"In my opinion, it will take many years for end-to-end technology to be applied to autonomous driving. However, as end-to-end models begin to be implemented, more powerful intelligent driving systems and even fully autonomous driving will emerge within a few years." Recently, in an interview with First Financial Times, He Xiaopeng, Chairman and CEO of Xiaopeng Motors, said that in the future, only a few companies will truly achieve fully autonomous driving, and there won't be too many players who can stand out.
On May 20th, Xiaopeng Motors held an AI Day to share the latest developments in Xiaopeng's AI technology and announced the full launch of the AI Dimensity system to users. Among them, Xiaopeng Motors has released the first end-to-end large-scale model for mass production in China. In addition, Xiaopeng Motors will invest 3.5 billion yuan in intelligent research and development, recruit 4000 new professional talents, and invest over 700 million yuan annually in computing power training in the future.
It is reported that the end-to-end large model can achieve a doubling of intelligent driving ability, a doubling of perception distance, and recognition of more than 50 target objects; Be able to recognize the waiting area and tidal lane, and read the road sign text to infer the intentions of traffic participants; The stuttering, stuttering, and secure takeover in complex scenarios have all been reduced by nearly 50%. According to the plan, in the third quarter of 2024, Xiaopeng Motors will be able to drive on every road in the country, fully achieving no map. By 2025, it will achieve a L4 level intelligent driving experience in China.
"The transformation of smart cars by big models is disruptive, not a proportional change, not a few tens of percent, but a tens of times change." He Xiaopeng said that with the support of big models, the Xiaopeng car intelligent driving system can achieve one iteration in two days, and within 12-18 months, it can achieve 10-30 times experience optimization of the intelligent driving system.
After Tesla announced that the FSD V12 will adopt an end-to-end big model, there has also been a wave of end-to-end big model applications in China. In addition to Xiaopeng Motors, car companies and suppliers such as Ideal Automobile, Yuanrong Qixing, and Millimon Zhixing have also started end-to-end research and development and mass production adaptation.
But He Xiaopeng believes that not all car companies can successfully transition from regulatory control to end-to-end.
In the past few years, Chinese car companies have intensively launched intelligent driving assistance systems, but almost all systems are based on traditional divide and conquer methods, dividing autonomous driving tasks into three independent models: perception, prediction, and planning, and then integrating them to complete the entire autonomous driving task. This modular model, based on rule control, makes intelligent driving actions prone to strong mechanical sensations.
The end-to-end large model integrates perception, prediction, planning, and other aspects, allowing AI to learn mature human driving behavior through a completely data-driven model. After inputting a scenario, the system will directly provide corresponding countermeasures.
"End to end can be considered as the ability to continuously personalize and respond in various complex scenarios; regulation will be more stable, but relying on regulation will never fully solve the Corner Case problem." He Xiaopeng said, "However, in the development of end-to-end, data needs to be recycled, and there is a great demand for computing power. Xiaopeng has invested $100 million in computing power this year, and it will further increase every year in the future.".
This year, Xiaopeng Automobile's R&D investment is expected to exceed 7 billion yuan, with investment in artificial intelligence, intelligent driving, and other areas reaching 3.5 billion yuan. He Xiaopeng believes that the current Chinese automotive market is moving from a knockout stage to the final "all clear competition". Without persistent investment, enterprises will not be able to stand out from the knockout stage.
Regarding the overall trend of the market, He Xiaopeng believes that although hybrid vehicles have developed rapidly in the past two years, the growth rate of pure electric vehicles will increase again from 2025 to 2026, and gasoline powered vehicles will still exist for a long time.
"Although it is widely believed in the industry that the Chinese car market has moved from the first half of electrification to the second half of intelligence, there are still no high-level intelligent cars on the market. It is in the stage of transition from functional machines to intelligent machines, and the training of big models and big data will accelerate the progress of this process." He Xiaopeng said that more universal and powerful artificial intelligence will definitely emerge in the future, but a fully universal AGI may still take a long time to appear.