Asservimento visivo basato su apprendimento per il controllo in retroazione di un robot soft biomedicale

Soft robotics is a growing field that focuses on building robots using soft and deformable materials, making them safer for patients in applications like minimally invasive surgery. Soft robots, built with flexible and deformable materials, offer innovative solutions for a wide range of applications, thanks to their unique capabilities and adaptability. They can have infinite degrees of freedom and can perform complex movements but controlling them is challenging. Machine learning methods provide a viable solution by training neural networks to learn a control policy for soft robots. This approach is particularly useful for artificial vision applications, where a camera captures 2D images to create an approximate model. In a thesis work, an object tracking algorithm is implemented using a learning-based approach, with a Multilayer Perceptron neural network on a robot, a single module of the the Stiff-Flop, actuated by pneumatic chambers. The network learns the appropriate response to reach a target position, and a feedback control system enhances the robot's positioning accuracy. Tests are conducted to track target points and validate the model's performance in the presence of external disturbances on the robot body, achieving a maximum positioning error of three pixels