A small cruising-type test-bed vehicle named “Manta-Ceresia” is developed to ensure performance of various control architectures for Autonomous Underwater Vehicles (AUVs). Although it is so compact (13 kg in weight) that one researcher can carry and handle the vehicle, it is equipped with sensors and actuators required for prototype AUVs. Pairs of elevators and thrusters make it possible to move in 3 dimensional space. It is realized to swim along walls of a pool keeping constant distance automatically making use of 6 channels of range finder. Taking advantage of these characteristics, adaptive controller which requires comparatively long swimming for adjustment can be examined in a small pool. The performance of the constant-altitude-controller utilizing neural-network, which can accumulate experience by adaptation, is examined with this robot. Switching structure of neural networks is introduced to keep experience which is apt to be forgotten through additional learning. It is demonstrated by the developed test-bed vehicle that introduced switching system represents underwater terrain more precisely and the controller is adjusted appropriately.