news

our reporter yin limei and zhang shuo reported from hefei

"intelligent driving will follow moore's law: the software experience will show an exponential improvement, reaching a level of 10 times improvement in two years, 100 times in four years, and 1,000 times in six years; the cost of intelligent driving hardware will also decline every two years. with half of the development trend, the cost of realizing noa (urban high-end intelligent driving) smart driving in a mapless city will be around 5,000 yuan by the end of 2025 or the beginning of 2026.”

on september 29, momenta ceo cao xudong said at the global intelligent vehicle industry conference (giv2024) that based on moore’s law of smart driving, momenta predicts that urban noa will experience explosive growth in the next five years, outpacing the development of electrification and new energy. faster.

at present, the intelligent driving level of mainstream models on the market is mainly concentrated at the l2 level. cao xudong said at the conference that the ultimate goal set by the company when it was founded was to achieve large-scale l4 autonomous driving, and to achieve large-scale l4 autonomous driving, the most critical point is safety. “we feel that we must achieve at least 10 times the safety of humans before we can achieve large-scale l4 autonomous driving. and to achieve 10 times the safety of humans, the most critical thing is to solve the problem of realizing large-scale l4 autonomous driving. millions of long-tail problems encountered.”

data-driven technology can solve most intelligent driving problems

according to the standards set by the society of automated engineers (sae), automobile autonomous driving technology can be divided into six levels (l0 to l5). l4 means that the vehicle can drive autonomously in most situations and under limited conditions (such as urban environment or highway) without driver intervention.

as an autonomous driving technology company founded in 2016, momenta's development path is to use ai to achieve scalable autonomous driving. what is large-scale l4 autonomous driving? cao xudong believes that large-scale l4 autonomous driving does not refer to hundreds or a thousand robotaix (self-driving taxis) performing demonstration operations in a city or several urban areas, but tens of millions or even hundreds of millions of vehicles can be operated on a large scale throughout china and even the world.

in cao xudong’s view, to solve the millions of long-tail problems encountered in realizing large-scale l4 autonomous driving, two profound insights are needed: first, it needs to be data-driven, which can automatically solve most problems. , and the rule-based method cannot handle the long-tail problem. second, mass-produced vehicles are needed to collect data for l4, and the l4 robotaxi fleet alone cannot meet the demand.

"achieving data-driven driving requires at least 100 billion kilometers of big data. the annual mileage of a household passenger car is about 10,000 kilometers, and 100 billion kilometers requires 10 million vehicles to run for a year. this is only to achieve large-scale l4 autonomous driving. the necessary condition is not a sufficient condition, and the sufficient condition may require more data." cao xudong said that based on these two aspects of insight, the company set up a data-driven "one flywheel" technical insight and approach at the beginning of its establishment. "walking on two legs" product strategy.

"walking on two legs" means that mass-produced autonomous driving provides data for fully autonomous driving, and fully autonomous driving feeds back l4-level autonomous driving technology to mass-produced autonomous driving, making mass-produced autonomous driving sustainable under the l4 architecture. upgrade and maintain market-leading competition.

according to cao xudong, the data-driven "flywheel" has now been iterated to the fifth generation. the automation rate of the first generation is 50%, and the automation rate of the latest generation, the fifth generation, has exceeded 99%. this means that every new 99 of the 100 problems can be solved through data-driven automation without human involvement.

“lidar and end-to-end technology are not inconsistent”

since this year, end-to-end, ai large models have become hot words in the automotive field and are profoundly transforming smart driving and smart cockpit experiences.

cao xudong revealed at the meeting that the momenta smart driving model can now effectively solve the long-tail problem of autonomous driving, support accurate prediction of the traffic intentions of vehicles or pedestrians in various complex road environments, automatically adjust vehicle speed, flexibly change lanes, avoid obstacles calmly.

it is reported that the momenta intelligent driving model can calmly deal with complex intersections or dynamic crossing scenarios, which can significantly improve driving safety and traffic efficiency. it can achieve precise parking even in extreme scenarios such as extremely narrow parking spaces and dead-end road parking spaces at night. currently, this technical solution has been delivered in mass production on multiple automobile brands.

automobile intelligence is a general trend. faced with fierce market competition, major companies have regarded high-end intelligent driving as one of their core competitiveness products.

at present, high-end urban smart driving is mainly installed on high-end models. cao xudong believes that by the end of 2025 to around 2026, high-end urban smart driving functions will become standard equipment for models priced above rmb 200,000 or even rmb 150,000. the driving factor behind becoming standard equipment is smart driving moore's law.

"there is a limit to the decline in the cost of smart driving hardware. it is estimated that the hardware cost may reach the limit of four to five thousand yuan. however, there is no upper limit for software's moore's law of 10 times growth in two years. zero-accident autonomous driving will definitely be achieved in the future. "driving." cao xudong said that it took five years for new energy vehicles to increase the penetration rate to more than 50%. it is expected that the development speed of high-end intelligent driving will be faster, and the penetration rate of high-end intelligent driving will increase to 70% in the next five years. %～80%.

in the field of intelligent driving, lidar and end-to-end are two different technical routes. at present, the industry faces some obstacles in terms of lidar technology: for example, the cost of lidar hardware is relatively high; the amount of data generated is large, and the system requires powerful computing power for real-time processing; hardware facilities that rely on high-precision sensors increase the impact on the environment. aware of hardware dependencies.

the end-to-end technology route is based on deep learning and neural networks, allowing the system to make full-process autonomous driving decisions directly from sensor input (such as camera image data) to output (such as vehicle control signals) without relying on intermediate modules. chemical treatment.

the end-to-end technology relies on low-cost sensors (such as cameras). compared with lidar systems, the hardware cost is lower, making it easier to commercialize on a large scale. at the same time, it is more flexible. through continuous learning, the end-to-end system can adapt to diverse driving scenarios and continuously optimize performance through data accumulation. however, the end-to-end technical system relies on large amounts of high-quality driving data for training.

in cao xudong’s view, lidar is not inconsistent with end-to-end technology. “first of all, lidar is scaling up very quickly and its unit cost is decreasing rapidly, so its price has become very competitive. secondly, the current industry consensus is that lidar is useful in some long-tail security scenarios, such as when a pedestrian crosses the road suddenly in a dark environment, the processing effect of lidar will be better. in addition, when entering and exiting a tunnel, it is easy for the camera to be overexposed. at this time, having lidar is better than not having lidar. more secure.”

cao xudong judged that from the perspective of industry trends, lidar will most likely become standard equipment in models with more than 300,000 yuan or more than 250,000 yuan. among models priced at 200,000 yuan or in the 100,000 to 200,000 yuan range, companies in the industry may benchmark tesla and choose the end-to-end technology route instead of lidar.

(editor: zhang shuo review: tong haihua proofreader: zhai jun)