The first in the country! Jiyue displays pure visual city NOA, facing off against Tesla FSD

On October 17, Jiyue announced that its high-end intelligent driving solution based on pure vision has implemented urban navigation assistance functions in the core urban areas of Shanghai. At the same time, the official also announced for the first time the Occupancy grid network technology jointly developed with Baidu. 

Based on our consistent judgment on JiYue, we believe that JiYue's first car, JiYue 01, will most likely be mass-produced as a purely visual urban navigation solution. 

Jiyue 01 is very likely to be the first domestic model equipped with a purely visual city navigation solution. 

In the past, all domestic models equipped with urban navigation functions have chosen solutions with multi-sensor fusion and lidar. Jiyue's route layout will be the first time that a domestic car company competes head-on with Tesla FSD on the smart driving technology route. 

1. The first domestic company to announce a purely visual city NOA, the research and development process of Jiyue

Jiyue Company was established in March 2021. In June and July of that year, the team completed the initial setup and started the research and development of vehicle models. Strictly speaking, it only took two years and four months for Jiyue 01 to go from product definition to imminent launch. 

When initially defining the product, Ji Yue internally wanted to take a purely visual smart driving solution route. 

At the earliest, Jiyue CEO Xia Yiping and Wang Liang, the chief R&D architect of Baidu’s intelligent driving business group, discussed the mass production route. However, due to limitations of the computing platform and algorithm performance, the plan at that time decided to add lidar and interact with the vision system. Independent, as a redundant safety net. 

But by the end of 2022 and the beginning of this year, as algorithms such as BEV + Transformer mature, switching to pure visual solutions will once again be put on the agenda. The important factors are the mass production of large computing power chips and the industry's technological engineering breakthroughs in large models. 

Xia Yiping said, "If it were last year, I thought it would take 10 years to achieve (fully) autonomous driving. This year, with the emergence of a series of particularly outstanding engineering practices in the industry such as GPT4, I assessed that 3-5 years would be achievable." 

Before embarking on the pure vision solution, Wang Liang met with Xia Yiping to discuss in detail the required data sets, the scale of computing power required for algorithm training, and the development cycle. 

Jiyue quickly invested in hundreds of test vehicles for pure vision solutions across the country to collect data and accelerate iterative improvement of algorithms. 

In the early stages of development, JiYue's smart driving solution adopted a highly unified architecture, such as AEB, parking and driving, using a unified large model for perception. This was very difficult to advance in the early stages of development, but JiYue and Baidu Apollo insists this is the way forward. 

Wang Liang concluded that the long-term evolution direction of intelligent driving systems should be "from greatness to simplicity". Hardware complexity decreases, model algorithms become more focused, and algorithms based on rules and assumptions become less common. 

2. "BEV+Transformer+OCC", a key breakthrough in achieving pure visual city NOA

We estimate that Jiyue is expected to launch mass production of a pure visual city navigation solution this year. In fact, this pure visual solution cannot be achieved in the short term. 

In 2019, Baidu decided to open up a new technology route and unbundle lidar from the original L4 technology stack due to its internal observation of some technological progress and product-driven needs. Therefore, Wang Liang led a small team at that time and began the development of a pure visual L4 solution. 

In the early stage of product definition, Jiyue's intelligent driving solution designs the visual system and lidar system to be independent of each other and redundant to each other. Starting in 2021, Jiyue will join forces with Baidu Apollo to reconstruct the previous purely visual solution based on the new BEV + Transformer technology, such as using a more centralized AI algorithm model to solve perception problems. 

The introduction of BEV + Transformer significantly improves the iteration efficiency of the algorithm. 

In the past, the architecture of model algorithms was composed of many hierarchical small models. For example, some small models will focus on the recognition of wheels, turn signals and lane lines. At this time, the algorithm will write some rules based on the results of the small model perception. For example, the wheels may hit the lane lines, and the vehicle in front may change when the turn signal is on. road. When there are more models inside, the complexity of the system will increase. 

After the introduction of BEV+Transformer, a central large model is used to replace many small models, and data infusion will bring about direct experience changes. This is much better than modifying many small models and rules in the past. 

Therefore, Xia Yiping said, "The quality and efficiency of data production are the key to future competition." 

Automated data production lines are a major advantage of Baidu. 

Baidu currently operates thousands of Robotaxi units, so it has built a very efficient data production line over the years. 4D data with time series cannot follow the manual annotation method in the past, and the previous generation of perception algorithms can be used to generate annotation data to train the next generation model.

Moving on to the next step, the biggest criticism of pure vision is the recognition of unknown alien obstacles. The Occupancy Network occupies the grid network ("OCC" for short), which solves the recognition problem of special objects through visual three-dimensional reconstruction. 

The advantage of lidar is accurate ranging, but purely visual solutions can provide richer semantic information and denser point clouds, and avoid the impact of time synchronization and calibration problems between heterogeneous sensors. 

"Compared with purely visual solutions, lidar algorithms consume less data because the latter is essentially geometric information." Wang Liang said. The connotation of geometric information is limited compared to image information. Therefore, after feeding a certain amount of data, the upper limit of the lidar algorithm is not as obvious as that of the visual algorithm. 

Tesla FSD's V11 is the world's first mass-produced OCC smart driving system. We interviewed some FSD users in the United States. From V10 to V11, after adding OCC, the perception of special targets in urban areas has been significantly improved, especially for close-range objects. 

Judging from the current progress, Jiyue is likely to become the second company in the world to mass-produce OCC. 

Not only in terms of technical benchmarking, but also in terms of commercialization, Jiyue has also begun to compete tit-for-tat. Many stores have shown special rights to attract Tesla owners-a "special" regret fund of 2,000 yuan, which will comprehensively strengthen Tesla's The atmosphere is full. ‍ ‍ ‍ ‍ ‍ 

3. This may be the most radical model in terms of intelligence this year.

In 2021, Robin Li invited Xia Yiping to come out. The two first had a fundamental consensus that intelligence will determine the development direction of future automobiles and become the core competitiveness.

At that time, Xia Yiping saw that one of the key elements was two chips, one was Qualcomm's 8295 and the other was Nvidia's Orin. With these two chips, we can carry strong intelligent capabilities, and the algorithms that were previously placed on more expensive industrial computers and ran on the cloud can be run on the car. 

Therefore, Xia believes that the core of today's competition is to maximize the capabilities of these elements. 

Many companies today use 8295 to migrate their systems from 8155 to 8295. Although this will indeed make the system smoother, it does not bring out the value of 8295. 

Jiyue has built a voice interaction system that is the smoothest in the industry, with excellent recognition and response effects, and has many innovative features. 

What’s more interesting is that Qualcomm’s current generation 8295 has two built-in NPUs, and most companies only use one of them, which has a computing power of about 30T. After JiYue completed the large speech model, its 30T computing power was basically exhausted, and it authorized another NPU to be unlocked, using the remaining 30T computing power for applications such as graphics rendering and games. 

The smart driving function, the core of intelligence, will also be a standard feature of the entire series. Jiyue has revealed that the city navigation function will be available in multiple cities including Shanghai by the end of this year. 

In terms of interaction, we are fully promoting the use of voice, canceling the shift lever, canceling the door handle, and laying out a purely visual solution for the smart driving system. Jiyue has many radical approaches. 

During the entire product development process, Xia Yiping himself admitted that he was "very anxious at one time" because of the difficulty of development. 

However, "I dare to say that no one will question Jiyue today if it is said to be the benchmark for intelligence in the industry."