AWS (Amazon Web Services) is to advance Aurora's autonomous vehicle (AV) system by serving as the preferred cloud provider for machine learning training and cloud-based simulation workloads.
AWS says the Aurora Driver system consists of sensors that perceive the world, software that plans a safe path through it, and a computer that powers and integrates Aurora’s hardware and software with any vehicle platform.
According to AWS, Aurora uses the cloud to process trillions of data points each day and is now scaling its training workloads in the cloud to complete up 12 million physics-based driving simulations per day by the end of the year.
This builds on the petabytes of data it collects during real-world road tests, the company adds.
The Amazon subsidiary describes autonomous driving as an “immensely complex technological challenge” that relies on cloud computing to enable breakthroughs in perception, embedded computing, machine learning, decision making and advanced sensor technologies.
Aurora CEO Chris Urmson says: “Aurora’s advanced machine learning and simulation at scale are foundational to developing our technology safely and quickly, and AWS delivers the high performance we need to maintain our progress. With its virtually unlimited scale, AWS supports millions of virtual tests to validate the capabilities of the Aurora Driver so that it can safely navigate the countless edge cases of real-world driving.”
Aurora can use data from one testing simulation it observes in the real world to inspire hundreds of permutations in the AWS-powered virtual testing suite. The virtual testing helps train the Aurora Driver to more quickly and safely navigate complex situations, such as road construction, jaywalkers, and unprotected left-hand turns.
For example, the Aurora Driver completed nearly 2.3 million left-hand turns in simulation before attempting an unprotected version on a physical road.
The offline components of the Aurora Driver software stack all run on AWS, including the virtual testing suite, high-definition road maps, machine learning models and software development tools.
Specifically, Aurora uses an AWS service called Amazon SageMaker to create, run and continuously refine the machine learning models that enable its driving simulations. With this service, Aurora accesses Amazon Elastic Compute Cloud (Amazon EC2) instance types like P4d, which deliver high performance for machine learning training in the cloud.
Aurora uses AWS before developing simulations to securely store and process the petabytes of data it logs during real-world testing, and then train its machine learning models on that data. The pre-processing workloads run on Amazon Elastic Kubernetes Service (Amazon EKS) and a service for processing data in the cloud using open-source tools called Amazon EMR. Aurora’s machine learning training workloads then rely on AWS-optimised deep learning frameworks, such as TensorFlow and PyTorch.
Finally, AWS claims that Aurora orchestrates and auto-scales its simulation workflows over hundreds of thousands of concurrent virtual central processing units and thousands of concurrent graphics processing units with Amazon EKS and Amazon EC2, which provides accelerated computing instance types.
Swami Sivasubramanian, vice president of machine learning at AWS, says: “Our reliable infrastructure and comprehensive set of cloud services, including industry-leading machine learning services like Amazon SageMaker, provide the ideal foundation for Aurora to gain insights from the trillions of data points it generates every day to continuously enhance its technology.”
