TuSimple’s proprietary algorithm serves as the “virtual driver” of our L4 autonomous trucks. The algorithm interprets the road environment and decides how to safely navigate each scenario. Our algorithm team is tasked with developing a world class virtual driver that is safer and more efficient than a human driver.
As part of our ongoing technology spotlight series, we sat down with Ersin Yumer—who was recently promoted to Vice President, Algorithm—for a brief interview to understand his stories and experiences that will help to ensure that the TuSimple Autonomous Driving System is as robust and groundbreaking as possible.
In this role, Ersin is responsible for developing the Algorithm Team’s technical capabilities, tools, and processes needed to ensure efficient design and deployment of our autonomous vehicles.
TuSimple has an innovative approach to the autonomous driving system because it is building its technology on a fused computer vision and artificial intelligence (AI) background rather than relying heavily on robotics. Outside of TuSimple, the self-driving industry as a whole has struggled through its early years due to its reliance on a robotics-first approach and has yet to pivot to a more realistic technology. I believe this robotics-first technology will not scale and solve the complex real world problems that a self-driving truck will face on the roadways in a reliable way, whereas a tech stack that is built with a machine learning and computer visualization core will be more capable of tackling this operational design domain (ODD). TuSimple has also kept its public promises in meeting milestones and deadlines, which is refreshing in this challenging industry and has the clearest path to commercialization with a solid go-to-market strategy.
The Algorithm Team was heavily involved in developing the “brain” of the vehicle that completed the Driver Out Program. This meant we had a very strong collaboration with the Hardware and Software Teams to make the brain seamlessly and reliably work with the whole platform to ensure maximum safety on public roads. This emphasis and ability to enforce safe development is one of TuSimple’s strengths. The industry over indexes for AI capabilities, but solving the real problems include ensuring that those technologies can work with the rest of the system and operate in a reliable, redundant, and safe fashion. Machine learning systems can fail, so for our real-world safety-critical application we must ensure that we provide redundant compute, brake, and steering units that work harmoniously with the “brains” of the truck to complete the mission or end it safely. This is the crux of that collaboration.
We had the opportunity to learn how to make a safety-critical system work with a primarily machine learning-driven brain while operating at highway-level speeds of travel. This must be done no matter what before any developer can start thinking about commercialization. With our experience and Driver Out learnings we are now very comfortable with repeating Driver Out in a systematic manner. We are so comfortable that our focus now is on improving the unit economics (cost per mile) to scale Driver Out towards commercialization. So, if we think about the entire process from zero to commercialization, it’s a linear trajectory, and along that trajectory you must complete Driver Out, demonstrate repeatability of Driver Out, and reduce cost along that line for commercial viability. Given our success with the Driver Out program, we have the experience and capability to keep going on that linear path.
Next up is the Driver Out ODD migration into high-density logistics routes, such as the Texas triangle, and further ruggedizing our system to bring the cost-per-mile down. Continuing on that same theme of commercialization in the Driver Out era, my team is playing a central role to bring the operational cost down to a level where it’s competitive in the logistics landscape. For instance, we have already started building virtual worlds with the data pouring in from our trucks in Texas, which is accelerating the pace of the development in the algorithm team’s daily work.
I think there are several motivations and driving factors for bringing L4 autonomous semi-trucks to the real-world: being able to help with public safety and to help solve the driver shortage that affects the logistics industry is a very fulfilling task. I also get to work with extremely capable, driven, and intelligent teammates every day. This is, I think, a factor because within the tech space, self-driving has the attention of the top talent. My job is really a geek’s dream because I get to play with sensors and compute units, most of which are pre-release gadgets!
Ersin has an outstanding academic background coupled with professional experience leading the development of computer vision, machine learning, and robotics in multiple industries. He was most recently an Engineering Director, serving as head of simulation machine learning at Aurora, a Sr. Staff Research Scientist at Uber ATG, where he built a research team. Before that, he led the perception machine learning team at Argo AI. He spent three years at Adobe Research where his work led to the development of innovative features in Adobe Dimension, Adobe Fuse, and various Adobe Photoshop mobile apps. Among his published major technical work, these are his favorites:
Ersin has a Mechatronics-focused BSc, a Machine Learning-focused MSc in Engineering from the Middle East Technical University in Turkey, and a PhD from Carnegie Mellon University where he developed real-time intuitive 3D design tools that leverage AI.
Ersin lives in San Francisco, California with his wife and two dogs. He is an avid triathlon competitor, and he loves going on training runs with his “very active” dogs.