Mobileye Inside —

This is Intel’s plan to dominate driverless car technology

Intel says its driverless tech will be ready in 2023—I'm skeptical.

A Udelv-branded van roams wet streets at night.
Enlarge / Udelv hasn't announced details about its forthcoming Mobileye-powered delivery robots—such as who will help Udelv build them—but they are expected to look something like this.

I've written before that Mobileye, an Intel subsidiary since 2017, is among the most formidable and underrated players in the self-driving sector. The Israeli company is the leading supplier of the chips, cameras, and software that power today's driver-assistance systems—a couple of years ago, Mobileye claimed 70 percent market share. The company is hoping to enjoy a similarly dominant position in the emerging market for fully self-driving systems.

On Monday, Mobileye announced that its self-driving technology stack would be branded Mobileye Drive. Mobileye says the technology will be "a turn-key self-driving system ready for commercial deployment at scale." A Mobileye Drive system will have 13 cameras, three long-range lidars, six short-wave lidars, and six radars. It will be powered by Mobileye's EyeQ 5 processors.

Mobileye

Mobileye says the technology will be ready for commercial use by 2023. That would be a big deal if true, but I can't help being skeptical. Over the last five years, a number of major self-driving companies have announced optimistic launch dates and failed to meet them.

If Mobileye can get the technology ready for market by 2023, it has the potential to become a major player in the emerging self-driving economy. Even if Mobileye misses its 2023 target, the company is still likely to be a significant player when it eventually enters the market.

Mobileye already has strong relationships with major automakers, and Mobileye plans to market Mobileye Drive to nonautomakers, too. On Monday, Mobileye announced a deal to sell Mobileye Drive to self-driving delivery startup Udelv. Udelv says it plans to deploy at least 35,000 delivery robots between 2023 and 2028 powered by Mobileye's technology. Mobileye hopes that this will be the first of many sales for the Mobileye Drive platform.

Unique strengths

Mobileye began testing in Munich in 2020.
Enlarge / Mobileye began testing in Munich in 2020.
Mobileye

Over the last five years, a number of companies have announced plans to commercially release self-driving technology in 2018, 2019, 2020, or 2021. Few (if any) have met their self-imposed deadlines.

But Mobileye says it's different. The company says it hasn't made unrealistic promises in the past and is on track to deliver full self-driving technology to customers before the end of 2023.

Mobileye does have a deep bench of engineering talent from its years of building hardware and software for driver-assistance systems. Mobileye is also in the process of rolling out one of the industry's most extensive testing programs.

Mobileye tested its technology in Israel and Germany last year, and the company is working to expand testing to Detroit, Tokyo, Paris, Shanghai, and possibly New York City this year. (Not coincidently, several of these cities are also the headquarters of major automakers.) Mobileye argues that testing in a variety of cities is important to ensure that its software doesn't overtrain on the quirks of one particular city.

That's related to another of Mobileye's strategic advantages: access to gobs of map data. Mobileye has signed deals with several of its automaking partners to collect data from cameras on customer cars as they navigate cities around the world. Mobileye says it has largely automated the process of transforming this image data into highly accurate, three-dimensional maps that cover a large portion of the world's roads.

These maps will not only contain information about road geometry—they'll also provide valuable information about driving patterns that will help Mobileye's vehicles blend more naturally into existing traffic.

A distinctive approach

Most self-driving companies use a technique called sensor fusion: they take input from cameras, lidar, radar, and other sensors and combine them into a single unified model of the world. This model is then handed off to a planning module that figures out what the vehicle should do.

By contrast, Mobileye is developing two completely independent self-driving systems—one driven by cameras and the other by a combination of lidar and radar. Once each of these systems has achieved a high level of performance separately, Mobileye will combine them into a single system. Mobileye believes this extra layer of redundancy will give the company higher confidence in the safety of its systems.

There's plenty to like about this strategy, but it's far from obvious that it will be successful. Mobileye continues to rely on a head-scratching statistical argument to help establish the safety of its technology. Mobileye claims that—if it can show its camera-based system can go 10,000 hours between crashes and its lidar-based system can separately go 10,000 hours between crashes—its combined system will be able to go 100 million hours (10,000 times 10,000) without a crash.

This argument implicitly assumes that the two systems' failure modes are statistically independent. That doesn't seem like a plausible assumption: situations that confuse one system are more likely to confuse the other one. So you can't just multiply the two probabilities together.

Ultimately, it's hard to see how anyone—even Mobileye insiders—could be confident that the firm's technology will be ready by 2023. By the company's own admission, Mobileye's camera-based system is more mature than the lidar-based system. The company is still collecting the data it will need to convince regulators and the public that its system is safe. Mobileye might still be on track to deliver fully self-driving technology by 2023. But we won't know for sure until it happens.

Mobileye inside

Udelv

Mobileye's relationships with automakers will give it an advantage selling its technology to automakers. But experts expect that a lot of self-driving vehicles won't be customer-owned. Instead, they'll be in fleets used for on-demand taxi and delivery services.

On Monday, Mobileye announced a deal to supply its forthcoming self-driving technology to Udelv, a startup that is building self-driving delivery robots. Under the deal, Mobileye will supply enough chips, sensors, and software to power 35,000 delivery robots between 2023 and 2028.

Udelv is unusual among self-driving startups in that it doesn't see self-driving technology as its primary differentiator. Udelv CEO Daniel Laury believes that self-driving hardware and software will eventually be a commodity that a company like Udelv can license from multiple vendors. So Udelv has focused on other aspects of the delivery business.

Udelv has been developing a new type of delivery truck that's optimized for self-driving operations. Rather than having discrete compartments, Udelv's truck features adjustable shelves and a two-dimensional iris-like door that gives customers access to one compartment at a time. This flexibility allows a single Udelv truck to carry a number of orders for different customers, with different item sizes, without worrying about customers taking the wrong items.

Udelv is also focusing on the logistics of a delivery business. Udelv is already operating a delivery network using prototype vehicles with safety drivers behind the wheel. One of its biggest clients is a Houston auto-parts supplier called XL Parts. Udelv hopes that XL Parts will be the first of many retailers who lease Udelv's trucks and technology to enable them to offer driverless delivery services to customers.

Mobileye wants to work with everyone

Nuro on Monday announced a pilot project to deliver Domino's pizzas in Houston.
Enlarge / Nuro on Monday announced a pilot project to deliver Domino's pizzas in Houston.
Domino's

One of the most fundamental questions in any industry is how work is divided up among different companies. For example: most smartphone makers license the Android operating system from Google and buy chips from Qualcomm or one of its competitors. By contrast, Apple is vertically integrated, making its own operating system and many of its own chips for the iPhone.

Self-driving companies have a wide range of theories about how the self-driving industry should be organized. Tesla is hewing to the Apple model: it's making its own cars and self-driving software, and it is planning to run its own ride-hailing network once the technology is ready. Waymo is planning to buy cars from traditional automakers, but it also hopes to develop most other components of its taxi service—lidar sensors, self-driving software, ride-hailing app—internally.

By contrast, Mobileye hopes to sell its hardware, software, and sensors to as many partners as possible. Its deal with Udelv is nonexclusive, and Mobileye is presumably hoping to sell the same technology to Udelv's competitors. We expect Mobileye to also court companies creating taxi services, shuttle services, long-haul trucking services, and any other kind of self-driving service you can think of.

There's an obvious attraction to this business model, since it lets Mobileye focus on what it does best—develop chips and software—and leave the messy business of operating delivery and taxi fleets to other companies. But the big question is whether Mobileye can actually build a self-driving stack that's suitable for use by a wide number of companies.

A big advantage of the vertically integrated approach is that it sometimes enables a faster pace of innovation, since different parts of the product can be tailor-made for a specific purpose. That's the approach being taken by Nuro, one of Udelv's leading competitors.

H-Town

Founded by two ex-Google engineers, Nuro is also building self-driving delivery robots. Nuro just announced a deal to deliver pizza for a Domino's restaurant in the Houston area. If the pilot program goes well, it will presumably expand to many more Domino's and, eventually, to many other kinds of stores and restaurants.

Nuro's vertically integrated approach means that it can carefully design its hardware and software for a robot designed for delivering pizzas and other short-range, low-speed trips. Nuro's robots don't need to go on the freeway. They also don't need to worry about keeping passengers safe—since the vehicles will never have passengers.

Developing a self-driving system that's customized for carrying goods at relatively slow speeds might just be an easier problem than developing a self-driving technology that works at all speeds and on all road types. Nuro can focus all of its energies on that task, while Mobileye is trying to design a system that works at freeway speeds and on a wide range of vehicle types. That might be more difficult and take longer.

Or the opposite could be true: maybe Mobileye's vast data-collection efforts will enable the company to master the general self-driving problem more quickly than the focused efforts of a startup like Nuro. And Mobileye's open approach might let the company expand its market share rapidly by selling its technology to many partners at once.

Channel Ars Technica