is promising in promoting highly adaptive, cost-effective robotic design across diverse real-world applications. A video was uploaded to https://youtu.be/LISP_DARQnM .