Picture this: A vehicle approaches an intersection to make a left-hand turn at the light.
The problem is, traffic is backed up. A human driver might consider two options: moving into the turning lane early, or bypassing the intersection and making the turn later.
This is the example Balajee Kannan, VP of autonomy at autonomous vehicle startup Motional, gives to illustrate how two different types of AV planning systems––rule-based versus machine learning-based––might respond to a common but slightly tricky driving scenario.
A rule-based system, he said, “is going to try to optimize…and try to get in at exactly the same spot in a small window within which it needs to come over. Which means it’s going to get stuck a lot more if that window is taken over.”
An ML-based system, on the other hand, “can learn that behavior, when you have a sequence of vehicles in front of you that’s starting to do a lane shift…blindly instead of waiting to do it.”
Scenarios like this reflect why Motional is transitioning from a rule-based system to an end-to-end ML-trained system.
“By embracing an ML-first approach,” the company said in a blog post, “Motional is building the foundation for the future: a fully end-to-end ML-based AV system.”
End-to-end: Traditionally, AV tech has relied on what’s known as a modular design that includes sensors, perception radar, and prediction capabilities that together represent rule-based planning, per Motional. This approach, according to the company, has disadvantages like “error accumulation” and “scalability bottlenecks.”
More recently, players in the AV sector have started to move toward a machine learning-based approach that involves training systems on large reams of data and then leveraging AI to dictate the vehicle’s behavior. Proponents contend that this approach makes self-driving vehicles behave more intuitively, and makes them more capable at handling edge cases that might confuse rule-based systems, as Wired recently noted.
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
“What it allows you to do at a macro level is to scale out much more quickly, because human driving behaviors have patterns that you can learn,” Kannan said. “What you learn in Las Vegas in terms of that behavior applies to Pittsburgh, applies to London, applies to anywhere else in there, too.”
“There are nuances between driving patterns; [the] Pittsburgh left is legendary,” he acknowledged. “But it translates much more [easily] across different cities, different dimensions.”
Moving forward: Motional is a Level 4 autonomy company that’s jointly owned by Hyundai and Aptiv. Headquartered in Boston, the company has operations in Las Vegas, Pittsburgh, and Singapore, and long-standing partnerships with Uber and Lyft. Motional’s commercial operations are currently paused while the company focuses on developing its technology.
“The key area that we are focused on right now…[is] introducing machine learning into the planning stack,” Kannan said.
Waymo also is using AI advancements to improve its tech by integrating large language models with visual language models to “create an end-to-end…system that’s multi-modal in its foundation,” Tekedra Mawakana, Waymo’s co-CEO, said at CES 2025. The Alphabet-owned company announced last year that it was developing a new training model called Emma built on Google Gemini, a multimodal large language model, The Verge reported.
Tesla, too, uses end-to-end machine learning in its AV tech, though its approach is somewhat more controversial because it doesn’t use some of the redundancies that its competitors do.
“This is the cutting edge,” Kannan said. “This is the last piece of the puzzle toward the ultimate goal that we all have, toward end-to-end machine learning-trained systems.”