• visual servoing
• visual recognition
• swarm robotics
• motion coordination

The goal of this thesis project is to provide a fish-shaped robot prototype the ability to coordinate its motion with other swarm members. At first, a theoretical analysis about how to pursue other robots (pure pursuit guidance) is presented; after that, an analysis is carried on about how to maintain a parallel linear formation (constant bearing guidance) to swipe larger coastal areas with a swarm. The realization of these techniques is strictly linked to a reliable method of image's object detection: the introduction of a convolutional neural network (YOLO, You Only Look Once) was fundamental to determine presence and relative location of other swarm members inside the field of view. All these information are held together by the development of a ROS code, divided in different packages. The most basic package has the role of controlling the robot's motion: storing sensor data, evaluating control action from a received desired command, sending pulse width modulation input to motors, changing LEDs color are just few of the low level operation that have been implemented. A second package implements the neural network, providing a message containing information relative to fish position on the current image grabbed from one of the three available cameras. These messages are used by some higher level scripts, that produce the desired motion taking into account visual information. An analysis of the obtained results is carried over, alongside critical aspects of the prototype's development.