• grasp
• human-like
• neural network
• reactive grasp
• sensing
• softhand

This paper presents the implementation of a reactive grasp using the IIT/Pisa SoftHand-2 on moving objects. The initial framework involved mapping a series of human-like primitives onto a robotic soft hand, based on the Feix et al. taxonomy, used for grasping objects recognized through a neural network. The gripper was chosen to be a soft hand to adapt the grasp to the shape of the object. The hand, mounted on a manipulator, reaches the object with a human-like planner and implements the grasp primitive associated with the object's shape recognized through an RGB-D camera.
To intercept the moving object and make the grasp robust, infrared proximity sensors were employed to enhance the approach phase and potentially set a pre-grasp strategy. To use the sensors, ring-shaped supports were designed with CAD and 3D printed, wearable on the fingers of the hand, connected to an Arduino board enclosed in a support positioned on the wrist of the gripper to avoid limiting the manipulator's movement. Once in contact with the object, the information from the infrared sensors is used to close the hand, increasing the probability of capturing the moving object. After grasping, the data from the infrared sensors and the currents provided to the SoftHand motors are analyzed to detect possible grasp failures and, if necessary, adjust the control values associated with the robotic hand to prevent the object from slipping, thereby making the grasp primitive more robust.
In future work, it will be interesting to implement grasp primitives that involve interaction with the environment, such as sliding or flipping, which can be implemented on the IIT/Pisa SoftHand-2 thanks to its soft nature, allowing it to adapt to the environment it contacts.