LoPECS: A Low-Power Edge Computing System for Real-Time Autonomous Driving Services

Jie Tang, Shaoshan Liu, Liangkai Liu, Bo Yu, Weisong ShiIEEE Access 2020 · DOI

Introduction

Autonomous driving systems typically require massive, power-hungry computing rigs that cost tens of thousands of dollars and consume thousands of watts. To operate safely, autonomous vehicles (AVs) must process massive amounts of sensor data in real-time for localization, perception, and decision-making. In standard prototype AVs, this requires a trunk full of high-end CPUs and GPUs, often costing upwards of $30,000. This computing payload severely limits the battery life and economic viability of ubiquitous, low-speed AVs like campus shuttles or delivery robots.

The DragonFly Pod, a low-speed autonomous vehicle used for testing the PI-Edge framework.

Furthermore, the software running on these robots—most commonly the Robot Operating System (ROS)—is resource-heavy, imposing high memory footprints and communication overheads that overwhelm constrained edge devices. This paper introduces $\pi$-Edge, a low-power edge computing framework featuring a lightweight OS, a heterogeneous task scheduler, and dynamic cloud offloading.

The $\pi$-Edge Framework

To squeeze maximum performance out of low-power hardware, the authors developed the $\pi$-Edge framework, which consists of three main pillars: $\pi$-OS, heterogeneous task scheduling, and dynamic edge-cloud offloading.

The PI-Edge system architecture, illustrating the interaction between the edge client device and the edge cloud.

$\pi$-OS: A Lightweight Middleware

The authors replaced ROS with a custom, ultra-lightweight middleware built on top of the Nanomsg networking library. $\pi$-OS uses shared memory for on-device communication and multicast for data broadcasting, eliminating the redundant data copying and deep network stack delays inherent to ROS. While ROS consumes 50 MB of memory just to start, $\pi$-OS has a memory footprint of just 10 KB.

Heterogeneous Task Scheduling

The system uses OpenCL to abstract underlying hardware (CPUs, GPUs, DSPs) and introduces a two-tier scheduler:

  • Inter-core: Uses a novel Diff-Min algorithm to assign independent tasks to different computing units based on relative performance differences.
  • Inner-core: Uses the HEFT (Heterogeneous Earliest Finish Time) algorithm to manage dependent sub-tasks within a specific computing unit.

By intelligently routing highly parallel tasks (like SLAM feature extraction) to the GPU and sequential tasks (like SLAM backend) to the CPU, the system maximizes the utility of mobile SoCs.

Dynamic Edge-Cloud Offloading

A coordinator dynamically decides if latency-tolerant tasks, such as object or speech recognition, should be offloaded to a local edge cloud. It calculates a cost function balancing energy savings against network latency, factoring in the vehicle's trajectory and the range of local cloud nodes. While SLAM requires updates every 5ms and must be local, recognition tasks can tolerate 100–500ms of latency.

Evaluation and Results

The system was evaluated on the "DragonFly Pod" hardware using an NVIDIA Jetson TX1. The evaluation utilized real-time sensor streams: a stereo camera (60 FPS), an IMU (200 Hz), and a microphone (44 KHz).

Key results include:

  • Extreme Power Efficiency: The entire software stack runs on the Jetson TX1 consuming only 11 Watts (3.5W for the GPU, 4.2W for the CPU).
  • Throughput and Latency: $\pi$-OS achieves 10 GB/s throughput with inter-node data transfer latency reduced to 0.18 ms.
  • Scheduling Speedup: The Diff-Min scheduler reduces total execution time (makespan) by an average of 10% compared to the standard Min-Min algorithm.
  • Battery Extension: Offloading speech and object recognition drops local power consumption to 5W, extending compute runtime by 2.5x (up to 5 hours on a 2200 mAh battery).

Conclusion

$\pi$-Edge demonstrates that real-time SLAM, object detection, and speech recognition can be achieved on affordable, low-power mobile hardware. While currently tailored for low-speed scenarios (~3 m/s), future work aims to support a wider variety of hardware accelerators (FPGAs, ASICs) and scale the framework for more demanding L3/L4 autonomous driving tasks.

Summary updated

BibTeX
@article{tang2020lopecs,
  author = {Jie Tang and Shaoshan Liu and Liangkai Liu and Bo Yu and Weisong Shi},
  title = {{LoPECS: A Low-Power Edge Computing System for Real-Time Autonomous Driving Services}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2970728},
}

Loading related content...

Figure