On September 5th, Tesla's AI team posted on social media platforms that Tesla plans to launch an advanced driving assistance system called Fully Autonomous Driving (FSD) in China and Europe in the first quarter of next year, and is currently waiting for regulatory approval. As soon as the news came out, it once again sparked a heated discussion about autonomous driving among people.
Industry insiders say that with the continuous iteration of technology and the acceleration of end-to-end technology mass production, upstream and downstream enterprises in the industrial chain are planning ahead, and the industrial and market landscape may undergo significant changes.
Tesla triggers catfish effect
The entry of Tesla FSD into China will definitely bring certain impact and stir up the domestic market. At the same time, technological progress will also bring new development opportunities for the industry, and Chinese enterprises have many development opportunities, "said Zhu Xichan, a professor at the School of Automotive Engineering at Tongji University. SAIC Saico CTO Yu Qiankun told reporters that instead of seeing Tesla as a competitor, it is better to expect more practitioners to work together to expand the autonomous driving market and promote the overall development of the industry.
With the dynamic of Tesla FSD entering China attracting people's attention, the end-to-end technology applied by Tesla has once again returned to the spotlight.
When it comes to the inspiration that Tesla FSD brings to the industry, it is that Tesla shows people the ability of end-to-end models: through more data collection and more model training, better performance can be achieved. This has made many companies firm in their technological path and choose to collect more data and build larger computing platforms to train end-to-end models. In fact, there have been studies on end-to-end technology in China for a long time, and this technology is not unique to Tesla, "said Zhu Xichan.
Accelerate the layout of related industry chain enterprises. The Aion strategic model released in July is equipped with an end-to-end advanced intelligent driving solution jointly developed by GAC and Momenta Intelligent Driving Global Co., Ltd. (hereinafter referred to as "Momenta"). The Momenta algorithm 5.0 applied in this solution is an intelligent driving big model that applies end-to-end models, and builds a data-driven R&D system and algorithm architecture for the entire process.
Algorithm 5.0 has been mass-produced and delivered on multiple car brands, continuously improving algorithm capabilities and product experience through the accumulation of massive data, "said Momenta CEO Cao Xudong. Through the passenger car data brought by mass-produced cars, it can better solve the long tail problem and achieve true scalable autonomous driving. Currently, many domestic host manufacturers have adopted the company's end-to-end image free NOA
Previously, domestic automobile manufacturers and intelligent driving companies represented by Xiaopeng Motors, Yuanrong Qixing, Shangtang Jueying, and Zero One Motors have invested in end-to-end system research and development, and disclosed their plans for mass production of vehicles to the public.
Increased demand for data and computing power
The development of AI cannot be separated from good models, big data, and high computing power. Autonomous driving technology has not yet fully found a safe boundary, "said Zhu Xichan. End to end model training requires a large amount of data and computing power support, and for enterprises that focus on end-to-end models, it is crucial to meet greater data and computing power demands in the future.
Liu Yudong, Executive General Manager of Chentao Capital, also pays attention to the demand for data and computing power. Liu Yudong stated that the importance of training data has reached an unprecedented level in the technical architecture of end-to-end models. The scale, labeling, quality, and distribution of data can all become obstacles to the development of end-to-end applications. Meanwhile, the demand for computing power in end-to-end training is rapidly increasing.
Liu Yudong said that in terms of technical roadmap, end-to-end technology has not yet formed a unified best practice, and there are certain differences. Meanwhile, traditional testing and validation methods are not suitable for end-to-end autonomous driving, and the industry urgently needs to develop new testing and validation methodologies and toolchains. From the perspective of resource allocation, end-to-end technology requires reshaping the organizational structure and investing more resources in data, which poses a challenge to the existing operational model.
At present, end-to-end technology is still in its early stages of development, and there are still many application challenges and pain points that need to be solved urgently. The new pain points and opportunities formed by these unmet needs will also become the direction for practitioners' future evolution and iteration, "said Liu Yudong.
Evolution of market and industry patterns
In the future, with the continuous iteration of algorithms and the accelerated increase in the penetration rate of autonomous driving, it will greatly drive the upstream technological progress, market and industrial pattern evolution of the autonomous driving industry. "The" End to End Autonomous Driving Industry Research Report "(hereinafter referred to as the" Report ") shows.
Zhu Xichan stated that currently, many end-to-end technologies have been implemented, but whether they are effective or not still needs to be considered. In terms of end-to-end technology, there will be at least two rounds of technological iterations in the future. The first stage is the coexistence of both fast and slow systems. Beyond the fast-paced end-to-end model, vehicles handle complex problems through slow systems resembling a 'brain'. The second is to solve problems that cannot be explained by end-to-end models through visual language models.
Xie Chen, founder and CEO of Lightwheel Intelligence, said that end-to-end models mean that the industry's focus shifts from algorithms to data. Traditional automakers are facing new opportunities, which means cost reduction and efficiency improvement opportunities for autonomous driving companies that already have mature mass production projects and data loops. At present, the core bottleneck of end-to-end technology is still in evaluating and verifying data capabilities, which may be solved through the large-scale application of synthetic data.
The report shows that in terms of technology, end-to-end implementation will accelerate the progress of upstream toolchains, chips, and other technologies. On the market side, the end-to-end improvement in autonomous driving experience will lead to an increase in the penetration rate of advanced assisted driving; Due to its strong generalization ability, end-to-end autonomous driving may also drive applications across geographic regions, countries, and scenarios; In terms of industrial landscape, the importance of end-to-end data and AI talent is further enhanced, which may give rise to new industrial division of labor and business models.
