This thesis presents the design and realization of two generations of robot elements that can be assembled together to construct amphibious mobile robots. These elements, designed to be individually waterproof and having their own battery, motor controller, and motor, have been used to actually construct a snake, a boxfish and a salamander robot. Central pattern generator (CPG) models inspired from those found in vertebrates have been used for online trajectory generation on these robots and implemented on their onboard locomotion controllers. CPGs proved to be an interesting way of controlling complex robots, providing a simple interface which hides the complexity of the robot to the end user. Online learning algorithms that can be used to dynamically adapt the locomotion parameters to the environment have been implemented. Finally, this work also shows how robotics can be a useful tool to verify biological hypotheses. For instance, the salamander robot has been used to test a model of CPG for salamander locomotion.