FPGA-based Deep Learning Acceleration for Visual Grasping Control of Manipulator

Abstract

The vision-based robotic arm control system is an important solution for intelligent production, and the robotic arm visual grasping system based on deep learning is an important branch. Aiming at the requirements of fast visual recognition speed, low power consumption and high precision of mobile visual grasping robot, a deep learning target detection scheme based on FPGA hardware acceleration is proposed. Use Vivado and Petalinux development kit to build the software and hardware system, then deploy YOLOv3 model in the system. Experiments show that the solution meets the demand of robotic arm visual grasping, and the real-time performance is better. The recognition speed is 18 times that of the CPU, the power consumption is 1/13 of the GPU, and the cost is lower.