We began our self-driving car program to create a universal and scalable autonomous vehicle technology. We aim to create a self-driving car that safely and effectively operates regardless of geography and traffic conditions. As part of our strategy to realize our self-driving vision, we decided to test our vehicles in diverse locations around the world that challenge our system.
Since we first introduced our self-driving cars on public streets in December 2017, we’ve conducted testing in five different locations across three countries: Moscow, Innopolis, and Skolkovo in Russia; Las Vegas in the United States; and Tel Aviv in Israel.
Combined, the driving conditions and challenges of these locations ultimately help advance our self-driving system to better operate in new environments. As we continue testing in these locations, we plan to collect more data by introducing over 100 cars by the end of this year.
We’re excited to continue our work building a scalable autonomous technology that can operate in any number of driving conditions. Here are just some of the many dynamics that are helping us improve our technology.
Rules and road infrastructure
Though traffic lights universally use red, yellow, and green to direct traffic, the signals and layouts of traffic lights vary by country. Self-driving technology should be capable of detecting different signals on traffic lights regardless of location. Our self-driving technology relies on computer vision, machine learning, and highly detailed 3D maps to identify the correct lights and signals.
Israeli traffic lights have the standard three colors, but to direct traffic at large intersections, separate traffic lights control every possible way through them. Our cars identify the correct traffic light while ignoring the signals of the other lights to proceed along the route accordingly.
In Russia, there is usually one traffic light to regulate the entire intersection. In addition to the standard red-yellow-green signals, traffic lights may also include green arrows for left and right turns. Traffic lights in the US can have five or more signals, creating much more variety compared to typical lights in Russia and Israel. US traffic lights are also usually located on the opposite side of the intersection, whereas in Russia and Israel, vehicles stop at lights just before the junction. In the US, our self-driving car had to learn to properly recognize a large number of signals, in addition to detecting the lights 25 meters (82 feet) farther from their location in Russia and Israel.
Intersections which are not controlled by traffic lights, or uncontrolled intersections, present unique challenges as there are often different laws around the world governing the right of way and passing.
At uncontrolled intersections in Russia, vehicles on the right always have the right of way. In the US, uncontrolled intersections are managed by a unique law that gives priority to cars that first reach the intersection. Vehicles in the US also come to a complete stop at intersections with four-way stop signs, which is not common in our other testing locations.
In Israel, the road network includes a higher number of roundabouts compared to the other two locations. Our platform is learning to manage multiple traffic situations by frequently driving through roundabouts that include one or two lanes and numerous exits.
By collecting data and challenging the platform to improve how it manages diverse uncontrolled intersections, we are increasing the platform’s scalability.
Besides the official traffic rules, each location has its own so-called social rules — the way drivers behave on the roads. Self-driving cars must be prepared to respond to any number of unpredictable human driving behaviors. As we collect more data from reacting to social rules on the road, our self-driving system becomes better equipped to handle various situations on the streets.
In Israel, for example, due to limited parking in Tel Aviv, drivers often park their cars wherever they can find a spot. Passing illegally parked vehicles, such as double-parked cars, requires high precision and can be challenging even for skilled drivers, due to the narrow, winding roads of Tel Aviv. Self-driving cars must be quick to respond in these situations, especially considering the fast-paced driving there.
The hectic streets of Moscow, where drivers frequently flout the rules, challenged our cars to learn how to navigate safely through chaotic road conditions and speeding drivers. In Russia, for instance, drivers tend to exceed speed limits by up to 20kph (12mph) because that is the minimum threshold to receive a ticket. Anywhere we operate our self-driving vehicle, we strictly follow the rules and the speed limits, so in Moscow, we must also adjust to driving in a traffic flow that goes 10–15kph (6–10mph) faster than the car.
When we brought our car to Las Vegas, it was comparatively easier to drive, as the road network is an orderly grid, and drivers are more likely to follow the rules. Over a period of just two weeks leading up to CES 2019, we taught our self-driving system how to drive on American roads for the first time. Our car had to adapt to the different driving style in America, and one example of a difference in social rules concerned the dynamics of lane changes. The many multi-lane roads in Las Vegas and the less severe traffic compared to Moscow meant that it was far more common for drivers to suddenly cut across many lanes to reach their turn or exit. Our self-driving car had to learn how to adapt to these new driving conditions, and how to react when cars would abruptly cut in front of it before passing to the next lane.
By testing in three very different locations, our self-driving car has demonstrated the ability to adapt and respond to some of the distinct driving behaviors that exist around the world.
Weather presents any number of challenges for self-driving cars. With our multi-country testing strategy, we’ve put our self-driving tech to the test in dramatically different climates, challenging our vehicles to adapt to varying local weather conditions. Winter testing in Moscow forced our cars to learn how to drive in snowy weather, while testing in Tel Aviv subjected our autonomous hardware to high heat.
We started our tests in Russia, where the climate provides us with various weather conditions throughout the year. One of the most challenging situations is heavy snow. It covers road markings, obscures signs, and adds a lot of noise to the data we collect from the car’s sensors. One of the biggest challenges caused by snow is the way it interrupts LiDARs, as the laser beams hit the snowflakes and don’t reach the objects behind them. We’ve been developing ways to deal with snow from the very beginning of our work, and we have gathered many real-world examples to teach the car’s neural network to create a 3D picture of the surrounding world with minimal noise. Thanks to our winter testing, even in heavy snow, our self-driving cars can detect objects around them, remain in the correct lane, and recognize crosswalks.
When we launched our program in Israel, we faced an opportunity to test our hardware in high temperatures. Overheated equipment can have lower performance, interruptions in operation, or it can altogether stop working. To prevent these possible problems, before we started testing in Israel, we set up our hardware and sensors to adapt to the higher temperatures than we experienced in Moscow. We also built custom cooling systems for both the interior and exterior of the car. We continuously test the reliability of our sensors and components to ensure our preventative measures are continuing to manage the heat as spring turns to summer. This May in Tel Aviv, temperatures reached as high as 40° C (104° F), and our car successfully operated throughout the city all day long without issue.
Traffic conditions are dramatically different in each city we test in, requiring a diverse approach based on several factors. We began testing our self-driving cars in Moscow, a city of 12 million people with some of the worst traffic in the world. By testing our cars in the grueling driving conditions of Europe’s second largest city, our self-driving tech has shown that it’s capable of navigating challenging, crowded urban environments.
Testing in Tel Aviv brought a new set of challenges. Though our self-driving car was used to busy city streets, Tel Aviv’s narrow, windy roads are densely populated by two-wheeled vehicles, much more so than in Moscow. Our self-driving car had to learn how to share the road with bicycles, scooters, and motorbikes, along with the standard dynamics of cars, trucks, and pedestrians. Managing traffic in Tel Aviv provides our prediction algorithm with a large amount of data on two-wheeled vehicles and their behavior patterns, so our self-driving tech can be deployed in other cities with similar vehicle dynamics on the road.
Driving in lighter traffic presents a new set of challenges. In Innopolis, Russia, where we operate our robo-taxi service, there is lighter traffic, with citizens freely crossing the street anywhere or often riding bikes in any lane (and in any direction). This location allows us to interact with pedestrians in various unexpected situations rarely seen in big cities.
These are just a few examples of the challenges and opportunities we’re encountering across our test locations as we advance our self-driving technology to be universal and scalable. By teaching our tech how to process numerous road hazards around the world, such as two-wheeled vehicles, pedestrians, erratic driving, and challenging weather, we are better preparing our self-driving car to operate in new locations.
Of course, there are still many challenges ahead. Stay tuned to learn more about our self-driving technology!