Manipulator arm construction has changed little over the decades and is unlikely to change radically in the near future. The mechanical design necessary to achieve dexterity results in a system with complex dynamic properties. However, many manipulator manufacturers choose to ignore this complexity, concentrating on the mechanical design aspects rather than the design of the dynamic controller. In most cases, simple fixed-parameter single-loop PID compensators are utilised. In spite of the fact that the compensators are implemented on programmable devices, there is simply not enough processing power available to implement an improved dynamic control strategy. A multiprocessor controller has been developed which allows all the hierarchical levels of a manipulator controller to be implemented. The major advantage of the new controller is its ability to handle complex and time consuming dynamic algorithms for positioning of the robot end effector. This has been accomplished by adopting a master/slave multiprocessor configuration comprising a 20 MHz IBM PC/AT (80386) with a number of DSP cards based around the NEC 77230 floating-point DSP chip. Analog and digital input/output interfaces are provided for reading position signals and providing command signals. Tile motivation for the provision of such a controller was the desire to implement linear and nonlinear self-tuning control strategies. Both centralised (multivariable) and decentralized (single-loop) control strategies are considered and the new controller caters for both schemes by virtue of (a) the master/slave configuration with individual DSP boards for each joint, and (b) inter-board communications, allowing joint interactions to be catered for. In the paper, some of the identification algorithms required to support the nonlinear self-tuning strategies are described and real-time results presented. These results demonstrate the operation of the new controller and indicate some of its capabilities.