A plug-in for "full map" in autonomous driving? A CP game between Xiaopeng and AutoNavi

A few days ago, AutoNavi Map held a press conference for the launch of its third-generation in-car navigation system and announced that Xiaopeng Motors became the first automaker to be equipped with AutoNavi's third-generation in-car navigation system.

The reason why this information caught my attention is that I had just experienced Xiaopeng NGP Beta not long ago, and the unique new interaction left a deep impression on me; combined with the new generation of in-car navigation, it can dynamically calculate lane and road condition changes and provide more dimensional information.

▽P7 "SR autonomous driving environment simulation display

This is also the first time that the two separate domains of infotainment and autonomous driving have been connected. Not only has the navigation been improved to the lane level, the perception, analysis, and decision-making processes of the autonomous driving system have also been conveyed to the driver in a visual way. Rich road information will be informed to the driver in advance, bringing a brand new experience both in terms of technology and driving.

▽High-precision map generated based on laser points

Why was Xiaopeng the first to land? Xiaopeng's hand in hand with AutoNavi made Xiaopeng a "map-opening" player in the field of autonomous driving. This is something that anyone who has played MOBA games knows: no matter how good your martial arts skills are, you are afraid of the sheep knife, not to mention the fog of war that opens up the entire map.

▽AutoNavi is the only map manufacturer in China that commercializes high-precision maps

Recap:

1. P7 hardware foundation

2. How does high-precision map support autonomous driving?

3. Why is it not popular yet?

4. Which is the best solution: single vehicle capability or V2X?

Handshake prerequisite: P7 hardware foundation

In several shopping articles, Dian Ge's advice is to buy the high-end model of Xiaopeng P7, because only the Zhizun version is equipped with the hardware of the XPILOT 3.0 system, which is the hardware foundation for experiencing the autonomous driving of Xiaopeng NGP.

The evolution from standard maps to high-precision maps has resulted in an explosive increase in computing power requirements.

The Xiaopeng P7 is equipped with 14 cameras, 5 millimeter-wave radars and 12 ultrasonic radars, combined with centimeter-level high-precision positioning and decimeter-level high-precision maps, to form a combination of hardware perception and map positioning. NVIDIA Xavier's computing unit and Bosch iBooster braking assist system are responsible for computing + control unit control, and on this basis, add functions such as real-time map management and fusion, behavior planning, and behavior/trajectory prediction.

The top-of-the-line XPILOT 3.0 hardware refers to NVIDIA's Xavier computing platform, an AI supercomputing chip released in 2018. The computing power of a single Xavier chip reaches 30TOPS (peak value is 32TOPS), which means that the chip can perform 30 trillion operations per second.

With cameras, radars and other sensors for perception and detection, and then analysis and calculations through autonomous driving chips, this is the basis for the single-vehicle assisted driving capability that we are most familiar with, and it is also the technical solution that Tesla has always insisted on. Xiaopeng's handshake with AutoNavi aims to achieve a dual perception solution, that is, Xiaopeng NGP will predict the behavior of surrounding vehicles while detecting, and use high-precision map information for assistance. Such a comprehensive architecture is of great help to autonomous driving.

Of course, more perception sources and more complex architectures will greatly increase the amount of data processing, and will also place higher requirements on algorithms and processors. High-precision maps are like an additional physical examination process in medicine, which means early detection and early treatment.

What will autonomous driving assistance become with high-precision maps?

With the support of AutoNavi high-precision maps, Xiaopeng NGP's automatic navigation assisted driving will be deeply integrated with the navigation function, greatly improving the intelligent experience. NGP currently supports automatic overtaking, optimal lane selection, automatic speed limit adjustment, automatic on-ramp and off-ramp, automatic switching of highways, automatic emergency avoidance when changing lanes, etc. Specifically, in navigation, it will be reflected in the form of pile bucket recognition, large truck avoidance, night overtaking reminder, congested road following, and faulty vehicle avoidance.

Specifically, most of the positions on the central control screen are used to display road conditions. The upper right corner of the assisted driving frame is a traffic sign, the lower right corner is a small map, and the upper left corner is the operating status of the assisted driving function, including whether NGP is activated or prompted to take over, etc. If there is a road section with data, real-time 3D modeling will be displayed around the building, and elevated roads and tunnels can also be displayed.

Xiaopeng NGP cooperates with AutoNavi to clearly mark the lane the vehicle is in. At the same time, the road markings and arrows can be displayed in real time, which is clear and intuitive.

Specifically for urban roads and highway sections, we will experience them separately.

After turning on NGP, we can see from the navigation interface that both lane lines and vehicles are recognized very accurately and can respond in time. Especially after entering the lane change preparation, if the road is found to be a solid line, the P7 will postpone the lane change operation. In addition to identifying other vehicles, NGP can also identify ice cream cones and water barriers, etc., to determine whether the lane is invalid (as shown in the picture, the water barrier is clearly visible).

When driving on the highway, NGP will give priority to the empty left lane. If there is a car in front, it will judge the speed of the car in front. If the speed is not enough, it will change lanes to overtake. Compared with the previous 1024 test version, NGP Beta's lane change strategy is more conservative, the lane change angle is smaller, and the timing of lane change is more cautious.

I can clearly see the lane selection and lane change process from the navigation, and after the voice prompts of the lane change intention, the feeling of "being in control of the whole process" is still very good. I can know the intention of the self-driving vehicle in advance, and get the same feedback as in reality in real time through the navigation interface, which makes the self-driving process more reassuring and the effect of freeing my hands is more obvious.

The most amazing operation is definitely the ramp switching. In the current version, NGP will limit the speed to about 60km/h, and will start to slow down about 300 meters before the ramp. It is more conservative than the previous test version. For the relatively fast speed of the following car, there will be a situation where "the car in front is still far away, but it just won't speed up, and there is a car behind it."

Of course, this strategy will surely be dynamically revised in subsequent versions. At present, we can clearly see the ramp selection and merging process from the navigation, and feel the perception and decision-making process of P7 throughout. After the high-precision map informs the vehicle of the road conditions ahead in advance, it will make complex operations such as ramp changes controllable. Dian Ge believes that as long as the data is complete, it will be completely feasible to enter and exit the gate in the future.

High-precision maps have a better sense of touch, which naturally brings better safety redundancy. If you encounter road conditions that the NGP function cannot be applied, through a clear boundary and responsibility division system, pre-prompts and take over according to the distance are used, including: I notification (voice broadcast); II prompt (voice broadcast + visual prompt); III warning (voice broadcast + visual prompt + warning sound); IV strong warning (voice broadcast + visual prompt + warning sound + seat belt warning), four levels of reminders, which become stronger step by step.

"The assisted driving system will exit after 200 meters", allowing the driver to prepare for driving in advance.

High-precision maps can greatly improve the ability to predict complex changing road conditions such as ramp switching, traffic flow changes, tunnels, etc., not only increasing safety redundancy for machines, but also for humans.

High-precision maps are so good, why haven’t they been popularized yet?

In addition to the high hardware requirements analyzed in the previous article, the entire process of standard maps and high-precision maps from collection to production to release is completely different. When we upgrade the road-level engine service to the lane-level engine service, not to mention the massive amount of data generated, how to process navigation and the perception of surrounding objects or the environment at the same time is very, very difficult.

To put it simply, high-precision maps are good in themselves but the cost of collecting them is very high. A lot of effort is spent on collecting data, and it is all in vain if the vehicle analysis capabilities are not sufficient.

▽Apollo uses Novotel's GPS and IMU combined positioning system

Furthermore, the in-car navigation functions of most current car models are very weak, with old versions and single functions. They are still not as good as the mobile phone navigation that is updated every month. Popularizing the third-generation navigation with high-precision maps in vehicles is itself a very high-threshold thing.

If the in-car navigation system is as real-time and easy to use as a mobile phone, and the greatest significance of the third-generation navigation system is that it can guide both people and cars, will anyone still use mobile phone navigation? Absolutely not.