Google has always said that it uses a combination of real road tests and virtual tests to train autopilot systems. In the end how the two fusion? What difficulties have you encountered in it? How to solve? The following few days, after Uber, Waymo's driverless test vehicle also crashed. Fortunately, it was only slightly injured. On May 4, in Chandler, Arizona, a manned vehicle slammed across the middle line to the Waymo test vehicle on the opposite side of the road, resulting in both vehicle damage and minor injury to the driverless safety officer. The police believe that Waymo vehicles and security officers are not responsible. It seems that the city’s police have adapted to the status quo that the driverless vehicle is one of the responsible parties. Despite its non-responsibility, it also made the public realize that in some cases, the response of unmanned vehicles to the unpredictable behavior of other vehicles is no better than that of humans. In the near-collision process, Waymo did not take amazing evasive action (perhaps keeping lanes sensible), but it was somewhat disappointing to be unresponsive. Driverless test vehicles did not even adopt routine measures such as braking. Waymo will thoroughly investigate the unmanned training log internally, but they will not be stupid enough to try to understand what the vehicle is thinking about. They will only push back from the results and find that the training system is not yet complete or even not yet involved. How to train a "black box" This led to Google's "Castle" plan exposure. Google has always claimed that it uses a combination of actual road tests and virtual tests to train autopilot systems. How the two merge, Google has been secretive. The new accident shows that Google still leads but is no longer unique. The unmanned system is manufactured just like a newborn baby with a "sensory" (camera, millimeter-wave radar, lidar) for sensing the outside world, and a high-speed brain (calculating unit, image processing unit). The function of the brain is still in the process of differentiation. It needs to be taught to identify all possible people and objects in the environment. It is also possible to teach it some basic countermeasures, but in practice how it makes decisions is not known to researchers. For humans, the AI ​​decision-making mechanism is a "black box." This is why many people are worried about this. Teaching unmanned systems to distinguish roads and everything else that may be encountered is the first step in training. It's like teaching babies to look at things. Because the image information is too rich to be modeled, deep learning seems to have a special advantage. Based on millions of years of human evolution, humans can often find solutions in complex situations with intuition. AI is learning this, just following another set of rules. Deep learning can be used for both perception and decision-making. For example, AlphaGo's travel network is a DNN training system. In the simplest terms, it is to make decisions based on the current state. Its designers and trainers do not teach it to make decisions (in fact, humans do not know how the system will make decisions), but teach it some basic knowledge. At this stage, identifying the environment is the core task. Objects, where to drive (no road shoulders and flower beds), legal driving routes, etc. First, basic features are extracted from the images of a large number of cars, such as the approximate geometry of the front and side of the car, allowing the system to distinguish between the left and right sides of the car (marked with different features). With continuous multi-frame images, the car's direction of travel can be distinguished based on continuous changes. And it can identify vehicles that are very small in the distance and look farther and clearer than any human being. How to evaluate their impact on themselves will be put behind training. Second, in the traditional image, the road shoulder and the road itself are indistinguishable in color, and stereo vision is also difficult to discern (after all, the elevation difference is too small). How can humans be easily identified? Rely on shadows. A shoulder of 10 centimeters will form a continuous narrow band of dark areas. System you learned? With continuous narrow-band shadows (individual sections that are disconnected, you can just do a high-risk reminder signal), combined with the identification of the road line, to outline the area that can be driven. It looks perfect, but sometimes there is no track, or because the weather track is difficult to identify. In this case, how does humans drive? Broken trees, street drains, and pedestrians on both sides can all be the basis for judgment. The system needs to extract human strategies from a large number of videos (actually multi-frame images) and optimize them. Waymo expects his AI system to take almost the same steps as a sensible human driver, but is quicker and more decisive than any human response. However, in the car accident on May 4th, this point was not reflected. What actually caused the problem? Limitations of the training system This shows that a large number of virtual environment training, at least some of them are not used in the actual scene. In other words, the problem is fused. When the unmanned system has the identification capabilities, it needs to face two types of scenarios: one is the real world, and the other is the digital world. After obtaining the “practical†experience from the former, various conditions are changed in the latter (such as moving obstacles and making pedestrians' behavior more unpredictable), and the coping strategies are continuously refined until they are optimal. Google experts admit that it is extremely difficult to simulate human behavior that is not reliable. Even at a simple crossroads, unmanned systems are confused by pedestrians and motorcycles that do not follow the signal. After experiencing a series of brakes, numerous vehicles tried to squeeze in from sideways, leading to a more chaotic situation. In the digital training system, Waymo once again simplified the road conditions. For example, two high-speed lanes in the same direction only involve two vehicles. A car carries an unmanned system and B car will act as an obstacle. When the A car goes straight on the inner road at a speed of 90 kilometers, the right B car suddenly overtakes and goes to the front of the A car, and immediately follows the brake. Can a car brake quickly and smoothly, but at the same time leave sufficient braking time for the rear vehicle? B blocks the A car in various ways and from different angles. The test of the braking process of the A car is repeated hundreds of times. The training system records the performance of the unmanned system, analyzes the failures, and optimizes the disposition of the latter. Then the situation is set to be more complicated: the city has multiple lanes, encounters vehicles in the driveway, suddenly appears rolling basketball on the road, or suddenly picks up pedestrians from the barrier and examines how unmanned systems will respond. Of course, the program cannot exhaust all input conditions. Programmers want unmanned systems to refine methods in tens of thousands of scenarios so that they can make sensible decisions in other contexts. The car accident on May 4 may be a "superclass" situation for the current Waymo test vehicle. The problem facing humans in the face of opposing vehicles is that they do not have enough observation and decision-making time to panic. But the unmanned system is not like this. Under millisecond-level sensor data calculations, the CPU accurately knows the immediate position of all the surrounding vehicles, instantaneously overshoots the instantaneous speed and acceleration of the vehicle, and predicts the continuous position of the opponent within a few seconds thereafter. And calculate that if you do not take the emergency brake + change direction, 1.5 seconds will endanger the violent collision of people in the car. What prompted the Waymo vehicle to make an indiscreet decision? Is there a lack of maneuvering space in the right lane, or is it unable to keep the vehicle stable after predicting the direction of change, or is it a result of sudden braking that does not change the result of the collision? Instead, it will cause the vehicle to roll over due to uneven road friction, resulting in more serious consequences? We cannot know the decision-making process. Waymo engineers reading data may solve some of the confusion. If their conclusions are the same as the decisions made by unmanned vehicles at the time—nothing is more favorable, then there is no problem. The problem is that such conclusions go beyond human cognition. In the face of the crisis, we must always do something. We quickly increase adrenaline levels, dilate pupils, tighten muscles, and increase blood pressure to meet the challenges. Comparing AI decisions with humans may not be appropriate. This, in turn, prompts people to think about the effectiveness of training systems based on human experience. The virtual world may be set too simple Waymo's experts boast that they are the only companies that use an "accelerated training system." In fact, Ford, Uber, and GM have all established similar training systems in Silicon Valley. Waymo is just the earliest one to start. Of course, they also have the most data. Virtual world training may run tens of millions of miles within 24 hours. Each minute can simulate the workload of 10 weeks ago two weeks. Some experts suggest that the ratio of simulated and real road tests should be 100:1. At the same time, the simulation part should cut off the boring place, focus on the interesting part (as complex as possible scene), and achieve the goal of speeding up the training. Some people think that once the number of unmanned systems in the virtual city reaches millions, their group behavior pattern is very close to that of the real super city. Behind it, enough physical vehicles and sensors must be deployed to build a highway database. An unmanned system trained entirely on virtual scenes may behave “differently†when faced with real road conditions. This reminds people that the model of a virtual city is too simple, and it will not lead the unmanned vehicles to control the complex situation of big cities. However, Chandler is a tourist destination and actually has a population of only a few hundred thousand. It is usually sunny and has a positive effect on the normal operation of the sensor. The environmental impact seems to be ruled out. Let's get back to the beginning. Although the training started with reference to the human driver's response, in the end artificial intelligence may adopt different strategies. With the deepening understanding of human behavior by unmanned systems, it promotes its own driving style. Surprisingly, the strategy adopted by unmanned systems in the face of another unmanned system is not the same as when it is confronted by humans driving a vehicle. And we haven't thought about the fact that the city is completely full of unmanned vehicles. This means that when an unmanned system dominates the entire city, it may spontaneously form a new standard of traffic. More efficient and more understanding. The guidelines that humans accumulate in the automobile era and are considered as guidelines are likely to be replaced. Car accidents reveal the confusion that may be caused by the mixing of people and unmanned vehicles, but it also allows us to look forward to new driverless traffic. By then, the unmanned system may be easier to work with. 8 Inch Tablet
Today let`s talk about 8 inch tablet with android os or windows os. There are many 8 inch tablets on sale you can see at this store. 8 inch Android Tablet is absolutely the No. 1 choice if you are searching a student online learning project. 8 inch windows tablet is more welcome when clients are looking tablet for business application. The most welcome parameters level is 2 32Gb with 4GB lite, 4000mAh, android 11 only around 60usd, price will be much competitive if can take more than 1000pcs. 7 Inch Tablet wifi only, android tablet 10 inch, Amazon 8 inch tablet is also alternative here. Except tablet, Education Laptop, Gaming Laptop , 1650 graphics card laptop, Mini PC and All In One PC are the other important series.
Therefore, you just need to share the configuration, application scenarios, quantity, delivery time, and other special requirements, then will try our best to support you.
Any other thing in China we can do, you can also feel free to contact us.
8 Inch Tablet,8 Inch Android Tablet,Amazon 8 Inch Tablet,8 Inch Tablets On Sale,8 Inch Windows Tablet Henan Shuyi Electronics Co., Ltd. , https://www.shuyielectronics.com