What is the mystery of Tesla’s self-driving technology?

Musk once said that Full Self Driving (FSD) will be launched by the end of 2019, but now it is 2020. After Tesla upgraded the system, in addition to adding the “Vehicle Summoning” function, it has not completely Open autonomous driving permissions. In this regard, Musk said that Tesla hopes to forcibly enter the city and drive by training a neural network to recognize every object that may be encountered in the city and take corresponding actions, which may take months.

It has to be said that although it will take several months, it will be amazing if fully autonomous driving in cities can be opened within this year. Recently, a Japanese media dismantled the Tesla Model 3, and even claimed that its self-driving technology is 6 years ahead of other automakers. What is the mystery?

What is “Hardware 3”?

Speaking of Tesla’s self-driving technology, what cannot be ignored is its self-developed chip, which is “Hardware 3” released in April 2019.

In fact, at the beginning, Tesla did not research this integrated chip completely independently, but cooperated with Mobileye, but later Tesla products experienced an accident in the case of assisted driving, which obviously did not meet the expected requirements. So the two parted unhappily. Then Tesla joined hands with NVIDIA to adopt NVIDIA’s chip solution, but obviously Musk felt that this was not enough, so in 2016, he invited “chip emperor” Jim Keller to serve as Tesla’s vice president of hardware development, and officially opened the The prelude to self-developed autonomous driving chips.

It can be seen from the finished FSD computer that it contains two self-developed chips. The two chips run independently and have their own power supply, DRAM memory, and flash memory, which means that when either one fails, the FSD computer will still drive the car normally for autonomous driving.

“Hardware 3” has 6 billion transistors, can perform 144 trillion calculations per second, and can simultaneously process images at 2,300 frames per second. Compared with Tesla’s previous generation processor (HW2.5), FSD has improved 21 times in performance, and can cope with the amount of data and computing power of the perception layer required for L5-level autonomous driving. At present, even with chips like NVIDIA The company is not at all disadvantaged.

Is “Neural Network” the point?

It is not enough to have a self-developed chip to be better than other car manufacturers in autonomous driving. Another key point why Tesla can have a good reputation in the field of assisted driving is the neural network.

Neural network (NEURAL NETWOTRK) is no longer unfamiliar to users now, because many familiar mobile phone products use this machine learning method, such as: Siri on Iphone. A mathematical model of how a neural network works is like a biological neural network. The main ability is to be able to rely on a large amount of data for self-learning, and Tesla has a huge user base and data volume. So Tesla can teach the machine to recognize lane lines, vehicles, pedestrians, traffic signals, etc. through a large amount of user data.

Specifically, Tesla will first place a seed data set in the neural network, randomly collect a large amount of data and then label it. When the neural network has inaccurate recognition in subsequent recognition or the driver in the car performs certain operations to trigger the feedback mechanism , the car will send the data back to the headquarters, and Tesla will verify and repair the data. After learning from massive data, the machine becomes more intelligent, the recognition is more accurate, and it is more conducive to subsequent path planning and decision-making operations.

According to a data released by Tesla before, using the fleet in the world, Tesla has collected more than 16 billion kilometers of real driving data, of which more than 1.6 billion kilometers of driving mileage have used the Autopilot automatic driving system, a large amount of automatic driving. Driving data is also where Tesla has an edge over other automakers.

How do vision sensors work?

After talking about the “brain” in Tesla’s autopilot, there is an indispensable part in order to complete the excellent autopilot function, that is, the “eyes” of the vehicle, that is, the visual sensor, commonly known as the camera.

In Tesla’s view, people realize the perception of the outside world through two eyes. The camera on the vehicle has a similar function. There are 8 visual sensors on Tesla products, and there are also millimeter wave radars, etc. Auxiliary devices work together to complete the vehicle’s perception of the outside world during autonomous driving.

Summarize

From the above, it can be concluded that the core of Tesla’s autonomous driving technology is visual perception + self-learning. And Tesla’s closed-loop improvement of “user data-autopilot-user data” completed through neural networks can also make Tesla’s autopilot technology more and more leading.

When it comes to when Tesla’s fully autonomous driving function can be decentralised to existing products, many industry insiders are optimistic that it will be completed within 2020. On the other hand, other car manufacturers, whether they are traditional brands or new car-making forces, are either imprisoned in the traditional car manufacturing logic and cannot escape, or they lack a large amount of user data available. Although these factors do not make them technically behind Tesla by 6 years, there is still some gap. According to reports, Tesla’s next-generation chips will also be available in two or three years. If they don’t catch up, will Tesla completely leave this era on the road of autonomous driving?