CEO DailyCFO DailyBroadsheetData SheetTerm Sheet

Inside Waymo’s autonomous trucks and A.I. systems

October 13, 2020, 3:42 PM UTC

If you’ve heard of Waymo, it’s probably for its self-driving cars. But the autonomous-vehicle subsidiary of Google’s parent company is also developing self-driving trucks that its executives hope will become the backbone of a major freight delivery business.

On Tuesday, Waymo revealed some previously undisclosed technical details about those trucks. The details are likely intended to assuage concerns about autonomous vehicle technology failing to progress as quickly as many people had expected.

There are multiple reasons for the delay, including the research being more difficult than originally believed and the high cost of perfecting and training A.I. systems.

Boris Sofman, Waymo’s director of engineering for trucking, sounded a hopeful note for his unit’s trucking initiative by saying much of the research into self-driving cars is also applicable to trucking. “The foundational systems are all the same,” he told Fortune, which implies that Waymo may be able to save on some costs.

In other words, Waymo’s autonomous cars and trucks generally have similar computer brains. “If I want to merge into a lane, whether I’m a car or a truck, the logic is actually quite similar,” Sofman said. “You look for an opening, you try to position yourself in the right way in order to interact with that opening, and if it doesn’t exist you create one.”

Of course, the significant difference in size between cars and trucks requires them to make different decisions when merging into traffic. For Waymo’s trucks, that means using two lidar sensors that remove nearly all blind spots except for what’s “100% behind the truck, which exists for humans as well,” Sofman said. 

Ultimately, Waymo’s biggest challenge to perfecting self-driving trucks is getting enough real-world driving experience to obtain the necessary data used to train the company’s deep learning systems. “It’s a data problem rather than an architecture problem,” Sofman said.

He declined to comment about how many self-driving trucks Waymo has or how many miles they’ve logged on the road since their debut in 2017. It’s therefore difficult to compare Waymo’s progress against competitors like TuSimple, Daimler, and Embark. Waymo still requires that human drivers be behind the wheel while trucks move, at least until the company’s next generation of trucks and computers, he said.

Sofman hopes that Waymo’s new truck testing grounds in Dallas will open the door to more long haul trials as the trucks shuttle back and forth between an existing testing facility in Phoenix. This is crucial for its autonomous trucking research because the long drives on the freeway could provide a wealth of much-needed data. Long drives would also mean that the self-driving trucks could experience more “edge cases,” or anomalies like cows running across the freeways; while these experiences are unusual, they are needed in order to create robust self-driving cars.

“It’s safe to say we know there’s still a lot in front of us and there’s a lot of unknowns, but we started to discover a lot through just our driving today.” Sofman said. 

Update: A Waymo spokesperson added that the company’s planned freight delivery business will rely on partnerships with original equipment manufacturers (OEMs), fleet, and freight players, but declined to name the partners. Although Waymo currently uses Peterbilt trucks as part of its self-driving truck research, Peterbilt is not one of Waymo’s OEM partners, the spokesperson said.

Jonathan Vanian 
@JonathanVanian
jonathan.vanian@fortune.com

A.I. IN THE NEWS

Still not enough. Members of The National Security Commission on Artificial Intelligence met last week in an online talk to discuss federal A.I. policy measures, which included a proposal for “the creation of a Malign Information Detection and Analysis Center, or MIDAC, staffed by a team of intelligence analysts,” according to a report by Nextgov. The commission’s chairman and former Google chief executive officer Eric Schmidt said: “I think everything you’re recommending we should do, and it’s probably still not enough. In other words, we’ve got to rethink how we’re going to deal with this because with the broad ability to do deepfakes, and fake texts and so forth, it’s only going to get worse.”

Think of the children(’s faces). Human Rights Watch has discovered that the city government of Buenos Aires, Argentina has been using facial-recognition to monitor children suspected of “criminal activity,” MIT Technology Review reported. “Buenos Aires first began trialing live facial recognition on April 24, 2019,” the report said. “Implemented without any public consultation, the system sparked immediate resistance.”

Facing a problem. As China continues its push to use facial-recognition software as part of its efforts to contain COVID-19, some Chinese residents are concerned about “data leaks,” in which sensitive data about their faces could get unintentionally leaked due to “sloppy mistakes,” The South China Morning Post reported. “As the use of facial recognition has exploded across China in recent years, the country has been hit with numerous data leaks related to the technology,” the report said. “But people without any legal expertise might not know how to fight back, especially when the installation of such systems are being pushed by local police.”

Am eye biased? Researchers from Birmingham University and other institutes examined a collection of data sets used to train A.I. systems that can identify eye disorders, and discovered that the eye data is predominantly comprised from patients in North America, Europe, and China, Wired reported. As a result of the biased data sets, the A.I. eye-scanning technology will likely produce poorer results on underrepresented racial groups. From the report: "The researchers found other problems in the data, too. Many data sets did not include key demographic data, such as age, gender, and race, making it difficult to gauge whether they are biased in other ways."

EYE ON A.I. TALENT

Twilio hired Jeremiah Brazeau to be the cloud communication company’s chief technology officer. Brazeau was previously a distinguished architect at Salesforce.

Yellowbrick Data picked Mark Cusack to be the enterprise tech firm's CTO. Cusack was previously a Teradata vice president for data and analytics. Yellowbrick specializes in data warehousing technology, which, as Fortune writer Aaron Pressman reported, is one of the hottest corporate technologies. Yellowbrick rival Snowflake recently went public in what is considered to be the highest valued software IPO ever.

Eye on A.I. emailed a couple of questions for Cusack about his thoughts on machine learning. His responses were edited for length. 

What distinguishes one data warehousing service from another?

The new players in the market are innovating in two key areas. First, they are investing in providing a “consumer cloud experience,” in which it is much easier to use and manage a data warehouse: easy to try, easy to buy, and easy to elastically expand capacity in line with business growth, and with low to no administration overheads. Second, the true innovators realize that the hardware is as critical to the overall performance as the database software, and they are developing offerings that span database operations across the hardware and software stack to deliver differentiating performance. 

Explain the importance of data warehousing for machine learning practitioners.

Data warehousing will become really important in the A.I./ML space. We’ll see the emergence of data warehouses as “feature stores”—repositories for the engineered features that feed ML models. There are several reasons for this. First, the data warehouse typically contains the most valuable and up-to-date data within an organization, and it is often this data that is needed to support model training and scoring. More than 80% of the time taken to develop ML models is spent in preparing the data. Data warehouses have in-built data transformation capabilities that can be used to wrangle the raw data into the features needed to plug into ML algorithms, and they can perform this task quickly and at massive scale.

Second, data warehouse feature stores promote the reuse of engineered features across different analytics projects. The same data modeling approaches that have been used for decades to support efficient BI [business intelligence] and reporting applications can now be used to overlay a schema across different feature sets, which reduces duplication and increases the quality of the data. Data scientists can use well-modelled feature stores in a data warehouse to find and reuse features created for past projects, further reducing the time to yield business value from analytics.

EYE ON A.I. RESEARCH

On artificial and real brains. Researchers from Stanford University, Princeton University, the University of Washington, and other institutions partook in an online conference last week by Stanford about the confluence of A.I., psychology, and neuroscience.

One fascinating portion of the event: Stanford assistant professor Chelsea Finn’s robotics research. Finn explained that a big challenge in A.I. is for researchers to create robots that are good at generalizing.

“Humans are good at generalizing—it’s not clear how we do it,” Finn said.

She explained that babies tend to learn by observing the actions of their parents. Using that basic premise of learning by observing, Finn and her colleagues showed a robotic arm (outfitted with deep reinforcement learning software) photos of various things the robot arm could potentially do to a nearby drawer, among other examples. Her team classified certain photos as “interesting” in order to prod the robot to do its own set of “interesting” actions.

After analyzing the photos, the robotic arm learned to "interact with the drawer and plays with it in different ways,” Finn said.

FORTUNE ON A.I.

IBM showcases latest A.I. advancements on Bloomberg’s “That’s Debatable” TV show—By Jeremy Kahn

Someone finally outright accused Huawei of collusion with the Chinese state—By David Meyer

What experts think of robots’ threats and benefits to humanity—By Brett Haensel  

Facebook tightens rules on political posts and ads as the presidential election nears— By  Danielle Abril

The tech startup trying to restore our faith in COVID-free air travel—By Vivienne Walt

IBM CEO will be on the hunt for acquisitions and new businesses after spinning off services unit—By Aaron Pressman and Jonathan Vanian

BRAIN FOOD

What’s up with patent law and A.I.? The United States Patent and Trademark Office (USPTO) released a report detailing how various members of bar associations, companies, academia, and trade and advocacy groups view A.I.’s impact on existing patent law. The rise of deep learning in recent years has posed some notable brain teasers for legal scholars. For instance, if an A.I. system can generate its own designs, who controls the patent?

According to the report’s findings, maybe these vexing patent dilemmas aren’t that complicated after all. The majority of the responders concluded that “current USPTO guidance, especially on patent subject matter eligibility and disclosure of computer-implemented inventions, is equipped to handle advances in AI.”

From the report:

The vast majority of public commenters asserted that current inventorship law is equipped to handle inventorship of AI technologies. One commenter went as far as to state that “there is no urgency to revise the law with respect to inventorship.” Many of these commenters suggested that assessment of conception should be fact-specific, as in the analysis done today. For example, one commenter stressed that there are different ways in which a natural person may contribute to the conception of an invention and that each contribution “should be evaluated on a case-by-case basis,” as is the law today.  A related view was that a data scientist carrying out the task of building and testing a use of an AI technology invention is doing nothing more than reducing the invention to practice. In the words of one commenter, “running [an] AI algorithm on the data and obtaining the results is unlikely to qualify as a contribution [to conception].”