A Scalable Embedded AI System for Autonomous Robotics Using Edge-Based Sensor Fusion

Abstract

Robotic systems that are centralised and rely on cloud networks are usually prone to high communication latency, bandwidth, and lower reliability, which limits their usefulness in autonomous applications that are required in real-time. In this paper, a scalable embedded architecture of autonomous robotics that uses an edge-driven multi-sensor fusion framework has been developed to enable overcoming these challenges. The presented system brings together heterogeneous sensors, such as vision, inertial, and ranging units, though the hybrid fusion approach, which will involve Extended Kalman Philtres to provide robust state estimation with a lightweight CNNLSTM model to improve multimodal perception. The architecture is sequenced to quantize models and prune models to a small size to be deployed on edge deployment platforms with limited resources to be provided to them to provide low-latency and energy-efficient service. Experimental verification on an embedded robotic platform shows better performance of perception in terms of accuracy, F1-score, mean Average Precision, and lower localization error and greater resistance to sensor noise than single-sensor and cloud-based methods. This system has a low latency, constant frame rates, and low power usage, and it shows that the system can be used in their deployment on the edge. Altogether, within the context of the study, there is a set of contributions to the modular and scalable embedded AI framework, hybrid classical-deep sensor fusion model, and a comprehensive assessment of the accuracy-latency-power trade-offs of the next generation autonomous robotic systems.