NIO reorganizes its intelligent driving R&D department, and merges perception and control into a large model team

DoNews reported on June 19 that Wanwan Auto learned that NIO’s R&D department has recently completed structural adjustments. Previously, NIO’s intelligent driving research and development department consisted of perception, control and integration departments. After the adjustment, the perception and control teams were merged into the large model team, and the integration team was reorganized into the delivery team. The merged large model team is headed by Peng Chao, the former head of NIO’s perception team.

The reorganized autonomous driving R&D department will still be in charge of Ren Shaoqing, NIO’s Vice President of Intelligent Driving R&D. After this adjustment, Ren Shaoqing conveyed to the team that the traditional paradigm of “perception-decision-control” that has been used in the industry for many years should be abandoned. This means that NIO will more clearly explore the use of end-to-end large models to achieve high-end intelligent driving.

Ren Shaoqing graduated from the doctoral program jointly trained by the University of Science and Technology of China and Microsoft Research Asia. In 2015, he published the residual network ResNet together with He Kaiming, Zhang Xiangyu and Sun Jian at Microsoft Research Asia, which became ImageNet. The winning model in the classification competition. In 2018, Ren Shaoqing participated in the founding of Momenta and served as partner and R&D director. In 2020, he joined NIO. Peng Chao graduated from Tsinghua University with a master's degree and worked as a senior visual algorithm engineer at Momenta.

Currently, NIO’s smart driving team has about 1,500 people, which is relatively streamlined compared to other companies. Huawei's smart driving team has more than 7,000 people, BYD has about 4,000 people, and Xpeng has about 3,000 people. After Ideal shrunk its smart driving team last month, it currently has about 800 people.

Tesla released the FSD v12 version at the beginning of the year, with stunning results. This has gradually made the end-to-end large model technology used by Tesla become an industry consensus, and more Chinese car companies have begun to try this route.

People drive by seeing with their eyes, judging by their brains, and then controlling with their hands and feet. The operating logic of the intelligent driving system is similar to this. It mainly consists of perception, planning and control modules: it relies on sensors such as cameras or radars to "see" the external environment, relies on the software system to determine how to drive, and finally controls the vehicle's steering system and accelerator. , brake, etc. to complete the driving action.

Before end-to-end, intelligent driving systems, especially the planning and control parts, needed to rely on a large number of rules to be programmed to cope with various scenarios. This is not yet the mainstream method of AI, that is, deep neural network .

The so-called end-to-end refers to the entire process from the input end sensing the external environment to the output end generating control instructions for components such as accelerator and brake, all completed by a deep neural network. This requires pre-training with a large amount of driving behavior data to create a “large model” that integrates perception and control.

Since last year, the main focus of the smart driving program has been to take the lead in realizing large-scale NOA (Navigate on Pilot, pilot-assisted driving) in urban areas. It can realize point-to-point assisted driving on urban road sections. The vehicle can autonomously overtake, change lanes, pass intersections, etc. It is close to what humans can drive, and intelligent driving systems can drive.

The main benefit of end-to-end is that it can accelerate the implementation of high-end intelligent driving functions such as NOA in urban areas. Because it can cover long-tail cases (corner cases) that cannot be fully covered by rule methods, it allows intelligent driving to better adapt to various environments and scenarios without the need to "open cities" one by one, reducing the cost of popularizing NOA in urban areas. , shortening the popularization cycle. A good end-to-end smart driving system can also better simulate the driving behavior of human drivers, making the experience more comfortable and smooth.

In the competition for NOA functions in urban areas, NIO, which had previously been relatively small in terms of the number of service users, has gradually caught up.

At the end of April, NIO launched global navigation assistance NOP+, which includes highway and urban NOA, to users (NIO calls the NOA function NOP). According to official information from NIO, the solution has covered highway sections across the country and urban sections in 726 cities, serving nearly 260,000 users.

The regulatory part of the plan that Weilai has embarked on is still based on rules. NIO has previously stated that it will launch end-to-end active safety functions (including AEB automatic braking, etc.) in the first half of this year. It is understood that NIO’s latest Banyan 2.6.5 version will be launched soon and will include end-to-end AEB functionality. NIO has not yet announced an end-to-end mass production node.

In terms of coverage area, Huawei’s urban NOA has made the fastest progress. In February, Huawei launched the ADS 2.0 solution. It is not a complete end-to-end architecture, but Huawei claims that the solution can already support users to activate point-to-point assisted driving on any road section across the country. High-end versions of Wenjie, Zhijie, Avita and other models have already used ADS 2.0.

Huawei expects to fully switch mass production to ADS 3.0 with an end-to-end architecture in August.

Xpeng is also one of the car companies that currently has the most extensive urban NOA functions. In May this year, Xpeng officially said that it had achieved the end-to-end network on the car. Xpeng’s urban NGP (Xpeng calls it NOA Function: NGP) has covered more than 300 cities and is expected to cover the whole country in the third quarter of this year. Xpeng's end-to-end currently consists of three models: XNet, a large perception model, XPlanner, a large control model, and XBrain, a large language model. There is still a difference between the perception and control processes completed by one model.

Ideal has opened the AD Max 3.0 picture-free NOA experience registration channel this week, saying that you can experience urban NOA and other functions nationwide regardless of road sections, and has recruited a total of 9,000 early adopter car owners. Lili official said that it is expected to launch a high-end intelligent driving solution based on the end-to-end large-scale model developed by Lili at the end of this year or early next year.

It is understood that Great Wall plans to achieve end-to-end mass production of smart driving on three models this year. The plan is to cooperate with supplier Yuanrong Qixing, which has also reached relevant cooperation with BYD. SAIC Zhiji has also pushed end-to-end architecture imageless high-end intelligent driving solutions to users. These solutions are in cooperation with the supplier Momenta.

From January to May, NIO delivered a total of 66,217 new cars, a year-on-year increase of 51%.

The above is the detailed content of NIO reorganizes its intelligent driving R&D department, and merges perception and control into a large model team. For more information, please follow other related articles on the PHP Chinese website!