ideal auto's lang xianpeng: the gap with tesla fsd is less than half a year, and it may even lead in china
2024-09-01
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
according to red star capital bureau on august 31, the 27th chengdu international auto show kicked off on august 30. ideal auto (02015.hk/li.us) announced the latest progress and future plans of the end-to-end model, vlm visual language model and world model autonomous driving technology architecture, and announced that the new generation of ideal intelligent driving based on end-to-end and vlm visual language model officially started recruiting 10,000 people to experience the group. in addition, ideal ota6.2 was officially pushed in full.
lang xianpeng, vice president of intelligent driving r&d of ideal auto, and zhan kun, senior algorithm expert of intelligent driving, accepted a group interview with the media including red star capital after the press conference.
lang xianpeng (right), zhan kun (left)
according to zhan kun, the so-called end-to-end refers to a research and development paradigm, which is to complete a task from the initial input to the final output without any other processes in between, and use a model to complete the whole process from input to output. when applied to the field of autonomous driving, it means that only one model can be used to convert the perception information collected by sensors such as cameras into vehicle operating instructions.
at the beginning of 2023, tesla (tsla.us) mentioned end-to-end. currently, many car companies are rolling up end-to-end, but their ideas and progress are different. in zhan kun's view, compared with modular end-to-end, integrated (onemodel) end-to-end is a more essential end-to-end.
lang xianpeng said, "our core idea for intelligent driving is end-to-end + vlm. we believe that this approach is a more promising intelligent driving solution that is closer to human driving."
in lang xianpeng's view, the end-to-end + vlm technology architecture is essentially an artificial intelligence solution. "from now on, we are truly using artificial intelligence to do autonomous driving." he believes that under this premise, the core competition in autonomous driving research and development is whether there is more and better data and the corresponding computing power to train the model, and training data and training mileage cannot be bought with money.
he revealed that ideal auto's current training computing power has reached 5.39 eflops, and it is expected to exceed 8 eflops by the end of 2024. ideal auto invests more than 1 billion yuan in training computing power each year, and will consume 2 billion yuan this year. "we believe that the training computing power required to achieve autonomous driving will eventually reach the level of 100 eflops, which is equivalent to an annual investment of more than 1 billion us dollars."
regarding the world model, lang xianpeng pointed out that in supervised (l3 level and below) autonomous driving, the end-to-end model and vlm visual language model play a greater role, "because under the needs of supervised autonomous driving, the end-to-end model is sufficient, and vlm only serves as a reminder and auxiliary." however, after unsupervised l4 level autonomous driving, this system has to independently handle all unknown scenarios and emergencies, and the number of model parameters increases dramatically, so the world model on the vehicle side is needed.
the takeover rate is one of the core indicators of system capabilities. according to reports, ideal auto has improved the takeover rate to once every 21 kilometers. in the future, it will be able to increase to once every 100 kilometers. however, if there is no takeover for a long time, people's mental attention will be distracted. ideal auto will use a new interactive experience to allow drivers to take over when they should, and push it to users in advance by evaluating the scene areas with high frequency of takeover.
compared with its competitors, ideal auto currently does not charge for advanced intelligent driving. lang xianpeng emphasized that standard configuration and free of charge are the strategies that ideal has formulated since the first day of entering the intelligent driving market. supervised autonomous driving is free for all ad max owners. "delivery volume is a very important metric. for us, it is not just about delivery volume, but also about providing more vehicle training mileage for autonomous driving. with good delivery volume and stable business operations, we also have enough resources to invest in intelligent driving research and development."
last year, lang xianpeng said that the gap between ideal auto's intelligent driving and tesla's fsd was about half a year. at this chengdu auto show, he said that "this year (the gap between the two sides) may be even smaller."
he explained that, first, in terms of technical architecture, ideal auto is not much different from tesla, and is even a little ahead, "because we have vlm and system 2, while tesla only has system 1, end-to-end."
second, in terms of training computing power and training data in china, "at least for now, we are ahead of tesla, because tesla is subject to constraints in data compliance and other aspects, and the deployment of training computing power in china still needs to be built. from this perspective, the gap between us and tesla in china may not be that big, and we also particularly hope that tesla can join in, learn from each other, and focus on improving itself."
red star news reporter wu danruo
editor: yang cheng
(download red star news and get a reward for reporting!)
