This article is reprint authorized by Chuxing Yike (WeChat official account ID: carcaijing), created by the transportation industry group of Caijing, written by Jingyi Wang Huaiyi Guo, and edited by Zhiliang Shi.
Tesla has introduced the perception, decision-making, and control concepts of autonomous driving into the production line, and each workstation is a “car” with autonomous driving functions.
In August this year, the 1 millionth Tesla electric car produced by the Shanghai factory officially rolled off the assembly line. At the same time, Tesla’s total production worldwide has exceeded 3 million cars. Also in August, Musk said he hoped that Tesla could achieve a production target of more than 100 million cars in the next ten years.
According to Musk’s plan, the increase in production capacity comes from the production capacity of new factories. He expects Tesla to have 10 to 12 factories in the future. On the other hand, the annual average production capacity of existing factories will also increase to 1.5 to 2 million cars.
As the Tesla factory with the highest current production capacity, when the Shanghai Super Factory was officially put into operation at the end of 2018, its designed annual production capacity was only 500,000 cars. By the second quarter of this year, this number had been increased to 750,000 cars by Tesla. Some media reported that the actual annual production capacity of Tesla’s Shanghai factory has exceeded one million cars.
With an area of only 860,000 square meters and without large-scale expansion of the factory and production lines, how did Tesla manage to double the production capacity of the Shanghai factory?
To explain this issue, we approached Allen Pan, who used to be responsible for the development of Tesla’s California unmanned factory, and delivered a complete solution for the capacity improvement and production line upgrade of Tesla’s Model 3 from 2017 to 2019. He founded his own start-up company Industrial Next (English Vision) last year, hoping to popularize the manufacturing techniques that have been verified at Tesla to more manufacturing industry companies.
Before joining Tesla, Allen was a L3/L4 system architect for GM’s autonomous driving and a core engineer for the CT6 autonomous driving. He was not related to factories. But Musk valued this in-person interview, “You don’t have any outdated production concepts in your mind, you can solve problems with different cognition.”## How does software define production at Tesla
Traveler One: From General Motors to Tesla, it seems like both are car companies, but autonomous driving and factories are two completely different fields. Why did you choose to work in Tesla’s factory?
Allen: I was also puzzled why they asked me, who specializes in software and autonomous driving, to work in the factory. I was recommended by a headhunter because they thought my hands-on experience with strong hardware skills matched Tesla’s recruitment needs. During the final interview with Musk, I asked about this and expressed my lack of experience in factories. Musk replied that Tesla’s factories are not like traditional factories, and “your mind doesn’t have the outdated production concepts of the automobile industry, so you can solve problems with a different mindset.”
In 2017, the California factory faced severe production capacity issues, only producing a few hundred cars instead of the planned 5,000 per month. This was far from the number of orders and Musk’s vision of solving production problems with high automation. However, the overall technology level of automation was not sufficient, and highly automated robots were not only less efficient than human workers but also lacked flexibility. Moreover, the cash flow was a major problem, and further investment in automation could affect the normal operation of the entire company.
By the end of 2017, Model 3 had begun mass production. Unlike traditional car manufacturers where once the product is finalized, it cannot be changed, this particular project had a significantly compressed time cycle from inception to mass production due to Musk’s agile development (DevOps) approach. This caused the Model 3 R&D team to iterate on product design continually. As it progressed to the production end, components started to change, and the automated production line could not keep up with the new changes. However, workers were smart and more flexible than machinery. They could quickly adapt to changes, but automation could not provide this flexibility as easily.
Musk had already figured out the general direction, and the next step was to recruit suitable engineers who could create a flexible automated production line to achieve mass production goals with close R&D- production integration.
Traveler One: What does Musk imagine the next generation of car production lines to look like?Allen: In Musk’s vision, the Tesla factory is like the Ford Motor Company in Henry Ford’s era, where batches of raw materials come in from one end and different versions of vehicles are produced from the other end. But there’s one difference: Tesla is using software-defined production. As the materials change with different product configurations and versions, we need to enable the production line to adjust adaptively.
Of course, software cannot directly define production. Software can define hardware, and by combining multiple functions and high integration together, the production mode is redefined.
Let me give you an example. When a change in a component is found, the material scheduling system will promptly notify the manufacturing execution system. The manufacturing execution system will then arrange for workers or production equipment to automatically skip the production step of that component until a suitable point in time to resume production.
This is actually autonomous driving technology. We introduce the perception, decision-making, and control concepts from autonomous driving into the production line, where each station is like a “car” with autonomous driving functions, which Tesla internally calls station control. In May 2018, our team’s proposed production solution was approved by Musk and began to be applied to the mass production line.
The appearance of station control does not change the four major steps of automobile production; it only becomes integrated into these four steps and optimizes the most basic production unit.
The production line is composed of individual workstations. If there is a problem with a workstation, the entire production line needs to stop. So, we thought about whether we could integrate production and quality inspection into the same workstation to solve problems on the spot. For example, during welding, when a problem is found during quality inspection, the workstation can adaptively correct the position and re-weld the problem.
Through intelligent transformation, every workstation on the production line can independently decide or revise the production process. This is like driving a car; if you turn a little too much, you can still turn back. Currently, only Tesla is doing this, while others set up additional quality inspection and rework stations in subsequent workstations.
Such workstations usually appear in critical welding and painting workstations, while final assembly workshops are more complex. To produce different cars on the same production line that require different components and fixtures, this is a problem that flexible production must solve, and it requires the recognition of different characteristic points of each component.
At the same time, when taking parts from the bin, they can’t always be in the exact same position every time, and there are always errors. For this reason, adaptive technology needs to be used to sense and make decisions – the core being image recognition in collaboration with robotic arms.The things just mentioned are all about specific workstations. From the perspective of the production line, each workstation has become an independent entity. If there is a change in the production process, materials no longer need to be sent to station A, but can be directly sent to B. Station A can also switch to other modules and be sent to other workstations. Regardless of product changes, the production line can adapt to these changes.
In the new production plan, the material scheduling system and MES (Manufacturing Execution System) play a core role. The former schedules materials around, sending various types of materials to workstations, while the latter organizes PLC equipment (Programmable Logic Controller, commonly used in automation devices) for production.
Behind this is a universal AI, not based on a specific scene, where it cannot be used outside of that scenario. Should Musk’s other companies require it, this universal AI can also be applied there.
Interviewee: How can workstations be upgraded to an intelligent level? Adding a camera to the workstation?
Allen: Two cameras, like how people have two eyes. The combination of the two cameras creates a 3D image.
In addition to the visual module, there are other perception modules (including cameras), edge computing modules, and intelligent gateway control modules (with autonomous learning capabilities), which form a complete solution for intelligentizing workstations.
Interviewee: Car companies always promote the automation level of their factories, claiming an automation rate of over 90%, dark factories, and so on. What is the difference between this and Tesla?
Allen: They are doing digitization. All commands are input, and after a series of operations, a result can be achieved without human intervention. The process of producing cola is similar to this highly automated way, but try producing a car this way. Cars are too flexible, too complex.
We are not building a digital unmanned factory. We focus on highly flexible unmanned factories. Workstations and production lines can adapt on their own, allowing for the smallest possible number of workers and achieving unmanned operations.
Interviewee: How can such a level of unmanned operations be achieved?
Allen: From a technical perspective, sufficient data is required for adaptive learning to be accurate. Otherwise, a worker could finish the job in one second, while adaptive learning may take 18 seconds, which would not be efficient. The more data you consume, the smarter the model becomes.
But some scenarios don’t have a lot of data to learn from. For example, placing a component into a fixed position using a robotic arm. If it deviates from the specified location, multiple incorrect attempts will damage the component. Production data is hard to come by, and there is not enough data to learn from, so we developed a set of algorithm logic through negative learning. With only a small number of erroneous data in training, we can reach 92-95% accuracy within 48 hours. After deployment, we continue to improve accuracy through online data.Traveler One: However, the level of automation and unmanned operation in Shanghai factory is not high, and many jobs are completed by workers.
Allen: At that time, I helped Musk plan the Shanghai production line, and I also asked this question. Musk said it was not necessary.
This is the difference between the Shanghai factory and the California factory. For the Shanghai factory, efficiency is the first priority, and rapid production of cars is the most important goal. Its positioning is an execution-oriented production factory. In contrast, at the California factory, the production line is also a product that needs to be upgraded. Musk’s requirement for the California factory is to create a research and development type production line, which not only completes the corresponding mass production tasks, but also connects the research and development end and the production end to achieve synchronous iteration of product and production line.
Tesla’s other Super Factories will also provide manufacturing and process-related information and data to the California factory. A factory with both mass production and production iteration functions is enough.
Tesla Values
Traveler One: Tesla has a concept of simplifying the production process and making cars that are easier to produce. How to understand this?
Allen: The production capacity in Tesla is not just measured in terms of the number of vehicles, but is more related to whether the parts have been simplified. If the parts are simplified, the number of vehicles that can be produced will naturally increase.
Nowadays, a car requires tens of thousands of parts, but if we simplify the components needed for a car into 10 parts, can the production capacity increase? There are many ways to reduce parts, such as integral die casting, high integration, and replacing the original hardware functions with software. When a multi-function component can replace several single-function components, the total number of parts will also decrease.
In order for a product to become easier to produce, constant iteration is necessary, and the iteration speed must be fast, which is highly demanding for the production line. If the product changes, it is best to iterate the production line without stopping; otherwise, the cost of updating the production line will be too high, and the competitiveness of the car will become weaker. This is a basic business requirement and a very core technology.
Traveler One: Musk once said that Tesla’s biggest advantage is its manufacturing.
Allen: In Musk’s companies, Tesla’s position is more like a production tool, not just a tool for making cars, but a company that manufactures and hones processes. Tesla is required to produce everything that Musk wants, and his ultimate goal is to go to Mars.
Although Musk will not directly say that everyone’s work at Tesla is for going to Mars, it can be summarized as this goal in the end. Therefore, it is also the reason why many devices in Tesla’s factories can also produce parts for SpaceX.
The glass used in rockets is produced at Tesla’s California factory. You may have noticed news reports about early Model 3s changing color when it rained – this was because the glass was originally intended for use in rockets and had to be adjusted for ultraviolet radiation.
While I was at Tesla, I had a lot of exchanges with SpaceX employees. Actually, it’s quite common for people to move around between Musk’s companies – not just between Tesla and SpaceX, but also Neuralink and Boring. Musk encourages cross-company exchanges.
Sometimes when the Tesla team encounters problems they can’t solve, the SpaceX team comes over to help – like when Tesla needed a line control system, which they took over from SpaceX. In fact, it’s quite common for parts to be interchangeable between Tesla and SpaceX, so many of Tesla’s parts are already aviation-grade.
The fundamental reason that Tesla is able to achieve software-defined cars and software-defined production is because of people. For the most part, it’s the same team who are writing the software programs across Musk’s companies. Because staff frequently move between the different departments of Musk’s companies, there’s not as much separation as you might expect.
Guest: You used to work for General Motors. What’s the biggest difference between GM and Tesla?
Allen: The difference between GM and Tesla is huge. The biggest change that electric vehicles have brought about is that software can define hardware. GM is a traditional car company that works to “follow the tracks” set out by existing products and production processes when making new energy vehicles. But Tesla is more disruptive – product lines and production lines can be defined by software, and so its organizational structure is very different from that of traditional car companies.
Another difference is reliance on suppliers. In fact, the majority of components used in automobile manufacturing come from suppliers. New entrants into the market see that the upstream suppliers are mature and therefore consider that they can enter the market by leveraging their own brands and channels. Tesla believes that car manufacturers should empower automotive component suppliers more.
Tesla also encourages challenges. For example, if your superior has a plan but you have a better one, you can challenge them and if you’re right, you can use your idea. So Tesla doesn’t look at seniority, but at who’s right. This is very different from traditional car companies, where people are judged by their seniority.
Guest: What’s the difference between starting your own business and working at Tesla?
Allen: My goal during the capacity-building battle when I assisted Musk at Tesla was the same as when I started my own business – to create a highly integrated, multifunctional (product high compatibility), and unmanned (minimal workers) production line. The difference is that now I’m doing it for the next generation of manufacturing, rather than for my own interests.Our company was established only last year, and is currently replicating previous products and processes with new technology, which takes time while also being an optimization process.
If we talk about vision, I hope to create a more flexible assembly line that can be compatible with different car models and even different products. There is still much work to do in terms of capacity improvement, such as the speed of product iteration, the speed of assembly line ramping up, and solving bottleneck workstations. In automotive factories, the assembly workshop is generally the most bottlenecked, and after we modify the assembly line, we can modify the stamping workshop, and even suppliers.
Traveler One: How is the financing and valuation currently?
Allen: In Q1 2022, we completed a $12 million Pre-A round of financing, with Lenovo Venture Capital as the lead investor, and Xiaomi Ventures and Boiling Point Capital as co-investors. We are currently doing a Pre-A+ round of financing.
As of June this year, the company has a valuation of $120 million.
Traveler One: Who are the customers?
Allen: SAIC Volkswagen, Ideal, Xiaomi, Lucid, Rivian, and more orders from the 3C industry, but they are still in the proof-of-concept (POC) stage.
From a product perspective, we mainly produce perception, decision-making, and execution module products. In addition, there are standardized workstation solutions, such as the paint inspection workstation, which costs approximately 2 million, as well as general assembly workstations, which are more difficult.
Traveler One: Tesla is your golden signboard.
Allen: Our team’s main members are former Tesla employees, distributed in China and the United States, with some small honors. But I think that the day Tesla is willing to use our products, invest in, or acquire us, that is the true signboard.
Now, Tesla’s technology is still running on the basis of what we delivered at the time, more than four years ago. Now, I am leading the original team, although we have just started, we have accumulated experience, and with new technology, we can catch up within 18 months. But even so, looking at other manufacturers and OEM automakers globally, they seem to have not yet started this effort, and their focus is still on the car itself, not on manufacturing.
Traveler One: What is the biggest gain at Tesla?
Allen: Three things: self-driving, crossing domains, and solving problems hands-on.
Traveler One: How do you understand first principles?
Allen: Believe in your instincts and don’t be afraid of failure. The most important thing is the process of going from point A to point B, which is the most difficult part.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.