Home > AI Solutions > Artificial Intelligence > White Papers > Dell Automotive Reference Architecture > Overview
Autonomous Driving (AD) was first implemented by the Department of Defense (DOD)-sponsored DARPA Grand Challenge[1] in 2004 for a $1M prize. The goal was to drive a 150-mile course in the Mojave Desert without a driver. More than 100 teams registered and the furthest distance any of the robotic vehicles achieved was 7.32 miles. Not one team finished the course[2].
In 2005, the DOD held the contest again with a $2M purse[3], with 195 teams registering. All but one team surpassed the 7.32-mile mark set by Carnegie Mellon’s Red Team in 2004. Five vehicles successfully completed the 132-mile course. What was the difference between 2004 and 2005? Many of the vehicles that competed on the course in 2005 were the same vehicles used in 2004, with many of the same sensors. What technological advances were made in just one year? According to Sebastian Thrun, Project Lead of Stanley[4], the 2005 Grand Challenge winner stated that the introduction of artificial intelligence (AI) was the difference, specifically, machine learning, probabilistic robotics, and distributed systems[5].
The era of autonomous vehicles was born, and every mobility company began grappling with the idea of how to make autonomous driving a reality. Autonomous vehicles (AV) require an increase and a diversity of sensors, including RADAR, LiDAR, video cameras, GPS, IMU, and SONAR, each having its own reference coordinate system. These sensor inputs must be combined or “fused” into a single coordinate system to formulate a 3D view for the self-driving vehicle through a process called “Sensor Fusion.” This computationally intense operation is performed at each defined time interval to create an updated 3D view for the AV. New capabilities and functions such as Sensor Fusion gave birth to a new more powerful kind of Electronic Control Unit (ECU) called a “Sensor Fusion ECU,” which is the brains of an AV.
The success of any self-driving vehicle lies in the ECU’s ability to interpret each sensor input and make appropriate decisions based on how well the AI models were trained to respond to the ego vehicle’s immediate environment. For human beings, this ability is called “situationally aware,” an ability or skill learned to always be aware of what is around and being able to respond to entirely new situations.
Achieving “situational awareness” for an AV is an enormous task. The Rand Corporation’s Driving to Safety study[6] states that for an autonomous vehicle to be equivalent in safety to a human driver requires driving 8.8 billion miles. It also states that “a fleet of 100 AV’s test-driving 24 hours a day, 365 days a year at an average speed of 25 miles per hour would take approximately 400 years.”
Driving for such an extended duration is not feasible and can only be achieved in a scalable simulation environment. Employing advanced High-Performance Computing (HPC) and AI methods to expand the environment to cover 8.8 billion simulation miles necessitates the use of tens of thousands of servers, over 100,000 GPUs, and an estimated 3.5 exabytes (EB) of storage[7].
DARA is an AI-based reference architecture tailored for the ADAS/AD market. It not only facilitates a complete AI model creation and development stack but also encompasses the processes of verification, validation, integration of AI models into an ECU or virtual ECU, data management, compute, storage, but also addresses the challenges of scale. This journey begins with a foundational starter kit that can be readily scaled to align with specific customer demand.
[1] https://en.wikipedia.org/ DellEMCPPCRSolution_14777wiki/DARPA_Grand_Challenge
[2] https://en.wikipedia.org/wiki/DARPA_Grand_Challenge%232004_Grand_Challenge
[3] https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2005)
[4] https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2005)
[5] https://www.youtube.com/watch?v=TDqzyd7fDRc
[6] https://www.rand.org/pubs/research_reports/RR1478.htm
[7] These values assume an SLA of 33 million simulation miles per day or 1 billion simulated miles per month. Reducing the SLA for simulation miles driven per day also reduces the infrastructure requirement.