Hardware acceleration for robotic perception

A significant portion of current and future space missions involves exploring unstructured environments, such as the surfaces of the Moon and Mars. These environments are characterized by complex and often unpredictable terrain, which presents unique challenges for autonomous robotic systems. Perception and mapping strategies play a crucial role in ensuring safe and efficient navigation in these environments. As a result, much research has been devoted to developing advanced techniques for perception and mapping in unstructured environments, and this is an active area of study in the field of robotics and space exploration. In particular, the increasing demand for real-time and reliable robotic perception systems has motivated the development of hardware acceleration algorithms. Hardware acceleration consists of the use of special-purpose hardware, which is specially designed to perform specific functions more efficiently than software running on a general-purpose CPU. Some of the advantages of hardware against software include speedup, lower power consumption, lower latency, and increased parallelism, at the cost of longer development times and reduced ability to update the designs after manufacturing. In the context of robotic perception, these algorithms aim to speed up the processing of visual and sensory data, allowing robots to make quick and accurate decisions in dynamic environments. For this purpose, hardware accelerators such as Graphics Processing Units (GPUs), Field-Programmable Gate Arrays (FPGA) and Application-Specific Integrated Circuits (ASICs) have been adopted. This Master’s thesis focuses on the implementation of a hardware acceleration algorithm for the calculation of the surface represented by a point cloud. The surface can be determined through the computation of its normal vectors, which provide valuable information about the surface shape. In particular, the thesis work is centered on the development of a hardware unit that exploits the Principal Component Analysis (PCA). Indeed, the PCA can be used to find the principal directions of a dataset that, in the case of a point cloud, return information about the vectors that are normal to the surface. The design of this computational unit was carried out to be implemented on a FPGA board. The results of this study demonstrate the feasibility of using hardware acceleration algorithms in robotic perception and provide insights into the trade-offs involved in the design of such systems.