Accelerating Robotic Intelligence: An FPGA -Optimized MLP Generator

• M. A. Pastrana UnB
• Lukas A. da Silva UnB
• Roberto de Souza Baptista UnB
• Daniel M. Muñoz UnB

Resumo

The rapid advancement of robotics has created a demand for intelligent systems capable of executing complex tasks with precision, adaptability, and efficiency. Multi-Layer Perceptrons (MLPs) neural networks have become a critical component in various applications, including object recognition, motion planning, and decision-making. However, deploying MLPs in real-time robotic systems presents significant challenges, such as stringent latency requirements, limited power budgets, and the need for high computational throughput. Field-Programmable Gate Arrays (FPGAs) provide a compelling solution by leveraging their reconfigurability, parallelism, and energy efficiency to accelerate MLP-based robotic intelligence. This paper introduces an FPGA-optimized MLP generator, a Python-based tool that automates the mapping of MLP models onto FPGA hardware. The generator incorporates advanced techniques, including custom precision floating-point arithmetic, to maximize performance. Experimental results demonstrate that 16-bit floating-point precision achieves an optimal balance between performance and resource utilization, delivering high accuracy with minimal power consumption. Furthermore, Hardware-in-the-Loop (HIL) validation on real-world scenarios confirms the tool's effectiveness, showcasing its potential for practical deployment in intelligent robotic systems.

Palavras-chave: VHDL, Codes, Accuracy, Service robots, Neural networks, Generators, Hardware, Robots, Field programmable gate arrays, Floating-point arithmetic, SoC FPGAs, Multi-Layer Perceptron, VHDL code generator