Cloud-Assisted Autonomous Driving over Wireless Network

Autonomous driving systems utilize machine learning models to identify the right strategy to control vehicles. Since these systems are safety-critical, high accuracy, especially in complex scenarios, is increasingly ensured by designing larger machine learning (ML) models that require higher execution time and memory. Such larger ML models are challenging to process on onboard systems, as specialized processors with large memories tend to be expensive. One possibility of utilizing such large ML models is to leverage cloud computing in real time. However, cloud services can only be accessed using the wireless network, which can vary depending on the location and network congestion. In this work, we design a system that adapts to such dynamic network conditions, where the amount of data sent depends on the available network bandwidth. This adaptation technique (i) tunes the quality of the sensor data and (ii) depends more on the LIDAR data for local processing when the network fails. Our system caches additional data, such as HD maps, from the cloud whenever it predicts that the network bandwidth is about to reduce. Our preliminary results using an existing autonomous driving simulator called Carla shows that our technique is effective in practice.

I. Introduction

While autonomous driving systems have significantly advanced, their promise of full autonomy has not been achieved yet. For example, The key reason is the presence of challenging edge cases, especially ones not commonly observed during training of the machine learning models [9], [23]. The standard strategy of improving the accuracy has been to design larger and more compute-intensive models [16]. Running such models requires specialized processors called graphical processing units (GPUs) with larger memories. However, such GPUs with large memories tend to be significantly more expensive, as shown in Fig. 1. Thus, a safer and more efficient strategy for enabling autonomous driving is needed.