Pneumatic Soft Grasping Robotic Devices Using Reinforcement Learning

Advances in soft grippers have achieved remarkable success in grasping tasks providing superior flexibility compared to rigid grippers. However, challenges still remain in limited loading capacity and understanding the working mechanism of the soft gripper due to its nonlinear behavior and deformation. In this study, by leveraging both soft pneumatic joints and rigid components, a hybrid grasping device is developed to provide superior dexterity compared to conventional electromechanical grippers. A grasping platform is established using a robot arm equipped with the designed hybrid gripper and a depth camera imaging system. Simple and difficult grasping tasks are conducted on multiple objects indicating different difficulty levels using imitation learning methods. Two model-based approaches including CNN and PointNet models are implemented and show good performance on grasping predictions. An average error of 5 mm towards desired grasping spatial locations is reached for the simple grasping task. It is envisioned that such platforms and reinforcement learning approaches can be applied to more complicated grasping problems including irregular shaped objects that requires tailored gestures and grasping trajectories.