The past 10 years have seen a major leap in autonomous vehicle development — what began with a few retrofitted prototypes has now proliferated into fleets of near-production self-driving cars and shuttles. In the next decade, investment in autonomous driving technology will accelerate even quicker.
The industry will significantly ramp up autonomous driving pilot programs. Some of these will include removal of the onboard safety driver. Autonomous driving technology will be applied to a wider array of industries, such as trucking and delivery, moving goods instead of people.
Production vehicles will incorporate the hardware necessary for self-driving, such as centralized onboard AI compute and advanced sensor suites. These new platforms must be open, and able to support an array of diverse and redundant algorithms to power Level 2+ AI assisted driving and lay the foundation for higher levels of autonomy.
A critical area that requires more focus, however, will be in the infrastructure used to train, test and validate production-level autonomous vehicles.
Safe self-driving begins in the data center
At the core of this infrastructure is high-performance compute. With the power of GPUs, developers can easily manage the millions of hours of driving data needed to train and test the deep neural networks (DNNs) that will run in the car.
The need for data to train autonomous vehicles is enormous. A car collecting data on the road can amass terabytes of information in just a day – roughly two petabytes a year – which is then used to train DNNs to recognize objects, learn traffic laws and more.
Managing this data and putting it into action for DNN training requires high-performance compute working around the clock. GPUs in the data center are critical for this type of processing and the only compute solution for efficient autonomous vehicle training.
Just one algorithm can’t accomplish self-driving capabilities on its own. An entire set of DNNs, each dedicated to a specific task, is necessary for safe autonomous driving. These networks are diverse, covering everything from reading signs to identifying intersections to detecting driving paths. They’re also redundant, with overlapping capabilities to minimize the chances of a failure.
There’s no set number of DNNs required for autonomous driving. And new capabilities arise frequently, so the list is constantly growing and changing.
Auto makers, suppliers, software startups, robotaxi companies and more are investing in this data center infrastructure to build out their self-driving development capabilities. However, training DNNs on driving data requires more than just raw compute power.
Advanced learning methods
Rather than collecting millions of hours of driving data — and spending millions of more hours combing through it — on their own, developers can employ new, advanced learning methods that apply deep learning to the training process.
Active learning makes it possible to cleverly select data using a dedicated DNN on already-labeled data. The network sorts through unlabeled data, selecting frames that it doesn’t recognize, finding data that would be challenging to the autonomous vehicle algorithm.
By picking out these unique scenarios, the DNN can find valuable training data without human curators having to sift through mountains of data.
Federated learning enables developers to take advantage of shared driving datasets without giving up any proprietary information. It works by repeatedly training the model at different sites rather than a single database.
The updated model is returned to the developer’s centralized server, keeping the dataset within its own secure structure. This allows autonomous vehicle developers to take advantage of shared data, rather than collecting and managing it all on their own.
Developers can significantly accelerate the DNN training process with transfer learning. This process takes a trained DNN and expands its capabilities using new sets of data, rather than starting from square one.
These advanced training methods are most effective when run on a robust, scalable AI infrastructure. With high-performance, energy efficient compute, developers can manage massive amounts of data in parallel, shortening the development cycle.
Why the test fleet of the future is virtual
Finally high fidelity simulation makes it possible to validate these DNNs with greater efficiency, cost-effectiveness and safety than what is possible in the real world.
Autonomous vehicles must be able to handle both routine driving and edge cases before they can operate on public roads. Simulation makes it possible to drive millions of miles in virtual environments across a near-infinite range of scenarios — ensuring DNNs experience rare or dangerous situations before they hit the road.
Effective simulation must accurately represent the variety and unpredictability of the real world. This can be achieved with a cloud-based, hardware-in-the-loop simulation platform that is open to a variety of environment, traffic and vehicle dynamics models.
Such a validation tool can run in the data center in real time, performing testing at scale, with multiple units running a variety of tests in parallel. With this level of efficiency, simulation can achieve massive amounts of driving experience, and more importantly, each mile driven contains events of interest — including rare or hazardous scenarios.
As investment in autonomous vehicle deployment continues to ramp up, it is increasingly critical that auto makers, suppliers, startups and others implement a robust AI infrastructure. With a solid foundation to train, test and validate this technology, safer and more efficient transportation is just on the horizon.
by Danny Shapiro, senior director automotive, Nvidia
In his role, Shapiro focuses on artificial intelligence solutions for self-driving cars, trucks and shuttles. He serves on the advisory boards of the Los Angeles Auto Show, the Connected Car Council and the Nvidia Foundation, which focuses on computational solutions for cancer research. He holds a Bachelor of Science in electrical engineering and computer science from Princeton University and an MBA from the Haas School of Business at UC Berkeley. Shapiro lives in Northern California where his home solar system charges his electric, AI self-driving car.
Interested in AV testing? Check out the Autonomous Vehicle Technology Conference in Stuttgart, Germany, which will feature a dedicated Autonomous Vehicle Test & Development conference track, bringing together expert speakers on the latest AV test and development technologies and practises.