Cloud technology and Data is the lifeline of Autonomous Driving development, and its management, reuse and analytics has a large impact on development cost and timelines.
By Rajiv Tandon
The automotive industry is going through a ‘Tech-Tonic’ change, especially with the large focus on Autonomous Driving (AD) development in Cloud. An autonomous vehicle (AV) processes the inputs (data) it receives from various sensors to make the driving decision. A typical autonomous vehicle (AV) generates upwards of 40 terabytes of data in a single day and for autonomous driving (AD) development this data is ‘Gold’!
The data generated whilst testing AVs must be stored and reused to train & refine the AD software and various Artificial Intelligence, Machine Learning and Deep Learning software components.
With increasing complexity in the development cycle of a vehicle, there is a keen focus on data management and dynamic-distributed computation, as it can have a critical impact and costs can rise exponentially if computation and data are inefficiently managed.
Figure 1: Autonomous Vehicles Generate up to 40 TB of data per day
Essentially all the data that is generated must be classified as useful – not useful so that it can be re-used for the development process to save time and cost. Additional data (Petabytes) is generated during the downstream development systems and process for e.g., when one executes validation processes like Software in Loop (SIL), Model in Loop (MIL) Hardware in Loop (HIL) and Processor in Loop (PIL). Each version of these validation tests must be stored and made accessible to the engineers so that the algorithms can be evaluated and fine-tuned. And this activity must be done at ‘Hyper-Scale’ for e.g., an AV must be tested in the virtual world (simulation) against millions of scenarios and if 1 virtual scenario is 2 minutes long it will take 1000 days to execute, but at ‘Hyper Scale’ one can execute this in 10 days.
The automotive industry is in the initial phase of adopting Cloud technology & techniques and there is lot that can be done in terms of application and innovations.
This article talks about three key questions around Cloud
1. What can be done with Cloud?
2. What are the typical challenges faced whilst implementing?
3. And how one can overcome these challenges?
Let’s look at each question in detail.
1. What can be done with Cloud?
The Tera bytes/Peta bytes of data that is generated constantly is ‘Curated’ to ensure that only the good quality and required data is captured for e.g., sometimes the data is not readable or is corrupt, such information is removed as it will take up space, and impact processing. Data is extracted and stored in correct formats in the right systems so that it that can be utilized by multiple downstream systems such as feature & function development, algorithm development, training, evaluation, and simulation etc. at Hyper-scale.
The activities to process the data are:
Capture of Raw Data from vehicles (through CAN, OTA, etc.)
Transforming captured into a structured database
Data Maintenance and Asset creation
Data filtering at point of collection. Large amount of data will require filter function quickly pull up relevant data
Data Cataloging is a process of creating meta data for each set of data which allows for efficient search e.g., finding of a Scenario where it is raining heavily on a crowded street at night
Ground Truth Creation and management
Scene Extract is the process of recreating/replicating real world traffic scenarios to virtual scenarios (synthetic data)
2. What are the typical challenges faced whilst implementing?
Capture and storage of data. This data can range in Petabytes and requires large storage infrastructure like ‘Cloud, or On-premises machines or servers’
Data Management and Modelling is the activity where a detailed structure and a strategy must be worked out for appropriate storage, classification, and data extraction
Meta Data of Data is the process where summary information must be created for each data set e.g., the date of test drive, data from camera or radar etc.
Analytics and Data mining is almost endless in the Cloud environment. For AD development some critical analytics should be enabled for e.g., engineers should be able to select multiple scenarios like weather, type of road (highway, tunnel), traffic etc. and asses the performance of their algorithms
Security of the data is paramount especially since data is provided by clients and multiple proprietary applications or codes are being deployed
Cost overrun is key challenge. In-correct estimation of required Cloud infrastructure & applications and poor data management techniques can lead to millions of dollars of additional costs.
3. How can one overcome these challenges?
Various options of computation. Cloud allows engineers to execute computation at hyper-scale. Cloud Infrastructure from multiple geographies across the world can be utilized to compute at scale. Additionally, multiple cloud platforms (Azure/AWS) and on-premises infrastructure can also be utilized simultaneously. For e.g., Virtual testing against millions of scenarios can be executed in a matter of 10 days – Hyper-scale
Various options for data storage. Organizations that were providing infrastructure in multiple industries have now directed their attention to automotive and different cloud services providers like Microsoft and Amazon provide a range of storage services.
Data Pipeline and addressing the management of sensor data which can be used by downstream systems for validation & verification. Cloud technologies allow engineers to organize/curate data so that they can pick and choose e.g., Data from Camera, Radar or Lidar, Driving on Highway at night.
Various Analytical and reporting engines are available in the Cloud environment for engineers to analyze, evaluate multiple scenarios
Application Monitoring is a key technique where engineers can monitor an application, or a cluster of applications are running across the globe and across infrastructure and to understand if downstream systems are running on the track or are any applications are down
Workflow Management systems help manage all the tasks listed above from approvals to scheduling to reporting and more
Figure 4: Cloud allows for multiple computation techniques and processes that can help OEMs advance towards Homologations
Cloud technology is truly a game-changer for the automotive industry. With the potential it brings in data management, hyper-scale computation, tremendous Analytics Cloud will help develop unique features, provide faster validations, and save huge costs and time in the development cycle.
Link: https://www.financialexpress.com/express-mobility/auto-technology/role-of-cloud-computing-in-autonomous-driving-development/2338925/?utm_source=pocket_mylist
Source: https://www.financialexpress.com