Automatic Driving Requires Better Assistance Driving

The answer to whether autonomous driving needs to make better use of existing technology to improve driving assistance is definitely yes. As far as car manufacturers are concerned in 2020, the focus should be on achieving more advanced levels of driving assistance. Currently, both traditional car manufacturers and new car manufacturers are aiming in this direction.

The price of the Tesla long-range version has been announced, and new car manufacturers have generally delivered products that can achieve a range of 500 km. Regarding the PK of the range, it seems to be approaching the ceiling of current battery technology.

The capabilities of smart driving assistance directly determine what features consumers can experience. The intelligence competition between car manufacturers is a major battleground. In contrast, new car manufacturers tend to be more aggressive in this area.

The preheated small Xpeng P7 will also be released in April. In addition to the official claimed 706 km range, the most important feature is the driving assistance. For this reason, we contacted Xpeng’s automatic driving product manager, Huang Xin, to understand Xpeng’s ideas on automatic driving.

Xpeng’s “Chinese-style” Autonomous Driving Route

Continuity Principle of Autonomous Driving

First is the L3 level. Since Audi gave up the L3-level assisted driving, it seems that the step-by-step route that car manufacturers insist on is disintegrating. However, Xpeng obviously doesn’t think so. As Huang Xin said, L3 is actually very abstract. It is only a definition in terms of engineering. Consumers may not care about what L3 is. Moreover, in the traditional sense, “configuration” is no longer the factor that blindly makes consumers spend money.

In other words, as cars have developed to the present day, users have been educated by the market. Car manufacturers not only need to solve the problem of having or not having the feature. The core issue is what the user experience will be like after the feature is implemented. Consumers are more concerned about what features you have implemented.

Under the engineering definition of L3, what capabilities you can achieve and how you can achieve them are what car manufacturers should consider. It should not be a case of calling it advanced driving assistance, but only be usable in an enclosed area.

So how does Xpeng differ from other L3-level cars on the market? What are the distinguishing factors?

Huang Xin also answered Xpeng’s solution from a consumer’s perspective.

First is the “continuity” of use. When the driving assistance is turned on, how can the owner intervene less and adjust it less?Understanding the “continuity” from the consumer end is crucial. People who have used assisted driving functions know that when the assistance is turned on, it will remain on as long as the road conditions are good and the driving order is good. However, if a warning sign or object is detected on the road, it will exit and let the driver take control of the vehicle. This is the standard practice of the manufacturer.

It is certain that this logic is correct. At this point, letting people take over can avoid some danger, but there is still room for improvement.

The first is the problem of road conditions in China. The situation of road repairs varies, but sometimes although there are warning signs, the real road conditions are very good when you drive past them. However, the assisted driving system will still exit, and in the case of continuous repairs, you will continuously enter and exit automatically.

This experience is very bad, and Xpeng will optimize it.

Huāng Xīn also gave us an example:

In fact, the current ACC (Adaptive Cruise Control), ALC (Automatic Lane Change), and LCC (Lane Centering Control) are standardized functions, but consumers may encounter some detail problems in actual use, and the vehicle cannot solve them by itself. Xpeng aims to make these standardized functions more “Chinese” by adapting to countless scenarios that actually occur.

For example, ACC has no problem in high-speed driving, but in congested traffic, even if the following distance is adjusted to the closest, it is still actually far. If you turn on the assisted driving at this time, you will be cut in and feel uncomfortable. There is also the issue of recognizing and judging minor lane changes by the adjacent vehicle while driving normally, so as to change the driving speed, etc., without the need to exit the assisted driving.

P7 can shorten the following distance and monitor the blind spot on the side of the car through the combination of the front camera and radar, as well as the side camera. It can simulate human driving and quickly speed up or slow down to follow the car in front safely and avoid the annoyance of “Chinese-style cut-in”.

Also, when using LCC, if there is a large truck beside it, the standard approach is to run in parallel as long as both vehicles stay in their lanes. But humans generally want to get away from it as soon as possible, so Xpeng P7 will avoid running in parallel with the truck, thus simulating human driving behavior with assisted driving.

The goal is not to stay put in dangerous situations, but to adjust the vehicle itself in a more humanized way when there is no danger under the current driving conditions.

To achieve this, XPeng’s hardware usage can be described as exaggerated:

In addition, XPeng P7 will also open NGP function in the fourth quarter.

What is NGP? It is a navigation assisted driving ability that integrates high-precision maps and positioning, similar to the NOA function of Tesla.

In practical experience, when the assisted driving is turned on, the vehicle will drive itself according to your designated destination, and will change lanes and overtake according to the road conditions. For example, if you are driving on a three-lane highway and driving in the middle lane, if the destination exit you designated is on the right side, the vehicle will change lanes to the right according to the road conditions.

In terms of automatic driving perception, XPeng is similar to Tesla, mainly based on visual perception.

Huang Xin also expressed the difference between XPeng and Tesla:

First of all, in terms of function realization, XPeng will not reduce compared to Tesla, and of course, it will not add arbitrarily either. But in terms of function usage, XPeng has its own considerations.

Huang Xin said, for example, identifying ice cream cones. It seems that the industry also regards the ability to identify ice cream cones as a manifestation of assisted driving ability. The problem is, what can be done after identifying ice cream cones? What kind of functional logic is behind this?

He said that Tesla’s visual perception ability is very strong, and identifying ice cream cones is also to avoid danger caused by warning objects. However, his experience with Tesla is that when encountering ice cream cones, the assisted driving will exit, and continuous manual intervention and automatic exits will occur when encountering consecutive ice cream cones, even though the driving conditions at that time were very good according to human judgment, with no other vehicles and wide roads (meaning people would not manually signal to change lanes and would exit).

Huang Xin said that this is actually a very typical Chinese road condition, which is what XPeng values. XPeng wants to focus on situations like this, where the system does not need to exit as long as it can automatically change lanes or slow down to avoid ice cream cones, as long as it is safe.

So this creates a “Chinese-style” demand for high-speed NGP products, and we want to do some specific research and development based on user needs during the use of such products.## Tesla’s radical practices

Furthermore, Huang Xin mentioned that he drove a Tesla equipped with HW3.0 and used its assisted driving feature when passing through a tollgate. The car followed the leading vehicle until it lost track due to the absence of lane markings, which prompted the instrument panel to sound an alarm warning the driver to take control of the vehicle. However, the Tesla continued to drive despite the alarm until it reached the height limit bar. Huang Xin commented that Tesla’s behavior was too radical for his liking.

On the other hand, XPeng Motors believes that assisted driving should provide functionality that is beyond users’ expectations but not beyond their psychological limits. Exceeding the latter may pose greater dangers than benefits.

Moreover, Huang Xin pointed out that XPeng Motors has optimized the recognition of road signs and warning devices, such as ice cream cones, for Chinese road conditions. XPeng has developed a much more extensive capability in road sign recognition than Bosch, although Huang didn’t specify any particular numbers.

In addition, we tested a Tesla alongside XPeng P7 in heavy rain. Although it was difficult to see the road surface with the naked eye, the P7 still traveled with assisted driving. Huang Xin explained that when the camera has been blinded, the P7 can rely on millimeter-wave radar and high-precision maps to function properly under extreme weather.

XPeng Motors has based its approach on prioritizing scenario functionality to better serve user needs over industry convention. Although it meets standard functionality at the same level as all other automakers in the market, many scenarios that are specific to China are optimized to fit user needs better.
In accordance with this principle, Huang Xin mentioned that P7’s assisted driving may not yet be on Tesla’s level. He acknowledges that although Tesla’s prototype testing may seem much better than the current P7 performance in intensive road conditions, P7’s software will be updated and enhanced to eventually catch up with Tesla’s.Question: Can the data from G3 be reused due to sensor changes?

Like Tesla, P7 can reuse the data from G3, which also collects real-time data through cameras. Although there are significant differences in the architecture between G3 and P7, the data cannot be fully restored to provide input as single data points, but it can provide group data for research purposes, such as vertical and parallel parking success rates, usage efficiency, and their proportion.

Question: Some automakers have developed many functions, but opening them requires certain conditions. Does XPeng have any related conditions that can cover how many scenarios?

First of all, XPeng will try to match the industry’s standard level in the first batch of open scenarios. Secondly, it will combine these scenarios with its research and development capability to achieve better coverage. For example, in tunnel scenarios without signal and GPS information, XPeng will not downgrade Next Generation Platform (NGP) but will ensure safe driving under these circumstances.

Question: Some functions have high error rates when activated. The car will generate warnings when there is actually no danger, which can significantly affect the driving experience. How does XPeng avoid incorrect and false alarms?

First, avoid generating alarms when there is no actual danger. When there is an important event, it must generate an alarm. Safety is the top priority. Secondly, XPeng will use data to analyze, for example, lane departure warning data. Some users turn it off manually, while some others do not recommend turning it off when they feel fatigued. Therefore, XPeng’s approach is to turn off this function by default, but it can be related to detecting driver fatigue through sensors. When fatigue is detected, this function will be turned on. This is the logic.

Question: Previously, P7’s video could correctly recognize multiple lines on highways. How was this accomplished?

First, it needs to determine whether the interfering lines appeared suddenly or were present from the beginning. If the interfering line is suddenly presented, the vehicle will calculate the position and timing based on the correct lane line and then remove the abnormal interference based on the regular pattern in the middle of the entire lane line. Therefore, the judgment logic is not based on color.

Conclusion:

XPeng’s approach to autonomous driving emphasizes safety and redundancy, with its hardware and architecture being top-notch. To meet the company’s future system matching and autonomy requirements, XPeng selected a self-developed vision perception system.

In terms of function, which is the core logic of XPeng’s autonomous driving, P7 will open functions in phases after being released. First, it will open the functions that exist on G3 and optimize them on P7. Then, NGP functions are expected to open in the fourth quarter of this year. Horizontally, XPeng’s open functions will mainly meet market mainstream requirements and be comparable with competitors.The last step is that XPeng will optimize its functions based on the “Chinese-style” demand. Simply put, the functions are standardized, and what distinguishes manufacturers is their sensitivity to “user” demands. After talking with Huang Xin, it is clear that XPeng’s focus will be on optimizing functions for the Chinese market.

P7 will further optimize its automated parking function, which is still a major highlight of the car. XPeng uses a visual parking recognition system that does not require the use of a side-view camera like traditional parking. Instead, it uses a unique camera located at the front of the car to detect the parking space ahead of time.

The camera ensures a high success rate when the function is enabled, and can model and record parking space data. For parking lots that have not been visited before, three data collections are generally sufficient for coverage of the area. The data can be uploaded to the cloud and shared with other XPeng car owners. This means that even XPeng car owners who have not visited the parking lot can achieve accurate parking through data sharing.

Written by Delu
Edited by Daji

Read these as well

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.