Cloud-Edge Hybrid Models for Autonomous Vehicle Data Processing

Abstract

Autonomous vehicles (AVs) generate massive volumes of data—from LiDAR, radar, cameras, and global positioning systems—necessitating comprehensive and efficient data processing architectures. Cloud-only models face latency, bandwidth, and connectivity limitations, while solely edge-based approaches may lack the necessary computational resources for complex AI processing. Hybrid cloud–edge frameworks balance these constraints by enabling low-latency, safety-critical functions to run onboard or at nearby edge nodes, while resourceintensive tasks like deep learning model training and system-wide analytics are handled in the cloud.

This paper investigates cloud–edge hybrid models tailored for autonomous vehicle data processing, with an emphasis on minimizing latency, optimizing bandwidth usage, ensuring energy efficiency, and safeguarding data privacy. The study explores frameworks such as the Cloud2Edge elastic AI inference engine—where model prototyping occurs in the cloud and deployment happens via hardware-in-the-loop testing on edge computing units—and systems like SAGE that dynamically offload computationally heavy AI modules to the cloud to reduce in-vehicle energy consumption and data transmission volumes. Additional models, like the PI-Edge low-power edge framework, demonstrate practical edge-cloud coordination for managing multiple autonomous driving services in real-world settings.

Through simulation-based performance assessments, the paper evaluates hybrid architectures in terms of latency, processing time, energy consumption, bandwidth usage, and scalability. Results indicate that hybrid models significantly outperform standalone edge or cloud systems in delivering real-time perception, decision-making, and vehicle control while maintaining energy efficiency and operational resilience even under intermittent network conditions.

The workflow details include cloud-based prototyping and large-scale training pipelines, edge deployment via automotive ECUs or embedded platforms, and runtime decisions dictating task placement depending on context and resource availability. The study sheds light on key design considerations and best practices for implementing hybrid cloud–edge architectures in autonomous vehicles, offering a viable roadmap to harness the benefits of both paradigms for enhanced performance and safety.