Autonomous driving is smarter, redundant sensing is not "redundant"

The driving culture of the "holy city" of Jerusalem is known for its wild and unrestrained nature. As the hustle and bustle of the city fades into the lights of the night, Jerusalem's roads are greeted by a confident and decisive self-driving car. The driverless vehicle autonomously completes the entire process of the ride-hailing taxi service, not only traveling to multiple destinations, but also automatically stopping where passengers need to be picked up and dropped off.

A few days ago, Intel subsidiary Mobileye released the latest self-driving video, this "smart" car equipped with a "truly redundant sensing system" is the protagonist. It is precisely because of the installation of this system that this self-driving car can successfully complete the ride-hailing service. Coincidentally, just after Mobileye released the video, the Jidu intelligent driving technology solution was exposed for the first time, and the pure vision and dual lidar solutions were fused on the car, and it is expected to achieve "true redundancy" after mass production. Previously, Xinna Sensing launched a new triple redundant sensor to escort the safety of automatic driving with accurate perception. At a time when Tesla's "pure vision" autopilot technology route is criticized, will redundant sensing be the future of autonomous driving?

The "eyes" of autonomous vehicles

Safely and successfully bypassing pedestrians who are jaywalking and making u-turns at intersections with multiple traffic lights... In the unedited video released by Mobileye, despite poor road lighting at night and complex road sign design, Mobileye's self-driving car successfully completed complex driving operations.

The system that gives Mobileye's self-driving cars "perception superpowers" is a truly redundant sensing system it is equipped with. According to Mobileye, the system consists of two separate subsystems, one of which uses a camera and the other uses a combination of lidar and radar.

The vehicle is equipped with a working sensor and also has a backup sensor system. The redundant sensing system that Mobileye calls is actually like people's two eyes. In an interview with the "China Electronics News" reporter, Gu Rongxiang, general manager of Xi'an Zhongxing Measurement and Control Co., Ltd., vividly compared it: people's left and right eyes can be assumed to be two image sensors. When one eye is obscured, the other eye can still distinguish objects normally, without affecting normal life.

Dong Bo, the algorithm engineer in charge of Xinna sensing, took the triple redundant IMU (inertial measurement unit) module as an example to further explain to the reporter of China Electronics News how the redundant sensing system plays a role in the field of automatic driving. Traditionally, he said, an IMU module has consisted of a three-axis accelerometer and a three-axis gyroscope. When any axis of the accelerometer, or any axis of the gyroscope, fails, the entire IMU module fails. The triple redundant IMU module of Xinna Sensing integrates three triaxial accelerometers and three triaxial gyroscopes, and the entire IMU module will only fail if more than one triaxial accelerometer or triaxial gyroscope fails.

In order to greatly improve the safety of high-end autonomous driving, redundant sensing systems are indispensable. Xu Tianyu, senior analyst of ccid consulting Internet of Things Industry Research Center, told China Electronics News that in the field of environmental perception of automatic driving, ultrasonic radar, millimeter wave radar, cameras, and lidar will form at least double 360 ° perception capabilities, that is, at least two sets of sensor systems are guaranteed, and each set can achieve 360 ° no dead angle detection around the car. In the event of a system failure, the other can automatically switch and function immediately, ensuring driving safety.

In general, video sensors and millimeter-wave radar sensors in autonomous vehicles must be redundantly designed for safety reasons, and multiple sensor solutions must be considered. Gu Rongxiang explained that this is because the road surface and driving environment have complex and changeable characteristics, and the driving state of the car itself is also prone to sudden changes, so the front end of the automatic driving system needs to "see six roads and listen to the eight directions".

The "promotion test" that cannot be bypassed

At present, Mobileye's self-driving taxi equipped with a "truly redundant sensing system" is planned to be put into small-scale commercial use later this year; Jidu automotive robots continue to move forward on the road of high-level automatic driving by adding redundant dual systems and dual lidar; previously, Xinna Sensing launched a new triple redundant sensor to escort the safety of automatic driving.

Redundant sensing systems are "getting on the car", but before installing redundant sensors in autonomous vehicles, it is not easy to "find" and "locate" redundant sensors. Gu Rongxiang said that when installing redundant video sensors, it is first necessary to comprehensively consider various factors such as direction, clarity, night driving, wind, rain, sandstorms, etc., and then determine the installation location of the sensor; when installing redundant millimeter-wave radar sensors, multiple factors such as direction, radar perception distance and resolution should be considered.

The successful installation of redundant sensing systems on vehicles is not the "end point" of the autonomous driving business. To achieve fully autonomous driving, redundant sensing systems must also go to market and achieve large-scale promotion and application. However, what needs to be seen is that due to the complexity of technology and the high cost, redundant sensing systems are facing a "promotion test" that cannot be bypassed.

The first is the test of technology. Xu Tianyu said that in terms of sensor miniaturization, only the size of the sensor is reduced to ensure the installation of redundant sensor systems. For sensor microdirection, MEMS technology can be used to integrate microsensors, micro-actuators, micro-mechanical structures, micro-power sources, micro-energy, interfaces, etc. on a chip, which greatly reduces the size of the sensor.

In terms of redundant system switching, it is necessary to consider how to make fault judgment, how to achieve no human intervention in the switching process, no interruption in the main and standby switching, and no user perception during the switching process. Xu Tianyu said that this requires the use of edge computing and edge controllers to achieve.

The second is the test from cost. Lidar costs nearly $1,000 per unit, and vehicle costs will increase significantly after considering redundant designs. In this regard, Xu Tianyu said that if in order to reduce the cost of vehicles, semiconductor manufacturing methods (MEMS, etc.) can also be used to mass manufacture to reduce the cost of a single sensor.

In Gu Rongxiang's view, sensors are related to driving safety, so they must take into account factors such as accuracy; sensors must have a very high level of safety, and they need to comprehensively and comprehensively consider external environmental factors, road environmental factors and the condition of the vehicle itself, and accurately perceive them in a complex and changeable internal and external environment. This is undoubtedly a very big technical challenge for sensors. In addition to improving performance parameters such as precision and intelligence, the reliability and environmental adaptability of the sensing system is also very important.

Sensor systems are moving in the direction of vehicle-to-road synergy

Going forward, is redundant sensing the only way to achieve autonomous driving? Dong Bo told the "China Electronics News" reporter that multi-sensor fusion is a necessary condition, but whether redundant sensing is necessary depends on whether a single (non-redundant) IMU can meet the requirements of reliability and the design of the entire automatic driving system.

There are differences between multi-sensor fusion and redundant sensing. Dong Bo explained to reporters that multi-sensor fusion is usually to achieve functions that cannot be achieved by a single sensor, or to overcome the inherent shortcomings of a single sensor; redundant sensing means that when one or more sensors fail, the entire system can still work normally, and the performance indicators are not affected.

In any case, in order to achieve true autonomous driving, "smart" cars also need to develop in tandem with "smart" roads. At present, the increasing safety requirements for autonomous driving are driving the continuous development of sensor systems in the direction of vehicle-road coordination.

Why is vehicle-to-road collaboration an inevitable trend in the development of sensor systems? In fact, this is because on-board sensors cannot fully and accurately perceive the road situation, and the installation of perception sensors on both sides of the road can provide a "God perspective" for autonomous vehicles, complementing the on-board sensor information.

Xu Tianyu further explained to the "China Electronics News" reporter that the perception range of the on-board sensor is actually more limited. Taking millimeter-wave radar and lidar as an example, the maximum detection distance of the former is 200 meters, and the maximum detection distance of the latter is 150 meters. In the networked state of multiple subgrade sensors, their detection distance can theoretically reach infinity, which can effectively expand the perception range of autonomous vehicles.

In addition, on-board sensors can only detect the environment and cannot detect themselves, while the application of subgrade sensors can effectively supplement the on-board sensor information.

The organic synergy of environmental sensors, road surface sensors, and on-board sensors can provide more reliable connectivity and accurate judgment for autonomous vehicles. Gu Rongxiang told reporters that the installation of corresponding sensors on both sides of the road can better solve some of the problems of automatic driving perception, but this is only one aspect of solving the coordination between the car and the environment, and cannot completely solve the problem of automatic driving safety and reliability.

Achieving safe and reliable autonomous driving cannot rely solely on redundant sensors, but must also be accurate and scientific planning on systems and algorithms.

In Gu Rongxiang's view, to truly solve the problem of automatic driving, it is necessary to achieve the following four points: first, with a comprehensive, real and reliable perception, which includes the perception of the car itself, the perception of the road conditions, and the perception of the surrounding environment; second, it has millisecond-level information transmission in the global range; third, it has intelligent computing similar to the human brain; fourth, millisecond-level accuracy and decisive execution control.