On August 25th, according to The Atlantic's report, a team within Google's parent company Alphabet has been secretly working on a crucial piece of software for autonomous vehicles. This R&D team, known as Carcraft, is reportedly inspired by the popular video game World of Warcraft. No journalist has yet laid eyes on it.
James Stout, the young engineer developing this software, sat with me in an open-plan office. On the screen, we observed a circular junction depicted in simple line drawings against a road texture background. The driverless Chrysler Pacifica was visible in medium resolution, alongside another car represented by a basic wireframe.
A few months ago, the driverless car team encountered a similar roundabout in Texas. The complexity and high speed confused the autonomous car, prompting the team to recreate similar physical channels on their test equipment. What I witnessed was the third step in this learning process—digitizing real-world driving scenarios. Actions like cars navigating roundabouts can be magnified thousands of times in simulations to explore the limits of the car's capabilities.
This simulation serves as the foundation for Waymo's powerful testing framework. Stout explained, "Most of the work we see is driven by simulations." Carcraft is a vital tool for accelerating the development of driverless cars. Since Waymo's separation from Alphabet's X lab in December 2016, Carcraft has played a significant role in reshaping the real world into a virtual one.
Initially developed as a "scene replay" tool for road driving, Carcraft has grown into a pivotal component of Waymo's autonomous planning. At any given moment, 25,000 virtual driverless cars are operating in cities like Austin, Mountain View, and Phoenix, reconstructed via Carcraft.
In just one day, Waymo simulates thousands of drives in areas with complex road conditions. Daily, their cars travel over 12.87 million kilometers in the virtual world. By 2016, their cumulative virtual distance reached 4 billion kilometers, compared to 4.83 million kilometers on actual roads. Importantly, these virtual miles are concentrated in "interesting places," providing valuable learning experiences.
These simulations are part of an intricate system developed by Waymo. While their autonomous cars have logged millions of kilometers on public roads, they also undergo "structured tests" at a secretive base in the Central Valley known as "Castle." This undisclosed system complements their road mileage, identifying areas requiring additional practice. They recreate these places in Castle to experience thousands of different scenarios firsthand.
To reach Castle, you drive east from San Francisco Bay, turn south onto Route 99, and head south along the Central Valley Highway to Atwater, near Fresno. It's 30 degrees Celsius hotter than San Francisco. Previously home to 6,000 employees for the B-52 project, it now sits on the northern edge of the Merced metropolitan area, where unemployment exceeded 20% in the early 2010s and remains above 10%. Here, 40% of residents speak Spanish.
Leaving Atwater, we crossed some railways and arrived at the old base, now housing the Merced Animal Control Center and Atwater Prison. My phone provided GPS coordinates instead of a specific address. Walking along tall, opaque green fences, Google Maps directed us to stop. There seemed to be nothing there, but my Waymo guide was confident. Sure enough, a security guard appeared, checked our documents, and let us pass.
Entering the bustling campus, many young people in shorts and hats roamed around. There was a mobile building, a dome garage, and parking spaces for driverless cars—the main goal of our visit. Several types of driverless cars were present, including the Lexus models commonly seen on roads, retired Priuses, and new Chrysler Pacificas.
Driverless cars are easily recognizable due to their numerous sensors. The most prominent are the laser scanners (commonly referred to as LIDARs) atop the cars. Near the side mirrors of the Chrysler Pacifica, there are smaller rotating LIDARs. Behind them are radars resembling Shrek's unsettling white ears.
When the car's sensors are activated, the rotating LIDARs emit strange noises even when parked, a sound somewhere between mourning and thumping. It was so novel that my ears couldn't automatically filter it out. Across from the main building, a more unusual car was parked. Wrapped in different sizes of red tape bearing the X logo, it marked the Class IV car.
The classification of autonomous vehicles was developed by the Society of Automotive Engineers. Most cars on the road fall into Classes One or Two, capable of intelligent cruise control on freeways. However, the red X car is entirely different. It is fully automated and cannot be driven manually, hence Waymo's decision to separate it from other vehicles.
Driving into the parking lot, we couldn't help but feel like we were stepping into a "Manhattan Project" startup. In a classroom-sized room in the main building, I met Steph Villegas, the driving force behind this magical place. Villegos wore a long, very fitted white T-shirt, sturdy jeans, and gray knitted sneakers—still stylish as when she joined Google’s driverless car project in 2011.
I asked, "Are you a driver?" Villegos responded, "I will always be a driver." She spent countless hours commuting between Highways 101 and 280, leading to San Francisco and the mountains. Her keen sense of what the car might encounter was invaluable in the driverless plan. Villegos told me, "After testing newer software and staying in the team for a long time, I started thinking about ways to challenge the existing system."
To this end, Villegos and some engineers began preparing to test new behaviors in a controlled manner. They occupied the Shoreline Amphitheater car park, ensuring only authorized Googlers could enter. Villegos said, "This is where we started. I and several drivers tried each week. We would come up with a set of things we wanted to test, load the truck with supplies, and drive to the parking lot for testing."
These became the first structured tests for the driverless project. Proving that the most challenging part wasn’t the "zombie doomsday" scenarios people imagined, but endless changes in normal traffic. Villegos began collecting props from anywhere she could find—dummies, cones, plants, children’s toys, skateboards, tricycles, dolls, balls, and other gadgets—and storing them in a prop box first in tents, then in the castle. Now there is a full storage room.
Walking from the main trailer office to her car, as we were about to leave, she handed me a map and said, "Like Disneyland, you can walk along the map." The map was well-drawn. In one corner, a Vegas-style sign read, "Welcome to the mythical castle in California." Different parts of the park even have their own naming conventions. The roads we passed were named after famous cars.
Passing through several pink buildings, former military quarters, one of which had been transformed. When Waymo employees couldn’t return to the Bay Area, they could rest here. Additionally, there were no other buildings in the test area. It was indeed a city of robot cars.
As an outsider, it felt like a video game scene without players. From boulevards to neighborhood streets, from cement driveways to suburban intersections, minus the buildings we associate with these places, it all seemed incredible. I continued to glimpse the road I had walked through and ended up on the huge two-lane circular road. In the middle, there was a white fence. Villegos said, "We had a multi-lane roundabout in Austin, Texas, and we installed this kind here. Initially, there was only a single-lane roundabout. At the intersection, it looks like a horse of different colors, thanks to Texas for inspiration."
Stopping at the new facility—a two-lane road and a bicycle lane built along a parallel parking lot adjacent to the turf—Villegos said, "I really like to build new facilities along parallel parking lots. Similar scenes have already appeared in the suburbs of commercial areas, such as Walnut Creek, Mountain View, and Palo Alto. People from stores or parks come out and walk between cars, and you can also carry things across the road." This path was very much like the debris in Villegos' memory, especially the memory embedded in asphalt and concrete, which would become more abstract forms helping robot cars improve their ability to adapt to the terrain of the home.
Villegos drove me back to the main office. We hopped into a driverless car modified by Chrysler Pacificas. Our "Left Seat" driver was Brandon Cain, who tracked the car's performance through the XView software on the laptop. Other test assistants, called "foxes" and derived from the word "man-made," drove cars, created traffic, acted like pedestrians, rode bicycles, and carried parking signs. They could be more or less called actors, and their audience was cars.
The first test we had to do was "simple overtaking and merging," but it only needed to be completed at high speeds, at about 72 kilometers per hour. We went straight on the wide road called Autobahn. When the "fox" stopped us, the Waymo car braked quickly and steadily, leaving a deep impression on me. Subsequently, Cain et al. examined the deceleration figures and realized that our brakes were not doing well enough. We must do it again and repeat it again and again. Another car used different methods to block us from different angles. They call this kind of test "coverage."
We experienced three other tests: high-speed parallel roads, encountering cars retreating in the driveway, and the third one dominated by the sight of the driverless car. When pedestrians threw basketballs on the road, the car smoothly slowed down to stop. Each test was impressive in its own way, but the blocking test was the most shocking to me.
When we decided to continue experiencing the driverlessness, Cain changed seats. He asked me, "Have you ever seen Pacific Rim?" It was Guillermo del Toro's movie. People inside were fighting monsters by synchronizing with giant robots. He said, "I try to keep pace with the car and we share some ideas."
I hoped Cain could explain what it meant to "synchronize with the car." He said, "I'm trying to adjust the difference in people's weight in the car. I often stay in the car and can feel what the car is doing. It sounds weird, but I can really feel it on the hip and I know what it wants to do."
Staying away from the dusty castle, we came to Mountain View’s comfortable Google headquarters. I visited Waymo's engineers. From a technical standpoint, they belong to the X department, Google’s long-term, high-risk research department. In 2015, Google was removed from its name when Google reorganized itself into the holding company Alphabet. In the year following the reorganization, X and Alphabet decided to split the driverless car project and establish an independent company, just as Google’s other previous projects also became independent businesses.
Waymo is like a Google kid, and its office is still inside the mothership. Although this sector consists of two small groups, they are slowly becoming integrated. The X/Waymo building is large and well-ventilated, with Project Wing’s drone prototype hanging. Here, I caught the clues of the company’s development of the Firefly car. Everyone seems to have Carcraft and XView on the screen, and black background polygons abound. These people created the virtual world that Waymo cars passed through.
The person waiting for me is Carst creator James Stout. He never talked about his project publicly, but his enthusiasm never faded. Carcraft is like his child. He said, "I was browsing recruitment information at the time and saw that the driverless car team was recruiting people. I couldn't believe that they were recruiting." Stout successfully entered the team and immediately began to develop this tool, now supporting driverless cars driving in the virtual world every day.
At that time, they mainly used this tool to observe what the car would do under difficult circumstances. In similar circumstances, human drivers had already controlled the car and they began to make various assumptions. Stout said, "Obviously, this is a very useful thing, we can do a lot of things in this area." Carcraft's space is expanding, even covering the entire city, the number of virtual cars is also increasing.
Stout finds Elena Kolarov, who is the person in charge of the "Situation Maintenance" team responsible for managing the controls. There are two screens in front of her. The left screen runs XView, showing what the car "sees". The car uses cameras, radar and laser scanning to identify objects in its field of vision. These objects are represented in the software by a simple wireframe shape and are easy to recognize. The green lines extending from the shape of the object represent possible ways of moving. At the bottom, there is an image bar showing the scene captured by a regular camera on the car. Kolalov can also turn on the data returned by the laser scanner, which is displayed in orange and purple dots.
We saw the replay of the true parallel scene at the roundabout of the castle. Kolalov converted to an analog version, they look almost no difference, but it is no longer a data log, but a new situation that the car must solve. The only difference is that at the top of the XView screen, it shows "simulation" with red letters. Stout said that they must give a supplementary explanation because it is difficult for people to distinguish the difference between simulation and reality.
They loaded another scene, this time in Phoenix. Kolalov enlarged the simulation to show that it is a city. Stout said that there are all kinds of lanes inside, there are connections between lanes, parking signs, traffic lights, restricted areas, lane center locations, and everything you need to know.
We magnified a 4-car intersection somewhere in the vicinity of Phoenix. Then Kolalov started putting on synthetic cars, pedestrians and bicycles.
Clicking on the hotkey, the object on the screen will begin to move. The car moves like a car, driving in the lane, turning. Cyclists are like real cyclists. After the driverless car team manipulated the driverless cars and traveled millions of kilometers in the real world, their logic had formed a fixed pattern. Behind all this, there is an ultra-detailed map of the world, as well as different physical models at the site, rubber and roads are modeled.
Not surprisingly, the hardest thing to simulate is actually the behavior of others. Waymetric's software director Dmitri Dolgov told me, "Our car sees the world and we understand the world. Then, for any object that plays a dynamic role in the environment, whether it is a car or not pedestrians, cyclists, or motorcycles, our cars can understand their intentions. It's not enough to track an object through a space, but you must understand what it is doing. This is to build a competent, safe The key to driving a car. This kind of modeling, this understanding of the behavior of other participants in the world, is very similar to the simulation of them in this task."
There is a key difference: in the real world, they must accept fresh, real-time data about the environment, translate it into understanding the scene, and then navigate. But now, after years of research, Stout and others believe that they can do this because they have run "a series of tests" and proved that they can identify a variety of pedestrians.
Therefore, in most simulations, they skipped the object recognition step, did not provide the car with the original data to identify pedestrians, but directly told the car that these were pedestrians. At the intersection, Kolalov was even more difficult for driverless cars. She clicks V, which is a hot key for the vehicle. A new object appears in Carcraft. Then she moved her mouse to the drop-down menu on the right hand side, where there are many different models, including my favorite bird_squirrel.
Different objects are told to follow the logic that Waymo has modeled them and need to move in a precise way in the Carcraft-built scene to test specific behaviors. Stout said, "There is a good spectrum, you can control a scene, then put objects in and let them go." Once they have the basic structure of the scene, they can test all the important variables it contains. So, imagine that for the intersection, you might want to test the arrival time of various vehicles, pedestrians and cyclists, the time they stopped, the speed they moved, and everything else.
They call this model "fuzzing." In this case, this intersection can produce 800 possible situations. It creates beautiful lace charts, and engineers can go in and see how the combination of different variables has changed the car's decision to take the path.
This problem has really turned into an analysis of all these scenarios and simulations to find interesting data that can guide engineers to better drive. The first step may be: Will the car be hit? If it is, it will be an interesting scene. The following figure shows this situation. In the real life of Mountain View, the intersection scene is more complicated. When the car was driving to the left, a bicycle approached, causing the car to stop on the road.
The engineers solved this problem and redesigned the software to correct it. The picture shows the real situation, and then simulate the operation. This means that in two different situations, you will see the simulated car continue to drive, and then the dotted box that says "shadow_vehicle_pose" appears. This dotted box shows what happens in real life. For the Waymo team, this is the clearest visual progress.
But they are not just looking for details when the car is hit. They may also want to look beyond the correct range for long-term decisions or the effects of sluggish brake response. If engineers want to learn from them, they will look for problems through simulation. Dolgov, director of software for Stott and Waymo Software, emphasized that there are three core aspects of simulation. First, they drive far more mileage than physical teams in the real world, and get more and better experience. Second, these cars focus on miles that are interesting and still interactive, and do not pay attention to boring miles. Third, the software development cycle can be faster.
Dolgov said, "This iterative cycle is very important to us and all the work we did during the simulation allowed us to dramatically reduce the time. In the initial phase of the project, this cycle may take several weeks, but now only It will take a few minutes.†I continued to ask him, are there any oil films on the road or punctures, strange birds, tiankeng and other crazy things that they can all simulate? Dolgov is optimistic about this, saying It's already possible to simulate, but the key is how high the fidelity of these simulations is? Maybe there are issues where you get better value, or you need to run a series of tests to confirm the scene in the physical world in the simulation.
The power of Carcraft's virtual world is not that they are real-world beautiful, perfect, realistic renderings, but that they reflect the real world in many ways. This is very important for driverless cars. It can Gain billions of kilometers more than physical tests. For driving software running in simulation, although it is determined differently from the real world, it is the same decision made in the real world.
This method is effective. The California Motor Vehicle Authority asked the company to report their annual unmanned mileage and automated tests that were carried out by the driver. Waymo not only has three orders of magnitude more mileage than anyone else, but the number of human driver interventions is also rapidly declining. From December 2015 to November 2016, the Waymo car drove 1.02 million kilometers in unmanned mode. In all mileage, the driver only participated in 124 times, and once every 8000 kilometers. In the previous year, these cars had traveled 680,000 kilometers automatically, but human drivers intervened 272 times, an average of 1,432 kilometers.
Although everyone has painstakingly noticed that these are not the same digital comparisons between Apple and Apple, let's take a look at the reality: these are the best comparisons we have, at least in California. In total, other vehicles traveled approximately 32,000 kilometers in automatic driving mode. Waymo's strategy is not surprising for outside experts. Chris Dixon, head of investment at venture capital firm Andreessen Horowitz, said: “Now, you can measure the maturity of an unmanned team almost by treating the seriousness of the simulation. Waymo is at the top of the list. It is also the most mature."
I asked Sunil Chintakindi, innovation director of Allstate Insurance Insurance, about the Waymo project. He said: "If there is no strong simulation infrastructure, we cannot build (higher-level) automated vehicles." Researchers who drive cars are also looking for similar paths. Huei Peng, director of Mcity, an automated and networked automotive laboratory at the University of Michigan, said that any system used for driverless cars will be "more than 99% simulation + well-designed structured testing + highway testing."
Peng Hui and a graduate student proposed a system that can speed up testing through simulation, which is not different from Waymo's implementation. Peng Hui said: "So what we're debating is just to cut off the boring part of driving and focus on the interesting part. This can make you speed up hundreds of times, and 1,000 kilometers into millions of kilometers."
What is surprising is the size, organization and intensity of the Waymo project. I described to Peng Hui the structured tests conducted by Google, including the 20,000 simulation scenarios designed by the Castle Structured Test Team. But he got it wrong and said: "The 2,000 scenes are impressive." When I interjected and corrected him, he replied, "Indeed, 20,000 scenes. It's really impressive." 20,000 scenes represent only a small part of all scenes tested by Waymo. They are all created by structured tests. There are more scenes from public driving and imagination. Peng Hui said: "They have done really good. They are far ahead of everyone else in the four-class field."
However, Peng Hui also emphasized the position of traditional car manufacturers. He said that they are trying to do something completely different, rather than aiming for a fully automated “moon shot (crazy and hard to achieve project),†they are trying to increase driver assistive technology, “make money,†and then gradually move toward full automation. Comparing with Waymo is unfair. It is rich in resources, installing a $70,000 laser rangefinder in a car, and car manufacturers like Chevrolet may see that $40,000 may already be mass market. The price cap used.
Peng Hui said: "GM, Ford, Toyota and other auto companies are saying: 'Let me reduce the number of car accidents and fatalities and increase the safety of the mass market.' They are completely different from Waymo's goal and we need to consider millions of vehicles. Cars, not just a few thousand." Even in a fully automated race, Waymo now has more challengers than ever, especially Tesla.
18 months ago, Chris Gerdes, director of the Stanford University Automotive Research Center, once said that Waymo has a deeper understanding of the depth of the problem. Compared with other people, we are closer to the solution to the problem. When I asked him last week if he still thinks so, he said: "A lot of things have changed. Automakers such as Ford and General Motors have been developing their own cars and have built road data sets. Tesla has now passed Deploying Autopilot collects large amounts of data to understand how the system operates under its customer experience. They can test algorithms in silent mode and quickly expand the number of test vehicles, making them an astonishing testbed.
In the simulation field, Geddes said: "I have seen many competitors with substantial projects. I am sure there is quite a lot of simulation capabilities, but I have seen many things that seem trustworthy. In this regard, Waymo doesn't look so unique anymore. Of course they started very early, but now many groups are using similar methods. So the most important question now is who can do the best."
This is not a low-risk demonstration of the "brain type" capabilities of neural networks, but a giant leap forward in the field of artificial intelligence, even for intra-company companies that have always been very active in adopting artificial intelligence. This is not Google Photos. If you make mistake here, the consequences are not too serious. And for an active and fully interactive system in the human world, it will understand our rules, convey its cravings, and let our eyes see them.
Waymo seems to want to use driving as a skill, speed and direction are some of them. They are now integrating driving into human social activities. Usually, how to drive is considered “legal� How does humanity teach artificial intelligence what it stands for? It turns out that the development of such artificial intelligence does not only require endless data and engineering techniques. These are necessary but not enough. On the contrary, the development of this artificial intelligence requires humans to synchronize with the car and understand the world. Although anyone can do it, the drivers in the castle know that as a car, they must observe and make decisions like humans. Perhaps there are also two ways: The deeper the human understanding of the car, the deeper the car's understanding of mankind.
The memory of Austin's roundabout became part of the castle, becoming a data log of driverless cars, turning into a scene in Carcraft, becoming a simulation network, and eventually returning to the Texas driverless car. New software. Even in the simulation, the AI is used to understand the polygon abstract image of the world, there are traces of the human dream, fragments of memories, the feeling of the driver. These elements are not mistakes, nor are they tainted by humans. They are the necessary components of systems that can radically change traffic, cities, and everything else.
AL/CU Universal Terminal Block
Cable Terminal Block,Universal Terminal Block,Waterproof Terminal Block,Low Voltage Terminal Block
ZHEJIANG QIANNA ELECTRIC CO.,LTD , https://www.traner-elec.com