Parallel Domain was founded in 2017 by a team of autonomous systems, simulation, and visual effects experts. The company offers a data generation platform capable of providing fully-labeled synthetic data and configurable, detailed, massively-scalable simulation environments. Toyota AI Ventures invested in Parallel Domain because synthetic data can play a major part in wide-scale deployment of autonomous vehicles (AVs).
This Q&A with Kevin McNamara, founder and CEO of Parallel Domain, looks at the data problem that exists in the world of AVs, and how Parallel Domain is providing new methods to safely and efficiently create that data. Kevin also describes the benefits of using synthetic data in AV testing, and why machine learning is the future of vehicle autonomy.
What were you doing before Parallel Domain?
I’ve always been fascinated by the possibilities that virtual worlds — which could be considered “parallel domains” — could enable. This fascination led me through my career at Pixar, Microsoft, and Apple Special Projects Group, where I contributed to animated movies, videos games, and some of Apple’s most cutting-edge autonomous systems projects. The common thread through all of this work has been a focus on computer graphics, always asking the question, “How can we push this technology into new and revolutionary use cases?”
Why did you decide to start Parallel Domain? What problem is the company solving?
Deep learning is driving a revolution in autonomy — from self-driving cars to delivery drones. This revolution relies very directly on massive amounts of high quality data. Collecting and annotating this data is currently a very manual, expensive and, at times, dangerous process. There is a huge opportunity to apply cutting-edge graphics and content generation techniques to alleviate this bottleneck, while helping deliver technology that fundamentally changes how we move through our world.
Parallel Domain is here to solve an acute data problem. In the world of autonomous driving, algorithm performance is very directly tied to the quality, quantity, and distribution of your data. We provide a data generation platform that enables our customers to generate labeled data sets, simulation worlds, and controllable sensor feeds in order to develop, train, and test their algorithms safely before they put cars on the roads. We are all about being the scaleable partner that helps companies achieve their autonomy goals without sacrificing quality, time, or public safety.
What types of customers are you working with?
We are working broadly with simulation, perception, and machine learning (ML) teams at companies across the autonomous vehicle ecosystem, including some of the world’s largest OEM’s, Tier 1 suppliers, full-stack autonomous vehicle companies, and broader tech companies to massively reduce the turnaround time, cost, and risk associated with collecting and annotating the data needed to feed ML models.
What do you think is the most challenging aspect of teaching AVs how to drive?
There tend to be three key pain points: iteration time when developing; obtaining the quality and quantity of training data; and robustly testing and validating algorithms across a spectrum of situations. We’ve found that synthetic data can significantly alleviate bottlenecks at all three of these stages.
As you know, AV stacks are complex and many companies opt to create everything in-house; i.e. vertical integration. Given that, what advantages do you see to collaborating with other vendors in a horizontal way?
Autonomy is a deceptively hard problem. Waymo’s director of engineering describes the hidden iceberg with the 90–90 rule: “When you’re 90% done, you still have 90% to go.” Our customers have found that in order to solve what seems like an impossibly hard problem, they need to focus on developing autonomy algorithms with all the data they can get. We focus on generating that data and, by learning across the variety of companies we work with, we can put customers in a winning position relative to what they would have built in-house.
What are some of the specific ways that simulated environments and synthetic data speed up the deployment and adoption of autonomous vehicles?
Using both simulated environments and synthetic data can accelerate all stages of autonomy development by allowing:
- Quickly prototyping of new sensor configurations
- Generating valuable data to improve perception models (i.e. help the car “see” better)
- Use of pixel-perfect labels in synthetic data to test and benchmark the performance of algorithms in very specific and targeted ways.
All three of these use cases address acute pain points that arise from the time, cost, and danger involved in putting real cars on the road for data collection and testing. We strive to complement the strengths and weaknesses of real data in a way that leads to significant market adoption as our users realize that the best approach is to combine both.
Any closing thoughts?
Machine learning is driving the most significant technological revolution we’ve seen in our lifetime. The demand for data that feeds these ML algorithms is huge — with new use cases for data emerging every month. At Parallel Domain, we’re deploying the industry’s leading platform to generate this data, and we can’t wait to see where it takes us over the next few years.
Visit the Toyota AI Ventures site to learn more about Parallel Domain and the rest of the TAIV portfolio.