The content of this article is sourced from 42how and has been translated by ChatGPT
At the beginning of the year, Great Wall Motor announced its 2021 sales performance: the annual sales exceeded 1.28 million vehicles, a year-on-year increase of 15.2%. This is the sixth year that Great Wall Motor has achieved sales of over one million vehicles, and it is also the best year in its history.
It is worth noting that most of the models currently sold by Great Wall Motor are equipped with L2 level intelligent driving assistance. As an enterprise with a wide range of product lines, whether it is Haval, Wey, pure electric brand Euler, or even the tough off-road Tank series and Great Wall Motor’s cannon, all have realized the application coverage of car networking. It can be seen that Great Wall Motor does not intend to sit idle in the wave of automobile intelligence.
In fact, Great Wall Motor has been laying out in the field of intelligent ecology for a long time, and 2021 has achieved frequent results in the field of intelligence: upgrading “Coffee Intelligent 2.0”, including a new electronic and electrical architecture, the third-generation intelligent driving calculation platform IDC 3.0, NOH smart navigation assisted driving system, self-developed cockpit system GC-OS and intelligent line control chassis and other advanced technologies.
In June 2021, Great Wall Motor announced that it will achieve global sales of 4 million vehicles by 2025, of which 80% will be new energy vehicles. If the transformation of energy form is the “main theme” of change, Great Wall Motor’s “Coffee Intelligence” is undoubtedly the first movement of change.
From Perception Intelligence to Cognition Intelligence
In 2019, Tesla launched the “Navigate on Autopilot” function, which realizes automatic entering and exiting the ramp, automatic overtaking, and automatic speed change according to the road speed limit after setting the destination navigation on the car map. From 2020 to 2021, NIO, XPeng and Ideal have successively followed up and launched this function. Since then, in everyone’s mind, “Navigate on Autopilot” has become a label of technology travel companies.
In November 2021, Great Wall Motor officially pushed the NOH smart navigation assisted driving function to Wey Coffee users. From the perspective of user experience, this system can provide point-to-point navigation and assisted driving capabilities in the scene of high-precision map coverage, breaking the global pattern of the “Big-Four” (Tesla, Rivian, XPeng, and NIO) dominating the Navigate on Autopilot function.
A month later, Great Wall Motor also demonstrated the urban NOH assisted driving system for Wey Coffee in a video. During the 34-kilometer test process, the test vehicle passed through 24 intersections, 27 pedestrian crossings, 22 traffic lights, 5 unprotected pedestrian crossings and two roundabouts. Throughout the entire process, the test vehicle equipped with the urban NOH system traveled a total of 11 kilometers without manual takeover.
So, how did Great Wall Motor achieve rapid research and development of its intelligent driving in a short period of time?The level of perception ability determines the lower limit of intelligent driving and is also the foundation of data collection. The core perception devices of Great Wall Motors’ intelligent driving are the lidar and cameras.
The cameras around the vehicle can achieve 360-degree coverage. The data collected by the cameras is first processed through a Resnet network to calculate the basic data, and then two branches are generated. One branch can track and extract the objectives with multi-layer features, including recognition of lane lines, stop lines, road edges, vehicles, and traffic lights. The other branch is used for generating the system’s Free Space and scene recognition.
For the point cloud generated by the lidar, HoloMation uses the PointPillar algorithm, which can greatly improve the calculation speed. The system first reduces the three-dimensional point cloud data to pseudo-two-dimensionality and then performs subsequent calculations using the general method.
The level of cognitive ability determines the upper limit of the actual performance of autonomous driving. From the macroscopic perspective of expressing driving behaviors in specific scenarios, there are several influencing factors: weather, road structure, traffic participants, traffic flow density, the main route of the leading vehicle, collision risk and collision time-to-impact. HoloMation mines and expresses these attributes from existing data and then performs clustering and classification to facilitate accurate identification of these special scenarios by the system.
At the microcosmic level of cognition, Great Wall Motors has refined the control actions of the vehicle. Taking the most basic start-stop action as an example, Great Wall Motors has detailed this process into four stages: steady-state following, front vehicle deceleration, front vehicle stop, and front vehicle start.
After the macroscopic scenarios and microscopic actions are digitized, Great Wall Motors annotates them specifically, allowing models to continue training in different scenarios. The vehicle-side execution layer takes the collected cases as guidance, simulates and learns the specific actions of each vehicle in different scenarios, namely end-to-end simulation learning.
Dual Platforms: Vehicle-side & Cloud-side
With the increase in the number of vehicle-side sensors, such as cameras and lidars, the amount of data collected has skyrocketed, posing a severe challenge to the computing power of onboard intelligent driving chips.
To address this, Great Wall Motors has launched the IDC 3.0 intelligent driving computing platform. The platform uses a combination of high-capacity SA8540P + SA9000P from Qualcomm, with a single-board computing power of up to 360T and can be upgraded to 1440T through board-to-board interconnects, making it currently the highest energy efficiency production-capable computing platform worldwide.As a high-performance computing platform for vehicle-end, IDC 3.0 can effectively support AI vision large-scale model computation, as well as the screening, cleaning, desensitization, and backflow of vehicle-end sensing data, greatly improving data recognition accuracy.
IDC 3.0 adopts the currently highest-speed PCIE 4.0 interface, which can effectively support high-speed bidirectional backflow of data between CUP and NSP, reducing system delays and response time.
In addition, IDC 3.0 supports 6 channels of gigabit Ethernet, with a board-to-board data transmission capacity of up to 6 Gbps. It can simultaneously connect up to 14 high-definition cameras with up to 8 million pixels, 8 high-resolution millimeter-wave radars, and 6 solid-state lidars. Such performance has also left room for future advanced L4\L5 all-scenario intelligent driving functions of Great Wall Motors.
Although the computing platform performance of the vehicle-end is strong, a small model of a vehicle-end generally only handles some perception tasks. Great Wall Motors has accumulated 4 million kilometers of actual road test mileage by installing intelligent driving systems on Wei Pai Mocha. Such massive data requires cloud computing platforms to solve the large amount of computation needed. Currently, Great Wall Motors’ own supercomputing center is also under construction, and the computing capability of the cloud can be improved in the future.
Great Wall Motors achieves full-task perception through a Transformer-based large-scale model Fundamental Model, which can identify recognition errors that occur in small models of vehicle-end in harsh weather conditions or missed detections, correct them, and then return them to the vehicle-end model for retraining, significantly improving data capture accuracy and model training speed.
The “cognitive intelligence” of Great Wall Motors’ intelligent vehicles is also reflected in the testing methods of data. The iteration speed of AI models is very fast, and the most effective way is to test their effectiveness in a simulated environment. However, the traditional simulation test efficiency is very low, and it is time-consuming and laborious from scene building, model setting to simulation testing, often allowing each person to only build 30 test environments per day.Longwei Automotive has developed a set of semantic scenario automated transformation tools and parameter generalization tools in CSS that can automatically transform scenario library description text into simulation test scenarios. Over 10,000 simulation test cases can be generated each day. Longwei has accumulated hundreds of thousands of typical intelligent driving scenarios, most of which are corner cases, to effectively verify and iterate performance in complex intelligent driving scenarios.
Longwei Automotive’s Director of Intelligent Driving, Zhen Longbao, pointed out that the core competitiveness of autonomous driving lies in two aspects: AI iteration cost and AI iteration speed.
CEO of HaoMo Intelligent Driving, Gu Weihao, believes that the reason why Tesla’s assisted driving can achieve today’s achievements relying solely on visual solutions is largely due to the closed-loop data. Furthermore, Tesla has actual mileage data of over 1 billion kilometers.
When it comes to the rapid iteration capabilities of Longwei Automotive’s intelligent driving, the most important thing is to quickly deploy coverage on a large scale. Objectively speaking, the many million-level annual sales models under Longwei have paved the way for its large-scale production of intelligent driving.
Safety is Essential
If technical solutions are the main theme of intelligent driving, then safety is its fundamental strength.
Longwei Automotive provides six-fold redundancy for intelligent driving, including steering redundancy, braking redundancy, perception redundancy, control redundancy, architecture redundancy, and power redundancy.
Through the design of double-winding steering motor, dual power supply, dual sensing, and dual communication, the intelligent driving system of Longwei Automotive can ensure that the minimum steering assistance under all scenarios is 50\%. Dual braking redundancy is achieved through ESP+IBooster design.
For perception, multiple heterogeneous sensor solutions are used. The design of the front-end perception alone integrates 3 LiDARs, 3 millimetre-wave radars, and 2 cameras, significantly improving safety compared to traditional fusion methods. In addition to the main chip of Qualcomm SA8540P+SA9000P, the safety redundancy chip is the Infineon TC397, which ensures that the controller takes over at any time in case of unexpected situations.
# Intelligent Driving System of Great Wall Motors
The intelligent driving system of Great Wall Motors ensures normal operation within 5 minutes after the main power supply fails and the switching between the main and auxiliary power supply only takes 500 microseconds, which is enough for the vehicle to enter a safe area or for the driver to take over. Currently, the law requires that the L3 level intelligent driving system can operate normally for only 10 seconds after the power supply fails. The safety redundancy of this system is indeed well-considered.
Future Prospects of the Intelligent Line Control Chassis
On June 29, 2021, Great Wall Motors globally launched the line control chassis technology with full independent intellectual property rights, which will bring about what kind of imagination for intelligent driving?
The line control chassis no longer requires a traditional steering transmission shaft between the steering wheel and the steering column, which can expand the seating space. At the same time, there is no traditional hard connection, and the steering mechanism is completely isolated from the vibration of the road, which is beneficial for NVH control and can simulate a richer steering feel.
As the line control chassis decouples people from the vehicle, in the future, through the cooperation of advanced intelligent driving, the central processing unit can independently coordinate and control the actions of various systems to achieve the overall level of autonomous coordination and control of the vehicle.
Expand the imagination, in the future, through signal control of steering, shifting, throttle, brake and suspension, and even at advanced levels of intelligent driving, the steering wheel can be retracted, and seats can be freely assembled to make the vehicle truly a “second living space”.
At the same time, Great Wall Motors has set up a safety backup for the line control steering and line control brake systems related to driving safety, keeping the overall failure rate of the system controlled at one in 10 billion per hour. In the extreme scenarios of steering failure, the turning movement can be completed by applying braking force to a single side wheel. In the extreme scenarios of braking failure, increasing the kinetic energy recovery force can generate a large drag force to reduce the vehicle speed.
When it comes to intelligent driving, people often associate it with new energy vehicle manufacturing companies such as Tesla, NIO, and XPeng. Any movement they make in this field will immediately be put under the spotlight, but the investment of Great Wall Motors in intelligent driving technology and the attention it receives is not proportional.
What is not known to most people is that Great Wall Motors has been researching and developing autonomous driving technology since 2009. From 2014 to 2016, the Great Wall Motors team won the China Intelligent Car Future Challenge for three consecutive years. At the 2015 Great Wall Motor Technology Festival, Great Wall Motors demonstrated its L3 level autonomous driving and the Haval H9 equipped with ADAS functions was launched in the same year.
Having accumulated over 10 years, the Great Wall Motors (GWM) intelligent research and development (R&D) team has expanded to 2,000 people in 2021. In just over two years from 2019 to now, GWM’s self-developed intelligent driving products have been mounted on cars at an astonishing speed, accumulating 4 million kilometers of actual road test mileage.
At the juncture of the automobile intelligence revolution, mastering the self-research ability means that GWM does not have to be “strangled” by the upstream supply chain, and also means that it has absolute voice in the relevant fields. In the long term, the full-stack self-research brings undeniable systemic advantages. However, self-research is never effective from the beginning. Before achieving scaled and commercialized effects, the road of technology iteration and improvement is arduous and long. There are not many traditional car companies that are willing to make massive investments and put them into practice.
GWM’s approach is also different from that of L4/L5 self-driving development companies that have been slow to land, and all of its research and development serve mass production. In 2021, GWM’s vehicles with effective data collection capabilities reached 250,000 units across all models. In this “data-driven” era, whoever obtains more road test data means faster iteration speed and more mature actual performance of intelligent driving products.
Currently, GWM has installed intelligent driving systems on five vehicle models, including WEY, Tank, and Haval. In 2022, this number will increase to 34 models with an installation volume of over 300,000 units, accounting for about 80\% of the overall listed models. The progress of hundreds of thousands of units each year also gives GWM’s intelligent driving iteration speed an advantage.
We are always generous with our cheers and attention for new things, but we sometimes overlook the enormous changes that occur to familiar roles. While everyone’s attention is focused on the intelligent driving process of new players in the vehicle manufacturing industry, GWM’s intelligent R&D is quietly advancing.