Focus: Tesla'S AI Ambition

      Tesla is not only a car manufacturing company, but also a technology company. A press conference last week made Tesla's image more vivid.

On the morning of August 20, Beijing time, Tesla AI day was held as scheduled. Tesla released FSD (fully automatic driving) pure vision solution, self-developed AI chip and Dojo supercomputer. The more technical details and faster technical iteration of related hardware disclosed by Tesla let the industry Marvel again that "Tesla's technology is really amazing".

However, for those who have been paying close attention to Tesla for a long time, neither pure vision solutions nor self-developed AI chips and Dojo supercomputing are "big news". What is most surprising is actually a "colored egg": Tesla announced that it will soon launch a humanoid robot, also known as Tesla bot.

For artificial intelligence technology, musk obviously has a greater pursuit. On August 21, Deng Zhidong, director of the visual intelligence research center of the Institute of artificial intelligence of Tsinghua University, told reporters of the 21st century economic report that the core of both FSD and consumer level personal robots is AI, and the two application scenarios of automatic driving and personal robots are subversive in the future economic and social development, and also contain huge industrial development space.

However, on the whole, the current application of artificial intelligence in the industry is not satisfactory. Even for Tesla, although its automatic driving technology has been relatively advanced in the industry, it is still far from mature. At present, it is only in the auxiliary driving stage which needs human supervision. Recently, some defects in its auxiliary driving function have attracted extensive attention.

Obviously, there is a big gap between the ambition and reality of AI landing.

Tesla's AI capability, its technology iteration speed, at least at present other new and old automobile enterprises in the world do not have. Visual China  

The soul of Tesla AI technology system

"Tesla's AI capability, its technology iteration speed, at least at present other new and old car companies in the world do not have. This provides greater expectations for the application landing and industrial development of autonomous driving and personal robots. " Deng Zhidong told reporters of the 21st century economic report.

At the AI day, Tesla demonstrated its achievements in AI field, including the development of FSD, neural network automatic driving training, D1 chip, Dojo supercomputer and so on. One of the highlights is the development of FSD.

Unlike most of the industry's autopilot solution providers, Tesla has always adhered to a pure vision solution, which uses cameras to sense the environment without using lidar or high-precision maps. Relatively speaking, the hardware cost of this scheme is relatively low, but the requirements for data and algorithm are higher.

According to the scheme introduced at the conference, Tesla obtains road surrounding information through eight cameras around the whole body, and stitches different images through a multi task neural network architecture. In order to make the stitched information more realistic and valuable, Tesla has developed a set of technology, which can draw 3D bird's-eye view through the information of camera, and form 4D space and time label "road network" to present road and other information, so as to help vehicles grasp driving environment more accurately.

The above-mentioned process, the professional term in automatic driving is perception, and perception is only one of the core technologies of automatic driving. With the massive data base, it is necessary to establish a powerful neural network to integrate and re analyze these data.

Deng Zhidong pointed out that large models of neural networks usually need the "violence" support of big data and big computing power. Based on crowdsourcing Internet thinking, Tesla has obtained the largest scale of real road traffic big data in the original shadow mode based on crowdsourcing Internet thinking. This conference also showed Tesla's achievements in automatic driving simulation or digital twin system.

According to reports, Tesla has established a data tagging team of 1000 people, which combines the delicacy of labor and the efficiency of machines to label object information. At the same time, Tesla also developed simulation scene technology to simulate the uncommon "edge scene" in reality to improve the training efficiency of neural network.

"This work can not only obtain more large-scale composite data, but also has special significance for decision-making planning research based on deep reinforcement learning, and how to solve long tail problems and marginal small probability events before application landing." Deng Zhidong said.

As the data needed to be processed began to grow exponentially, Tesla was also improving the computational power of training neural networks. As a result, Dojo supercomputer was born. As a matter of fact, musk has repeatedly "revealed" the existence of Dojo supercomputer. This time, AI released the key unit of Dojo supercomputer - D1, a neural network training chip independently developed by Tesla.

As early as April 2019, Tesla produced the self-developed FSD chip in mass production. The D1 chip displayed this time has greatly upgraded its architecture and performance. It is said that the D1 chip, with a distributed structure and 7-nanometer technology, carries 50 billion transistors and 354 training nodes. The internal circuit alone is 17.7 kilometers long, realizing super computing power and ultra-high bandwidth.

However, Dojo supercomputer was originally built from 5760 NVIDIA A100 graphics cards, but now it is completely self-developed. It is composed of 3000 D1 chips or 120 training units. The total computing power reaches 9pflops (900 million times). It can adapt to massive video big data and realize AI big model training.

Musk has said that Dojo supercomputing will eventually be provided to other companies that want to use it to train neural networks, which means that Tesla may extend AI applications to other fields other than autonomous driving, and this robot's appearance proves this possibility to a certain extent.

It is reported that Tesla BOT is 1.72 meters high and weighs 56.6 kilograms. The screen on its face can display information. It has human level hands and strong feedback. It can achieve balanced and agile movements. It will use the training mechanism of Dojo supercomputer to improve its functions.

"There will be no shortage of labor in the future, but manual labor is just an option," Musk said. Tesla bot can perform dangerous, repetitive, and boring tasks. " The project is on the agenda, and Tesla BOT is expected to launch its first prototype next year.

Deng Zhidong pointed out that the current AI, which can empower the industry, is actually data intelligence with learning ability, or large-scale neural network. From the perspective of AI algorithm, Tesla obviously grasped the core architecture of multi camera visual neural network and tried to use the inherent learning ability of neural network, To solve all the problems of autonomous driving and even personal robot scene.

"Self learning to solve the challenges of perception, prediction and regulation can give full play to the ability of machines to surpass human beings, which is the soul of Tesla AI technology system and the most advanced place in concept of Tesla." Deng Zhidong pointed out.

AI landing still challenges

This time, Tesla aims to attract more professionals in AI field.

Although Tesla is the "leader" in the automatic driving industry in a certain sense, it is still a long way from the real unmanned driving. Even the auxiliary driving function which has been implemented now is not perfect. From this point of view, AI technology to further play a role needs to be further improved.

Recently, the National Highway Traffic Safety Administration (NHTSA) has launched an investigation into Tesla. It is reported that the 11 accidents investigated by NHTSA were all related to Tesla's autopilot or other automatic driving functions. Among them, 7 accidents caused casualties, with a total of 17 injured and 1 dead.

Specifically, these accidents occurred between January 22, 2018 and July 10, 2021, across nine different states. Most of them happen at night, and there are some objects in the scene after the accident, such as emergency lights, flares, luminous arrow plates and road cones. From the scene, when the road rescue personnel stop the vehicle to carry out rescue tasks, Tesla's automatic assisted driving function fails to recognize these objects and vehicles, and then a collision occurs.

When analyzing the causes of such accidents, a planning and control engineer of SAIC Technical Center pointed out that the cone barrel + static special-shaped vehicle is a typical edge scene, and it is difficult to cover this scene with a simple camera or even a combination of cameras and millimeter wave radar sensors“ First of all, the cone barrel shape is relatively small, which makes it difficult for millimeter wave radar to scan at a long distance, while stationary targets (including vehicles) are easily filtered out. And if the camera does not have targeted training, can not identify obstacles will become blind. Even if the vehicle is identified after approaching, it is impossible to avoid collision due to high speed and close range. "

Tesla, which adheres to the pure vision solution, has deployed the camera training program, but obviously has not solved this problem so far. For Tesla, this functional defect needs to be dealt with as soon as possible, which inevitably requires it to further strengthen the AI related team. This conference shows its technical path and reserves in detail, and Tesla is throwing olive branches to like-minded industry people.

From Tesla's point of view, musk, the company's helmsman, is now showing greater ambition. The humanoid robots presented at this conference show that Tesla not only needs to be an intelligent automobile company with autonomous driving ability, but also an artificial intelligence robot company covering more aspects, which puts forward higher requirements for the construction and reserve of relevant technical teams of Tesla.

From a meso perspective, in recent years, the global artificial intelligence industry has achieved rapid development, the number of artificial intelligence enterprises and financing level has increased rapidly, and the talent competition around the artificial intelligence industry has become increasingly fierce.

According to a report by the yiou think tank, in the past 10 years, the number of new AI enterprises in major countries around the world peaked around 2016. Among them, the absolute value of the number of new AI enterprises in China and the United States is still significantly ahead of other countries, which are the leading regions of global AI enterprises; From the perspective of introducing funds, the number of global AI enterprises' financing continues to grow, showing a geometric growth trend after 2016. By 2018, the global AI enterprises have raised a total of 78.48 billion US dollars, with the United States taking the lead.

However, AI talents have been in serious shortage for a long time. A number of research institutions have reported that there is a huge gap in AI jobs around the world. According to the AI jobs report released by uipath, a U.S. AI machine processing automation technology developer, there are 7465 job vacancies in the United States in 2018; The 2019 global AI talent report released by JF Gagne, CEO of element AI in Canada, also shows that the global AI talent pool is growing, but the demand still exceeds the supply. The latest global AI talent report 2020 shows that the demand for "new roles" has been stable despite a certain decline in demand last year.

According to the latest report released by JF Gagne, talent is the constraint of AI development, and AI industry at this stage needs more than talents who master software algorithms“ Whether AI's full potential is over hyped remains to be discussed, but we can say that AI's true success requires more than high-level experts and correct data algorithms. The AI industry initially focused on very advanced experts, because only they could manage new technologies and apply them to new areas. But now people realize that this new technology needs more than just engineers and people who can build good models to deploy it effectively. "

      The report further explains that AI is a new generation of software, which is coded with data rather than logical rules. In contrast, traditional software is static, and AI needs a new infrastructure ecosystem, which not only needs to be built, but also managed after deployment. Therefore, in order to make AI play a large-scale role, engineering, infrastructure construction, and New business model development and target monitoring and other fields need a large number of new talents.