How AWS fuels Zoox’s autonomous robotaxis – News

Beginning last month, visitors to Las Vegas were treated to a sight that’s unusual even by that town’s standards: boxy, two-sided vehicles looking a little like toasters perched atop overgrown roller skates.

They’re called Zoox (pictured), and if Amazon.com Inc. has its way, they’ll be ubiquitous on the streets of major cities in a few years. The retail giant acquired startup Zoox Inc. five years ago, but the vehicles began taking on passengers only last month. Visitors to Las Vegas can ride for free to select destinations on the Strip. Zoox also has pilots in San Francisco, Seattle, Austin, Miami, Los Angeles, Washington, D.C., and Atlanta.

Amazon Web Services plays a central role in powering Zoox’s futuristic robotaxis. The Amazon subsidiary relies on the cloud provider’s infrastructure to train, test and refine the machine learning models that handle navigation and safety. But Zoox is just another customer to the cloud giant, albeit a preferred one. “Our role is to give Zoox the tools and infrastructure to get it right,” said Paul Roberts, AWS director of technology for strategic accounts.

Smart endpoints

Autonomous vehicles are essentially endpoints in a complex distributed network. Because latency isn’t an option at 60 miles an hour, each vehicle is essentially a rolling data center, using its on-board infrastructure to navigate while capturing vast amounts of real-world information about everything from traffic patterns to weather conditions and pedestrian behavior. That data is streamed to AWS, where large-scale simulations predict how all Zoox vehicles should respond to dynamic situations such as someone stepping into a crosswalk or a reckless driver weaving in traffic.

Once trained and validated in AWS, updated models are deployed to Zoox’s vehicles. Each car can operate independently, but cloud-based learning and edge-based execution enable continuous improvement without sacrificing safety or latency.

“There’s a tight interlock between where they’re trying to go and how we can best support them,” Roberts said, “but safety is job zero.”

Unlike Waymo, the robotaxi service owned by Google LLC parent Alphabet Inc., Zoox vehicles aren’t retrofitted passenger cars. Designed from the ground up for autonomy, they can move forward, backward and even sideways without turning.

There is no driver seat, pedals or other conventional controls. Passenger seats face each other, permitting riders to converse or even play card games while in transit. Passengers hail a Zoox via mobile app the same way they would an Uber or a Lyft.

Massive backend

Despite the vehicles’ diminutive size (currently 11.8 feet long and 6.2 feet tall), they require a ton of cloud computing power. The backend is based on Amazon’s Elastic Kubernetes Service, which can scale up to thousands of compute instances for model training. “They don’t have to worry about setting up their own clusters; it’s all managed,” Roberts said.

The data center interconnect is AWS’ Elastic Fabric Adapter. “It allows us to use all the network paths available, some with up to 3.2 terabits per second of throughput,” Roberts said. “If you tried to do that in a traditional data center, it would take forever.”

Amazon’s FSx for Lustre managed distributed file system feeds training data from S3 storage to computing clusters delivering “hundreds of gigabytes per second across thousands of instances,” Roberts said.

The combination allows Zoox machine learning models to analyze petabytes of sensor data and retrain in hours rather than weeks. For more advanced workloads, Zoox taps into AWS EC2 Capacity Blocks, allowing it to reserve up to 2,000 graphics processing units in a single cluster for simulations scheduled as far as eight weeks ahead.

Plug in and go

Getting data from the street to the cloud required some creative engineering. A single vehicle can generate up to 4 terabytes of raw data per hour. Cellular data networks can’t handle that kind of volume, so each Zoox pulls over occasionally to plug into AWS Data Transfer Terminals at secure, physical locations to upload datasets directly at speeds of up to 400 gigabits per second.

Amazon – and the half-dozen other companies building robotaxis around the world – see the autonomous taxis as a way to address the massive inefficiencies of personal vehicle ownership. Roberts noted that only about 4% of human-driven vehicles are in use at any given time, while drivers spend 95 billion hours behind the wheel in the U.S. alone each year.

“When people begin taking rides in Zoox, they say it’s weird at first to have no steering wheel, but over time they get used to it,” Roberts said. “Just like Uber or Airbnb, it’s only a matter of time before it feels normal.”

Photo: Zoox