One Metre to the Right

You are in the back of a Waymo, scrolling your phone, when the car pauses mid-lane. A cyclist ahead wobbled left, a dog trotted off the kerb, and a delivery truck began a three-point turn unexpectedly. All happening at the same time, and you did not notice because the car had adjusted so smoothly. It slowed, shifted a metre to the right, and resumed speed before you looked up.

What's new is how the car decided to move. Right now, Waymo's vehicles send their sensor data through a chain of separate software systems, each trained on a different task, such as one finding objects, with another guessing what they will do, and another one picks a lane. Each system adds processing time and computing cost. A recently published patent from Waymo describes a different approach, where a single machine learning model takes fused sensor data and predicts where the car (and everything around it) will be in the next few seconds.

Waymo completed over 15 million rides in 2025 and is targeting one million per week by the end of 2026. As it adds cities and countries, the computing bill for each car becomes a lot larger than when the fleet was small.

Waymo, based in Mountain View, California, is Alphabet's autonomous driving subsidiary. The patent describes a way to close the gap between what the car's sensors see and what the car decides to do next.

A Waymo vehicle carries five types of sensors: (1) laser scanners that measure distance, (2) radar that tracks speed, (3) sonar for close-range objects, (4) cameras for images, and (5) infrared detectors for heat. Today, data from each of these sensors goes through a chain of separate software programs. The patent replaces most of that chain with a single program.

The system takes readings from all five sensor types and merges them into one digital map of the car's surroundings. Think of a spreadsheet laid over the road, where every cell holds a number. The number tells the system what occupies that spot, such as a lane, a sidewalk, a traffic light, a person, a truck, or a piece of debris. The numbers carry detail, too. A cell for a pedestrian can include which way their head is turned. A cell for a traffic light can include its current colour. A cell for a parked car can include whether a door is open.

What happens next?

That map goes into one AI model, which outputs a prediction, being a list of points describing where the car will be over the next few seconds. For example, point one might say "the car will be here in half a second," point two "here in one second," point three "here in 1.5 seconds." The points, connected together, form a route.

The model can produce several possible routes at once, each with a confidence score. The car's driving system picks the route with the highest score and adjusts the steering, brakes, and speed to follow it. If two routes score similarly, the system can choose the more cautious one.

How does it keep up with a changing road?

The car builds a new map of its surroundings every 0.1 to 0.25 seconds, which is four to ten times per second. The model keeps the last several maps in memory, so each new prediction takes into account what happened a moment ago. A pedestrian who was walking slowly in the previous map and is now stepping off the kerb in the current map gives the model a motion trend to work from, rather than a single snapshot.

The same model predicts where other road users will go. When the car's predicted path and a cyclist's predicted path would cross at close to the same time, the driving system can brake or steer before the conflict arrives.

If the system works as described, one program does the job of several, which could reduce the computing cost and response time on each vehicle.

Before a self-driving car can decide where to steer, its sensor data has to pass through a series of separate software systems. Each of these systems was built, trained, and maintained independently, and each one has to finish its work before the next system in the chain can start. Only after all of them have run does the data reach the models that actually plan the car's path.

The patent states the problem plainly, that running all of these pre-processing systems and their individual models uses a large amount of the processing resources on the vehicle's onboard computers. The computing hardware on a single autonomous vehicle can cost US$2,000 to US$20,000, and the power it draws can reduce the car's driving range. When you operate a fleet of thousands of vehicles and plan to scale to tens of thousands, the cost of the computers inside each car shapes the economics of the whole business.

The patent's proposed fix is to skip the chain of separate systems entirely and feed sensor data into a single model that handles detection, prediction, and trajectory planning together.

The category is end-to-end autonomous driving, where a single model (or a tightly coupled pair) handles perception and prediction together instead of relying on a long sequence of handoffs.

Waymo's patent sits in this end-to-end space, but with one key difference from its competitors, that it keeps the full sensor suite (lidar, radar, sonar, cameras, infrared) and fuses them into a single model. Waymo is betting that sensor diversity and model simplicity can coexist in a financially viable way.

Tesla is the closest philosophical match. Tesla's FSD system uses a vision-only, end-to-end neural network trained on billions of miles of fleet data. Tesla launched a robotaxi service in Austin in June 2025, and its purpose-built Cybercab (with no steering wheel or pedals) joined the Austin fleet in September 2026. Tesla's approach skips lidar and radar entirely, relying on cameras alone. That lowers the hardware cost per vehicle but limits the system's depth perception in certain conditions, such as heavy rain or direct glare.

Wayve, a London-based startup, has built an end-to-end AI platform that generalises across cities without HD maps or city-specific programming. Wayve drove autonomously in over 500 cities across Europe, North America, and Japan in 2025, using the same underlying model. In 2026, Wayve partnered with Uber for robotaxi trials in London and signed licensing deals with Mercedes-Benz, Nissan, and Stellantis for consumer vehicles. Wayve's bet is that a camera-first model can transfer to new environments more cheaply than a sensor-heavy stack.

Mobileye, an Intel subsidiary, provides ADAS and autonomous driving chips to over 50 automakers. Mobileye's approach combines cameras with a separate mathematical safety model called Responsibility-Sensitive Safety (RSS). It functions as a supplier, which gives it reach across many brands but limits its control over the full driving stack.

Nuro was founded in 2016 by two former Waymo engineers and originally built small autonomous delivery pods. In late 2024, the company pivoted to licensing its Level 4 autonomous driving software, the Nuro Driver, to automakers and mobility platforms. In July 2025, Nuro announced a partnership with Lucid and Uber to deploy over 20,000 autonomous Lucid vehicles on the Uber platform over six years, with the first commercial robotaxi service targeted for late 2026. Nuro has also expanded to Germany as a base for European operations. Its system runs on automotive-grade hardware built on the NVIDIA DRIVE platform and includes a backup parallel autonomy stack as a safety layer.

Baidu's Apollo Go operates robotaxis in several Chinese cities and reached 250,000 weekly rides in late 2025. Apollo Go uses a sensor-heavy setup similar to Waymo's and is expanding into the UAE and Europe through local partnerships.

The robotaxi market is small and growing fast. Grand View Research valued the global robotaxi market at US$612.4 million in 2025, projecting it to reach US$147.3 billion by 2033 at a CAGR of 99.1%. MarketsandMarkets estimates US$45.7 billion by 2030 at 91.8% CAGR. The spread comes from how each firm defines "robotaxi" and whether it counts vehicles with safety operators.

The broader autonomous vehicle market is larger and steadier. Mordor Intelligence estimated it at US$231.5 billion in 2025, projecting US$747.7 billion by 2030, because it includes Level 1 and Level 2 driver-assistance systems already shipping in millions of cars.

Waymo is the dominant player in the US robotaxi segment. The company was completing over 400,000 paid rides per week across six cities as of early 2026, and has expanded to at least ten metropolitan areas through the year (TechCrunch). Weekly ride volume increased tenfold in less than two years. The company targets one million weekly rides by the end of 2026 (Forbes).

At an estimated US$20 per ride, one million weekly rides would put Waymo's annualised revenue near US$1 billion (Forbes). Waymo has a fleet of over 3,000 robotaxis equipped with its 5th generation self-driving system and is transitioning to its 6th generation hardware on Zeekr and Hyundai IONIQ 5 platforms. Its Arizona manufacturing plant is set to produce 2,000 vehicles in 2026, with capacity for tens of thousands per year (SiliconANGLE).

The patent's relevance here is computational efficiency. As the fleet scales and enters international markets (London and Tokyo are planned for 2026), a model architecture that reduces computing cost per vehicle could compound into a significant operating advantage.

Capital is flowing into autonomous driving at a pace that suggests investors believe the commercialisation window is opening.

Waymo closed a US$16 billion funding round in February 2026 at a US$126 billion post-money valuation, led by Dragoneer, DST Global, and Sequoia Capital (CNBC). Alphabet remains the largest shareholder. Other participants included Andreessen Horowitz, Mubadala Capital, Bessemer, Silver Lake, Tiger Global, T. Rowe Price, Fidelity, Kleiner Perkins, and Temasek. The previous round, a US$5.6 billion Series C in October 2024, valued the company at US$45 billion. The valuation nearly tripled in 16 months.

Wayve raised US$1.2 billion in a Series D in February 2026 at an US$8.6 billion valuation, led by Eclipse, Balderton, and SoftBank Vision Fund 2, with participation from Microsoft, NVIDIA, and Uber (Wayve). A US$60 million follow-on from AMD, Qualcomm, and Arm landed in April 2026 (CNBC). Wayve plans to launch commercial robotaxi trials with Uber in London in 2026 and license its AI Driver to automakers from 2027.

Zoox, acquired by Amazon in 2020 for US$1.3 billion, launched paid rides in Las Vegas and is expanding to Houston and San Diego. Its robotaxis will be available on the Uber app in Las Vegas in summer 2026 (TechCrunch). Zoox operates purpose-built pods with no steering wheel, which took the NHTSA exemption route for regulatory approval.

Nuro closed a US$203 million Series E at a US$6 billion valuation (Crunchbase News).

The capital is concentrating in companies that already have vehicles on public roads. Waymo's valuation jump from US$45 billion to US$126 billion in 16 months tracks the growth in its weekly ride count over the same period.

The patent describes a system that predicts where a car and everyone around it will be in the next few seconds. When that prediction is wrong, it could have deadly consequences. That doesn't even begin to get into the crucial decision, where there are circumstances that could involve preserving the life of someone inside the car versus someone outside the car.

Waymo's existing systems have already shown where prediction gaps can appear. In December 2025, Waymo issued a voluntary software recall after its robotaxis illegally passed stopped school buses in Austin at least 20 times during the 2025-2026 school year (CBS News). In one incident, a car moved forward while a student was still crossing the road in front of it. The Austin Independent School District asked Waymo to stop operating during school pickup and drop-off hours. Waymo refused (CBS News). NHTSA opened an investigation, and the National Transportation Safety Board launched its own probe in January 2026 (TechCrunch).

Waymo said its software had initially slowed for the buses but then continued driving. The specific failure was in recognising the stop-arm as a persistent command, which sits squarely in the kind of multi-object, multi-signal reasoning this patent addresses. A simpler model may be faster, but a simpler model also has fewer internal checkpoints. If one fused model replaces several specialised ones, the system loses the cross-checks that separate subsystems provide.

And what if swerving to avoid hitting the child getting off the bus will hurt the driver? Should the car's systems favour the driver or the child? Why do companies like Waymo get to decide that? This example can be extrapolated to more contentious situations than an adult driver almost hitting an innocent child… and all of a sudden autonomous vehicles are making morally heavy decisions. And if autonomous vehicle companies are aiming for less layers of decision making, where do these kinds of considerations even fit in to assessing the risk profile?

If Waymo deploys a model like this, your next robotaxi ride could be a few hundred milliseconds faster at noticing the cyclist on your left, because the car stopped running the same data through five systems in sequence. That small speed gain, multiplied by 500,000 rides a week, could also mean fewer onboard GPUs per car, which matters when you are building thousands of vehicles a year.

The larger implication may be architectural. If a single fused model handles prediction well enough, other companies building autonomous stacks may face pressure to consolidate their own pipelines, which could reshape how the industry buys and builds computing hardware for vehicles.

This week's patent is US 2025/0162618 A1, titled "Trajectory Prediction from Multi-Sensor Fusion," published by Waymo LLC.

Hit reply and tell us what you think. Vote in the poll above, and find us on Instagram and LinkedIn.

•  Motion Forecasting for Autonomous Vehicles: A Survey with arXiv: Covers the full landscape of trajectory prediction methods, including sensor fusion, graph neural networks, and transformer architectures.

•  Trajectory Prediction for Autonomous Driving: Progress, Limitations, and Future Directions with arXiv: Compares Tesla's vision-only approach with Waymo's multi-sensor strategy, with detailed analysis of each company's research output.

•  Waymo's Skyrocketing Ridership in One Chart with TechCrunch: Shows the tenfold increase in Waymo's weekly paid trips over less than two years, with fleet and city expansion data.

•  GM Halts Cruise's Robotaxi Development with Fortune: Examines why GM shut down Cruise after US$10 billion in investment, with context on the business model challenges facing all robotaxi operators.

•  Wayve Series D Announcement with Wayve: Read Wayve's own account of its end-to-end approach and commercialisation timeline, including OEM licensing plans for 2027.