In August of this year, Ren Shaoqing, the former R&D director of Momenta, joined NIO as the vice president of assistance and directly reported to Li Bin. At the same time, Weilai also announced plans to issue more than 1.7 billion U.S. dollars in stocks, and the funds raised will be used for research and development of autonomous driving technology. And this morning, it was reported that Weilai plans to independently develop autonomous driving computing chips and will invest several billion yuan. Although the project has not yet been discussed by the board of directors, Li Bin seems to have a clear intention and Weilai has established an independent hardware team. “Smart HW” is responsible for this project.

Hearing this news, the author suddenly understood the doubt that the author had been unable to answer before.

A while ago, the author participated in the Beta version of the NIO pilot assisted (Navigate On Pilot, NOP) experience activity. Pilot automatic driving assistance system, navigation system, and high-precision map enable the vehicle to realize automatic lane change, overtaking, automatic on-ramp and main road switching on most highways and urban elevated roads according to navigation path planning.

From a functional point of view, this set of NOP of Weilai is completely based on the Tesla pilot assistance(Navigate On Autopilot, NOA) as a template The system built; in terms of experience, there are still some differences between Weilai NOP and Tesla NOA.

Radical or conservative?

The NOP system is actually based on L2 automatic driving, adding automatic lane change and road (lane) planning functions, that is, integrated navigation on the basis of the original adaptive cruise, lane keeping and steering lane change functions Route and high-precision map guidance to complete semi-automatic driving on applicable roads.

The actual experience can be summarized in one sentence: smooth lane changes, clear planning, excellent interaction, but not strong practicality.

Automatic lane change is the core function of the NOP system. In the NIO system, the setting of “Whether the lane change requires manual confirmation” can be adjusted. Like Tesla’s NOA, if it’s set to “manual confirmation required”, then when the NOP system determines that a lane change is needed, it will remind the user to turn on the turn signal to confirm the operation.Then the vehicle starts to change lanes according to the surrounding vehicle conditions and road environment; if it is set to “No manual confirmation”, the system will directly change lanes when it determines that a lane change is needed.

I personally believe that if “manual confirmation” is required to change lanes, it is essentially the same as the original automatic driving assistance system, because the user can directly call on the basis of lane keeping and adaptive cruise. Turning to change lanes, so NOP will not bring any experience improvement at this time.

So the whole test is based on the “no confirmation” setting for lane change.

The automatic lane change of the NOP system, from the perspective of the lane change process alone, the steering is relatively smooth, the occupants in the car feel more natural, and there will be no discomfort or abruptness.

At the same time, when the system decides to change lanes, the surrounding vehicle conditions and road environment are also in a relatively safe situation.

Because NIO’s NOP system uses Baidu’s high-precision map, it can plan lanes far before exiting the ramp, and merge to the right in advance, so that the exit ramp is relatively smooth; Thanks to high-precision maps, some multi-level ramps can also be handled well by NOP.

Of course, the process of entering the ramp is essentially an automatic lane change process, so the experience is relatively good.

Speaking of this, in fact, some people will inevitably have questions. So, NIO’s NOP functions are all very good, how can they not be practical?

The reason lies in usage scenarios.

L2 level automatic driving is used when driving on high-speed roads or loops. The intention is that the vehicle takes over most of the accelerator brake and steering wheel tasks, thereby saving energy; if the user is not in a hurry, or on congested roads, auxiliary systems such as L2 It is enough to use and can bring a very big experience improvement. But the problem is that if the user is in a hurry and wants to save energy at the same time, whenever he needs to change lanes to overtake or off-ramp, he needs to take over manually, which greatly affects the overall experience, and it is easy for the user to feel “I might as well do it myself” .

The NOP function is to meet this type of scene. If the vehicle encounters a slow car, it will automatically change lanes and overtake, approach the off-ramp, and automatically merge to the right, thus smoothly entering the ramp. In essence, except for users who are early adopters and novelists, those who are willing to use the NOP function are all to save energy and save time compared to L2 automatic driving.( And don’t miss the ramp).

If you say that NOP is still competent in relatively good car conditions, but in a slightly more car environment, NOP has exposed certain problems.

For example, when changing lanes or entering a ramp, if there are too many vehicles, because Weilai NOP exitsDue to “safety” considerations, automatic lane changes will only be completed when the distance is very wide. If the vehicle is not taken over manually, it will be difficult for the vehicle to merge successfully after almost all the ramp entrances are driven.

(Do it yourself at the end of the NOP ramp)

There are two reasons for this: One is the problem of safe distance between cars. In consideration of actual braking safety and related laws and regulations, NIO set relatively harsh conditions for lane change judgment, which led to the automatic NOP core function. There are many restrictions on lane change; the second is hardware computing power or software issues, and the system takes a long time to consider “safety or not”, which leads to missed lane change “window period.”

(Safe distance determination for automatic lane change)

At this time, users accept that the core difficulty of the autonomous driving system is the question of confidence.

As the most critical foundation of the NOP experience, the safety and decisiveness of the automatic lane change actually determine the user’s confidence in it. Although Weilai, in order to enhance user confidence, when the automatic lane change fails, NOMI will give a voice prompt “Oh, NOMI is a little scared”, prompting the user that the vehicle cannot execute the lane change due to safety reasons such as the distance between vehicles, so that the user can fully understand the vehicle very clearly The surrounding driving environment; but if in order to achieve “absolute” safety, leading to multiple failures in lane changes, it will greatly weaken the user’s desire to use.

In contrast, in Tesla’s NOA function, the vehicle changes lanes very decisively and aggressively. For example, when changing lanes to the left, if the vehicle is blocked, Tesla will accelerate or slow down to find the gap and complete the lane change. . This experience will cause discomfort to passengers a little, but for drivers, as they gradually understand the upper limit of NOA’s ability to change lanes, they will become more and more convinced that this system can even partially replace themselves to complete overtaking operations, thereby making users The desire to use NOA is rising.

This kind of experience difference will not be unknown to NIO’s technical staff, but NIO still chooses this strategy for other reasons.

If it doesn’t work, it’s OK

As the best-developed Chinese new car force, it is also a car company that has successfully opened up the domestic luxury car market. Weilai wants to follow Tesla’s footsteps. Autonomous driving is inevitable. After the road. So now that Tesla has already taken a direction on its autonomous driving technology route, it is understandable that Weilai has followed the benchmark with “good faith like flow”.

The problem is that the route can be learned, but the hardware and software are the core.

In 2014, the first generation of Tesla’s autopilot hardware (Hardware 1, HW1) used Mobileye Q3 chips, and then in 2016 and 2017, Tesla adopted NVIDIA’s PX 2 and PX 2 upgraded versions, and finally Upgrade to self-developed chip FSD in 2019. The reason for these changes, on the surface, is a different technical route. Mobileye focuses on visual processing, supplemented by deep learning (neural network), while NVIDIA focuses on deep learning, supplemented by computer vision, and Tesla thinks Deep learning is everything. Ordinary computer vision processing does not require, or even HD map and lidar data processing. Therefore, some chips on the market are expensive because they cannot be used. A lot of deep learning is needed by themselves. The computing power is not enough, so I can only research on my own.

However, in this process, Tesla entered the market earlier and was able to ownHave some voice. And just after Tesla parted ways with Mobileye in 2016 due to an accident, Intel acquired Mobileye at a high price of $15.3 billion in 2017, hoping to become a chip manufacturing giant and take a share in the field of autonomous driving.

NIO installed the Q4 chip on the ES8 before and after this juncture. At first glance, Wei Lai is developing in the footsteps that Tesla has taken. At least the direction is not wrong. Using the chips and development kits provided by Mobileye as a starting point, it began to develop a vision-based autonomous driving system.

One thing to understand here is that the main function of the Mobileye chip is to analyze the surrounding environment through the data returned by the camera and radar, and give the driving area. The decision-making level is in charge of NIO(NXP chip and self-developed software layer). But there is a huge problem in this. In order to ensure its own interests, Mobileye does not know how the data enters the chip and the car company does not know (like a black box), just to get a data result, car companies can only fine-tune, that is to say, the core part of the software processing algorithm, car companies do not have any dominant power.

The problem caused by this is that if car companies want to add special independent functions, it will cost a lot of money and time. In addition, upgrading and optimization are also restricted by the speed of suppliers. This is also why Weilai’s self-driving team in North America experienced multiple resignations and layoffs, and finally decided to move all the research and development focus back to China.

Not only that, at the same time, BMW, Volkswagen, Nissan, Ford and other manufacturers were also cooperating with Mobileye. Compared with Weilai, these big car companies were obviously more “hard-hearted” when negotiating terms with Mobileye. Can’t get any exclusive dividends.

So if Weilai wants to continue this path, he will score two steps.

First, hug Mobileye’s thighs first, consolidate the foundation, and use the existing hardware architecture to hone the software development capabilities as soon as possible. Therefore, in 2019, Weilai announced a strategic cooperation with Mobileye. In addition to being the first car company equipped with Q5 chips, the two parties also collaborated on the development of autonomous driving systems and fleets above the L4 level; later, Weilai accelerated the NOP system in the same year. Research and development, and invested a lot of financial and manpower for this.

Secondly, even if it can be equipped with Mobileye Q5 chips in the first batch, as a whole, the computing power of its chips still has a gap with NVIDIA and Tesla. In addition, the “black box” data processing in the research and development stage, Weilai wants To gain a foothold in the field of autonomous driving, or to further build NIO’s “intelligent” label, it is necessary to “walk on two legs.” That is, while cooperating with Mobileye to conduct research and development and loading into the car simultaneously, on the other hand, you need to design and develop your own self-driving chip.

Although this matter will consume a lot of money and time cost, since the function points of the product can be realized by the cooperation of Mobileye in a short time, even if it seems to be a thankless thing, NIO is very likely Desperately enter the self-developed autonomous driving chip.

In addition, Mobileye will not be uncooperative because they also have vested interests. REM(Road Experience Management) is a very important part of Mobileye’s technical route. Data is collected through the on-board camera, and then the data is marked and Compressed and recorded; in addition, these data can be integrated together to form a “Roadbook” high-precision map, which can be used for Mobileye’s autonomous driving research and development.

Currently, Mobileye’s REM process in China is very slow. There is an urgent need for car companies to drive a large number of vehicles equipped with their chips on the road. With NIO’s NOP system, both parties can obtain the benefits they need.

So in essence, although the current NIO NOP is still a little far away from practicality, it is the only way for the entire large-scale framework planning; at least many users will use it out of novelty, so that NIO can Collect a large amount of data for future research and development; the cooperation with Mobileye can not only ensure the competitiveness of products at hand, but also serve as a “sparring exercise” to continue to build the foundation for Weilai’s autonomous driving; eventually, Weilai is likely to move towards self-developed chips. It is possible to guarantee their long-term competitiveness in the market.