Nuro Begins Testing Autonomous Vehicles in Tokyo

Table of Contents

• 1. Nuro tests autonomous vehicles on Tokyo streets
• 2. Nuro’s International Expansion into Tokyo
• 3. Testing Autonomous Vehicles on Tokyo’s Streets
• 3.1 Use of Toyota Prius for Testing
• 3.2 Involvement of Human Safety Operators
• 4. Challenges Faced by Nuro in Tokyo
• 4.1 Adapting to Left-Side Driving
• 4.2 Navigating Unique Traffic Rules
• 5. Nuro’s AI Strategy for Autonomous Navigation
• 5.1 Zero-Shot Autonomous Driving Approach
• 5.2 Simulation and Closed-Course Testing
• 6. Future Prospects and Business Model Shift
• 7. The Future of Autonomous Vehicles in Urban Environments
• 7.1 Navigating Complex Urban Landscapes

---

Nuro tests autonomous vehicles on Tokyo streets

• Nuro has started public-road testing of its self-driving software in Tokyo using Toyota Prius vehicles with human safety operators.
• The Tokyo program is the Silicon Valley startup’s first overseas expansion, after opening offices in the city last August.
• Nuro says Japan introduces new driving styles and rules, including left-side driving, dense traffic, and different signage and lane markings.
• The company is leaning on an end-to-end AI “foundation model” and a “zero-shot” approach meant to work in new places without prior local training data.

Nuro Begins Tokyo Road Testing
– Confirmed (reported): Toyota Prius vehicles equipped with Nuro’s self-driving software began testing on public roads in Tokyo last month, with a human safety operator behind the wheel. (TechCrunch, Mar 11, 2026)
– Confirmed (company statement): Nuro frames Tokyo as the start of “compounding benefits of global deployment,” and has suggested more international expansions. (Nuro blog)
– Not disclosed: fleet size, specific routes, and any timeline for removing the human safety operator.
– What this test is (today): a validation program in a new driving environment—not a driverless commercial launch.

Nuro’s International Expansion into Tokyo

Nuro, the Silicon Valley autonomous vehicle startup backed by Nvidia, Uber, and SoftBank, has begun testing its self-driving technology on public roads—its first international deployment. The move marks a notable step for a company that, until now, had been associated primarily with U.S.-based development and a delivery-focused origin story. (TechCrunch, Mar 11, 2026)

The Tokyo tests began last month, according to Nuro, and follow the company’s decision to establish a local presence: it opened offices there last August. Nuro has not disclosed key operational details of the program, including fleet size and any timeline for removing the human safety operator. While Nuro did not disclose how many vehicles are involved in the Japanese fleet, it framed the rollout as more than a one-off experiment. In a blog post announcing the program, the company suggested additional international expansions could follow.

That ambition is tied to a broader thesis: that autonomy systems should improve as they encounter more diverse environments. “Our autonomous operations in Tokyo are the beginning of the compounding benefits of global deployment,” Nuro wrote, positioning Japan as an early proof point for a strategy built around generalization rather than heavy, location-specific tuning.

Tokyo is also a symbolic choice. It is a dense, complex urban environment with driving conventions that differ from the U.S.—a setting that can quickly expose weaknesses in perception, prediction, and planning. For Nuro, demonstrating credible performance there supports its post-pivot identity: not as an operator of a bespoke delivery fleet, but as a supplier of autonomy software that can travel.

Tokyo as a High-Signal Testbed
Tokyo is a high-signal test environment for autonomy because it combines:
– Different driving conventions (left-side driving) that can flip “default” assumptions in planning and lane behavior.
– Dense, mixed road use (cars, pedestrians, cyclists) that stresses perception and prediction.
– New road “language” (signage and lane markings) that an AV must interpret correctly to be compliant and predictable.
For a company pitching generalizable autonomy, Tokyo is less about easy miles and more about quickly surfacing edge cases.

Testing Autonomous Vehicles on Tokyo’s Streets

Nuro’s Tokyo program is not a driverless launch. It is a structured public-road testing effort designed to validate how the company’s software behaves in a new geography, under real traffic conditions, while maintaining a conservative safety posture.

The company says the vehicles are being tested on public roads, with a human safety operator behind the wheel as backup. Nuro has not said when—if ever in the near term—it expects to remove the safety operator in Japan. That omission matters, because it signals the current phase is about learning and verification, not commercialization.

Nuro has also described a workflow that blends on-road exposure with safeguards. Once vehicles are on the road, they can be manually driven while Nuro’s software runs in “shadow mode,” producing what it would do without sending commands to the vehicle controls. Nuro then checks those results to decide whether the system is ready for autonomous operation on public roads.

Nuro Autonomy Testing Pipeline
A practical way to read Nuro’s testing pipeline (as described by the company) is:
1) Release candidate built: a new version of the “universal autonomy model” is prepared.
2) Closed-course validation: repeatable maneuvers are tested in a controlled environment before any public-road exposure.
– Checkpoint: does it pass predefined scenarios without safety-critical failures?
3) Simulation for edge cases: rare/dangerous scenarios are stress-tested at scale.
– Checkpoint: do regressions appear versus the prior release?
4) Public-road “shadow mode”: a human drives; the autonomy stack runs in parallel and outputs what it would have done.
– Checkpoint: do the model’s proposed actions match safe, lawful, and socially compliant driving in Tokyo conditions?
5) Limited on-road autonomy (only if ready): the system may progress toward sending commands to vehicle controls within a defined operational design domain.
– Checkpoint: performance is consistent enough to justify expanding conditions/routes.

Use of Toyota Prius for Testing

Rather than testing with a custom-built vehicle, Nuro is using Toyota Prius cars equipped with its self-driving software. The Prius choice is pragmatic: it is a mainstream passenger vehicle platform, widely understood and well-suited to incremental testing in a dense city.

Using a conventional vehicle also aligns with Nuro’s current business direction. If the company’s goal is to license its autonomy stack to automakers and mobility providers, proving it can integrate with and perform on common vehicle types becomes strategically important. The Prius tests in Tokyo serve as a visible demonstration that Nuro’s software is not confined to a proprietary delivery robot form factor.

Nuro has not provided technical specifications of the Tokyo vehicles beyond the software integration and the presence of a safety operator. It also has not disclosed fleet size, routes, or a timeline for expanding the operational design domain within Japan.

Involvement of Human Safety Operators

A human safety operator is behind the wheel during Tokyo testing, acting as a backstop. This is a standard practice in many AV testing programs, especially in early deployments in new environments where the system must contend with unfamiliar signage, lane markings, and local driving behavior.

Nuro has been explicit that it is not “disregarding safety” even as it promotes a more generalizable AI approach. The safety driver is one layer; the company’s broader validation process is another. Nuro has not announced any plan or date to transition to fully driverless operation there.

Just as importantly, the safety operator setup fits with Nuro’s “shadow mode” methodology: the car can be driven manually while the autonomy stack generates decisions in parallel, allowing the company to compare what the model would have done against what actually happened—without taking control in situations where the system’s readiness is still being assessed.

Challenges Faced by Nuro in Tokyo

Nuro has acknowledged that Japan introduces “a number of new challenges and different driving styles and rules.” Tokyo is not simply another large city; it is a different driving culture, with different conventions and road communication systems. For an autonomy stack, those differences can surface in everything from lane positioning to interpreting markings and signs.

Among the challenges Nuro has highlighted are left-side driving, dense traffic, and differences in road signs and lane markings. Each of these factors can stress an AV system’s ability to perceive the environment correctly and to plan safe, socially compliant maneuvers.

The company has not detailed specific incidents or performance metrics from the Tokyo tests, and it has not said how quickly it expects the system to progress from shadow-mode evaluation to more direct autonomous control on public roads. But by choosing Tokyo for its first overseas test, Nuro is implicitly betting that its approach can handle meaningful variation without requiring a long, location-specific retraining cycle.

Adapting to Left-Side Driving

One of the most immediate differences between Japan and the United States is that vehicles drive on the left side of the road. For human drivers, this is a cognitive adjustment. For autonomous systems, it can be a deeper architectural test: lane selection, turn behavior, intersection handling, and the “default” expectations about where traffic flows all invert.

Nuro has pointed to left-side driving as a concrete example of what makes Japan a new challenge. The company’s broader claim is that its autonomy stack—built on an end-to-end AI foundation model—can generalize to such differences without prior training on Japanese driving data. In other words, the system should be able to interpret the scene and produce appropriate driving behavior even when the “rules of the road” differ from what it has previously encountered.

Tokyo’s density compounds the issue. Left-side driving is not happening on empty roads; it is happening amid heavy traffic, frequent interactions, and tight urban geometry. That combination makes the Tokyo tests a more stringent check on whether Nuro’s software can behave safely and predictably when the baseline driving convention changes.

Navigating Unique Traffic Rules

Beyond which side of the road to drive on, Nuro has emphasized that Tokyo brings different driving styles and rules, along with different road signs and lane markings. These are not cosmetic differences. Signs and markings are part of the language an AV must read to understand right-of-way, lane boundaries, permitted movements, and constraints.

Tokyo’s dense traffic adds another layer of complexity. In heavy urban flow, an AV must not only follow formal rules but also navigate the informal rhythm of city driving—merging, yielding, and responding to other road users in a way that is safe and does not create confusion.

Nuro has not publicly broken down which specific Japanese rules or signage patterns have been most challenging, nor has it described how it is measuring success in Tokyo. What it has said is enough to outline the core problem: deploying autonomy internationally is not just a mapping exercise; it is an interpretation problem, where the system must correctly read and act on a new set of road cues.

Nuro’s AI Strategy for Autonomous Navigation

Nuro’s Tokyo expansion is tightly linked to how it says its autonomy stack works. The company describes its system as being built on an end-to-end AI foundation model that can learn as it drives. This is a notable positioning choice in an industry where many systems have historically leaned on extensive, location-specific data collection and tuning.

The company calls its approach “zero-shot autonomous driving,” and it has presented Tokyo as evidence that the strategy can work: Nuro says its software was able to autonomously navigate in Tokyo without any prior training on Japanese driving data. As described by Nuro, this capability is evaluated through a staged process that includes closed-course testing, simulation, and on-road “shadow mode” before any move toward direct vehicle control. It also points out that this does not mean it is ignoring safety; instead, it pairs the model with closed-course testing, simulation, and shadow-mode evaluation on public roads.

Nuro is not alone in pursuing an end-to-end AI approach. The U.K.-based startup Wayve has taken a similar strategy, also emphasizing end-to-end learning for self-driving software.

Key Terms Behind Tokyo Test
A plain-English map of the key terms Nuro is using:
– “End-to-end AI foundation model”: a model that takes in sensor inputs (what the car sees) and outputs driving actions (what the car would do) as one learned system, rather than stitching together many hand-tuned submodules.
– “Zero-shot” (in this context): the system can begin operating in a new geography without first being trained on a dedicated local dataset.
– What “zero-shot” does not mean: no validation, no safety driver, or instant readiness for driverless service.
– “Shadow mode”: the software generates decisions in parallel while a human drives; outputs are logged and compared, but not executed.
Why this matters in Tokyo: it’s a real-world test of whether the model generalizes to left-side driving and new road cues before the company expands autonomy beyond evaluation.

Zero-Shot Autonomous Driving Approach

“Zero-shot autonomous driving” is Nuro’s label for a system designed to operate in new environments without being trained first on local driving data. In the Tokyo context, that means handling left-side driving, different signage, and different lane markings without a Japan-specific dataset as a prerequisite.

Nuro’s claim is not that the system is finished the moment it arrives in a new country, but that it is broadly capable enough to begin operating—and improving—without the long lead time typically associated with collecting and labeling local data before meaningful on-road performance is possible.

This framing matters for Nuro’s business goals. If the company wants to license autonomy technology to automakers and mobility providers, it needs a stack that can be deployed across regions with less friction. A “zero-shot” narrative supports the idea that Nuro’s software could be integrated and rolled out faster than approaches that require extensive pre-deployment localization.

At the same time, Nuro has been careful to pair the claim with safety language and process: the company describes shadow mode and evaluation steps intended to ensure the system is ready before it is allowed to control the vehicle.

Simulation and Closed-Course Testing

Nuro says it conducts closed-course testing of each new release of its universal autonomy model and evaluates performance and edge cases using simulation. This is the less visible side of the Tokyo story: the work that happens before and alongside public-road exposure.

Simulation is particularly relevant when a company is trying to validate behavior across many rare or dangerous scenarios—“edge cases”—that may not appear frequently in real-world driving but still must be handled safely. Closed-course testing, meanwhile, allows controlled repetition and verification of specific maneuvers and system changes.

Once on public roads, Nuro says the vehicles can be manually driven while the software runs in shadow mode. In that setup, the foundational AI model produces what the system would do, but those commands are not sent to the vehicle controls. Nuro then checks the results to determine readiness for autonomous operation on public roads.

Taken together, these steps describe a pipeline: controlled testing and simulation to validate releases, then shadow-mode comparison in real traffic, then—only if performance is sufficient—progression toward more direct autonomous control.

Future Prospects and Business Model Shift

Tokyo testing is also a milestone in Nuro’s corporate evolution. Founded in 2016 by early Google self-driving project engineers Dave Ferguson and Jiajun Zhu, Nuro initially focused on developing and operating a fleet of low-speed, on-road delivery bots. That early pitch attracted major capital, including a $940 million investment from SoftBank Vision Fund in 2019.

But the company’s trajectory changed. Nuro faced high development costs and an industry wave of consolidation, prompting staff cuts and a reassessment of its model. In 2024, it ditched the low-speed bots and shifted toward licensing its technology to automakers and mobility providers, including ride-hail and delivery companies.

The Tokyo program fits that licensing narrative: it demonstrates Nuro’s autonomy stack on a conventional vehicle in a major global city, rather than on a specialized delivery robot in a limited operational context. It also supports the company’s argument that its AI foundation model can generalize across geographies—an attractive proposition for partners who operate across multiple markets.

Nuro has continued to raise capital to support this strategy. Last year, it raised $203 million in two tranches in a Series E round that included existing backer Baillie Gifford and new investors Icehouse Ventures, Kindred Ventures, Nvidia, and Pledge Ventures. Uber also participated, after saying it would make a “multi-hundred-million-dollar” investment in Nuro as part of a broader deal with electric car maker Lucid.

What remains unanswered is the timeline from testing to product. Nuro has not disclosed when it might remove the human safety operator in Tokyo, nor how large it intends to scale its Japanese fleet. Still, the company’s own messaging suggests Tokyo is a starting point—an attempt to turn international deployment into a compounding advantage rather than a bespoke engineering project each time it crosses a border.

The Future of Autonomous Vehicles in Urban Environments

Navigating Complex Urban Landscapes

Tokyo is a reminder of what makes urban autonomy hard: dense traffic, constant interactions, and a road environment filled with signals—signs, markings, conventions—that vary by country. Nuro’s test is significant not because it declares victory, but because it treats complexity as the point of the exercise.

If autonomy is to become broadly useful, it must work in places that look like real cities, not just simplified corridors. Nuro’s approach—testing on public roads with safety operators, using shadow mode, and validating releases through closed courses and simulation—reflects an industry reality: progress is iterative, and the hardest environments are often the most informative.

Tokyo also highlights the operational challenge of scaling: companies must build systems and processes that can absorb new rules and driving styles without restarting from scratch. Nuro is explicitly trying to prove that its stack can do that.

The Role of AI in Autonomous Driving

Nuro’s Tokyo deployment is, in part, a referendum on its AI strategy. The company is betting on an end-to-end AI foundation model and a “zero-shot” capability that can begin operating in new geographies without prior local training data. It is a vision of autonomy that emphasizes generalization and learning, rather than heavy pre-mapping and bespoke tuning for each city.

At the same time, Nuro is pairing that vision with a safety narrative grounded in process: closed-course testing for each release, simulation for edge cases, and shadow-mode evaluation before allowing the system to control the vehicle.

Whether this combination becomes a template for broader deployment will depend on what the Tokyo tests ultimately show over time—especially as Nuro expands beyond initial routes and conditions. For now, the company’s first overseas test underscores a central tension in the AV industry: the push for faster, more scalable AI-driven deployment, and the need to prove safety and reliability in the most complex places humans drive every day.

Viewed through the lens of building and scaling technology in regulated, multi-stakeholder environments, Martin Weidemann (weidemann.tech) focuses on the operational signals that matter most at this stage—what’s been validated on-road, what remains in shadow-mode evaluation, and what the company has (and hasn’t) disclosed about the path from testing to deployment.

This article reflects publicly available information and company statements about Nuro’s Tokyo testing as of March 2026. Some operational details—such as fleet size, routes, and plans for removing safety operators—have not been publicly disclosed and may change as testing evolves. References to terms like “zero-shot” reflect Nuro’s stated approach and should not be read as independently verified performance claims.