Smart driving players, start the breakout race on the cloud

Today’s autonomous driving has reached the stage of competing in “digital infrastructure”.

Since "BEV+Transformer" became the recognized algorithm architecture, algorithm barriers are actually being gradually broken down, and the once mysterious end-to-end paradigm is also gradually unveiling its technical veil.

New car companies represented by Wei, Xiaoli , and Li Auto have entered the market, and suppliers such as Huawei and Bosch have also unveiled their end-to-end intelligent driving models .

It can be seen that each company has reached the stage of implementing the end-to-end model. At this time, the competition is about who iterates faster and who can provide an on-board experience more like that of an "experienced driver."

From a technical perspective, this points to the ability to build a closed data loop beneath the iceberg.

NVIDIA Wu Xinzhou, head of autonomous driving, believes that data closure has become the only way to achieve high-level intelligent driving, and no car company can bypass it.

The underlying logic of the so-called data closed loop is:

The vehicle and the cloud form a closed loop. The vehicle feeds data back to the cloud, where the autonomous driving model is centrally trained and simulated. The model data is then sent back to the vehicle for OTA deployment and updates. The two ends collaborate and complement each other, thus completing the iteration and evolution of intelligent driving technology.

The end-to-end technical architecture, from perception to regulation and control, is connected into an entire AI neural network . To make AI think and make decisions like humans, it means that the training data in the cloud is large and precise enough, the computing power is large enough, and the algorithm is strong enough. Otherwise, generalization, accuracy and recall rate will be difficult to guarantee.

This also means that in the era of AI-defined cars, the focus of intelligent competition is concentrated on the data closed loop from the car to the cloud.

Looking at a deeper level, the implementation of a closed data loop requires strong engineering capabilities to support it, that is, to form a complete technical system and tool chain for the entire process from data collection, transmission to storage, training, and optimization.

Therefore, around this bottom-up methodology, intelligent driving players have begun to start with infrastructure construction and closely bind with cloud service providers in a "complementary" relationship.

For example, Bosch recently announced that it would deepen its strategic partnership with Tencent to further cooperate in the fields of public cloud and private cloud for autonomous driving.

01. Taking the “ladder”, Bosch came later but arrived first

In the past two years, the fertile soil of domestic intelligent driving has produced many intelligent driving players with strong technology and mass production capabilities.

If latecomers want to enter the market, they must achieve differentiated results in a short period of time.

This is undoubtedly a huge challenge for Bosch.

First, the market competition is fierce. In the domestic smart driving and smart cockpit fields, it faces at least 200 competitors;

Second, it is easy to fail to adapt to the local environment. This old company known for its hardware has entered a new track where it competes with software and needs to re-establish its advantages.

However, as a global auto parts giant, Bosch still showed the courage and determination to "turn the elephant around."

For example, at the organizational structure level, in order to adapt to the development trend of "full stack of software and hardware", Bosch established the Intelligent Driving and Control Division. At the beginning of the year, it was integrated into the "Bosch Intelligent Mobility Group", officially getting on track in the software field.

Of course, Bosch is not starting from scratch. The engineering thinking accumulated over a century has already built up mass production capabilities. What it needs to do is to first build a high-level intelligent driving solution system based on the adaptation of software and hardware.

In terms of underlying capability building, Bosch found Tencent and built a stable and complete development tool chain for it. The key point is that relying on the cloud and compliance integration solution provided by Tencent, the data compliance problem was overcome.

For this German company, compliance issues have always been a part of the entire chain of data collection , processing, training, and storage, and relevant regulatory requirements have become increasingly stringent. Therefore, leveraging Tencent is an effective way.

In this regard, Tencent has its own map provider qualifications to back it up. It has also exclusively opened autonomous driving cloud zones in East China and North China, creating an end-to-end, fully compliant data closed-loop service specifically for autonomous driving development.

Wu Yongqiao, president of Bosch Intelligent Driving Control China, said that Tencent is acting as a policeman, helping it complete compliance tasks in data processing, screening, labeling and other work, greatly improving development efficiency.

Currently, Bosch has achieved the first phase of results, and has spent 18 months to mass-produce high-end intelligent driving solutions on Chery Xingjiyuan models.

According to Bosch's plan, the next stop is to continue along the model trajectory, mass-produce the two-stage end-to-end solution, then transition to a one-stage end-to-end solution in 2025, and finally complete the construction of the world model in 2026.

In this process, the demand for data, algorithms, and computing power has become increasingly stronger. In other words, Bosch needs a more powerful underlying tool chain for stable support.

Wu Yongqiao believes that the key factor in determining how fast and how far end-to-end big models will go in the future is high-quality data.

For smart driving suppliers, data resources are extremely valuable. Therefore, it is necessary not only to collect and accumulate large amounts of useful data, but also to make efficient use of this data to meet the needs of rapidly iterating algorithm training.

This means that Bosch needs to build up the management capabilities for these data assets from the underlying tool chain, that is, how to effectively classify, schedule and train the data.

Tencent's integrated storage and computing solution provides corresponding development tools. Data is classified into hot spots and non-hot spots in advance, and then stored in different programs, which can be quickly dispatched and used by the system.

In this way, it not only ensures the ultimate response latency and cost optimization, helps training tasks to fully release the computing power of the GPU, but also guarantees high bandwidth and ultra-large capacity requirements, making AI training reliable and efficient.

Bosch also uses Tencent Cloud Vector Database to quickly and accurately process massive amounts of unstructured data such as images, videos, and point clouds.

For example, in a scenario where the autonomous driving system judges the reflection of a car’s taillights as an obstacle, making it impossible to drive, a simple screenshot can be used to extract key features, and a large amount of relevant data can be searched almost instantly from the Tencent Cloud vector database based on image retrieval, greatly improving the efficiency of the Bosch algorithm team’s algorithm optimization.

One point that needs to be emphasized is that Tencent focused on playing the role of Tier 2 in its cooperation with Tier 1 Bosch.

Liu Shuquan, vice president of Tencent Smart Mobility, said in an interview that Tencent Cloud tends to provide general technologies, such as blurring sensitive information at the data cleaning level, while Bosch is closer to the business field and has a deeper understanding of simulation training and other aspects. The two sides always cooperate with each other.

Under the iceberg, the rapid iteration of the data closed loop accelerates the performance improvement of the model above the iceberg.

Based on this, Bosch can quickly keep up with the domestic end-to-end technology wave and make a key leap in high-end intelligent driving through its 1+1 greater than 2 partnership with Tencent.

02. The evolution of the identity of “image merchants” under the trend of no images

Under the general theme of vehicle-road-cloud collaboration, map data also constitutes a core part of cloud data.

Therefore, map providers have also become a key role in the field of intelligent driving, and have evolved their identity along with the evolution of intelligent driving.

In the era of high-precision maps, map vendors often only need to deliver a "black box" solution for car manufacturers and intelligent driving suppliers to mass-produce and install in vehicles. However, this has high costs, low freshness, and low coverage, and is no longer suitable for intelligent driving in urban areas.

Moreover, with the evolution of technical architecture from "BEV+Transformer" to end-to-end, the vehicle-side perception capability is continuously strengthened with model training, which means that the dependence on high-precision maps will be greatly weakened.

Of course, this does not mean that the presence of "image vendors" will be reduced.

Because the fundamental task of autonomous driving will not change, that is, driving from point A to point B, which is bound to be inseparable from maps.