Xiaomi SU7 City NOA has been pushed for half a month, and there are few evaluations about this intelligent driving assistance system. The reason is that it only covers ten cities such as Beijing, Shanghai and Guangzhou, and it still needs to brush up on 1000km of safe and intelligent driving mileage before it can be used. Moreover, it will be pushed one after another from June 6, instead of OTA if you have a car; After the system is updated, you can start using it only after the 1000km mileage mentioned above is enough.
In short, Xiaomi SU7's urban NOA function wants to be used, and the threshold is relatively high.
Then, it's about how easy it is to use this set of intelligent driving. Because at present, the actual use information about the corresponding urban NOA function of Xiaomi SU7 is very limited, but it can also be seen whether the urban NOA of this intelligent driving function is good enough to use. Let's give a summary first,XiaomiSU7The urban NOA function after the update,The degree of completion is quite high,From the overall strategy, it is a conservative strategy,I want to ensure that the whole process is completed by the system as much as possible,Reduce the driver's involvement。
What is the level of this set of Xiaomi SU7's intelligent driving?
Decisions are conservative, and takeovers are normal
After half a month of successive pushes, only a small number of car owners can use this set of urban NOA functions in the ten cities covered by the current urban NOA, probably because the delivery volume of the Pro version and Max version of Xiaomi SU7 is not enough, or the condition of 1000km intelligent driving mileage has not been reached by some people; Or, there are car owners who have not been "successively" pushed to update the NOA function in the urban area.
Referring to the feedback given by some car owners who have started to use the NOA function in urban areas, it is "still a little conservative" and "under complex road conditions, it will still prompt to take over, or take the initiative to take over".
To give a specific usage level analysis, give a few examples. First of all, at the right-turn/left-turn intersection, there are crosswalks, lidar + cameras and other perception hardware, which completely achieve the accurate identification of pedestrians, bicycles, electric vehicles and other objects, and then the decision of the software strategy is to wait for all pedestrians/bicycles and other objects to completely pass through the front of the vehicle, and continue to drive according to the original route, rather than using a detour strategy (that is, going around the rear of non-motorized vehicles and other objects to pass through the intersection).
The decision in this working condition is a bit conservative, and you need to wait for the pedestrian to pass completely before continuing to drive; Then, when a non-motorized vehicle suddenly rushes to the side of the front of the vehicle, the vehicle will actively brake and exit the urban NOA function.
In other words, the driver is passively triggered.
After that, it's time to merge into the lane. Merge in and out of the lane, this is actually a high-frequency scene when driving on urban roads, especially in the working condition of merging from the ramp to the main road and from the main road to the ramp, Xiaomi SU7's strategy is mainly to let the behavior be the main one. Specifically, if there are vehicles in the adjacent lanes continuously, the Xiaomi SU7 will continue to maintain the execution of merging; There is no particularly aggressive operation that merges into the adjacent lane at once (not in the inside version, nor after OTA).
In this case, in fact, there is a distance between vehicles, which is enough for the vehicle to merge into the lane under the premise of the driver's operation, but the algorithm does not execute the instruction to merge. After that, it is about the scene of exporting or overtaking on the right, under the premise that there is space in the side lane to merge and overtake, and the intelligent driving system has turned the front of the car and driven into the side lane, after the car coming from the side and rear, it will still brake and wait for safety to complete the merge again.
Of course, we only make a judgment on this set of Xiaomi urban NOA through limited test information. At least for now, the strategy of NOA intelligent driving in urban areas is more conservative. And more often, it is necessary to "human-machine co-driving" to solve the driving problem.
Why, is the tuning more conservative?
The models that get OTA push urban NOA are all two NVIDIA Orin-X chip solutions, with a computing power of 508TOPS, theoretically speaking, it is not difficult to cope with the challenge of information collection + decision execution, and for the computing power of 508TOPS, it is not yet possible to reach the upper limit of the performance of the two chips.
Although the use of perception hardware and chip hardware has been the mainstream choice in the industry, the actual function of NOA in the urban area that we finally see still maintains a relatively conservative implementation strategy.
What is the reason?
Xiaomi's intelligent driving solution, the underlying logic, includes zoom BEV technology, super-resolution OCC occupancy network and road model three sections to achieve the final decision-making + execution. Zoom BEV and OCC occupy the network, which needless to say too much, is a function that perceives the environment and collects information.
The main thing is to talk about the road model. Xiaomi's road model mechanism is a road model built by using big data and learning technology, which can help vehicles understand the structure of the road and traffic rules, and can also predict the behavior and possibility of other participants on the road.
Relatively speaking, the function of the road model has a certain learning ability, a bit like the CNN network in Tesla's FSD system, which constantly feeds the driving video of human drivers to learn and improve the maturity of intelligent driving; But the difference is that the data processing in FSD mainly relies on sensor capture, while Xiaomi's intelligent driving still relies more on high-precision maps at the data processing level.
However, the underlying learning logic is somewhat similar, and it still needs to be supported by a large amount of data. So, after figuring out the underlying logic, now let's go back and look at two questions, why there is a relatively high threshold for conservative tuning and the use of urban NOA functions, which is easy to answer.
Let's talk about the use threshold first, you can only use it after brushing enough 1000km mileage. Perhaps this limitation is to collect more high-quality intelligent driving learning materials, so as to help the maturity and ease of use of the intelligent driving system. So,About the conservative tuning style,Personal feeling has nothing to do with the hardware,After all, it's all the configuration of the top.,The core is still the setting of the algorithm.。 Conservative strategy, maybe not enough time to learn?
In fact, in some working conditions, we have also seen the highlights of the system, such as the detour of temporarily parked vehicles on the side of the road, the attitude and response speed of braking at intersections, etc., and the first version of urban NOA can come up with a high degree of completion of the use experience, which is rare in itself.
From an objective point of view, the Xiaomi SU7 urban NOA system will be significantly improved after the next few versions of the update. The version update rate will depend on the collection of materials and the learning ability of large models. If you take Xiaomi SU7's current urban NOA function as the most comparison, compared to Huawei's ADS 2.0 system, it may only reach the level of 80%, the latter is honestly iterative after multiple versions of the current anthropomorphism is very good; However, compared with the urban NOA function of these products of Ideal and Weilai, it is basically equal, and after a few iterations, it is not a big problem to catch up with Huawei ADS 2.0 version.