A Learning Framework for Visual Depth-based Hardness Sensing in Soft Robotic Grasping

This thesis presents a novel framework for hardness sensing in soft robotic grasping, employing a single embedded depth camera. Leveraging convolutional neural networks, it captures intricate hardness features from depth images, achieving a commendable mean absolute percentage error (MAPE) of 0.46% for trained shapes and hardness. The innovative GripDepthSense3DNet outperforms existing networks with significantly fewer parameters and shorter training time. A dynamic tuning strategy enhances generalization and robustness, enabling seamless integration of new shapes and hardness levels. Demonstrating adaptability across various objects, this advancement holds great potential for automation and robotics applications.