Zebrafish brain science integrates autonomous driving, Musk's Chinese disciples make new progress, not picky about cars or chips
End-to-end has reached the next level?
Tesla is on fireEnd-to-endthe industry is scrambling to shift away from rule-driven and embrace data-driven.
Some people even say that data-driven end-to-end is the ultimate goal of autonomous driving.
However, just recently, a company founded by the backbone of Tesla AutopilotNullmaxpublicly proposed a new paradigm:
Intelligent drive.
Combining brain science with autonomous driving, we try to explore the next end-to-end stage.
This direction is also attracting well-known OEMs and suppliers to invest in research, such as Ideal and Horizon.
The current situation, which is mysterious from end to end, has become even more indescribable.
End-to-end has reached the next level?
Nullmax recently launched the latest technical solution:Nullmax Intelligence (hereinafter referred to as NI), based on an end-to-end architecture, adoptsMultimodal large modelandSafety Brain.
First, let’s look at the large multimodal model, which has billions of parameters.
As a startup derived from Tesla, Nullmax is also a believer in pure vision, and its solution does not use lidar.
But the NI model input is not only visual information, but also includes sound, text, etc.
The output includes driving actions, visualization results and scene descriptions.
After the model outputs the results, Nullmax will also useTwo-level arbitrationMechanism to judge whether the vehicle will collide with other traffic participants.
First-level arbitration is a common method in the industry. On this basis, Nullmax also introduces second-level arbitration.“Safe Brain”.
The so-called “safe brain” is a neural network with 100,000 parameters, which corresponds toZebrafishThe number of neurons.
Zebrafish isThe “little white mouse” in the field of brain science has a brain structure that is highly similar to that of the human brain and can be seen throughAfter treatment with fluorescent molecules, the relationship between neuronal activity and fish movement can be observed.“Internal state dynamics shape brainwide activity and foraging behavior”already atNaturePublish.
co-authorLi MengDoctor, yesHarvard UniversityHe is currently a postdoctoral researcher at the Yansi Brain Research Institute.
Nullmax is working with it to combine brain science with autonomous driving.
As the industry changes,idealandhorizonTogether with Nullmax, we are trying to explore this new field.
Ideal and Tsinghua University are trying to integrate human fast and slow thinking into autonomous driving to form System 1 and System 2.
Horizon Intelligent Driving SolutionsSuper DriveThe core of decision-making planning“Interactive Game Algorithms”making the car seem to have a “brain”, and the thinking motivation behind it is also human “slow thinking”.
However, Nullmax's “brain-like” model is designed based directly on the zebrafish brain structure, which is different from the inspiration sources of the two companies. Therefore, Nullmax proposed “intelligent drive” based on end-to-end data-driven.
Moving towards a new stage, “intelligent drive” also existsdataandtrainTwo major bottlenecks.
Nullmax believes thatHigh-quality data remains criticalHowever, high-quality data is hard to come by, so Nullmax has established a data platform that includes both real collected data and virtual data generated by videos. The combination of the two speeds up algorithm iteration.
In addition to high-quality data, another bottleneck istrain.
Including how to input data, in what order, as well as debugging models and adjusting parameters.
The above is the latest progress of Nullmax upper-layer technology, which will be implemented early next year.
The key technology supporting Nullmax's commercialization is its self-developed middleware platform:
Not picky about cars, chips or cameras.
Nullmax's commercialization
Nullmax's commercialization is currently mainly based on1 platformand3 product solutions.
1 platform, namely the middleware platformMaxOSIt tokenizes multimodal inputs, decouples software algorithms, domain control, and chips, abstracts the software layer as much as possible, supports different computing platforms, and adapts to any sensor, such as customized cameras for car companies, and can even adapt to different locations.
MaxOS makes it possible for Nullmax to adapt to different car models and different hardware configurations, thus making it possible for the product to be launched.
Currently, its products can be divided into the following three types according to computing power requirements and implementation capabilities:
1V2T, one camera, plus a TI 2TOPS computing chip, achieves basic L2 assisted driving. 5V/6V8T, 5/6 cameras, using 8TOPS computing power, achieves high-speed NOA and memory parking, and can detect obstacles during parking. 11V, achieves L2+, including high-speed NOA and autonomous valet parking.
Nullmax said that the future intelligent driving solution will beLess than 100TOPS of sparse computing power to achieve full-scenario NOA.
To sum up, the commercialization of Nullmax shows two major characteristics:
High flexibilityandHigh cost performance.
Why did Nullmax choose to create such a product?
Nullmax believes that the industry has developed a variety of solutions using chips with different computing power scales, but no company has yet developed a solution that can adapt to different chips and cover the 2TOPS-2000TOPS computing power range.
There is a gap in the market, which is Nullmax's opportunity and the reason why the product is highly flexible.
On the other hand, although the industry is currently very busy with high-end intelligent driving, the excitement only benefits a small number of people. Only high-end models can enjoy high-end intelligent driving. However, there is also a demand for high-end intelligent driving for a wider range of cost-effective models. Therefore, Nullmax has developed a set of cost-effective solutions.
The highly flexible and cost-effective solution has won Nullmax a number of well-known partners, including leading OEMs.BYD, Chery, SAIC and Fordetc., as well as well-known Tier 1mainlandandDesay SVwait.
Founded eight years ago, how did Nullmax, starting from Silicon Valley, gradually establish its business model and win the trust of its partners?
The creation and development of Nullmax
Founder, CEO and CTO of NullmaxXu Leigraduated with a bachelor's degree in computer science from the University of Science and Technology of China, and later received a Ph.D. in computer science from the State University of New York at Farrow.
After graduation, Xu Lei worked at Qualcomm for three years and then joinedTesla,becomeCore member of Tesla Autopilot team,Report directly to Musk.
During his time at Tesla, Xu Lei led the construction of the Tesla Vision deep learning network from scratch, and used it to successfully replace the Mobleye vision system in the first-generation products. Later, the solution installed in the Model X was also developed under the leadership of Xu Lei.
In 2016, Xu Lei left Tesla and became a senior manager of Tesla's supply chain and product department.Song Xinyu,existSilicon ValleyThey jointly established Nullmax, which is one of the first companies in China to obtain a California autonomous driving road test license.
In early 2017, Nullmax’s headquarters was established in Shanghai.
Two years later, Nullmax quickly released its L3 autonomous driving solution.Max 1.0which enables high-speed designated driving and autonomous parking, and drives directly onto the Shanghai Central Ring Road.
However, after this press conference, Nullmax rarely spoke out to the public until four years later.
In the winter of 2023, when Nullmax returned to the spotlight again, it had already delivered products to vehicles and obtained780 million yuanof new financing, with a total amount of over1.066 billion yuan.
Founded eight years ago, Nullmax, which originated in Silicon Valley and flourished in Shanghai, has grown quietly in silence and has successfully implemented its business model.
In terms of technology, we continued to explore and, at a time when the industry was scrambling to shift gears, we once again held a technology press conference to speak out and propose a new paradigm.
It is not a simple matter of replacing the original rule module with the model, but some interesting explorations have been carried out, such as installing large multimodal models on the car and introducing brain science into autonomous driving.
After the press conference, Xu Lei also expressed his feelings, “It is impossible to find the optimal solution all at once only through continuous exploration.”
On the road of exploration, industry competition is inevitable.
Who are the players that create the most competitive pressure in the industry?
There is no doubt that it was the man who had worked for Xu Lei in Silicon Valley.
Facing competition, when talking about how he views Tesla's FSD entering China, Xu Lei made an interesting analogy, showing his carefree and expectant attitude:
Only high-level competitions can produce high-level players.
Only by introducing high-level solutions can the industry become stronger.