In the economic downturn, how can intelligent driving break out of the siege?

In the field of intelligent driving, "integrated driving and parking" seems to have become a popular practice among major suppliers this year.

What is "all-in-one travel and parking"?

Driving and parking were originally two systems on the same car. Driving means driving, and parking means parking. The so-called "driving and parking integration" means that driving and parking share a domain controller, realizing deep multiplexing of sensors and sharing of computing resources. The direct benefit of this is that R&D integration will greatly reduce costs, which will ultimately allow mid- to low-end models to achieve high-end intelligent driving functions while still remaining affordable.

However, at present, the level of "all-in-one travel and parking" solutions provided on the market is uneven.

The "all-in-one travel and parking" displayed by most suppliers is somewhat fragmented. It seems that the hardware is made of rows and berths integrated into a controller box, but in fact there is no chemical reaction between rows and berths. It is a pseudo-row and berth integrated architecture. The computing resources are not truly shared, and communication delays are prone to occur. Conducive to data fusion processing, the intelligent driving performance of the car has not been improved. Some are unable to fully process the data from each sensor due to limited computing power. In addition, the "all-in-one travel and parking" solutions provided by some suppliers cannot iterate and customize new usage scenarios later.

These are all problems that need to be solved urgently with "all-in-one travel and parking" products.

“There are a lot of algorithm optimizations, including bottom-level optimization, hardware-level optimization, etc. Only when these tasks are truly solid can we really squeeze the computing power harder, so that the integrated travel and parking can not only ensure safety, but also To achieve good computing power on medium and low computing power platforms," ​​Guo Yang, chief product architect of Baidu's intelligent driving business group, said at the ANP2.0 communication meeting.

At present, ANP2.0 launched by Baidu has also entered the integrated travel and parking market. Backed by the data and scenarios collected by Baidu’s Robotaxi autonomous driving for ten years, including data on highways, as well as Baidu’s huge funds and mature smart driving team, it seems to be a bit of a dimensionality reduction attack.

How to achieve high power with low computing power

ANP2.0 is Baidu's all-in-one smart driving solution for parking and parking. It is currently in mass production and is mainly focusing on parking lots, highways and urban loop scenarios, and can achieve point-to-point autonomous driving.

How does Baidu lay out its hardware? First of all, ANP's hardware sensors and ultrasonic radars share a relationship with AVP's hardware sensors. The computing unit ACU also shares the hardware part. In terms of sensing hardware, 9 visual sensor cameras, 5 millimeter wave radars, and 12 ultrasonic radars are used. Coupled with GPS+IMU+steering wheel angle signal, wheel speed signal, etc. to achieve inertial navigation, it finally achieves all-round perception of driving and parking with high cost performance.

In terms of computing power, dual TDA4 chips are used, which refers to two TDA4VMs from TI (Texas Instruments). The single-chip AI computing power is 8Tops, and the two chips are 16Tops. The dual TDA4 platform achieves 360° omnidirectional perception of the entire scene.

"The cost of a single TDA4 is about US$50, and the 2.0 version comes with two, which adds up to about US$100. Baidu is the only current supplier that can use such a low-computing power platform to achieve high-end integrated travel and parking capabilities." Guo Yang said.

In fact, other suppliers using similar chip solutions include DJI, which also uses two 8Tops TDA4, but its hardware layout is a forward-looking binocular camera, four surround-view fisheye cameras, and a forward-facing millimeter wave radar. and 12 ultrasonic radars. This solution is installed on Wuling Baojun’s 2023 KiWi EV for the first time.

“In comparison, although the chip computing power is the same, due to the different hardware layout, DJI is lacking in high-end intelligent driving functions. It has neither high-speed piloting nor high-end functions such as memory parking and valet parking. function." A user who has driven KiWi EV told Auto Business Review.

Excellent driving experience

Baidu organized an ANP2.0 experience event this time, and the set route was the expressway from Apollo Park in Yizhuang to Daxing Airport, a total distance of 75 kilometers.

For the author who has tested the NGP and cross-level memory parking of Xpeng P7, I am amazed by Baidu's performance on highways and parking lots.

When you reach the highway, double-click to turn on the assisted driving function, and the central control interface becomes a human-machine co-driving map. It is said that this is an interface unique to Amap and Baidu Maps. After the vehicle takes over, the vehicle drives smoothly and switches to the expressway on its own. If a vehicle enters from the left, the red area will be displayed on the map until it becomes a safe distance. When it is displayed in blue, the vehicle will automatically cut in. The whole process is smooth and there is no shaking. . The driver can also direct the vehicle to change lanes by turning on the turn signal without exiting the smart driving mode.

ANP2.0 is set to change lanes 2 kilometers away from the ramp. In actual experience, it is true that vehicles begin to change lanes to the right within this range. During this process, when a car temporarily joins or there is a car in the side lane, Or in solid line situations, the vehicle responded very well. After reaching the ramp, the vehicle will slow down on its own. In the case of multiple intersections, because of the high-precision map, the vehicle can enter the correct fork in the road and make smooth turns.

The speed on the highway and on the ramp can be adjusted manually without affecting the assisted driving mode. When exiting the assisted driving mode, you can return to the normal state by turning the steering wheel or clicking the start button. The whole process felt smooth.

According to Baidu, ANP2.0 has consistent performance even in harsh environments such as rainy days and stagnant water. At the same time, more than 300,000 kilometers of generalization testing further ensures the stability of ANP2.0.

Another one is memory parking. This function cannot be regarded as Baidu’s original creation. Many manufacturers such as Xiaopeng and Tesla have implemented it long ago. But the difference is that Baidu ANP2.0 can accurately park into a mechanical parking space in a very short time, and the live demonstration does not exceed 30 seconds. The difficulty of parking is that the car can be parked into a narrow parking space by visually identifying only two iron chains. In contrast, Xpeng does not seem to have demonstrated this feature.

In general, ANP2.0 provides excellent parking and parking services at a very competitive cost.

The logic behind high cost performance

"Among products with the same performance, ANP2.0 has a lower cost." Guo Yang revealed.

The current general consensus in the industry is that the integrated travel and parking system is on the eve of large-scale mass production, gradually becoming a standard feature of new models in the Chinese market, and the price of equipped models has dropped from the previous high-end to the price range of 100,000 yuan to 250,000 yuan. , and the core of competition lies in cost performance.

It is understood that most of the models currently targeted by Baidu ANP2.0 are in the range of 150,000 yuan to 250,000 yuan. For next-generation platforms, hardware costs will be significantly reduced over time.

As a major Internet company, Baidu's unique advantage in smart driving is that ANP2.0 shares a data center with its RoboTaxi model, which reduces the product's dependence on hardware computing power. Through the ultimate optimization of computing power and CUP, ANP2.0 can achieve high-end parking integration on the TDA4 low computing power platform, with the characteristics of high computing power density and higher technical capabilities per unit of computing power. In addition, ANP2.0 can also achieve omnidirectional perception, making vehicle planning and decision-making safer and more autonomous. This results in the lower cost of ANP2.0 among products with the same performance.

“We believe that in the future, as computing power continues to increase and sensor prices continue to drop, everyone will definitely hope to develop more scenarios such as three-domain integration. But at this stage, we still need a platform with medium and low computing power to meet the needs of the short term. Guo Yang believes that in the configuration of ANP2.0, lidar is optional. In Beijing, Shanghai, Guangzhou and other places, the visual experience without lidar can also run very smoothly in the city. . If lidar is added, the performance and safety will be even greater.”

Baidu has its own high-precision map, which is a natural advantage over the maps sourced from other smart driving suppliers. ANP2.0 can enjoy high-quality image terminal services, improve efficiency, and reduce joint debugging costs. At the same time, when using the dynamic data of Baidu mobile map, drivers can have a more calm experience during intelligent driving, especially when they need to judge the decision-making position on ramps and other forks. The car is not easy to shake. Obviously ANP2.0 has better Performance.

In addition, Baidu's "all-in-one transportation and parking" solution can be continuously iterated. It does not become a "black box" after being delivered to the OEM and cannot be upgraded.

This ANP2.0 can be understood as a dimensionality reduction application for L4 level autonomous driving capabilities. Previously, Baidu's smart driving products have been mass-produced in cooperation with many mainstream car companies such as GAC Eon and Great Wall. In the future, the number of smart driving products empowered by Baidu will maintain an average annual growth rate of 100%.