Improvement of Visual Perception in Humanoid Robots Using Heterogeneous Architectures for Autonomous Applications

Humanoid robots find application in a variety of tasks such as emotional recognition for human-robot interaction (HRI). Despite their capabilities, these robots have a sequential computing system that limits the execution of high computational cost algorithms such as Convolutional Neural Networks (CNNs), which have shown good performance in recognition tasks. This limitation reduces their performance in HRI applications. As an alternative to sequential computing units are Field-programmable gate arrays (FPGAs) and Graphics Processing Units (GPUs), which have a high degree of parallelism, high performance, and low power consumption. In this paper, we propose a visual perception enhancement system for humanoid robots using FPGA or GPU based embedded systems running a CNN, while maintaining autonomy through an external computational system added to the robot structure. Our work has as a case study the humanoid robot NAO, however, the work can be replicated on other robots such as Pepper and Robotis OP3. The development boards used were the Xilinx Ultra96 FPGA, Intel Cyclone V SoC FPGA and Nvidia Jetson TX2 GPU. Nevertheless, our design allows the integration of other heterogeneous architectures with high parallelism and low power consumption. The Tinier-Yolo, Alexnet and Inception-V1 CNNs are executed and real-time results were obtained for the FPGA and GPU cards, while in Alexnet, the expected results were presented in the Jetson TX2.