Uber puts Israeli AI startup in driver’s seat for Munich robotaxi test run

Looking for path to scale up service, US ride-hailing firm teams up with Autobrains for low-cost, vehicle-ready tech that developers say outpaces competition by using AI agents working in concert

Ride-hailing app Uber has tapped Israeli autonomous driving startup Autobrains and US chipmaker giant Nvidia for a pilot program set to bring driverless robotaxis to the streets of the German city of Munich.

The project marks the first major testing ground for Autobrains, which develops self-learning artificial intelligence tech for assisted and autonomous driving, as Uber looks to enter a rapidly expanding competition for scaling up self-driving operations.

For at least a decade, tech titans, including Tesla’s Elon Musk, have made grand promises to make self-driving, autonomous cars a reality for the public.

While self-driving cars are not yet being offered to consumers, robotaxis have been rolled out in several US cities, led by Google’s sister Waymo, with a fleet of 3,000 self-driving vehicles giving some 500,000 rides a week. Its main rival, Tesla, currently only available in a few Texas cities, has plans to expand to several more metro areas around the country.

Europe, in contrast, has lagged far behind, due to tight safety rules and regulatory bottlenecks, alongside a culture that relies on public transport. But as the Continent pushes toward allowing AI-powered cabs to roam its streets, ride-sharing app Uber is positioning itself as among the first in line to tap into the market.

In early June, after several years of delays, Germany, Italy, France and 15 other European Union countries signed a joint declaration to coordinate autonomous vehicle testing across borders, intended to facilitate the regulatory path and adoption of the technology across Europe.

Uber has placed bets on a few AI autonomous technology companies that could compete with the likes of Waymo and jump-start a larger-scale commercial rollout of autonomous vehicles, which has been harder than anticipated due to high costs, safety issues, and tight regulations. Among other ventures, it has introduced robotaxis with Chinese AI firms WeRide in Spain and Pony.Ai in Croatia, as well as other pilots in the Gulf.

Earlier this month, the San Francisco-based company teamed up with Autobrains, the developer of a self-learning artificial intelligence technology for assisted and autonomous driving, for its latest venture, this time in Germany.

Autobrain’s autonomous driving software system takes a different path than other self-driving technologies, relying on several specialized AI agents for the various processes involved in driving, which is meant to create a system that’s both safer and more cost-effective.

“For automakers and autonomy developers, the challenge is not just building autonomous vehicles — it’s bringing them into a commercial network where they can reliably serve riders at scale,” said Sarfraz Maredia, Uber’s global head of autonomous mobility and delivery. “This program creates a new path to do that by combining vehicle-agnostic autonomy, leading AI compute, and Uber’s ride-hailing platform.”

By cutting down on the compute power required to operate the self-driving vehicle, a major barrier to scaling up autonomous driving, the Uber robotaxis planned for Munich will need only an affordable, off-the-shelf mass production Nvidia chip and six cameras that cover the same field of vision as a human would have, according to Autobrains founder and CEO Igal Raichelgauz.

“Autonomous driving will not scale by relying on a single model to solve every driving scenario,” said Raichelgauz. “Our autonomous driving system uses AI agents, similar to an array of specialized drivers, that are being called on the fly for assessing different traffic situations on the road and making decisions in real time, whether it is getting into a roundabout, overtaking a truck, or driving on a rainy day on a highway.”

Unlike many of its competitors, which rely on a single, monolithic AI model that handles all driving tasks, Autobrains uses an array of specialized AI agents that work together to guide the vehicles using standard ​automotive ⁠sensors. One agent senses and evaluates what is happening around the car, while another uses reason to assess risks involved in any move, and another selects responses in real-time.

Other systems, like Mobileye, Israel’s leading maker of self-driving car technologies, use machine learning, teaching and training computers to recognize objects and scenarios by feeding them millions of images and data. The process is labor-intensive and needs very high computational power.

Raichelgauz argued that Autobrains’ orchestrated system can handle road uncertainty more accurately and safely than the other model.

“Even the most expensive and advanced system, like Waymo, cannot reason beyond what it was shown or trained, and that’s the gap that we are closing with agentic AI and the common sense embedded in those agents,” said Raichelgauz. “Our system is designed to learn and improve through reasoning and thinking, not heavy data, or compute.”

Mor Kaspi, head of the Shlomo Shmeltzer Institute for Smart Transportation at Tel Aviv University, cautioned that the issues involved in scaling up autonomous vehicles went far beyond the inner workings of the self-driving system.

“Even if the technology significantly advances the brains of the autonomous vehicles, it’s not enough to get these vehicles on the roads — the vehicles need to be cheap enough for services to use them, or for the government to invest in them,” he said.

“In addition, there are also psychological, social, moral. regulatory and legal aspects, and they all need to be resolved in order for us to see autonomous vehicles at large scale on our roads,” Kaspi added.

Raichelgauz said the Autobrains system does not require a specialized vehicle outfitted with custom sensors, and also provides other cost savings.

“Our autonomous driving technology can be installed in a regular car, doesn’t depend on a fleet of custom vehicles, and removes the need for heavy and expensive sensor stacks, or any prerequisite to do infrastructure work like high-definition mapping, which are cost barriers for commercializing robotaxis at scale,” he said. “For the mapping, we use AI technology that matches aerial satellite imagery with vehicle camera data.”

Autobrains was founded in 2019 by Raichelgauz as a spinoff from the Cortica Group, an autonomous AI tech firm he co-founded with Karina Odinaev, a specialist in brain sciences, alongside computer vision and neuroscience professor Josh Zeevi.

“The idea was to develop AI for the physical world based on the human brain,” said Raichelgauz, a graduate in electrical engineering from Haifa’s Technion – Israel Institute of Technology and a veteran of the Israeli army’s elite 8200 tech unit. “One of its key characteristics is that it’s not one neural network, but a lot of specialized networks, and there is an orchestration process going on, and so the whole inspiration for our technology is modelled on neuron activity and learning mechanisms occurring in the brain.”

With operations in Tel Aviv and Munich, Autobrains employs about 100 AI experts coming from backgrounds in neuroscience, quantum physics, autonomous driving, and software development. To date, the startup has raised about $140 million from investors including Temasek, BMW i Ventures, Toyota Ventures, Continental, Knorr-Bremse, and VinFast.

The Munich pilot is meant to serve as a testing springboard to develop a model for the rollout of autonomous robotaxi fleets in several dozen cities worldwide, pending regulatory approvals, safety certifications, and assuming the tech works and is economically viable.

“We have a plan that by 2028 to scale to 20 cities and Munich is a good benchmark,” said Raichelgauz. “The focus for us will be heavily on Europe and South Asia.”

He said the company had applied for permits to test out the driverless taxis in Israel as well.

Although Uber and Autobrains did not provide a specific timeline for when passengers can start ordering autonomous robotaxis in Munich via the ride-hailing app, Raichelgauz said that the process will be “short and aggressive,” once authorities provide the necessary regulatory green lights.

“We got approval from local authorities after a lot of tests to run the system with a test driver, but still not with commercial passengers,” said Raichelgauz. “The next step would be to have the system integrated into the Uber ride-hailing fleet and scale the number of cars, to drive passengers, but still with the human test driver for a certain period of time.”

“The final step is to take the human test driver out of the loop,” he added.

Kaspi noted that even if robotaxis and autonomous freight become more commonplace, it will be a while before most cars on the road are driving themselves.

“According to the latest projections, the deployment of commercial autonomous vehicles would be feasible in a decade or less,” he said. “But a large-scale calibration, meaning a considerable percentage of autonomous vehicles on the road, this is at least three or four decades away.”

Reuters contributed to this report.