In recent years, reconfigurable soft grippers have gained significant attention in robotics. Reconfigurable grippers can adjust their shape and function based on task needs. This adaptability allows for in-hand manipulation, which enhances the versatility of the gripper to a great extent. In this project, A versatile robotic gripper system has been developed, which allows the fingers to be reconfigured into various poses, including scoop, pinch, and claw, utilizing the base of the hand. Since each finger operates independently in this design, it has advantages not only in grasping diverse objects but also in performing intricate in-hand manipulations.
The previous finger design is inspired from the Fin Ray® Effect which is a mimic of the fish fins. It has a overall structure of triangle form, and reinforced with rigid beam, but the joint of the beam is soft so that this structure can easily bend and conform to the shape of the object.
Using Fin Ray Effect® for the gripper finger design have several advantages. This design alllows for a compliant and adaptive gripping mechanism. When the Fin Ray finger is pushed against an object, it conforms to the shape of the object, ensuring a secure and stable grip. Also this distributes the gripping force evenly across the whole surface, reducing the risk of damaging delicate objects. This is particularly important in the daily cooperation with human. Plus The design is light-weight and cost-effective, since the finger can easily be manufactured by 3D printing with a relatively low cost. Thus incorporating the Fin Ray Effect® in robotic grippers offers a balance of adaptability, efficiency, and practicality, making it a valuable tool in advancing robotic manipulation.
In recent years, the development of machine learning makes Graphical Neural Networks become one of the most effective tool in the graph processing domain due to its ability to represent complex patterns, the adaptability of the models. GNNs operate through a mechanism known as message passing between the nodes of the graph specifically designed for graph processing, Graphs are data structure that symbolize objects(or nodes), and their interconnections(or edges). This flexible model can represents various systems across diverse domains such as social networks,cellular structures, chemical compounds and more.
In a recent study , a comparison between Convolution Neural Networks and Graphical Neural Networks in tactile recognition is carried out showing that the accuracy of GNNs is 99.53%. The prediction accuracy is slightly higher than that of CNNs, but the training efficiency is much better than that of CNN. Which is because, the GNNs has better interpretability than CNN, as it can extract key information between node and edges but the CNN need to process the entire tactile images which contains less important information like background.
In conclusion, Using GNNs in this project has advantages in its Computational Efficiency Flexibility and potential.