How cloud computing contributes to autonomous driving – a thought experiment.
During his guest lecture at Stanford University, Dean Paron (Microsoft Azure) mentioned a fascinating argument about the importance of cloud computing which stayed present in my mind until now: The amount of data generated by an autonomous car is so high that cloud computing is one of the enablement factors which make autonomous driving in its current fashion possible.
In this context, this blog post aims at creating an understanding of the storage of data required by an autonomous vehicle, how a cloud solution for autonomous driving is in general designed and which challenges of implementation exist for the future.
A thought experiment – the number of laptops an autonomous vehicle needs for data storage
To better understand the data storage requirements of current autonomous driving solutions, let’s imagine a hypothetical autonomous vehicle with a hard drive of the size of your laptop and no access to cloud storage (adapted from Dmitriev, 2017). In this thought experiment, we assume a laptop storage capacity of 500 GB.
To estimate the number of laptops required to store the vehicle data for one day, we need to understand the basic information an autonomous vehicle needs to store. Those kinds of data highly depend on the technology applied by the autonomous vehicle. As the management consulting firm McKinsey&Company (2017) points out, the most recent autonomous driving technology mainly processes three different kinds of data in a so-called hybrid approach: Camera, light detection and ranging (LIDAR) and radar data. All those technologies capture different elements of the environment and result in a combined data pool. This data is not only processed but needs to be stored for model training and simulation purposes (Liu et al., 2017). To be even more precise those three data classes need to be extended by ultrasonic sensor data and vehicle motion data (Dmitriev, 2017).
In this context, Dimitriev (2017) assigned data storage requirements to each of those data classes. However, a detailed analysis would extend the range of this blog post, so we use an aggregated number provided by the management consulting firm Accenture (2018). As pointed out in their industry report “Autonomous vehicles: The race is on”, they estimate the data storage requirements resulting from radar, LIDAR and camera to be between 4-10 TB per day depending on the number of sensors.
Coming back to our example: If this autonomous car would store all those data on a laptop hard drive of 500 GB, it would need to carry up to 20 Laptops per day of driving.
The design of current cloud solutions
As a MIT design shows, cloud solutions for autonomous driving basically consist of an information flow between the vehicle and a data processing platform within the cloud (Kumar et al., 2012). There are also solutions in which the cars interact with each other but this would extend the scope of this blog post (Kumar et al., 2012). Within this data flow all the sensor information are send to the cloud platform in which two major processes occur:
- Within the planning module, the sensory perception and localization data are mapped to plan the path for the car. An emergency module screens changes in those data to adjust the path or initiate an emergency brake (Kumar et al., 2012).
- The request module sends the data back to the car where the controller executes the received data (Kumar et al., 2012).
Image source: Kumar et al. (2012).
An outlook for future challenges
The cloud-based autonomous driving as described in the previous paragraphs requires a tremendous amount of data. For the future, this brings a range of challenges. Looking only at the storage will not be sufficient. As pointed out by Intel another major challenge is the data transmission: If one million people are driving in an autonomous car, this equals the data transfer of three billion people on their phone according to Intel (Bouland, 2017). Thus, in addition to providing the storage capacities, the data transmission capacities also need to be provided; not only locally but with high geographical coverage.
Furthermore, the information could be simplified for the autonomous car to reduce the amount of data to be screened visually. Just think about speed limits: Why should a car screen a speed limit if it is possible to transmit it via a wireless sensor?
In my personal view, the process of implementing autonomous cars on a large scale will still take several decades. This is the case not only because of the technology but especially because of a lack of surrounding infrastructure. I am looking forward to your thoughts!
Accenture (2018). Autonomous vehicles: The race is on; accessed from https://www.accenture.com/t20180309T093025Z__w__/us-en/_acnmedia/PDF-73/Accenture-Autonomous-Vehicles-The-Race-Is-On.pdf#zoom=50; 10th July 2018.
Bouland (2017). Autonomous car: ‘a huge data transmission challenge’; accessed from: https://newmobility.news/2017/09/19/autonomous-car-huge-data-transmission-challenge/; 10th July 2018.
Dmitriev, S. (2017). Autonomous cars will generate more than 300 TB of data per year; accessed from: https://www.tuxera.com/blog/autonomous-cars-300-tb-of-data-per-year/; 11th July 2018.
Kumar, S., Gollakota, S. & Katabi, D. (2012). A Cloud-Assisted Design for Autonomous Driving.
Liu, S., Tang, J., Wang, C., Wang, Q., Gaudiot, J. (2017). Implementing a Cloud Platform for Autonomous Driving.
McKinsey&Company (2017). Self-driving car technology: When will the robots hit the road?, accessed from: https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/self-driving-car-technology-when-will-the-robots-hit-the-road; 10th July 2018.
Users who have LIKED this post: