OTA, one of the hallmark technologies of intelligent vehicles

An intelligent vehicle, through continuous OTA, can constantly bring value and surprises to its users, while a vehicle that cannot be updated OTA would be labeled as a “traditional car”.

The use cases and functions of vehicles are varied and complex. During the early stages of development, product managers and engineers often lack the resources to cover everything.

OTA allows businesses and users to connect and interact through products. User feedback during actual usage becomes an important source for OTA updates. The feeling of an effective response from the company to the “voice of the customer” forms a meaningful connection between the brand and the user.

Of course, important new features often come as surprises to users like Easter eggs.

On June 7th, 2021, Xpeng Motors announced a new feature through a poster. Xpeng will soon release the latest version 2.6.0 of Xmart OS. In addition to optimizations to NGP navigation assisted autonomous driving, in-car display systems, and voice systems, the most important update is the launch of VPA (Valet Parking Assist) automatic parking with memory recall.

Xpeng says this is the “first mass-produced and non-reliant on parking facility modification” automatic parking feature.

Xpeng claims this is the “world’s strongest automatic parking” and welcomes friendly comparison and evaluation.

What is VPA?

On June 9th, 2021, Xpeng introduced the OTA update during their smart experience event, where we had the opportunity to test-drive the VPA function.

In addition to regular scenarios, we also encountered extreme conditions during the test drive (see section 4 below)

Undoubtedly, VPA is the highlight of this update, and Xpeng dedicated a special introduction to VPA during the start of their smart experience event.

Simply put, this feature allows the vehicle to assist the driver in driving along a user-set route, from the starting point to the endpoint of the set route, and automatically park near the endpoint in a location previously saved by the system.

Under normal circumstances, similar to NGP, the driver only needs to supervise the vehicle without any hand or foot operations.

The most common use case for this feature is that every time the user returns home or enters an underground parking lot beside their workplace, if they pass by the set point, the vehicle can drive to the designated parking spot and park by itself along the established route. Essentially, users have access to a “private chauffeur” in the parking space, saving time and effort for searching for a parking spot.

Xpeng says that VPA was launched because ultra-low speed scenarios are one of the three important autonomous driving scenarios planned by Xpeng, in addition to high-speed scenarios and urban scenarios. These three scenarios form a complete cycle.Before, NGP ADAS has been launched for high-speed scenarios. The Valet Parking Assistant (VPA) was introduced by XPeng Motors as a starting point for parking in ultra-low-speed scenarios.

What makes XPeng’s VPA different?

Prior to the introduction of XPeng’s Valet Parking Assistant, many automakers and top suppliers were researching parking functions for the last kilometer in parking lots, especially for automated driving in underground parking lots.

The GPS signal in underground parking lots is weak, and the vehicle’s positioning ability is poor, sometimes with no signal at all. In addition, there is currently no map service provider that provides navigation maps for underground parking lots. Without positioning and navigation map support, automated driving is difficult to achieve for vehicles.

Many of the previous research and products by automakers and suppliers primarily hoped to use parking lot end devices as perception and positioning assistance, to help vehicles achieve positioning and path planning. For example, Mercedes-Benz S-Class had previously launched this feature, but requires parking lot end devices as perception support.

The problem with this approach is the inability to promote on a large scale:

Firstly, the cost for the vehicle end is high, requiring additional equipment to be installed to interact with the equipment at the parking lot end. The high cost means only luxury cars such as the Mercedes-Benz S-Class can bear the burden.

Secondly, it relies on parking lot end intelligence, requiring equipment operators and parking lots to negotiate on a case-by-case basis. The ability to promote on a large scale is poor, and the time and cost of this approach is undoubtedly astonishing.

As a carmaker that is exceptionally sensitive to cost and scale, the most desirable approach must be the intelligent vehicle end, and to realize this in a low-cost way.

The biggest feature of XPeng Motors’ VPA is that it does not increase any hardware costs and does not rely on the renovation of parking lot sites.

How is VPA implemented?

To achieve single-car intelligence, XPeng has made a new plan for the perception system and algorithm under the VPA scenario.

The first change is in the perception system.

The GPS signal in underground parking lots is poor, and the millimeter-wave radar is almost unusable due to strong interference from metal objects such as walls, ceilings, and vehicles in this environment.

In high-speed scenarios, vehicles need to see far, but in low-speed scenarios such as parking lots, the field of view (FOV) of vision is more important, and vehicles need to see widely.

XPeng Motors says that for VPA perception, visual input is the main input method, achieving functions such as goal recognition and ranging. VPA uses all the cameras on the P7, including front, side, and surround cameras, to form 360° panoramic coverage.

Moreover, when the vehicle is driving and the chassis pitches (such as passing speed bumps), it will cause changes in the camera’s field of view, changing the target’s direction in the screen and causing judgement errors. To eliminate errors, the vehicle will fuse inertial navigation data and vehicle attitude data for correction.

The algorithmic improvements are even greater than the changes to the perception system.To achieve vehicle positioning and path planning, XPeng Motors has developed a “semantic map + matching algorithm” method to replicate the remembered route.

The semantic map refers to the collection and recognition of parking lot environmental elements and vehicle location information, and the fusion to generate an aerial view of the environmental information.

In the process of route memory, the semantic map will identify “landmark elements” (such as lane markings, pillars, intersections, etc.) on the route, and obtain relative positional information such as the distance between the vehicle and the “landmark elements”.

When using the memory parking function of the parking lot, the vehicle continuously adjusts the driving route based on the “landmark elements” in memory, matches and overlaps the position of the remembered elements with the current ones, so that the vehicle accurately replicates the remembered route. At the same time, the fusion of inertial navigation data further helps the vehicle to complete positioning and route guidance.

In addition, XPeng has conducted special training for the system to predict difficult behaviors of vehicles and pedestrians in complex parking lot environments. When using the memory parking function, the system makes predictions about their behaviors based on the special training.

The semantic map and the memory path help solve the problems of vehicle positioning and path planning.

These two features are why XPeng Motors boldly claims that this is currently the “first mass-produced and not dependent on parking lot reconstruction” “last kilometer” parking function. Although it has added a long modifier, the almost 0 cost and vehicle-side intelligence are indeed the lowest-cost and most widely applicable solution for implementing the “last kilometer parking” function.

In addition, XPeng Motors also stated that the VPA algorithm is completely self-developed by XPeng Motors, independent of suppliers, so the update and iteration speed is faster, and welcomes other companies to experience and compare the functions.

How is the performance and experience of XPeng VPA?

This VPA experience is located at the Huiju Shopping Center in Daxing, Beijing, which has 6652 parking spaces and has been awarded the “largest underground parking lot” Guinness World Record. Therefore, it can almost meet the size of most parking lots for testing.

Firstly, the usage of VPA parking memory function is divided into two steps:

1. Set up the route (learn the parking route): The user takes a certain position on the same floor of the parking lot (such as the entrance of the parking lot) as the starting point, drives the vehicle manually to a fixed parking space, and manually parks it in place.

2. Activate the function (use memory parking): When passing through the starting point of the set route next time (slow down the speed), the system pops up a memory parking function card and activates this function.The experience of the VPA function, combined with test drive and engineering Q&A exchanges, is as follows:

3. Currently, it only works for underground parking lots with clear parking lines, and does not support ground parking lots. In the future, ground parking lots and parking buildings will be added, and eventually full-scenario coverage will be achieved.

4. The current applicable model is the XPeng P7 intelligent flagship edition and Pengyi edition equipped with XPILOT 3.0 system.

5. Due to safety considerations, the version currently pushed only supports routes on the same floor. However, the engineering demo version under development can recognize cross-floor routes and will add this feature in future versions.

6. For cross-floor routes, the algorithm determines which floor the vehicle is on by judging whether it is going downhill, the height difference, and through fusion algorithms.

7. Only one route is memorized in a single parking lot, and the maximum length of the longest route memorized is 1 km. A single vehicle can learn 100 parking lot routes.

8. During the first learning of the vehicle, the speed of manual driving cannot exceed 15 km/h, and the route cannot have duplicate lines.

9. After completing the initial route learning, the vehicle needs to return to the ground, and the GPS signal needs to be restored and repositioned. After entering the parking lot, the GPS signal is lost and all positioning is done based on the semantic map.

10. The current Xavier chip’s computing power is sufficient to support all the cameras’ operation under all VPA scenarios, but the specific computing power required has not been disclosed.

11. When starting the VPA, it is not necessary to return to the starting point of the memory map by 100%. The VPA can be started from any position on the route, but the vehicle speed must be slow.

12. The normal driving speed while the vehicle is in automatic driving mode is 5-7 km/h.

13. During the driving process, the vehicle can automatically adjust its speed, turn automatically, bypass obstacles, meet other vehicles, avoid pedestrians while parking automatically, and avoid entering and exiting vehicles. If the designated parking space is occupied, the vehicle will stop near the space.

14. In the parking SR interface displayed on the vehicle screen, various elements in the parking lot can be displayed, including parking spaces, pillars, speed bumps, and other static elements, as well as dynamic elements such as vehicles and pedestrians. Currently, it cannot recognize objects such as shopping carts, pets, and bollards, but this feature will be added in the future.### 小鹏 VPA 试驾体验：场景限制、停车泊车效果等改进

13. Vehicles have higher safety authority over pedestrians. In complex scenarios such as encountering children, pedestrians pushing shopping carts, and intersections, vehicle speed slows down, and human attention may be reminded through voice and visual cues to prepare for taking over. Users can intuitively understand the system status, driving path, target parking space, next step status, and usage boundaries.

14. Regarding the issue of route sharing, XPeng will not share memory routes and parking spaces until the issue of user agreement is resolved.

15. The current version has limited parking space left for the system near the end of the road, which cannot be covered temporarily. Training scenes have not yet covered parking spaces without parking lines.

16. To park in non-fixed parking spaces, users can currently remember to park near the parking space, and then use automatic parking to connect to the parking space. However, parking in non-fixed parking spaces will be launched soon.

17. The system cannot currently recognize ground arrows, but ground arrow recognition and reverse-driving problem solutions are under development and will be added in future versions.

18. VPA combined with XPeng’s super-charging parking space autonomous navigation function is also being planned. It can provide car owners with the driving path of the super-charging parking space, and will realize this scenario in the future. For example, autonomous navigation to XPeng’s super-charging parking space in an unfamiliar parking lot.

19. When the parking memory function is turned on, based on safety and regulatory restrictions, people cannot get out of the car. Memory parking is not autonomous driving and cannot completely replace the driver.

20. There is currently no active call function for parking lots, but plans to launch autonomous recall and autonomous parking in the future.

During this test drive, we encountered an extreme case: at an intersection, the vehicle needed to make an unprotected left turn, but there was a car coming from the opposite direction and the oncoming vehicle happened to be parked on the left side of the vehicle. Due to limited left-turn space, the vehicle slowed down and moved slowly at the intersection.

In this scenario, we can see that the vehicle is traveling at a very low speed and is in a hesitant state. However, unfortunately, due to the urging of the car behind, we took over manually to complete the turn, and VPA can be reactivated after manual takeover. In the end, we are not sure if it can complete this turn.

Overall, XPeng VPA drives at a low speed with a very cautious and conservative driving style, similar to that of a novice driver. However, when finding a parking space and using automatic parking, the parking efficiency is very high, and many times it parks perfectly in one go.

From the practical experience and official introduction of XPeng VPA, XPeng Motors is gradually covering all scenarios of autonomous driving, with many practical functions such as autonomous recall and super-charging parking space active navigation already under development.

In addition, XPeng Motors has independently developed the entire autonomous driving stack without relying on suppliers, which is why many functions are already launched at a higher speed and efficiency in terms of data collection and version changes.

What other minor upgrades are there?The new 2.6.0 system has made new upgrades and optimizations in NGP, in-car systems and cabin interaction, except for VPA.

For example, under the new NGP status, the driver can awaken XPeng to say “Help me change lanes to the left” and “Please overtake the car in front,” and control the vehicle to achieve voice-controlled lane-changing in both the left and right directions, which is the industry’s first function of this type achieved through voice control.

The new functions added under the NGP status include keeping away from large vehicles, NGP self-recovery, optimizing NGP’s response to cutting-in vehicles, and displaying instrument guide lines under the NGP status, among others. According to XPeng’s latest data, the latest version of NGP has made further improvements in handover frequency, lane-changing and overtaking, entering and exiting bends, and tunnel passing compared to the current version.

Inside the cabin, the overall music experience has been further optimized for voice interactions, which support voice playbook song lists and albums, and have optimized the return results of voice search. XPeng can be awakened by saying “Play Jay Chou’s album”, “Play jazz song list”, or “Listen to Guangdong radio”.

In addition, the camera on the steering wheel is turned on for continuous monitoring of driver status security during NGP and LCC operations, and reminders are given when distracted or closed-eyed behaviors are detected.

XPeng stated that only facial status is monitored, and there is no face recognition or any form of data storage.

Other optimizations include driver instrument display optimization, including adding lane display on the navigation card, instantaneous power display on the energy consumption card, and optimizing the 3D model of P7 in the instrument; additional small functions include using the steering wheel switch for air conditioning, optimizing SayHi lights, adding daytime light sword mode, and adding permanent ambient lights to the door handle when parking.

Among the new players in the car industry, XPeng is the one that uses “Smart Badge” as its ace. Not only is it fully self-developed but its autonomous driving route is also very clear and precise.

Currently, it relies on OTA to constantly implement its plan. According to the plan, Xpilot 3.5 for the urban area scene of NGP will be released next year, and XPeng’s autonomous driving is even more anticipated.

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.