On-line adaptable controller system for underwater robot

This paper describes the concept and the example of application of an on-line adaptable controller system which is based on “SONCS: self-organizing neural-net-controller system” proposed by Fujii et al. (1991). The SONCS´ basic concept is to adjust a controller using back-propagated error signals. For the error signals are calculated as the differences between target value of the central and sampled motion data from the actual robot, there is idling time in the system´s operation. Namely, the adaptation mechanism should wait until sampling of all the necessary data has been finished to adjust the controller. In order to adjust the controller more quickly to enhance the flexibility of the control system, the operations for controller adaptation should be executed parallel with the time-dependent processes, such as control operations data sampling, etc. In this paper, a new concept of on-line adaptable controller system is proposed. The system has two independent parts in which necessary processes for robot control, data sampling and controller adaptation are executed parallel with each other on a robot´s computer system with parallel processing capability. The proposed system is implemented onto the versatile test-bed robot “Twin-Burger” which has a transputer based multiprocessor system. And the performance of the system is examined through tank tests. The experimental results show that the controller is adjusted smoothly and becomes suitable for operating the robot in a short time even if the dynamics of the robot happens to be changed