Computer architectures for robot control: a comparison and a new processor delivering 20 real MFLOPS

The author considers the resources required to implement robot trajectory, kinematic, dynamic, and servo algorithms at rates high enough to ensure that the computation rate does not limit performance. He examines current computer architectures in general and those intended for robot control. He then describes a very-long-instruction-word (VLIW) processor designed for robot control. This processor, called JIFFE, was designed to run close to its peak performance even on the random scalar operations characteristic of robotics algorithms. JIFFE performs nearly 20 MFLOPS (million floating-point operations per second) on practical scalar robotics code, while maintaining high performance on poorly structured code and executing some matrix code at 40 MFLOPS. Because JIFFE was designed to be a good research tool, it can be easily programmed in C. A variety of compiler features help generate efficient code, even from voluminous machine-generated expressions