[Paper Review] PI-Edge: A Low-Power Edge Computing System for Real-Time Autonomous Driving Services

Jie Tang, Shaoshan Liu+2 more|arXiv (Cornell University)|Dec 29, 2018

Autonomous Vehicle Technology and SafetyEngineering31 references17 citations

TL;DR

PI-Edge is a low-power edge computing system that enables real-time autonomous driving services on affordable embedded hardware by leveraging a heterogeneous runtime, a lightweight OS, and dynamic cloud offloading. It achieves full support for multiple autonomous driving workloads using only 11 W of power on an NVIDIA Jetson platform, marking the first complete production-ready edge system for autonomous vehicles.

ABSTRACT

To simultaneously enable multiple autonomous driving services on affordable embedded systems, we designed and implemented π-Edge, a complete edge computing framework for autonomous robots and vehicles. The contributions of this paper are three-folds: first, we developed a runtime layer to fully utilize the heterogeneous computing resources of low-power edge computing systems; second, we developed an extremely lightweight operating system to manage multiple autonomous driving services and their communications; third, we developed an edge-cloud coordinator to dynamically offload tasks to the cloud to optimize client system energy consumption. To the best of our knowledge, this is the first complete edge computing system of a production autonomous vehicle. In addition, we successfully implemented π-Edge on a Nvidia Jetson and demonstrated that we could successfully support multiple autonomous driving services with only 11 W of power consumption, and hence proving the effectiveness of the proposed π-Edge system.

Motivation & Objective

• To enable multiple real-time autonomous driving services on low-cost, low-power embedded systems.
• To optimize energy efficiency in edge computing for autonomous vehicles through dynamic task offloading to the cloud.
• To design a lightweight operating system that efficiently manages service isolation and inter-service communication.
• To fully utilize heterogeneous computing resources (CPU, GPU, TPU) in low-power edge devices.
• To deliver a complete, production-ready edge computing stack for autonomous vehicles.

Proposed method

• A custom runtime layer was developed to dynamically schedule and utilize heterogeneous computing resources (CPU, GPU, TPU) on low-power edge devices.
• A minimal, microkernel-based operating system was implemented to manage multiple autonomous driving services with low overhead and strong isolation.
• An edge-cloud coordinator was designed to monitor system load and energy consumption and offload compute-intensive tasks to the cloud when beneficial.
• The system uses lightweight inter-process communication (IPC) mechanisms to ensure low-latency coordination between services.
• The entire stack was deployed and evaluated on an NVIDIA Jetson platform to validate real-time performance and energy efficiency.
• Task offloading decisions were based on real-time monitoring of local resource utilization and energy consumption metrics.

Experimental results

Research questions

• RQ1 Can a low-power edge system efficiently support multiple real-time autonomous driving services simultaneously?
• RQ2 How can heterogeneous computing resources in embedded systems be fully utilized to meet real-time performance demands?
• RQ3 What is the optimal strategy for offloading tasks between edge and cloud to minimize energy consumption?
• RQ4 Can a lightweight operating system effectively manage service isolation and communication in autonomous driving workloads?
• RQ5 Is it feasible to build a complete, production-ready edge computing system for autonomous vehicles on commodity embedded hardware?

Key findings

• PI-Edge successfully supports multiple autonomous driving services on a single NVIDIA Jetson platform with only 11 W of power consumption.
• The system achieves real-time performance for all services under dynamic workloads, demonstrating low-latency task execution.
• The edge-cloud coordinator reduces client-side energy consumption by offloading compute-intensive tasks when beneficial.
• The lightweight operating system achieves low overhead and efficient inter-service communication, enabling stable service isolation.
• The heterogeneous runtime layer effectively utilizes CPU, GPU, and TPU resources, improving overall system throughput.
• This work presents the first complete, production-ready edge computing system for autonomous vehicles, validated in real-world embedded deployment.