Open-Source Educational Platform for FPGA Accelerated AI in Robotics

Abstract

Artificial Intelligence (AI) using neural networks is growing rapidly in the area of robotics and many tools have been developed in the last few years to utilize these networks. However, these tools are very abstract and do not provide deep knowledge on how the neural networks perform their computations. This makes it difficult for roboticists to understand and fully harness the power of AI. In this work, we present an open-source framework for designing and implementing a simple neural network targeting edge computing platforms. The framework goes step-by-step through the training, synthesis, and hardware implementation on a Zynq platform. The final hardware implementation is evaluated against a classical implementation in software. The platform was used in the Embedded Systems Course at the University of Southern Denmark.