The integration of Edge AI in Autonomous Vehicle Navigation for Ultra-Low-Latency Applications

Abstract

Autonomous vehicle navigation must involve perception, localization, prediction and motion planning, all of which must be accomplished in real-time with tight timing. Cloud-centric architectures can offer large amounts of computing power but cause network propagation delay, queuing delay, and transmission delay, which are not suitable for manoeuvres that are critical for safety. This paper discusses the possibility of using Edge Artificial Intelligence (Edge AI) for ultra-low latency autonomous navigation, which involves onboard inference, roadside edge servers, and vehicle-to-everything (V2X) communication. A hybrid edge-cloud architecture is proposed by having latency-critical perception and local planning tasks on-board, whereas the computationally intensive tasks and cooperative perception tasks are selectively offloaded to nearby multi-access edge computing (MEC) nodes, and the non-urgent model updating and map enrichment tasks in the cloud. A deadline aware task-placement policy is based on the urgency of a task and available edge capacity, link quality and vehicle mobility. The simulation results show that edge-first execution can significantly decrease end-to-end latency over cloud-only processing under various network latency conditions, and that selective offloading can maintain the accuracy gain of a shared perception. The results indicate that in case of retaining critical functions locally, Edge AI can handle navigation functions with latency in the tens of milliseconds, while the capacity of MEC systems, wireless connectivity, and the additional load of model management still pose significant challenges. The paper introduces a structured architecture, an offloading policy and an evaluation framework for creating dependable Edge AI navigation systems.