How Autonomous Vehicles Get Safer, Faster
It’s easy to get excited about the possibilities of fully autonomous vehicles. They promise to redefine cities, save lives, and unlock millions of hours of currently unproductive time.
But the road to fully autonomous vehicles (AV) remains littered with challenges. The perception models that power self-driving cars must sort through a seemingly infinite array of edge cases, be it novel traffic patterns or unpredictable behavior from drivers or pedestrians. Testing and training these models to handle such cases is a time-intensive process, with estimates from RAND Institute projecting that self-driving cars would need to travel 11 billion miles before delivering on the promise of being safer than a human driver. Deploying AV’s to train in the real world while they are still learning to drive creates fatal risks.
Still, the promise to add trillions of dollars of economic value remains, and self-driving car companies continue to pour billions of dollars into AV R&D, racing to be first to market.
That’s why we’re thrilled to announce our investment in Parallel Domain. Parallel Domain accelerates the time to safety for self driving cars by creating a dynamic, powerful software platform that builds better autonomous vehicle simulations faster.
Training and testing complex, risky tasks in a simulated environment is not a new concept. NASA famously built a synthetic reproduction of the moon to simulate the first moon landing, and flight simulators are used by every airline and air force in the world today to train their pilots.
But as we move from training and testing humans to training and testing machine learning models, new challenges emerge. Machine learning models only work as well as the training data they are given — low fidelity data is insufficient. Models also require massive quantities of new training data to learn how to handle edge cases. Both quantity and quality are essential.
This is what makes Parallel Domain’s technology so exciting.
What you see pictured here is how their platform can generate a highly detailed, realistic virtual section of a city in less than a minute — a task that would have been programmed by hand and taken teams of visual artists weeks in the past. Simulation teams can use Parallel Domain’s software to quickly spin up a new virtual world and manipulate any variable within it, from the number of lanes to the condition of the asphalt.
The potential need for this type of simulation platform is not limited to self driving cars. Anything autonomous that uses computer vision and machine learning to improve decision-making will require high-fidelity virtual worlds to test and prove safety: Robots, drones, and other autonomous systems.
But what made us most excited about Parallel Domain’s approach is the unique background and experience of its founder, Kevin McNamara. Kevin started his career building virtual worlds at Pixar (you may have seen his work in Brave) and at Microsoft Games Studio (Sunset Overdrive, Crackdown 3). He then joined Apple’s Special Projects Group, leading a team focused on simulation for autonomous systems. His experience leading teams at world-class engineering organizations makes him uniquely qualified to build large scale simulation software at Parallel Domain. He’s also as scrappy and entrepreneurial as they come, securing a large enterprise customer in NIO prior to raising a dime.
We’re thrilled to be partnering with Kevin and his rapidly growing team to deliver on the promise of safer self driving vehicles and autonomous systems.