Autonomous transport is one of the most exciting and important technologies of our generation, but it’s capital-intensive and typically not scalable. A new approach that can reduce costs while still ensuring that self-driving is safe and reliable is needed, said Waabi founder Raquel Urtasun. The 20-year AI veteran spoke at the recent TransformX AI conference on artificial intelligence (AI) and machine learning (ML).
“Self-driving is one of the most exciting and important technologies of our generation. It is really going to change the way that we live [from a safety perspective, and it will change the] landscape of our cities.” —Raquel Urtasun
To be able to drive anywhere accurately and safely, Urtasun explained, an autonomous vehicle must know where it is, identify the other objects or people around it, and understand how likely those others are to act. Then it must plan an appropriate and safe path.
At any point, the vehicle must be able to perceive changes in its surroundings and the reasons behind those changes, then update its planned path accordingly.
To do so, a vehicle must identify where in the physical world it is. It must sense its environment via radar, lidar, and other means. Once the vehicle perceives its environment, it localizes itself within a precision of a few centimeters.
Next, Urtasun said, the vehicle’s AI makes predictions about the potential future paths of surrounding objects and plans its own path accordingly. It makes decisions about where to steer and how fast to move using all the information it has received, which is constantly being updated, with the data pipeline repeating every fraction of a second.
Thus far, commercial deployment of autonomous vehicles is limited to “very small and very simple operation domains,” Urtasun said. This is largely because the technology requires complex decision making in an almost infinite number of potential situations. “Many of these situations appear very, very rarely, and we still need to be able to handle all of them,” Urtasun said.
Solutions so far are imperfect, although there has been much improvement in sensors, algorithms, and other areas since the 2004 DARPA Grand Challenge, which was designed to accelerate development of autonomous vehicles for military use.
Urtasun said that many of today’s solutions, which rely heavily on manual tuning, are costly and typically not scalable. “We need to instead have a much more holistic approach that really is designing the entire system from the top down,” she said.
The interfaces between the different modules that make up an automated vehicle are usually very small, she said. As a result, little information is passed around, and “the vehicle is more and more blind as it has to do more and more sophisticated processing.” Further, “if you have a mistake at the beginning of the [software] stack, it's going to be impossible for the self-driving vehicle to actually correct that mistake,” Urtasun said.
Every time you want to make a change in the software, “you need to tune every single module one at a time until you end up at the end of the stack.” As a consequence, every time that you change something, “it takes you almost a quarter or more to basically land it in production, even if it's a very, very small change. And this is because of the lack of automation and the fact that you need to do this one module at a time by hand.”
AI and deep learning might help, but neural networks are black-box approximators that can lack interpretability, Urtasun said. AI also often requires a lot of data examples to learn from.
Waabi provides a new model for self-driving technology, Urtasun said. The company uses an AI-first approach, meaning the system is trained and tested as a whole, instead of training and tuning separate modules that are then chained together.
Training is done via a 3D, real world, high-fidelity simulation that allows for training on rare scenarios and reduces the need to collect data from the real world. In other words, there’s no need to drive around for “millions of miles” and create potentially dangerous situations or even collisions, Urtasun said.
It's very expensive to have a fleet of hundreds of vehicles out driving, and that can be dangerous, she said. Instead, Waabi uses an AI approach “that is able to learn from fewer examples,” and can scale, Urtasun said.
The current industry approach for simulation testing requires artists to hand-design 3D assets and for engineers to implement rule-based simulation behavior. Instead, Waabi automatically generates virtual worlds and multi-agent systems based on actual human driving behavior.
Through the combination of probabilistic graphical models, deep learning, and complex optimization, Waabi created “a new generation of algorithms that can do very, very complex reasoning, similar to how humans actually reason in the world.” And she said, this technology “is much more affordable because it can actually be developed at a fraction of the time and a fraction of the cost.”
Waabi uses a “breakthrough simulator” that can test simple scenarios or scenarios that happen all the time, as well as the rare cases, Urtasun said. Waabi uses AI and physics to model sensor noise, which results in perception outputs that behave similarly in both simulation exercises and the real world.
For now, Waabi is focusing on the long-haul trucking segment of the transport industry. There are two reasons for this, Urtasun explained. As the recent shortage of some consumer goods in the United States has shown, “there is an acute driver shortage that is just getting worse and worse.” Second, truck driving is one of the most dangerous professions in North America. So there is a need for increased safety that could come from automation.
With Waabi’s new technology, Urtasun said, “we believe that we are building the right technology to help us really get there in a safe manner.”
For more details about Waabi’s self-driving technology, watch Urtasun’s talk, “An AI-First Approach to Enable Self-Driving at Scale,” and read the full transcript here.