Traditional Culture Encyclopedia - Weather forecast - Zero intervention driving? Destroy the ghost detector? This is the closest we've come to unmanned driving?

Zero intervention driving? Destroy the ghost detector? This is the closest we've come to unmanned driving?

Tesla is more radical than anyone else on the road of autonomous driving.

10 year 10 20th, Tesla pushed FSD (full? The latest version of self-driving) FSD? Beta, according to Musk, this update will realize "zero intervention" and "full perception" autopilot function for Tesla.

As early as June this year, Musk publicly announced that Tesla's FSD will usher in a qualitative change, FSD? Beta is an infrastructure rewrite, not a gradual code adjustment. The rewritten algorithm can extend the perception ability of FSD from 2D to 4D, and add the functions of depth prediction (bird's eye view mapping network) and time prediction (RNN processing occlusion), which not only means that vehicles can perceive the real world more clearly, but also make more accurate predictions of position, direction and speed.

Prior to this version, automatic driving assistance driving based on FSD must meet some basic conditions before it can be enabled, including clearly identifying lane lines, identifying vehicles in front, and exceeding 30 miles per hour. However, in FSD? In the Beta version, the enabling conditions are greatly reduced, and the function can be activated simply by setting the destination in the navigation.

Judging from the videos shared by some early bird users, FSD? Beta can realize the identification of complex road conditions around the island, autonomous driving independent of lane lines, autonomous overtaking (or bypassing distractions), traffic light identification at intersections, timely left turn, and the ability to cope with various weather conditions, including night and rainy days.

You know, just "driving independently of the lane line" is enough to beat 90% competitors in the market.

Qualitative FSD? beta

Next, let's take a look at FSD? What are the highlights of Beta?

Because this version is currently only pushed to some users (professional users or media) in the United States, most of the materials and materials in the article are shared by American early bird users.

The driving test video from the network can be found, FSD? Beta redesigned the UI interface, which can be displayed differently according to the lighting environment during the day or at night. On the left side of the screen, 1/3 is the vehicle status area, and on the right side is the navigation map. The vehicle model is also upgraded from the lower part of the area to the middle part of the area.

The left area displays road markings, traffic signs, intersection structures and cars on both sides of the driveway, and is marked with lines of different colors, which can instantly display information according to the driving state of vehicles, as if demonstrating "the world in the eyes of the fire department". At the same time, the visual model is improved from 2D to 4D, and the viewing angle is changed from the first viewing angle inside the car to the third viewing angle behind the car.

Generally speaking, the information displayed is richer and more detailed than before, which seems to have a stronger sense of science and technology. But compared with the previous version, this time FSD? The UI design of the Beta version is a bit rough, and it looks like an engineer-oriented version, which is quite hard-core. Moreover, too much information display may attract the attention of drivers, so it is estimated that it will be adjusted in the later official version.

FSD now? What function can Beta achieve?

Let's continue to look at:

(1) Identify traffic lights

The camera in front of the vehicle can accurately identify the traffic lights at the intersection and decide the driving of the vehicle accordingly. This feature was previously available in some parts of the United States and is now fully open:

At the same time, the identification of traffic signals at night is also very accurate:

(2) Identify stop signs

In American traffic regulations, a stop sign means that the vehicle needs to stop and observe for about 3 seconds before moving on. Usually at intersections, FSD? Beta can also well identify and cooperate with the operation:

(3) Judgment of crossroads

If you want to turn right at the intersection, FSD? Beta can judge whether there is a car coming to the left or in the opposite direction, and then move on:

(4) Automatic traffic around the island

The most surprising thing is, FSD? Beta can now take the initiative to pass around the island, and in the process, judge whether there is a car coming in the nearby lane, and decide the lane to drive out according to navigation:

Hold it completely even when driving at night:

(5) Call for more wisdom.

FSD? Beta enhances the application scenario of the summoning function, and can judge the driving trajectory according to the actual road conditions of the parking lot, so as to realize complete unmanned driving in the area:

(6) actively avoid non-motor vehicles

In a narrow road section, if it is recognized that there are non-motor vehicles driving on the roadside, the fire department? Beta will look slightly to avoid to the other side:

Even can accurately touch the ground obstacles and avoid:

(7) Ghost Probe Early Warning

"Ghost probe" refers to pedestrians or obstacles that suddenly rush out in the blind area of vision, which often exceeds the response of human drivers and is too late to avoid. What about FSD? Beta now has occlusion prediction, which can detect and brake in time:

In addition to the above, FSD? There are many new functions of Beta, so I won't list them here. Friends who feel dissatisfied can search the real car demonstration video.

How to realize these functions?

Brand-new fire department? Beta did achieve a technological leap, as Musk said. So, how are these functions realized?

Four words can be simply summarized: hardware+algorithm.

Tesla's FSD hardware is factory installed and there is no subsequent upgrade. We use the latest? Models? Take three models as examples:

Models? There are 8 cameras, of which 3 cameras are responsible for the front view, and the other 5 cameras are responsible for monitoring the side and rear of the vehicle.

In the latest FSD? In Beta, the images collected by these eight cameras will form an image, instead of each camera working independently and analyzing independently. This is a great progress.

Here, why a camera instead of a lidar solution? Because Musk has repeatedly stressed that Tesla's goal is to realize an automatic driving system that can be used in a wider range and road scenes, unlike Waymo, although the sensor is armed to the teeth, it can only drive in the designated area. The low cost of the camera can support this scheme.

Back to the sensor, model? 3 1 forward-looking radar with enhanced processing capacity is also integrated to provide additional environmental data for vehicles, and at the same time, it plays a role of safety redundancy in rain, fog, dust and other weather. There is also 12 ultrasonic sensor. Although the working radius is short, they can work stably at any speed, mainly for the blind area of vehicle control.

Besides sensor support, hardware? 3.0 also has a heart-FSD chip. It uses two self-developed SoCs, two GPUs, two neural network processors and a lock-step CPU. In order to improve the access speed and computing power of the neural network processor, 32MB cache is also integrated in each FSD chip. According to the official data released by Tesla, Autopilot? HW? 3.0 can process 2300 frames of images with 8 cameras working at the same time every second, and the final power reaches 144TOPS.

Under the condition that the hardware foundation remains unchanged, the key to realize the iterative updating of functions lies in Tesla's algorithm. But the algorithm is FSD? The core technology of Beta is also Tesla's secret, which cannot be developed here. We can only capture some key points from the information revealed by Musk.

At last year's autonomy meeting? On the same day, Musk revealed that Tesla had a supercomputer project codenamed "『Dojo』". According to Musk's description, Dojo is another killer after Tesla's FSD chip. It can be implemented with FP32 1. Exaflops calculation. FP32 is a floating-point number, which is more accurate than FP 16 with16 bits. Exaflop refers to how many floating-point operations a computer can handle per second. 1? Exaflop means one billion times per second. By comparison, the strongest supercomputer in the world is 0.4 15? The speed of Exaflop

This supercomputer can input a large number of video data uploaded from vehicles, create a huge database, thus establishing a driving model, and perform unsupervised algorithm exercises through the training server. These data include lane signs, traffic conditions, obstacles, traffic signs and so on. These videos will be screened, cleaned and marked by the system, and hundreds of skilled marking engineers will process them at the same time to achieve highly accurate machine learning.

Surely these are fsds? Beta can fundamentally rewrite the key to the stack.

Because of such technical support, Musk boasted at the 2020 World Artificial Intelligence Conference. He said that Tesla is very close to L5-level automatic driving, and will finally realize L5-level automatic driving through existing hardware and constantly improving software.

Some technical questions

However, although FSD? Beta's actual measurement function is amazing, or many geek fans have found the problem, and the most controversial one is FSD? Does Beta use high-precision maps?

Tesla said that Autopilot realizes automatic (assisted) driving based on pure vision algorithm, which is completely different from the laser radar route commonly used in the industry. Musk also has absolute confidence in his visual technology, insisting on 100% visual recognition, and even boasting that "a high-precision map based on GPS is a bad idea, which will make the whole system fragile".

However, a blogger questioned on social media that Tesla may have adopted a "high-precision map". An example is shown in the figure below. In fact, the visibility of the leftmost intersection on the road is very low, but FSD? Beta clearly shows the picture of the crossroads:

So some people speculate that FSD? Beta is pre-installed with road information, which is not a real visual recognition driving.

But there is another saying, FSD? Beta adopts SLAM drawing method. SLAM, which means at the same time? Localization? And then what? Map drawing, mainly used to solve the problem of artificial intelligence positioning and map construction in unknown environment, was first proposed by NASA in 1989.

The basis of this speculation is FSD? Beta is collecting huge user feedback data, and another blogger on Twitter said, "After receiving FSD? Beta updates his model within two days? 3 uploaded 2 1.09GB data ",which may contain massive road information.

The famous foreign Tesla hacker (@Green) also analyzed this, and thought that Tesla used the data uploaded by vehicles to draw 3D maps of the city, providing reference for other users' navigation. Map data is pre-loaded, even if it is not a high-precision map, it is enough for the system to predict in advance.

At present, Tesla officials have not responded to these questions and will follow up this issue in the future.

How much is Tesla ahead?

In recent years, the technology of neural network and visual learning has developed rapidly, and Tesla has contributed greatly. From above, FSD? According to the detailed explanation of the Beta function, Tesla is not only radical in the field of autonomous driving, but also technically hard enough. If there is no accident, Tesla will continue to expand its advantages.

Why?

At present, there are three main types of forces in the field of autonomous driving, one is the Internet technology giant, the other is the new energy vehicle enterprise, and the other is the traditional vehicle enterprise that emphasizes intelligent transformation.

In the final analysis, the automatic driving system is realized by artificial intelligence, which needs a lot of data as "nourishment" to evolve. Because of this, for autonomous driving technology, it is necessary not only to master the software and hardware integration of autonomous driving, but also to conduct deep learning with a large amount of artificial intelligence data.

From this point of view, Tesla has an absolute advantage. Although Google's Waymo has algorithms, it lacks real user data; The general Cruise has the same problem, and the data volume is much smaller than Tesla; Not to mention the traditional car companies, no algorithm can compare with the former two. On Tesla's side, is it just a model This car can sell nearly 300,000 vehicles in 65,438+0 years, and most of them are equipped with shadow mode.

The so-called shadow mode means that when automatic driving is turned on, the system will detect the data around the vehicle driving road in the background, learn the driving operation of human drivers, and finally send the data back to the server for modeling operation, thus optimizing the driving algorithm and OTA it to the vehicle. Since 20 15 Tesla began to carry Autopilot in cars, shadow mode has traveled more than 3 billion miles in the background. This scale is beyond the reach of other autopilot systems tested on closed roads.

How far are we from unmanned driving?

Analysis here, there should be many people think that we are about to realize unmanned driving? Sorry, there's still a pot of cold water to pour. Even though we have been making progress in theory, I am afraid there is still a long way to go before we are completely unmanned.

First of all, from a technical point of view, the corner case of technology (corner? Case) cannot be effectively solved. In the automatic driving system, the vehicle collects data through radar or camera and uploads it to the machine for learning. However, in actual driving, it is inevitable that there will be some road conditions beyond the experience of the machine. These are corner cases (such as the Taiwan Province model? Three examples of collision with overturned trucks).

According to the research report of the Electric Vehicle 100 Committee, today's driverless technology can handle 90% of normal road conditions, but the remaining 65,438+00% corners have a huge impact and need to be solved within 90% of the time.

Tesla also understood this, so he gave it to FSD? Some tips for public beta users: First of all, it is only pushed to some professional users, indicating that there are still uncertain factors whether this system can be fully opened; Secondly, Tesla also mentioned FSD in the update instructions? Beta is only a beta version, so you need to pay special attention when using it. You may make the wrong choice at the most dangerous time, so drivers must put their hands on the steering wheel and constantly observe the traffic conditions on the road.

In addition to the technical level, other issues include but are not limited to:

Attribution of legal responsibility. The subject of responsibility is a crucial concept in any law. However, autonomous driving technology blurs the division of this concept. If an accident happens to a self-driving vehicle, is the driver responsible? Is it a technology supplier? Or the brand of the vehicle? These are all unsolved problems (such as the example of Uber causing pedestrian death in the United States).

Right of way and formulation of road rules. Do unmanned vehicles enjoy the same right of way and accept unified management as manual vehicles? Do they drive in the same lane and apply the same traffic rules?

Changes in product attributes. Automatic driving will greatly improve the utilization rate of vehicles, thus reducing the stock of vehicles on the whole road, because at this time, people no longer need the ownership of vehicles, but only the right to use them. Can consumers accept the change of product nature?

Such as technical ethics. The famous tram problem will reappear. Suppose a self-driving vehicle can avoid and protect pedestrians when facing the pedestrians who suddenly rush out of the road ahead, but it will sacrifice the safety of passengers and other vehicles on the road, otherwise it will hurt pedestrians. How should AI judge and choose?

There are still many problems that autopilot has to face, and it is impossible to list them here. Because of this, enterprises committed to autonomous driving technology should not be too hasty, and some unrealized function points will be publicized in advance, which may easily lead to public misunderstanding and even lead to major safety accidents.

China, a 20 13 automobile supplier, conducted a survey on autonomous driving. The results show that 66% of Americans think that "self-driving cars scare me" and 50% think that "this technology cannot run reliably". By 20 18, the data of the two survey results increased to 77%. The reason may be that the traffic accidents of Tesla, Uber and other companies in the autopilot test have affected the public's confidence in autopilot.

The final form of autonomous driving

Now we understand that autonomous driving means "handing over the control of vehicles from people to cars", but the author thinks that the control of vehicles should be "handing over from people to roads".

Assuming that one day in the future, all people's autonomous driving will come true, then vehicles driving on the road will definitely adopt autonomous driving solutions from different companies, and the technology will be mixed, and the judgment standard will be thousands of people. Moreover, no matter how powerful the automatic driving of bicycles is, it is impossible to predict the actions of other vehicles on the road, and there must be unexpected problems.

The real way to achieve complete unmanned driving is to integrate the whole city traffic into a traffic network. Every car on the road is connected to this big network through 5G and Internet of Things technology, and coordinated by the artificial intelligence center. Every car knows its relative position in the system and the route it will travel, so as to achieve the most efficient transportation efficiency and maximize the use of road capacity. At that time, people only need to choose the starting point and the end point, and the system will automatically assign you a vehicle and drive to the destination.

Of course, this may be just an unrealistic fantasy, but the technology giant represented by Tesla may really bring us closer to science fiction movies.

(photo/text/photo: all-electric? Tang Ke)

This article comes from car home, the author of the car manufacturer, and does not represent car home's position.