news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

at the chengdu auto show media day a while ago, two leaders of ideal auto’s intelligent driving r&d were interviewed. they were vice president dr. lang xianpeng and senior algorithm expert zhan kun. they explained in a simple and easy-to-understand way how ideal does intelligent driving and the logic of the entire large model optimization.

there were constant "big news" throughout the process, revealing that ideal intelligent driving's urban noa usage rate has exceeded that of highways, and its end-to-end + vlm solution has surpassed tesla, and it is currently aiming for the number one position in the industry.

to be honest, ideal has been very low-key in the past few months. suddenly it started to "make tough statements", which made many people feel uncomfortable. what level has ideal's intelligent driving achieved? let kungfu auto show you.

(1) from being usable to being easy to use, ideal is on the fast track

on september 6, ideal auto released its smart driving travel report for august, which mentioned several important data. the total number of ideal smart driving users has reached 908,000, of which 44,000 were newly added in august. judging from this data, the vast majority of ideal users have entered the smart driving circle.

among them, the proportion of urban noa usage exceeded that of highways for the first time, which is definitely a very surprising thing and even contrary to traditional cognition.

the reason is not complicated. compared with closed roads such as highways and elevated roads, the biggest feature of urban roads is congestion and openness. there are not only cars around, but also pedestrians, bicycles, food delivery vehicles, and various temporary objects.

therefore, most car companies, including ideal, started with noa (navigation assisted driving) on ​​closed roads, after all, the scenario is relatively simple, such as adaptive cruise control, automatic entry and exit ramps, automatic lane change, and automatic overtaking.

consumers' awareness of this "hands-and-feet-free" technology also began with closed roads. for example, when noa became popular two years ago, many people's first impression was that it could only be used on highways, with wide roads and few cars. if you often travel between the two cities on weekdays, it would be the perfect scenario.

as for city following, although many car companies have equipped the congestion assistance function, the reviews are generally not good. the reason is that the system is too "cautious". the space reserved for following the car is too large, which makes it easy for the car to be squeezed in. it also often asks to take over. most of the time, it is indeed "unusable".

the ideal urban noa is strong in that its driving style is very close to that of a "three-year new driver". it can easily make unprotected left and right turns at traffic lights, give way to pedestrians and avoid other vehicles, and can autonomously change lanes and overtake in all road conditions. as long as the lane lines are clear, the vehicle can achieve intelligent speed control and following vehicles, and even take the initiative to detour.

although it is not yet realistic to be as fast and smooth as an experienced driver during rush hours in the morning and evening, being able to go shopping more comfortably and commute to work without fatigue is already a huge improvement.

city noa is becoming more useful, and it is normal that the usage ratio exceeds that of highways. after all, city driving requires more intelligent driving, and driving is tiring.

(2) surpassing tesla, relying on “one model”?

at the chengdu auto show, lang xianpeng said that ideal's end-to-end + vlm solution has surpassed tesla, and the core part of it is a large network called "onemodel".

the ideal is to have an integrated onemodel end-to-end, that is, the system receives the input signal from the sensor, then the model is inferred and gives the trajectory planning, and then the vehicle is controlled to drive according to the planned path, with no other steps in between.

compared with other solutions that divide the model into two models, use signal bridge in the middle, and separate the perception model and the control model, ideal intelligent driving has no loss of intermediate information and no manual information digestion process. it is significantly more efficient and has a higher upper limit.

of course, there are disadvantages to this approach, which is that training is very difficult. this can be seen from the description of the technical path. what comes in is raw data, and what comes out is a trajectory line. it is a one-step process in between. you can imagine how difficult the entire ai analysis process will be.

therefore, unlike other car companies, ideal does not use the "shotgun method" to train a large amount of meaningless data, but rather selects the data accurately. among the existing 2.2 billion kilometers of training data, ideal only selected data from some experienced drivers, that is, those who have good driving habits, rarely accelerate or decelerate suddenly, rarely trigger aeb, but still have high commuting efficiency.

ideal has selected about 3% of drivers and used their data for training. this is probably why many users commented on ideal's intelligent driving, saying that it feels like driving like an experienced driver.

simply put, ideal has a question bank of simulated exams, which comes from users. ideal will verify the system's capabilities in each version, and then continue to accumulate to ensure that users can drive more and more easily.

as for whether it is better than tesla, that depends on personal opinion. after all, tesla only has a basic version of intelligent driving in china, while fsd sells for 64,000 yuan, but it has never been introduced.

(3) striving to be at the top, what does this ideal depend on?

recently, an interview with xia yiping, ceo of jiyue automobile, went viral. he said that baidu's investment in intelligent driving is at the level of hundreds of billions, while huawei's investment in intelligent driving is at the level of tens of billions, so baidu's intelligent driving is more convincing.

of course, in reality the gap is not that big, as the investment of both is in the hundreds of billions.

in fact, this also puts an invisible pressure on the market. baidu, huawei, or tesla are all giants. ideal's investment in intelligent driving cannot catch up with these three companies.

so why does ideal dare to shout out the slogan of striving to be the best in the industry? the answer is a more focused strategy and being closer to users, which is also the "magic weapon" that ideal always has at the table.

just like now, many car companies, regardless of size, are working on robotaxi. it seems that if they don’t work on robotaxi, they don’t understand intelligent driving. but ideal believes that robotaxi is just a product or a form of business, not a technology. it should be a business form that is automatically generated after the technology develops to a certain level.

watching a lot of american tv series instead of memorizing words in order to learn english is putting the cart before the horse. the former is indeed beneficial to learning, but in the final analysis, the data collected by the millions of robotaxi devices has very little impact on ordinary intelligent driving.

it’s not that it can’t be done, it’s just that there’s no need to focus all your attention on it.

why has ideal intelligent driving received such high praise in the past two years? there are actually two reasons. first, it chose the right strategic direction and did not do much useless work. second, it focused on the car usage scenarios of ordinary family users.

ideally, intelligent driving is meant to be used, not to show off.

ideal's ota upgrade frequency is very high, almost monthly, in order to continuously optimize and improve the user experience. for example, functions such as free parking space drawing and low-speed automatic emergency braking were all explored step by step by ideal.

the concepts of other leading car companies or consulting firms are indeed grand enough, but the final results may not be better than ideal.

always focus on actual experience and always put user needs first. this is why ideal can remain a lasting new force.

(4) kung fu clap

on september 3, ideal auto ota 6.2 was officially released. the main changes are the integration of lane-level dynamic information from amap and the strengthening of the driving assistance function in construction road scenarios. users generally reported that the update was not so aggressive and it was also very easy to drive in the city.

when it comes to "arguing", ideals are not strong, and often even to the point of being "bullied by everyone". but when it comes to "pleasing" consumers, there are really few car companies that are so attentive. working down-to-earth on technology, not doing so many "high-level" things, and serving consumers well is the ideal and invincible kingly way.