Nearly two years ago, Motional was at an autonomous vehicle crossroads. 

The company, born from a $4 billion joint venture between Hyundai Motor Group and Aptiv, had already missed a deadline to launch a driverless robotaxi service with partner Lyft. It had lost Aptiv as one of its financial backers, prompting Hyundai to step up with another $1 billion investment to keep it going. Several layoffs, including a 40% restructuring cut in May 2024, had whittled the company from its peak of about 1,400 employees to less than 600. Meanwhile, advancements in AI were changing how engineers were developing the technology. 

Motional was going to have to evolve or die. It paused everything and picked option No. 1.

Motional told TechCrunch it has rebooted its robotaxi plans with an AI-first approach to its self-driving system and a promise to launch a commercial driverless service in Las Vegas by the end of 2026. The company has already opened up a robotaxi service — with a human safety operator behind the wheel — to its employees. It plans to offer that service to the public with an unnamed ride-hailing partner later this year. (Motional has existing relationships with Lyft and Uber.) By the end of the year, the human safety operator will be pulled from the robotaxis and a true commercial driverless service will begin, the company said.

“We saw that there was tremendous potential with all the advancements that were happening within AI; and we also saw that while we had a safe, driverless system, there was a gap to getting to an affordable solution that could generalize and scale globally,” Motional president and CEO Laura Major said during a presentation at the company’s Las Vegas facilities. “And so we made the very hard decision to pause our commercial activities, to slow down in the near term so that we could speed up.”

This meant shifting away from its classic robotics approach to an AI foundation model-based one. Motional was never devoid of AI. Motional’s self-driving system used individual machine learning models to handle perception, tracking, and semantic reasoning. But it also used more rules-based programs for other operations within the software stack. And the individual ML models made it a complex web of software, Major said.

Meanwhile, AI models originally built for language began to be applied in robots and other physical AI systems, including the development of autonomous driving. That transformer architecture made it possible to build large and complex AI models, ultimately leading to the emergence, and skyrocketing use, of ChatGPT. 

Motional searched for ways to combine these smaller models and integrate them into a single backbone, allowing for an end-to-end architecture. It has also maintained the smaller models for developers, which Major explained gives Motional the best of both worlds.

“This is really critical for two things; One is for generalizing more easily to new cities, new environments, new scenarios,” she said. “And the other is to do this in a cost optimized way. So for example, the traffic lights might be different in the next city you go to, but you don’t have to redevelop or re-analyze those. You just collect some data, train the model, and it’s capable of operating safely in that new city.”

TechCrunch got a first-hand look at Motional’s new approach during a 30-minute autonomous drive around Las Vegas. One demo can’t provide an accurate assessment of a self-driving system. It can, however, pinpoint weaknesses and differences from previous iterations, and gauge progress. 

Progress is what I saw as the Hyundai Ioniq 5 I rode in autonomously navigated its way off Las Vegas Boulevard and into the pickup and drop-off area of the Aria Hotel. These bustling areas are notorious in La Vegas and my experience was no different as the autonomous vehicle slowly nudged its way around a stopped taxi and unloading passengers, changed lanes, then back again, passing dozens of people, giant flower pots, and cars along the way.

Motional previously operated a ride-hailing service in Las Vegas with partner Lyft using vehicles that would autonomously handle portions of a ride. Parking lots and hotel valet and app ride pickup areas were never part of those operations. A human safety operator, always behind the wheel, would take over to navigate parking lots or the busy pickup and drop-off points of hotel lobbies.

There is still more progress to be made. The graphics displayed to riders within the vehicle are still under development. And while there was never a disengagement during my demo ride — which means the human safety operator takes over — the vehicle did take its time to nudge itself around a double parked Amazon delivery van.

Still, Major argues Motional is on the right path to deploy safely and cost effectively. And its majority owner Hyundai is in it for the long haul, she said.

“I think the real long-term vision, you know, for all of this, is putting Level 4 on people’s personal cars,” Major said, referring to a term that mean the system handles all driving with no expectation of human intervention. “Robotaxis, that’s stop number one, and huge impact. But ultimately, I think any OEM would love to also integrate that into their cars.”