Design, modeling and control of a manipulator with bio-inspired soft robotic gripper

Abstract

Safe manipulation of delicate and irregularly shaped objects remains a major challenge for conventional rigid robotic systems due to their limited compliance and adaptability during physical interaction. To address this issue, this project presents a soft robotic manipulation system that integrates a 4-degree-of-freedom (4-DoF) rigid manipulator with a bio-inspired soft gripper, enabling adaptive grasping while supporting both manual and autonomous control modes. The system is developed using a structured, model-based design approach, beginning with theoretical kinematic and dynamic analysis, along with payload torque calculations to guide actuator selection and mechanical configuration. Based on these analyses, the manipulator and gripper are designed and evaluated in Autodesk Fusion 360, including static stress and motion studies, and subsequently fabricated using 3D printing. For modeling and control, a URDF-based robot description is implemented in a ROS2 and MoveIt2 environment, enabling collision-aware motion planning, workspace analysis, and repeatable end-effector positioning in both simulation and hardware. Numerical workspace sampling shows that the manipulator achieves an asymmetric reachable volume of approximately 0.77 m³ under joint and self-collision constraints. Autonomous perception is achieved using a vision-based object detection pipeline, where four deep learning models - YOLOv11-m, Faster R-CNN, RF-DETR, and RTMDet-m, are trained and evaluated on an 11-class manipulation dataset. Among these, YOLOv11-m provides the best overall balance of accuracy and efficiency, achieving 94.7% precision, 94.3% recall, and a mAP@0.50:0.95 of 0.86. The complete perception, planning, control pipeline is validated through simulation and real-world experiments using a ROS2 distributed architecture. Control performance is evaluated across randomized target poses, in simulation autonomous control achieves a mean positioning accuracy of 98.99% with an average execution time of 4.06s, compared to 70.08% accuracy and 46.72s for manual control. Physical experiments on fragile objects demonstrate grasp success rates of 33% in fully autonomous mode with an average execution time of 23s and 66% positioning accuracy under manual teleoperation with an average execution time of 51s. The results confirm the feasibility of the proposed soft robotic manipulation system while highlighting current hardware and actuation limitations. This project provides a practical foundation for low-cost soft robotic manipulators and offers clear opportunities for future improvement in autonomous performance, making it suitable for applications in industrial automation, agriculture, and service robotics where safe and compliant interaction is required.