BYD talks about the progress of smart driving for the first time: mass production of domain control with large computing power within the year, large-scale sensing model on the car

Not long ago, HiEV exclusively reported BYD’s latest mass production plan in the field of intelligent driving. 

BYD will launch a high-end intelligent driving solution based on Horizon Journey 5 in the third quarter of this year. The first high-end intelligent driving function is high-speed navigation assisted driving DNP. The first model is likely to be Han. 

Wang Chuanfu has said on many occasions that the first half of smart electric vehicles is electrification, and the second half is intelligence. Starting this year, BYD has accelerated the pace of intelligent layout. 

Last weekend, at the Beijing Zhiyuan Conference, Han Bing, assistant to the president of BYD Planning Institute and director of the Electronic Integration Department, revealed BYD's latest progress in developing intelligent driving. This is also the first time BYD has publicly discussed the path and progress of developing intelligent driving. 

Han Bing's current role is equivalent to the person in charge of BYD's smart driving research and development. 

Han Bing graduated from RWTH Aachen University in Germany and holds a master's degree in electrical engineering and information technology. According to 36Kr’s previous reports, Han once worked at Delphi and was engaged in the development of middleware and underlying systems. The Electronic Integration Department he led in the past was mainly responsible for operating systems, domain controllers, middleware, etc., but in a recent adjustment, Han Bing's team also integrated the business of intelligent driving algorithms and chip research and development. 

Han Bing’s sharing mainly focused on “AI large models”. 

Han Bing believes that this year's combination of BEV sensing and other large-model technologies will be an opportunity for BYD's high-end smart driving to form "overtaking in corners." 

And by combining intelligent driving with the Yi Sifang platform, BYD is developing some distinctive high-end driving assistance functions. 

For what Han Bing shared, you can see the short video below.

1. Data and large models: BEV sensing will be implemented this year

Large model applications are inseparable from the most critical "data". Only with big data can we develop large models, and at the same time, we need a platform layout with large computing power on the car side. 

Let's take a look at BYD's progress from big models, big data, and big computing power. 

Han Bing revealed that BYD has currently established a research and development fleet with more than 300 vehicles. 

In terms of data, BYD has currently accumulated more than 150PB of data, and will add 1PB of data every day. These data are used for downstream training tasks, and most of them are automatically annotated. The annotation automation rate exceeds 95%. 

Han Bing predicts that 600 million kilometers of data will be accumulated this year, and exponential data reserves will be achieved in the next few years through research and development vehicles and mass production fleets to solve the long-tail problem of smart driving. 

Data is the foundation, and large models are tools to achieve experience leaps. 

Han Bing said that BYD is developing large-scale data-driven models. 

BYD's current intelligent driving perception model development has been 100% data-driven, and has internally developed a multi-camera fusion BEV perception model (for more information on BEV perception, please also see our recent series of articles and live broadcasts on BEV perception) , and plans to achieve mass production this year. 

The perception model can integrate multiple tasks into a large model and cover the entire process of continuous monitoring, fusion tracking and prediction. 

However, planning decisions are still based on rules. In the future, BYD also hopes that regulation and control will shift to deep learning as the mainstay, supplemented by rules. 

At the same time, in the cloud, BYD has also developed a large multi-sensor and multi-task Transformer model to serve various perception verification and labeling tasks. Large models can also be used for automatic annotation of ground truth systems. 

Han Bing believes that algorithms based on large models such as BEV may be an opportunity for BYD's high-end intelligent driving to achieve overtaking in corners. 

2. Large computing power platform: mass production of 508 Tops domain control this year

To achieve high-end smart driving, the car needs a large computing power platform. This is also one of the areas that the Electronic Integration Department specializes in. 

Han Bing showed a solution here: with 508 TOPS computing power and 64GB of memory. Obviously, the high-power computing chip uses NVIDIA Orin-X. 

Han Bing also said that BYD will develop all key software and hardware such as in-vehicle operating systems and domain controllers in-house. 

This platform will soon be installed on BYD's flagship model, most likely the U8. 

At the same time, the development of the entire intelligent driving software and hardware platform is based on a modular division of labor, which not only ensures controllability, but also enables parallel collaborative development with external suppliers. 

The advantage is that the product is more reliable and R&D costs are reduced. High-quality codes and modules are used in mass-produced products. This division of labor can achieve efficient cooperation. 

To some extent, this reflects BYD's current smart driving research and development strategy: the underlying software and hardware are self-developed and controllable, and the upper-layer algorithm modules and applications can be provided by suppliers to speed up mass production.

3. Outlook for the next step: Deployment and occupied network

In the next step, BYD will further implement a large data-driven model for the entire process of perception, prediction, and decision-making planning. 

For occupancy modeling, recognition and detection of road signs, and decision-making planning modeling, BYD will use massive data to support more accurate perception, achieve decision-making planning closer to experienced drivers, and provide users with a better experience The product. 

Han Bing said that BYD will establish a strong infrastructure and build its own model production line for decision-making and planning data. In this way, approximately 14 million kilometers of data can be automatically produced every day, and based on the computing power platform, algorithm models for decision-making and planning can be quickly iterated. 

The infrastructure here may refer to BYD Supercomputing Center. 

Of course, Han Bing also said that the data-driven part of BYD's smart driving system is still limited to the level of perceptual intelligence. 

The decision planning module is more of a rule-based algorithm. This is also the bottleneck for the next stage of smart driving development. Its complexity increases exponentially compared to the perception problem, and it is not unique. 

The next step is to drive data-driven decision planning. 

With the application and development of large decision-making planning models, BYD also hopes to evolve from perceptual intelligence to a higher level of cognitive intelligence.