With applications already sent in for 52 Chrysler hybrid minivans, Waymo is putting the pieces in place to have the U.S. first driverless taxi service by year’s end.
By Murray Slovick, Contributing Editor
The autonomous car testing game is heating up and not just with the arrival of warmer weather. Unless you’ve spent the past few years in a cave, and possibly even then, you know that Google’s self-driving car company Waymo announced that it will launch the first driverless commercial taxi service in the U.S. in Phoenix later this year. The cars can be summoned using an app and customers will have to pay just like any other taxi service.
There’s no script for how to do this—it’s virgin territory for everyone involved. But a bigtime effort is being made involving a number of large-scale initiatives. Recently, one of these organizations, Waymo, lifted the veil a bit on its technical acumen courtesy of a Team Google blog.
Following last month’s Google IO developer’s conference, Dmitri Dolgov, CTO and VP of Engineering, gave us a rare glimpse into how Waymo was using artificial intelligence (AI) and machine learning (ML) to make self-driving cars a reality. Dolgov noted that whileperception is the most mature area for deep learning, the company is also using deep nets for everything from prediction to planning to mapping and simulation. ML, he added, enables the Waymo autonomous driving system to navigate difficult situations such as maneuvering construction zones, yielding to emergency vehicles, and giving room to cars that are parallel parking.
Dolgove explained that Waymo has trained its ML models using lots of different examples accumulated during the 6 million miles driven on public roads and the hundreds of millions of interactions observed between vehicles, pedestrians and cyclists.
The Importance of Infrastructure
It takes more than good algorithms to be able to put self-driving vehicles on the road, Dolgove noted. Infrastructure plays a key role in training and testing Waymo’s ML models. The company employs Google’s TensorFlow ecosystem, including TPUs (tensor processing units) to train its neural networks.
TensorFlow is an interface for expressing ML algorithms, and an implementation for executing such algorithms. A computation expressed using TensorFlow can be executed with little or no change on a wide variety of heterogeneous systems, ranging from mobile devices such as phones and tablets up to large-scale distributed systems of hundreds of machines and thousands of computational devices, such as GPU cards.
The system has been used to conduct research and deploy ML systems into production across more than a dozen areas of computer science and other fields. These include speech recognition, computer vision, robotics, information retrieval, natural language processing, geographic information extraction, and computational drug discovery.
Google announced the TPU last year and followed up with a detailed study of its performance and architecture. In short, the company claims that the TPU delivered 15X-30X higher performance and 30X-80X higher performance-per-watt than contemporary CPUs and GPUs.
With TPUs, Google says it can train neural nets up to 15X more efficiently. They also rigorously test ML models in simulation, driving the equivalent of 25,000 cars all day, every day.
Testing and Remote Operation
In April, California’s Department of Motor Vehicles (DMV) started accepting applications for fully driverless testing. Waymo’s application, obtained using public record laws and first reported by IEEE Spectrum, gives insight about how the planned Waymo taxi service would work. The application sought permission for 52 fully driverless Chrysler Pacifica hybrid minivans, 27 registered in California and 25 with Arizona plates. Waymo intends to test its vehicles in an intensively mapped geofenced area of about 50 square miles near Mountain View, Calif.; the company’s software will not create a route that travels outside of this zone
California requires fully self-driving vehicles to have a remote operator in case the car fails or suffers a collision. Thus, Waymo will have two teams monitoring each car while it’s in service. Fleet Response Specialists possessing valid driver’s licenses will be responsible for monitoring the status of all Waymo vehicles in real-time using a virtual tool. A separate Rider Support team provides customer support and is available to communicate with passengers at any point. Waymo specified that it has trained 70 Fleet Response Specialists and 23 Rider Support team members.
Waymo reports that its vehicles can handle most roads and parking lots, and speeds of up to 65 miles per hour (mph). They can also cope with fog and light rain, and night-time driving. If Waymo’s minivans encounter heavy rain, snow or ice, flooded roads, or off-road terrain, they will seek a “minimal risk condition” (stopping in or by the side of the road). Waymo vehicles will also halt operations if they detect a failure, hit something, or sense that their airbags are deploying.
According to the company’s DMV filings, the autonomous taxis are programmed to pull over once it detects a police car’s flashing lights behind it. When this happens, the Waymo driverless taxi will unlock its doors and roll down a window to enable an officer to communicate directly with the Rider Support team. Waymo’s DMV application also includes a law enforcement “interaction protocol,” which provides information for paramedics, police officers, and firefighters.
The financial services company UBS predicts 12% of cars sold in 2030 will be for driverless taxi fleets. The firm further expects that demand for self-driving taxis will take off around 2026, depending on public acceptance of the technology and regulatory approval.