Robot manipulator control is highly complex because of the intensive computations of inverse kinematics and inverse dynamics and the non-linearity associated with these. With the advent of the Field Programmable Gate Array (FPGA), the compact and efficient realizations of embedded complex structures were made possible.


This paper addresses the methodology for efficient and dense implementation of the real-time inverse solution in FPGA using backpropagation neural networks, exploiting the parallelism in all the stages of the backpropagation algorithm, and hence the improvement of speed compared to the software implementation of the algorithm.


The methodology enables the improvement in performance-to-cost ratio and the order of magnitude ratio with fine usage of FPGA resources. This implementation can be adopted as a step for building inverse solutions for other manipulators. The special implementation is designed here for the reuse as well.