Low-Power, Logic-Only FPGA Implementation of a Deep Reinforcement Learning Locomotion Policy on a Quadruped Robot
Résumé
Artificial Neural Networks (ANN) and Machine Learning (ML) have become valuable tools for robotic motion, but their high computational cost typically demands complex and energy-intensive hardware. Field-Programmable Gate Arrays (FPGA) offer a promising alternative for efficient online Artificial Intelligence (AI) model inference. This work presents an FPGA-based architecture capable of handling both sensor data processing and Artificial Intelligence model inference, entirely within the Programmable Logic (PL) of the device. Applied to the SOLO12 quadruped robot, the proposed design can efficiently replace the conventional software AI algorithm implementation on Raspberry Pi and its interconnection with a dedicated Master Board. Furthermore, the FPGA implementation of the ANN has been validated in simulation by connecting the FPGA to a computer running the SOLO12 environment, as illustrated in Fig. 1. This paper focuses on the hardware implementation and optimization of ANN on a small-scale FPGA, highlighting its potential as a compact and energy-efficient solution for robotic control.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |