Progressives break the wall: Tesla in the west, Chimo in the east! Using data intelligence to drive the history of autonomous driving into the city

Assisted driving is the only way to autonomous driving, and the progressive route wins the end faster.

On the 1020th day of its founding, at the sixth HAOMO AI DAY scene, this attitude was clearly expressed and the conclusion was spoken loudly. At this time, here and now, when the practical results and trends can’t be clearer, there is no need to argue over the years of autopilot route battles.

The route pioneered by Tesla started on the road with the FSD public test abroad, and was pushed to the test time by faster Wallfacers in China, and it was the start time of a larger data closed-loop model.

This wall-facer who overtakes cars on the straights in China is Zhimo Zhixing, and now he has clearly put the true meaning of industry axioms on the public screen: following data intelligence, relying on data intelligence, and practicing data intelligence, this is the first principle that drives autonomous driving to the end.

And as the software and hardware related to autonomous driving enter the mass production period, the exploration of the industry has come to the moment of large-scale urbanization – a new watershed, which is equally clear.

The entire autonomous driving will enter the 3.0 era driven by data intelligence.

The road is simple, the methodology is clear, and there is no wall for autonomous driving

Practical results are the only criterion for revealing truth

The wall-faced achievements of Momozhixing are shown through practical results.

Since its establishment, Momo Zhixing has just passed its 1,000th day, but it has quickly secured the first place in China’s mass-produced autonomous driving.

Mass production of autonomous driving corresponds to unmanned driving, which refers to the practice of applying autonomous driving capabilities to the landing of mass-produced vehicles.

In the field of passenger cars, Momo Zhixing has launched three generations of passenger car assisted driving products, HPilot, from scratch in just over two years, with 6 OTA upgrades in two years and mass production of more than ten passenger models. Landed, and concurrently developed 30 projects asynchronously.

The listed models, including Wei brand Mocha, Wei brand latte, Wei brand macchiato, tank 300, tank 500, Haval animal, latte DHT-PHEV, etc., have all been installed on the car. In addition, Mocha DHT-PHEV lidar version, Euler Lightning Cat, Euler Ballet Cat, and a new generation of Great Wall Cannons are being delivered one after another.

As a result, as of September 2022, the mileage of assisted driving by users at the end of the mill exceeded 17 million kilometers, ranking first in the mass-produced autonomous driving camp in China.

And this is just the beginning of a large-scale implementation. Mimo Zhixing said that by the end of 2022, HPilot is expected to carry nearly 30 models, and the number of models to be carried in the future will reach millions.

At the same time, the clarion call for the history of autonomous driving in the city has also been blown.

Just at the Chengdu Auto Show, Wei Pai announced that the new Mocha DHT-PHEV lidar version equipped with Mimo Zhixing City NOH will be mass-produced in September, and will be released within the year, and will be delivered upon launch.

Urban NOH (Navigation On HPilot) is the core feature of Mimo HPilot 3.0, which plans to enable passenger cars to achieve end-to-end intelligent driving under open urban road conditions.

Once delivered, it also means that Noh City will become the first mass-produced urban navigation assisted driving in China, and will once again set the record for mass-produced autonomous driving and even the entire Chinese autonomous driving track.

At the scene, the industry-university-research gurus not only expressed unanimous approval of the speed, mode and achievements of the mill, but also believed that the mill represents the direction of the industrialization of autonomous driving.

Zhang Yaqin, academician of the Chinese Academy of Engineering, professor of Tsinghua University, and dean of the Institute of Intelligent Industry (AIR) of Tsinghua University, said that the achievements made by Haomo within 1,000 days of its establishment are remarkable, and the persistently held HAOMO AI DAY is also a rare event in China that focuses on AI autonomous driving. The technology feast has built an industry technology exchange platform.

Jia Yangqing, vice president of Alibaba Group and a well-known AI framework leader, believes that Momo is promoting a new AI technology engineering paradigm in the field of autonomous driving, especially the creation of data intelligence systems and autonomous driving supercomputing, which may become influential in the entire industry system.

Chen Tianshi, the founder of Cambrian, a well-known AI chip company, shared the computing needs and trends that are being displayed in the cloud and on the vehicle end, and believes that the speed of the minute is on the side of the trend.

So what’s the secret to millimove speed? Chairman Zhang Kai gave a conclusive answer this time straight to the point

Firmly follow the progressive development route. In the 3.0 era of autonomous driving, assisted driving is the only way to autonomous driving.

Zhang Kai said that at present, China has become the main battlefield of smart cars in the world. It is expected that by 2025, the loading rate of high-level assisted driving will reach 70%. The era of intelligent driving is inevitable.

The reason why it is confirmed that assisted driving is the only way to autonomous driving is because of the central role of data-driven in it.

Zhang Kai emphasized that the progressive route is the best path for data accumulation, which is increasingly recognized by the industry and has become the general development direction of autonomous driving companies.

The millimo mode and millimo speed are the result of the firm practice of the progressive route and the source of the fastest 1,000-day speed in the autonomous driving industry.

The next question is how to accelerate and sustain?

Attention large model accelerates the end of autonomous driving?

Relying on MANA (Snow Lake), the first autonomous driving data intelligence system in China established by the self-research.

It includes multiple sub-modules such as data acquisition, transmission, perception, calculation, verification, etc., which can realize the iterative closed-loop of data from input to output, which is also the guarantee for fast iteration and continuous acceleration.

On HAOMO AI DAY, Momo disclosed the latest data of MANA. As of September 2022, MANA has learned more than 310,000 hours, and the virtual driving experience is equivalent to 40,000 years of human drivers.

Gu Weihao, CEO of Maomo Zhixing, also revealed the secret of how MANA continues to evolve.

One is the force from product and landing challenges.

The second is to continuously absorb the most cutting-edge innovative technologies.

For example, in the implementation of urban scenes, there are 4 types of scene problems and 6 major technical challenges.

Among them, the scene problems mainly include “frequent urban road maintenance”, “intensive large vehicles”, “narrow space for lane changes”, “various urban environments” and so on.

Correspondingly, there are six major technical challenges:

  • How to apply large models in the field of autonomous driving?
  • How to make data play a greater value?
  • How can real-space understanding problems be solved using re-perception techniques?
  • How to use the interactive interface of the human world?
  • How to make the simulation more realistic?
  • How to make the autonomous driving system move more like a human?

Under the 4 major scenarios and 6 major challenges, Momo Zhixing has made targeted upgrades and adjustments to MANA’s perceptual intelligence and cognitive intelligence.

First, data annotation. By using the self-supervised learning method of unlabeled data of mass-produced vehicles, the effect of the model can be effectively improved. Compared with training with only a small number of labeled samples, the training effect is increased by more than 3 times, and the training can be completed more efficiently and better adapted to perception need.

Second, incremental data learning methods. For the newly added data, extract some training data to form a mixed data set, rather than the method of distinguishing between the old and new data, and pursue the fitting of new data and the output of the new model to align the output of the old model, which can save 80% of the overall computing power. The speed is increased by 6 times, and it can also avoid the inability to take into account the scale and efficiency after the large-scale production of mass-produced vehicles.

Third, pay more attention to perception than maps, and bid farewell to high-precision map dependence. By using the time series Transformer model to do virtual real-time mapping in the BEV space, the output of the perception lane line is more accurate and stable, so that the urban navigation assisted driving does not need to rely on high-precision maps – this is actually a faster and lower threshold landing city A necessary capability for navigation-assisted driving.

Fourth, vehicle signal light recognition. Through the upgrade of the vehicle-end perception system, the status of the vehicle’s brake lights and turn signals is specially identified, so that the driver can be safer and more comfortable when dealing with the sudden braking and emergency cut-in of the preceding vehicle.

Fifth, simulation system evolution. For the most complex scene in the city, the intersection, introduce high-value real traffic flow scenes into the simulation system, cooperate with Alibaba Cloud and the Deqing government, introduce the most complex scene of the intersection in the city into the simulation engine, build an automatic driving scene library, and automatically The real simulation verification of driving has higher timeliness and more realistic micro-traffic flow, effectively solving the “big difficulty” problem of urban intersections.

It is worth noting that this is China’s first large-scale autonomous driving scene library based on vehicle-road collaborative cloud services, and it is also China’s first autonomous driving scene library generated using traffic data. The external release and application also marks China’s autonomous driving. came to a new stage.

Finally, anthropomorphic cognition. In the face of urban road conditions, how to make driving decisions more human-like is a recognized experience problem. The method of Xiaomo is to deeply understand the massive human drivers covering the whole country, learn common sense and anthropomorphize actions, so that the system can choose the optimal route according to the actual situation to ensure safety, and feel more like an old driver.

The above is the feedback force of the data and scene obtained after the system is mass-produced and put on the car.

At the other end, Momo continues its true nature and regards the absorption of the most cutting-edge innovative technologies of AI as another guarantee for progress.

Nowadays, it is the consensus of all players in the industry that large models and Transformers are used for autonomous driving, but those who are familiar with the industry may still have some impressions.

In terms of the latest cutting-edge judgment, Gu Weihao particularly emphasized the surprises brought by the large model under the Attention mechanism.

The mechanism behind the Attention model is mainly to solve problems with the idea of ​​a unified model, saying goodbye to the paradigm of using specialized models for different AI tasks. In fact, this mechanism has been proposed as early as 2014, but it is mainly used in the field of NLP. It has also made significant breakthroughs in the field of computer vision since 2020. From Google’s VIT to Microsoft’s SwinTransformer, it has easily topped the major rankings.

The Transformer structure based on the Attention mechanism performs amazingly in various general tasks, showing the potential of an effective general AI model paradigm.

Moreover, the structure of the Attention mechanism is simple, and the basic units can be infinitely stacked to obtain a huge parameter model, and the effect is also improved as the parameters are improved.

Gu Weihao believes that based on the Attention model, the large-scale human-machine co-driving data obtained by assisted driving can be converted more efficiently. With the delivery of mass-produced vehicles and the road, the amount of data is not only large but also diverse enough, and it can be faster. Arriving at the autopilot endgame.

This is also the underlying confidence in the technology that Momo believes that “assisted driving is the only way to autonomous driving”. At this stage, no data of sufficient scale and diversity can be accumulated more efficiently than assisted driving.

But if you want to enjoy the benefits of the Attention model, you have to solve the difficulty of its landing.

The core is the demand for computing power from super parameters: high demand, high cost, and high difficulty of landing, making Moore’s Law no longer effective.

Gu Weihao revealed that the method at the end of the mill is to reduce the training cost through low-carbon supercomputing, and realize the landing of the vehicle by improving the vehicle-end model and chip design. Both the cloud and the terminal are used simultaneously, and both are optimized.

So the Momo Supercomputing Center was officially unveiled, becoming the first autonomous driving company to build supercomputing.

Mimo also revealed that the goal of the Mimo supercomputing center is to meet the large model of 100 billion parameters, the training data scale is 1 million clips, and the overall training cost is reduced by 200 times.

The 3.0 era of autonomous driving?

Mass production, scale, data intelligence…

This is the most talked about word and the most emphasized word, and it is also a summary of the latest understanding of the development stage of autonomous driving.

In the exploration of autonomous driving, there have been route divisions, such as the ultimate group represented by Waymo and the progressive group represented by Tesla; there have been sensor camps, such as lidar camp and pure vision camp; even commercial The mode is the benchmark, and there are To C, To B and To G divisions.

But if starting from the first principles, is there a unified standard and review?

Momo Zhixing believes that there is, and there is only one standard: data.

According to the scale of the data, the road of autonomous driving exploration can also be classified into three eras:

In the 1.0 era, hardware drivers are the main drivers, and the scale process is about 1 million kilometers. The main perception method is lidar, and cognition relies on artificial rules.

In the era of 2.0, software drivers began to play a role, the scale can be accumulated to 100 million kilometers, and perception began to merge, but it was still the result of separate outputs from different sensors. In terms of cognition, artificial rules still dominated, and small-scale and small data began to be used to achieve better. Forecast and plan.

In the era of 3.0, data-driven is the core, hardware and software are unified here, perception also realizes the unified output of multi-modal sensors, and cognitively, it is possible to rely on large models and big data to have interpretable scene driving knowledge , capable of driving data iterations of more than 100 million kilometers.

In fact, according to the three eras proposed by Xiaomo, not only many past autonomous driving phenomena can be explained, such as why Robotaxi, which is stacked with lidar in the 1.0 era, still has low-level accidents, and for example, the high-speed loop on mass-produced cars in the 2.0 era The differences in the experience of road navigation assisted driving products…and the experience of Tesla AutoPilot and FSD are indeed in the past and continue to be optimized.

Therefore, what is more important is that the division of data dimensions proposed by Maomo can truly allow the autopilot genre and the evolution of the Ten Thousand Buddhas Dynasty, both inside and outside the industry, to have more objective coordinates and references.

In the past, to measure the technical development level of autonomous driving, there was the dimension of VC approval, the dimension of self-reported MPI, the dimension of somatosensory experience, and the dimension of road test license plate…

But without exception are biased towards the subjective dimension.

Only the data dimension based on mass production is close to the principle of AI transition and is a more objective dimension.

And this is also the result of the first stage of autopilot racing, and the reason why mass production is used as an autopilot midfield post.

Interestingly, with the introduction of data-intelligent autonomous driving 3.0, the field of autonomous driving has nowhere to go. No matter which route, no matter which camp, the core competitiveness has been put on the bright side

How large is the data? How efficient is the ability to acquire, train and utilize data?

It’s about iteration speed, but also about power consumption, cost, and acceleration to win the endgame.

The ability of data intelligence is an indicator to measure the core barriers of autonomous driving companies.

The ability of data intelligence is the watershed in the new stage of autonomous driving.

In fact, this watershed effect has been shown in the more concerned route battle before.

The value of Tesla was fully recognized after the production capacity problem was solved in Shanghai. The stock price and market value skyrocketed. Musk personally became the richest man in the world. The technical capabilities of AutoPilot and FSD became stronger and stronger. The more you get on the road, the larger the scale of data obtained, the richer the scenarios, and the iteration and evolution of this capability will continue.

Correspondingly, Waymo, the pioneer of this wave of autonomous driving, has been lowered again and again in its valuation, and its landing and advancement speed has been delayed again and again.

However, the above-mentioned watershed effect has been seen more as a “private grievance” between Tesla and Waymo, which conceals the essential problems reflected behind it.

Now, Momozhixing, who is crossing the river by touching Tesla, based on the practical results of daring the world, shouts the truth of progressive and assisted driving faster to win the end of automatic driving, and uses the discriminant of Autopilot 3.0, Validate new laws under the axiom of autonomous driving.

This means that the final battle of autonomous driving has kicked off, and it also means that it is time to reshuffle and reassess the ranking of the entire autonomous driving arena.

At least, it’s time to ask the question. 

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *