Baidu finally found its place in the car circle

At the 2023 Baidu World Conference that just passed, the two-and-a-half-hour smart car sub-venue, Baidu's latest thinking and practice on the smart car business, were all displayed to the outside world.

To sum up briefly, the focus of this year is the launch of large models , mass production projects, and clarifying the position of Baidu’s smart car business in the automotive supply chain system.

Technically, the Wenyan Yixin large model and the perception large model are represented to empower smart cars, and eventually move towards the direction of cabin-driving integration that efficiently utilizes hardware capabilities;

Mass production projects, in addition to the Jiyue 01 that has been disclosed to the outside world and uses Baidu's complete set of smart cockpit and smart driving technologies, and the Xinlantu FREE, which is equipped with Baidu Apollo Highway driving Pro for the first time, Baidu has also previewed more mass production cooperation projects in advance. , such as China’s first integrated cabin-driving model that will be launched soon;

Finally, Baidu's market position has gradually become clearer. Baidu believes that the traditional Tier1/Tier2+OEM automotive industry chain is being deconstructed, and a new model of platform+ecology is taking shape. Baidu's role will emerge as a system solution provider. Still left to the car companies.

In short, there is a lot of useful information and the key points are highlighted. We will interpret them one by one.

01. What did Baidu do when the big model was launched?

In Baidu's words, "all products need to be redone with large models." This of course also includes automotive-related businesses with smart driving and smart cockpits as the core. Judging from Baidu's on-site display, this sentence The words are not lies.

In terms of intelligent driving, according to Su Tan, general manager of Baidu Intelligent Driving Business Group and Intelligent Car Business Department, based on Apollo Highway Driving Pro, which has been installed on Xinlantu FREE, the underlying logic code-based algorithm is deconstructed and replaced through a large perception model. , Baidu will launch an upgraded product, Apollo City driving, in the future.

The biggest upgrade of this product is that it can use mature sensor combinations and code lines to realize some city-leading intelligent driving capabilities, such as commuting mode, on the existing medium- computing hardware platform, highlighting an ultra-high cost performance.

The technical basis for realizing all this is Baidu's latest Apollo Efficient BEV Model perception model. This perception model is characterized by its simplicity and efficiency, which can break through the technical difficulties in deploying BEV on computing power platforms in the past.

Su Tan introduced that Baidu engineers reconstructed the underlying operators of existing products based on large models, and at the same time conducted in-depth lightweight optimization of the models, replacing multiple small models based on logical combinations with a single model, making the existing Mature mass-produced platforms can support the latest algorithm models.

The advantage of the new large model that brings upper-level functional experience is that the system can have second-level mapping capabilities, so that the commuting mode can be learned and used in one trip, and the same route can also be shared with the same car model.

In addition, with the new large model of Apollo City driving, the existing capabilities will also be strengthened to a certain extent. For example, the parking distance can be extended to 2 kilometers.

Finally, the signal revealed by Baidu this time is that it has firmly chosen the autonomous driving route of pure vision + high-precision maps. None of the products exposed this time have lidar . Considering that current car companies generally pursue the demand for low-cost smart driving solutions, this is actually easy to understand.

At the same time, on the technical side, Baidu said that based on the Apollo lite pure visual solution, it has developed a second-generation pure visual perception system, using 4D BEV Transformer and OCC network to reconstruct the three-dimensional environment, and its scene three-dimensional reconstruction resolution can also exceed the high-end Count lidar.

As for lidar itself, Baidu does not have an obvious line. Pure vision is more based on the current lidar performance and large model perception effects, and also takes into account the needs of car companies.

Sutan believes that the current battle between pure vision and lidar is not one of pure vision and lidar, but lidar is fighting against itself.

Perhaps in the near future, lidar can achieve a balance between performance and cost. It will not be surprising if Baidu adds lidar to the plan again. After all, this is also the majority opinion in the industry.

Baidu also gave an in-depth explanation of the application of large models in smart cockpits.

In this regard, Baidu believes that based on core capabilities such as large model understanding, generation, reasoning, and memory, cockpit products will be reconstructed. For example, the interaction mode will change from the past command-and-click style to natural conversational interaction; in-car navigation, Demand scenarios such as car use will be reconstructed, and long-tail demand for in-car entertainment and services will be further released.

Of course, the most important thing is the deep integration of the large-model-based automotive AI native application development model with the localized vehicle terminal. Based on the above changes, users and automobile manufacturers will also establish long-term relationships.

Specific to the technical level, Baidu's large cockpit model is based on the Wenxin large model, which deeply integrates the interior environment and car usage scenarios. It is also equipped with a smart cabin development tool chain, which allows car companies to combine their own application scenarios with a low-threshold personality. Develop your own cockpit functions.

At the same time, Baidu also provides car companies with a series of underlying technical support, such as training and evaluation data sets accumulated by Baidu in the past, as well as smart cabin data flywheels and ecological development plug-ins that can quickly iterate models.

Finally, based on the application and attempts of large models in the field of intelligent driving and cockpit, Baidu will set its next goal to integrate the cockpit. Su Tan believes that with the centralized evolution of automotive electronic and electrical architecture, the integration of the cockpit is an inevitable trend and is also the most critical node in the process of realizing the final centralized architecture. This trend will change in the next year. Land quickly.

Currently, Baidu is developing a cockpit chip that simultaneously implements basic smart driving functions and cockpit capabilities including AEB, LCC, etc.

This cockpit-integrated computing platform is named Apollo Robo-Cabin. It adopts a cabin-cabin partition isolation design to isolate hardware, communication links and resource scheduling to ensure the independence of business, communication and operation.

In the future, the ultimate goal of this core-software integrated solution is to realize urban commuting + autonomous parking 2.0 and smart cabin capabilities in terms of intelligent driving functions. Of course, judging from the deployment plan of Baidu's smart driving solution, the underlying computing chip should be at least a medium -power chip by then . The current mainstream 8295 is still a bit stretched.

02Mass production project is accelerating

The usual high-profile technology and lackluster mass production capabilities may be the industry's impression of Baidu's smart driving products for a long time in the past, but this time it is different. Baidu has disclosed a number of mass-produced projects and future products. The specific rhythm of landing.

The first thing I want to talk about is the Apollo Highway Driving Pro, which was first installed on the Xinlantu FERR. If you don’t count the WM car that has basically failed, this should be Baidu’s first high-end smart driving model.

According to Su Tan, so far, this solution has performed ideally in actual use. The number of high-speed pilot takeovers per 100 kilometers is less than 1, the success rate of lane changes is >99.2%, and the success rate of entering and exiting ramps is greater than 98.7%. At the same time, APA parking The success rate of the car is greater than 99%, and the parking time is less than 35 seconds.

In addition, according to data given by Lantu CEO Lu Fang at the scene, the current activation rate of Lantu users’ high-end smart driving is 83.74%, and among users who have activated the function, the penetration rate (understood as the open rate) is 99.06%. This means that user acceptance and satisfaction levels are relatively high.

At the same time, Baidu's high-end intelligent driving solution will be installed on nearly 10 models before the end of the year. Currently, it is relatively certain that Geely will release the Galaxy flagship pure electric model E8 at the end of this year.

In addition, Jiyue 01, which is deeply empowered by Baidu, will also be launched in the fourth quarter, when Baidu's full set of intelligent driving and cockpit solutions will also be unveiled.

It is worth noting that the Apollo City driving Max installed on Jiyue 01 can be regarded as Baidu’s most comprehensive solution for mass production of intelligent driving technology at present. Its pure visual + high-precision map technical route will eventually usher in market testing. .

In addition to what has been confirmed, Baidu's other products also have clear mass production time points. Among them, Apollo City driving, which will be upgraded on Apollo Highway driving Pro, will be officially OTA in the second half of next year.

The first model equipped with the Apollo Robo-Cabin, an integrated cockpit computing platform, finally appeared at the venue as an easter egg. Sutan said that it will soon be known to the public.