The secret behind Baidu’s accelerated implementation of autonomous driving: Wenxin large model and high-precision map automation

With the large-scale implementation of autonomous driving, the scale of data will explode. How to find more valuable data and how to efficiently use data to improve intelligent driving capabilities have become the key to continuous learning and scale-up of autonomous driving. Baidu today gave its own solution.

The commercialization of Baidu’s smart driving business is accelerating.

Today (November 29), at Baidu’s online Apollo Day technology open day, Baidu Apollo announced that it will continue to expand its business scale and will strive to build the world’s largest fully unmanned autonomous driving operation service area in 2023.

Baidu Q3 financial report data shows that in the third quarter of 2022, the order volume of Luobo Kuaipao exceeded 474,000, with a year-on-year increase of 311% and a month-on-month increase of 65%. Among them, the number of fully unmanned autonomous driving orders in Wuhan and Chongqing has grown rapidly; in addition, in the three first-tier cities of Beijing, Shanghai and Guangzhou, the average daily order volume for bicycles is more than 15, approaching the average daily order volume of traditional online ride-hailing services.

At the technical level, Baidu ranks first in the world with 3,477 applications for autonomous driving patent families, and its total autonomous driving test mileage exceeds 40 million kilometers.

After the development of multiple AI technologies, Baidu has quickly achieved a transition from fully unmanned operations to large-scale commercial operations. Currently, the delivery time for new city technologies is only 20 days.

This series of technologies includes autonomous driving system perception, predictive decision-making, planning and control, as well as full-chain technical solutions such as data closed-loop, maps, and computing power.

In addition, high-precision mapping is currently a key technology for realizing high-end driving assistance and autonomous driving. In the past, it was believed that the cost was high and it was impossible to be completed by one manufacturer alone. Baidu is making this impossibility gradually evolve into possible.

This is the Wenxin large model.

01.

Wenxin large model makes AI more intelligent

For L4 autonomous driving, to achieve a success rate of more than 99.99%, maps are an indispensable underlying capability.

Whether it is map generation, road environment perception, or data processing, today they are all inseparable from AI. In order to improve AI processing efficiency, Baidu has also evolved AI again.

This is large model technology.

As a key factor in the evolution of AI efficiency, large models have been mentioned more and more frequently in the past two years.

Baidu autonomous driving technology expert Wang Jingdong said:

Baidu launched the Wenxin Large Model - a weakly supervised image and text pre-training model. Backed by the Wenxin Large Model's ability to recognize thousands of objects, it has greatly expanded autonomous driving semantic recognition data, such as: special vehicles (fire trucks, ambulances ) recognition, plastic bags, etc., the efficiency of solving long-tail problems of autonomous driving has been improved exponentially; in addition, thanks to the Wenxin large model-autonomous driving perception model with a parameter scale of more than 1 billion, small models are trained through large models, and autonomous driving perception Generalization ability is also significantly enhanced.

Huang Jizhou, an expert on autonomous driving technology at Baidu, said that currently, at the level of high-precision maps, AI is the core driver of cost reduction and efficiency improvement. The automation rate of Baidu’s high-precision maps has reached 96%, which has greatly solved the problem of high application costs.

Apollo autonomous driving map can integrate vehicle-side sensing data and multi-source maps to generate online maps in real time to meet the needs of real-time updates during the autonomous driving process and ensure the safety of autonomous driving.

In addition, in order to improve the comfort of autonomous driving, the Apollo autonomous driving map is based on Baidu Map's 12 million kilometers of road network coverage and massive spatio-temporal data, combined with the driving knowledge accumulated by hundreds of millions of drivers, to build a driving knowledge map at the entire road network level.

In terms of data processing, Baidu's autonomous driving technology expert Li Ang proposed a closed-loop data design concept of "high purification and high digestion" to comprehensively strengthen the data alchemy of autonomous driving.

According to reports, the data purification path of this solution uses small models on the car and large models on the cloud to achieve high-efficiency data mining and automated annotation; the data digestion architecture realizes automated training, has the ability to jointly optimize and understand data distribution, and effectively utilizes high purity Data further enhances the overall intelligence level of the autonomous driving system.

The accumulation of this set of technologies gives Baidu more confidence to enter the passenger car market.

Baidu's current route is the L4/L2+ technology symbiosis route. Wang Liang, Baidu's autonomous driving technology expert, said that at the technology stack level, it has achieved unified visual perception solutions, unified technical architecture, unified maps, data integration and integration of L4 and L2+ smart driving products. Infrastructure sharing.

L4 will continue to provide advanced technology migration for L2+ smart driving products, and L2 data feedback will also help improve L4's generalization capabilities.

But he also emphasized that high-precision maps are necessary to ensure the high safety and good experience of L2+ city-level smart driving products.

Therefore, Baidu’s high-precision map automation rate of 96% is of profound significance to the commercialization of smart driving functions and cost reduction.

02.

The dawn of self-research on hardware

In addition to technology releases, the integration of software and hardware has also become an area that Baidu must explore when facing an uncertain external environment.

In addition to Baidu's full-chain technical solutions that were mentioned many times throughout the press conference, Baidu's self-developed AI chip Kunlun Core 2 is also particularly worthy of attention.

At the Apollo Day Technology Open Day, Kunlun Core Technology CEO Ouyang Jian revealed that Baidu’s self-developed AI chip Kunlun Core 2 has completed end-to-end performance adaptation for driverless scenarios. Although it is only a short sentence, it contains a huge amount of information.

Kunlun Core (Beijing) Technology Co., Ltd., formerly known as Baidu Smart Chip and Architecture Department, completed independent financing in April 2021, with a first-round valuation of approximately 13 billion yuan.

It is the first AI chip company in China to lay out the field of AI acceleration, with more than 10 years of deep experience in the field. At the same time, it has profound accumulation in architecture, chip implementation, software systems and scenario applications.

Kunlun Core 2nd generation AI chip is the second generation cloud general artificial intelligence computing processor. It will be mass-produced in 2021. It is the first general AI chip in China to use GDDR6 video memory. It uses 7nm advanced technology and GDDR6 high-performance video memory.

This is one of the few pieces of information currently revealed online. Of course, using the 7nm process, its production is still subject to non-mainland production capacity.

According to Ouyang Jian, in the end-to-end performance test of the driverless perception module, the performance of Kunlun chip 2 was better than the mainstream solutions in the industry.

With the acceleration of automobile intelligence, the demand for chips in automobiles is rapidly becoming high-end, and Baidu is also silently overcoming the difficulty of chips.

Today, Baidu comprehensively demonstrated its intelligent driving technology roadmap for the first time, introducing to the outside world Baidu's full-chain technical solutions in autonomous driving system perception, prediction and decision-making, planning and control, as well as data closed-loop, maps, and computing power.

Wenxin's large model and high-precision map automatic generation capabilities can be said to have helped Baidu open up the two channels of Ren and Du for the commercial implementation of smart driving business, reducing costs and improving efficiency.

Today, the commercialization process of autonomous driving in China is gradually connecting the dots into a line. On the consumer side, Chinese consumers are becoming more and more accepting of autonomous driving, and Baidu is obviously a powerful promoter of this process.