Identification Robot

In the midst of the capital winter, five rounds of financing were completed in just over a year;

The L2 perception solution based on Journey 5 was launched in less than 2 months, and the intelligent driving system based on Journey 5 was launched in less than 5 months;

In less than 3 months, we successfully obtained OEM designation and delivered more than 500,000 smart driving products in the next 2 years

In the intelligent driving track where fast fish eats slow fish, intelligence robots are like cheating. In two years, they have gone through the path that many companies have completed in five years, and with the autonomous driving paradigm and differentiation based on strong AI The company's binocular stereo vision products have become a new star company with great identity and competitiveness in the smart driving market.

But behind the "Shuangwen" general entrepreneurial journey, there is actually no unique secret recipe, just because of correct business strategies, leading technology construction, and like-minded partners and teams.

Common ideas, get on the road quickly

In August 2021, a group of people with the dream of universal robots jointly founded the Intelligence Robot Company. They chose to start with intelligent driving and first enter the field with the strongest market demand and the largest market space in the next 10 years

However, at this time, the domestic smart driving industry is at a turning pointfrom blue ocean to red ocean. Some autonomous driving companies established as early as four or five years ago have begun to enter the high-end intelligent driving pre-assembled mass production market as suppliers and seize the opportunity. This means that for Jianzhi, which has just been established, it must do better what others can do; it must also do what others cannot do; and what other companies have completed in three years must be completed in a few months.

Advanced intelligent driving needs to cope with complex and changeable road environments. It not only requires powerful sensor and chip computing power support, but also requires efficient and iterative algorithm models to make full use of hardware performance and achieve continuous upgrades of intelligent driving system capabilities. Get the best balance between performance and cost

Therefore, software and hardware collaboration is the key to breaking the deadlock and the best way to achieve efficient mass production deployment

However, at almost the same time as Jianzhi was established, Horizon released Journey 5, bringing the first domestically produced 100 TOPS high-performance, large-computing power car-grade smart driving chip to the smart driving industry. At the same time, Horizon has always been a staunch practitioner of software and hardware collaboration, building an efficient open technology platform with "chip tool chain" as the core and a complete and mature development environment. It is equivalent to providing developers with one-stop development support from hardware reference designs, tool chains, basic middleware to rich reference algorithms.

Thus, an unexpected rush to each other kicked off

. Jianzhi has established a close cooperative relationship with Horizon within two months of its establishment, and has received all-round support from research and development to commercial implementation. It is equivalent to

starting from the very beginning, on the "foundation" laid by the horizon, and quickly "getting on the road". In order to achieve efficient collaboration, we need to propose a plan within two months

In April 2022, Jianzhi became one of the first companies to obtain Journey 5. Using Horizon's tool chain, the team quickly began the quantitative deployment of the model

Quantitative deployment is the first step after migrating the model to a new platform. The goal is to use lower computational precision to represent the computational behavior of the algorithm model while ensuring the prediction effect of the original model, thereby reducing the memory and computing resource requirements during algorithm operation, and ultimately achieving efficient operation on embedded devices.

Li Bo, vice president and head of software technology of Jizhi Robot, is one of the persons in charge of this work. As early as 2007, when he was a graduate student, he came into contact with general parallel computing and has now become a senior technical expert in the industry. At the same time, he was also fortunate to witness the entire process of the Journey series from 0 to 1.

The entire development environment, both in terms of friendliness and ease of use, has made a qualitative leap and can be used out of the box

. "Whether it's a new employee or an old employee, they can quickly get started after getting it."

Regarding the self-developed dynamic obstacle detection multi-task model, the team’s expectation is that after quantification, while maintaining the functional correctness of the model

, the performance

will reach 10~ 20 FPS, Accuracy reaches a level comparable to the original model. Function correctness verification Usually it is necessary to manually export the data of the quantitative model and the original model, and then compare them separately, which is very cumbersome. With the help of the Verifier tool of the tool chain, the consistency of the results of the two models can be directly verified after the model conversion, and the entire process is highly automated.

FPS is a key indicator of model performance. Higher FPS means faster perception and lower delay. Compared with theoretical computing power, it can better reflect the safety and security of the intelligent driving system. Driving efficiency is an important criterion for measuring the real computing performance of a chip. Previously, the team used a computing platform with the same computing power as Journey 5, but the actual FPS did not reach the expected 10~20 FPS. However, on Journey 5, it finally achieved an unexpected 30 FPS. The team also uses the pyramid and other modules provided by Horizon to enable some processing processes that originally relied on the CPU to achieve hardware acceleration through Journey 5’s heterogeneous computing system.

Maintaining quantization accuracy is quite challenging. Initially, some obstacle detection accuracy was excellent, while others left gaps. "The training of multi-task models is inherently difficult, because the data distribution and task characteristics between tasks will interact with each other. If more data is added to one task, the results of the other task may lose accuracy. ”

In order to solve the accuracy problem, the teams of both Jianzhi and Horizon conducted multiple rounds of intensive discussions. At this time, the tool chain of Journey 5 was also rapidly iterating, and the QAT (quantification in training) solution was soon provided. "QAT not only provides quantitative training capabilities, but also provides a wealth of debugging tools and tuning suggestions. When faced with accuracy loss problems, you can quickly and directly locate the problem through layer-by-layer comparison. Combined with horizon tuning suggestions, it can be solved very efficiently. question."

"These tools, like bridges and assistants, can help us locate performance bottlenecks in the software system more quickly and make targeted adjustments." Li Bo said, " Horizon has put a lot of effort into supporting us in making good use of Journey 5. We were able to quickly transplant our past work, largely due to the close cooperation between and both parties and tool chain maturity.”

Jianzhi’s development team is very impressed with the depth of this support. As a solution provider, Jianzhihui is exposed to many different platforms. However, compared with companies on some other platforms, sometimes emails sent by the team do not get a reply for three or four months. This is in stark contrast to Horizon

Efficient and intensive development collaboration allowed the Jianzhi team to release the L2 autonomous driving perception system solution based on Journey 5 in less than 2 months. In early August of the same year, Jianzhi further launched an L2 autonomous driving mass production solution based on the Single Journey 5, becoming the first batch of companies to implement autonomous driving system solutions based on the Journey 5.

Endless optimization, achieving multiple firsts

It’s time to test the effectiveness of the plan made in 4 months. In August 2022, during a media event, Horizon gathered a circle of friends to discuss the innovative paradigm of intelligent driving. The intelligence robot participated, and the media was also invited to experience the ride.

On the highway with heavy traffic, the Jianzhi vehicle's smooth traffic intersection assistance, automatic lane change with lever, automatic overtaking and lane change, etc. have still won praise from the media. In September, Horizon founder and CEO Yu Kai also personally experienced the Jianzhi solution and was quite impressed with its effectiveness.

The intelligence team did not stop there. In the following months, they devoted themselves to promoting the deployment practice of

BEV perception

and exploring the upper limit of Journey 5's capabilities. BEV is a new paradigm of intelligent driving perception technology. As early as 2021, Jianzhi proposed the self-developed

BEVDet

paradigm, whose inference speed can reach 4 to 15 times that of similar algorithms. However, BEV requires the support of a lot of computing power. In order to make full use of the BPU performance of Journey 5, this team "selected the best" from a large number of basic reference algorithms in the Horizon Toolchain. After some attempts, we referenced Horizon's self-developed VargNet model structure for optimization, which significantly improved network performance, reduced memory access time, and achieved computational compression, while ensuring operating efficiency and task performance. Next, the rapid migration and secondary development of its own algorithms were completed. In the end, BEVDet achieved high-timeliness BEV perception on Journey 5, and the model operating efficiency reached 50 FPS, becoming a strong guarantee to ensure mass production performance.

After fixing the BEV perception, the Jianzhi R&D team also carried out in-depth optimization of forecasting

and

planning to achieve the ultimate NOA experience. However, it is extremely challenging to utilize the computing power of BPU for prediction and planning tasks. After many ideological collisions, the two parties established the overall optimization idea of ​​" increasing revenue and reducing expenditure

", making full use of all computing power as much as possible, such as replacing intent classification operations that are suitable for AI with learning-based methods. rule-based method to deploy it to the BPU. Finally, Jianzhi handed over the

prediction performance optimization 5 times and the planning performance optimization 5 times answers. The team also made full use of other heterogeneous resources in Journey 5, and transplanted some algorithms that originally ran on the CPU to the DSP

. Although DSP programming is more difficult than CPU and requires certain accumulation, with the support of the Horizon team, heterogeneous computing can still be achieved, further improving computing power utilization.

In this regard, Jianzhi’s R&D staff also recalled, “As early as the Horizon first-generation chip launch conference, Horizon co-founder and CTO Dr. Huang Chang said, Through the combination of software and hardware optimization, continuous optimization can be achieved To fully tap the potential of the chip, this is actually the core of the Horizon chip that can efficiently support various advanced high-level intelligent driving algorithms."

These software and hardware collaborative optimization and engineering practices have helped Jianzhi set many "firsts".

Between the end of June and July this year, Jianzhi successively launched the first fully self-developed standard sensing product PhiVision based on a single journey 5 within the horizon ecosystem, andFirst High-speed NOA system solution based on Zhengcheng 5 TC397PhiGo Pro. This solution does not require an external CPU, reducing the overall cost of high-speed NOA to less than 3,000 yuan, and has a strong cost-effective advantage.

[Intelligence Robot releases PhiVision, the first fully self-developed standard perception product in the Horizon ecosystem based on a single Journey 5 (from left: Horizon Chief Ecological Officer Xu Jian, Inspection Robot co-founder and CTO Du Dalong, Tian Liu Junchuan, Vice President of Quasi-Technology, Liang Zhujin, Vice President of Intelligence Robot and Head of Sensing Product Line)]

[Intelligence Robot invites industry leaders to discuss the value of standard sensing packages: (Wu Wenguang, CEO of Meixing Technology/President of Meixing (Shanghai), Liu Junchuan, Vice President of Tianzhun Technology, Zhang Hongzhi, Vice President of Horizon and Smart Car Business Head of business development of the department, co-founder & CTO of Du Dalong Jianzhi Robot, Vice President of Liang Zhujin Jianzhi Robot & head of sensory product line]

With its brand-new solution, Jianzhi has obtained several leading OEM appointments.

"The fact that we can come up with these solutions also proves that Journey 5 has a lot of potential that can be tapped. Through collaborative optimization of software, hardware, and algorithms, high-speed NOA can achieve very good results." Dr. Du Dalong, CTO of Intelligent Robot, mentioned, "Optimization is never-ending. The key is to discuss and collide with each other to simplify complex problems and solve them together with the help of everyone's strength." ”

A group of people walking can make towering trees grow

Years of experience in the industry have made Jianzhi’s core team well aware of whether a set of chip platforms can become mainstream. Hardware is the foundation, and software tool chains and ecology are the key to growing towering trees.

Powerful hardware architecture can attract developers. Of course, we cannot simply pursue peak computing power, but also increase effective computing power to effectively meet the computing power needs of specific applications.

Easy-to-use software tool chains can retain developers. Simple and easy to use can lower the learning threshold, and good scalability can quickly adapt to changing development needs and fully mobilize the enthusiasm of developers.

A healthy ecosystem can mutually benefit and help developers. Really go into enterprises and campuses, and while answering difficult questions for developers and partners, you can also listen to their opinions and feedback, and continuously feed back the iteration and optimization of software and hardware.

In the eyes of the core team of Jianzhi, Horizon, as the new generation leader of China's AI chips, has devoted itself to building and polishing it bit by bit with developers and partners. In just a few years, whether it is The maturity of the hardware architecture, the friendliness of the tool chain, and the prosperity of the ecosystem have all reached the top level in the industry.

This experience of cooperation and co-creation is an indispensable and important process, whether it is for product iteration and ecological construction, or for making plans and deciding on a fixed point. In the domestic smart driving circle, everyone wants to run faster than others. As a supplier, Jianzhi can run faster and think further than the OEM. Although you will undergo tremendous pressure and tests, as long as you have the support of your partners and team, you can win every battle.

The speed of intelligence appraisal is the result of mutual support and mutual trust with Horizon, just like two partners who have the same path, connected hearts, and united strength. Horizon's insistence on implementing the open and collaborative ecological cooperation model can be said to have created a more cohesive soil environment for the development of China's smart driving industry. Nowadays, Intelligent Robot has established R&D cooperation with overseas car company customers, and the "In China For Global" strategy has taken solid steps. As representative companies in the rise of China's hard technology, Intelligent Robots and Horizon will further deepen their cooperation and gain more results in the international market.

"A person can go very fast, but a group of people can go further. The power of partners and teams is to go longer in the market. Under the wave of fierce competition in the industry, An important asset that will stand firm and lead to success.

”

The above is the detailed content of Identification Robot. For more information, please follow other related articles on the PHP Chinese website!