The glare of the sun, icy wind, dust from the construction site of a new metro line. The Parking lot leaves a white-red car full of fancy electronics. Nobody’s driving the car. Taxi dutifully waiting passengers. Sounds futuristic, in the spirit of Cyberpunk 2077? But it is a reality. We swept through Moscow at the drone of «Yandex», and that’s what he knows.
In the driver’s seat sits a specially trained instructor (steering wheel and pedals, it does not touch), we — who where. The autopilot set the route and are allowed to move. Turn on your Blinker, the car gently creeps to the surrounding area through the filling, blind turn and a tricky entrance to aminyevskoye motorway. The interactive display displays more information than can be seen in the Windows here and all detected objects (360 degrees!) and high-precision map with the estimated trajectories and the behavior of the AI.
On-Board system draws a General route options driving.
To choose actions is impossible, the system only visualizes options: to rebuild the strip «just right» to avoid the slug on the old «Oka» or skip the spit to the left. It also displayed the estimated speed of movement, barriers to the maneuvers of robotaxi recognized signs, pedestrian crossings, dangerous areas… to Understand the intentions digital driver easily without instruction — it is sufficient to observe the behavior of the drone a couple of minutes.
The system is smarter than it seems at first glance. It is, for example, knows that staying on the right lane public transport — not just a obstacle, but still a great generator of sudden pedestrians, running to the landing. Live drivers are also able to predict this situation, but not every time insure. In the same way the AI detects small animals on the road. Squirrels, rabbits, dogs and cats the car will try to skip safe for yourself and passengers.
In the first minutes of the trip scary and suddenly the digital brain will begin to kink?
But then some mild anxiety was replaced by a stone tranquility: the machine goes to Moscow perfectly. A person is able to provide a little more comfort in terms of ride, but with a probability of 99% inferior to the system in the speed of reaction to unusual situations. And, I confess, is not so disciplined.
On the road robotaxi holding confidently. But how this miracle of technology works? A tour inside the car we had Artem Fokin — Director of business development drones of «Yandex».
Iron and the brain
The neural network «Yandex» there is absolutely no difference in what the machine to function. Importantly, there was a technology Drive-by-Wire digital throttle control, brakes and steering via CAN-bus. In fact, this is one reason why everyone runs in on the Toyota Prius. Firstly, the model uses not only the Russian IT-giant — and a successful international experience, so it’s easy to learn. Secondly, these cars powerful mains and large trunk.
The first version of autopilot consume obscene amounts of energy and taking up the entire trunk of the station wagon. The current system is a server unit, hiding in a niche for a spare. In a slim metal body — power supply units, a pair of server Xeon processors, three NVIDIA graphics cards and many gigabytes of RAM. But even this configuration is largely redundant and not used at 100% system Autonomous movement. Stock performance founded for developers — allows you to connect additional analytic modules on the fly to copy data from vehicle sensors, to monitor the implementation of the code of the autopilot. And also to be able to expand it without resorting to hardware resources, not to waste time on upgrades.
Today run are a few «Priuses» different versions differ as the machines themselves and the equipment on Board, its location relative to the housing. Experts are trying to choose the best option: I want to have maximum visibility and it does not suffer from adverse factors like bad weather and interference from other vehicles. To collect car from ready modules now available for a few days.
The question price — about 6 million rubles.
The price includes pre-owned Toyota Prius, shipping, customs clearance, mountain expensive server equipment and the necessary electronics: radar, laser rangefinders, wide-angle camera with the function of night work and so on. A year ago, a drone was worth almost 20 million So don’t be afraid of the prices, the iron is rapidly getting cheaper.
Now «Yandex» is building the second generation of robotronica. Took the new Hyundai Sonata. The engineers didn’t even have to crack the side IN representation of the Korean brand itself has proposed documentation, and implementation of Autonomous control. In the end it worked for CAN-bus: program taught you to drive an unfamiliar car, explained how to accelerate and gently brake.
Risk and human factor
The main feature of the autopilot — 100% discipline. And we are not talking about unobjectionable observance of traffic rules. No one is able to steer with such a concentration, as it is able AI. Its a physical thing: our brain is not engaged in the tasks in parallel, he quickly switches the contexts. A vision is only a small area of focus, all the rest is developed the work of the neural network in the head, according to extended picture. The autopilot also captures all 360 degrees around you, keeps track of hundreds of factors that can see in complete darkness and — most importantly — always calm.
When we launched the unmanned taxi in Innopolis, regularly encounter the following traffic situation. A large bus that brings people from Kazan Innopolis is on the right side, the road was two lanes. He just stands right and all. The pedestrian crossing in front of him, but sometimes people run out of the bus. Thus the bus — it’s big and iron, even radar does not sweep, so it’s a blind spot. Therefore, we realized two things. First, it is necessary to program the car so that it is different from blind spots. Secondly, to teach them to be prepared for unforeseen consequences: from blind suddenly can see other road users.
The result of this debug is extremely responsible behaviour of AI near public transportation. Passing another blind area, the car slows down just enough during emergency braking not to offend potential pedestrians. Live driver, of course, also realize that the road could pop up another «pin» — but how many times do you pull over? And if the cabin still plays the music is cheerful and all thoughts of the Scorcher is the upcoming football stream? A key difference between man of the autopilot. AI always acts according to the algorithm «blind spot — reduce», and no matter what plays on the radio.
The second advantage of digital brain — no tunnel vision characteristic of people. The system sees equally well in all directions, reads dozens of times per second the velocity vectors of all objects within a radius of 250-300 metres. For her there is simply no sudden scenarios: there are only unaccounted behavior in the code. If the drone does not work to handle the unusual situation of General methods, experts will teach his particular decision. Actually, for the sake of the machine and wound millions of kilometers in the metropolis. By the way, in 90% of cases the whole «Unistat» solved a reduced speed to a safe (when it is possible to urgently stop without accident).
When running autopilot in Innopolis many interesting cases have arisen. The system under test, the engineer sits on the right, the car is driving itself, there’s nobody driving, and the road is two lanes in one direction. Suddenly a car suddenly changes lanes from the left lane to the right. The engineer was drawn to get out of the glove compartment of the keyboard to make a mark «Unmotivated rebuild, we need to understand». But after some time realizes that the lane that he was moving, going towards the two cyclists. On the opposite, without lights! He did not see them, unlike lidar and radar in the minds of digital data is combined from all sensors. Taxi evaluated the threat, calmly regrouped and went on.
The software handles millions of such cases, including multiple faults. The airborne system is able to read the number of errors that knows how to classify them. Something can be ignored, warning the user behind the wheel. Other problems require either a manual confirmation from the man or the machine translate transport mode emergency stop is activated the relevant alarm, AI safely resets the speed, stops, and activates manual mode.
If you don’t like the traffic situation or the behavior of the robot to «pick up» the wheel easily at any time — the autopilot immediately stops any interaction with the car, is the driver to influence the authorities. For emergency episodes emergency button, hardware and disables all the system of Autonomous driving. Bang! and you have the usual Toyota. Well, except that strange thing on the roof and a server in the trunk. However, this is an emergency system in casual mode it will hardly take advantage of.
Learning digital brain
One of the most important elements of Autonomous driving / navigation system with maps high-definition. To rely only on a GPS is impossible. The signal is able to get lost in the skyscrapers, the gap in the tunnel, it, in the end, you can drown out or replace. And the positioning accuracy is insufficient. Therefore, AI relies on a comprehensive solution: satellite systems and cellular base stations provide a General understanding of where the car is located. And digital vision, gyroscopes and accelerometers help the autopilot to use 3D maps of «Yandex».
Unfortunately, the market for high-precision three-dimensional maps now all difficult. Worldwide there are about 200 companies, claiming that they make the cards high definition. The caveat is that if you don’t develop the autopilot, you will not be able to create a suitable map. Get some abstract product, which is simply no applications.
To create three-dimensional maps of «Yandex» enough equipment that is already installed on the transport complexes of self-control. Any car is not only able to travel independently, but also works as a scanner space. Single passage through the streets with a live driver creates an array of data sufficient for further navigation. Any subsequent pass (with human driver or without) clarifies and complements the scheme of orienteering.
Unmanned vehicles «Yandex» test not only in Moscow and Kazan. In addition to Russia, domestic robots toured the United States and Israel. This diversity of routes is only for the benefit of the system. Engineers receive Analytics on various occasions on the road and teach the neural network to adequately respond to an atypical situation. For example, in tel Aviv is very much a two-wheeled vehicle (bikes, scooters, motorcycles). But the behavior of people behind the wheel there is not much different from Russia. In Nevada, the average drivers much more disciplined, but at night the car harder due to the cameras flash.
In January of this year, we showed their car at CES. To go to the exhibition decided later — at the end of October 2018, i.e. for only two months. Initially, didn’t even plan to travel there without the man behind the wheel. We just arrived and realized that the resolution to superpose: you need to fill in a web form. After two weeks in the mailbox lay a special room.
During this time the engineers bought an ordinary Prius, the necessary hardware, all equipment for orientation in space. The car is actually assembled by one person. Yes it took a bit to finalize. In America traffic rules are different from international, and vary from state to state: for example, have their own signs and layout features. The second issue tossed Las Vegas — in the metropolis on the backdrop of neon signs and colorful showcases hard to see traffic lights.
Began to travel and realized over the three day test drive that the electronics itself copes with everything. Then we decided that we will be in Las Vegas to show the technology still without a driver behind the wheel, and, say, put him on the right. I called the DMV Department that performs the function of regulation. They say: well, you have people in the car is — put it where you want. So during CES, we rode without a human driver — local laws allow. In Moscow yet so impossible.
The human factor
Train the robot to go to the landfill — is simple enough. Special zones there are in the suburbs, and in Innopolis, where drones clocked millions of kilometers, driving on empty tracks round the clock. In their digital brains and then modeled the various conditions seen in the real world: the machines understand how to respond to abnormal situations. The same instructors teach the living test, controlling the system behavior on the roads.
We have created a polygon near Moscow. There are trained drivers. We have a special mode, it’s called Crazy Mode when we paratextual all of our testers approximately every two weeks. The point is that we are producing cars on the road, and periodically it starts to kink. The task of man is to maintain control and time to intervene in the management. Just so that people do not relax. Many of our drivers noted that when you sit in your own car, you don’t understand why she’s not going herself. That’s how relaxing the brain and atrophies driving skill when using the autopilot. They used to go there, nothing happens. It is necessary to periodically arrange a fair shake, because the cases are different. And is responsible for the safety of the people. Yet.
A live tutor is needed to teach digital brains to properly handle the difficult situation in the city. The lion’s share of the team’s effort is spent on programming the machine to adequately predict human behavior and to act on it. Whether it be other drivers or pedestrians. Human stupidity is no simulation do not predict, therefore, difficult moments you just have to practice is to ride through the streets.
A trivial example with restless pedestrians: is at a marked crosswalk to a man, rumpled, kind of like going to the other side, and the active actions are not taken. In traffic the driver is obliged to skip it. In practice, the man behind the wheel assesses the scenarios are either passing or waiting. The car is required to see such a situation, the number of times that the neural network is indirectly learned to define the further algorithm of actions of other drivers, including undisciplined. The same is true of construction equipment, transport security services, and the like.
We have a number of tasks, which we call the long tail. I’ll explain what it is. There are problems that need to be addressed before the car will be safe to cope with any situations on the road. Basic support for orientation, recognition of traffic situations, following the route of clear, gross concept. After that taxi ride to ride. At this point issues go into the nuances, in the long tail.
The difficult conditions have not put the system to a standstill. Electronic intelligence is largely reacts almost like a living, though it is not specifically taught. For example, there is heavy rain or snow impairs visibility. People in such circumstances reduces the speed to minimize the risks and to have more time to make decisions. So also does the autopilot. Reduced predictable distance? It is necessary to reduce speed, to slow down in case of an emergency situation. Safety is paramount.
When we solve the problem of security, we begin to work on optimization. How to maneuver if moving convoy of equipment? How to maintain a safe distance to the vehicle ahead? How to skip an ambulance? That is, the points that are important but not enough to be able to switch off the machine in time. This includes communication between the autopilots in the stream. Learn to drive yourself — make a module that allows the car to agree in advance on the maneuvers to be predictable for each other.
The results and experience
Many of us daily travel on streets, whether it is a personal car, cars of car-sharing, taxis or public transport. And among all these drivers in city traffic can be a robot. Today it is a rarity. But, as shown by our little test drive, the UAV kept on the road much more predictable and safer than 90% of the «bomb» on the yellow cars with bombs on the roof.
Yes, Yandex is over what to work, however, it works great economies of scale. The more robomachine on the roads, the better they understand us and our world. Learn smoothly to get under way, maneuver in heavy traffic and gently slow down. Today one thing is clear. If you are a taxi driver or your profession is related to driving, it’s time to master something new, it is too late. It is now the autopilot — something expensive, futuristic and strange.
By 2025, it may leave you without a job.