He Xiaopeng helps Musk debug: FSD drives randomly without navigation–Fast Technology–Technology changes the future

He XiaopengGo to the United States, giveMuskDebug.

Recently, He Xiaopeng, who was on a business trip in the United States, personally tested Tesla's FSD. He drove through the city and on the highway, and praised the performance of FSD many times along the way.

Of course, FSD also exposed its shortcomings in actual testing.

What is more outrageous is that compared with the statements of domestic competitors that “can be driven nationwide” and “can be driven with navigation”, FSD has already achieved“You can drive without navigation”.

What’s going on? Has FSD become a spirit?

He Xiaopeng personally tested FSD and praised it as smooth as a veteran driver

He Xiaopeng's evaluation this time isFSD V12.3.6Version, the latest V12.4.1 is still in the internal testing stage, and I haven’t borrowed the pushed car.

The public evaluation video is over half an hour long and covers various road conditions in Silicon Valley.

At the beginning, He Xiaopeng started by praising FSD, saying it was “very good” and “can be turned on at a very low speed.”

When starting at the intersection, He Xiaopeng also felt that the performance was good, “The acceleration was very comfortable, neither too fast nor too slow.”

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

Regarding this, He Xiaopeng also announced that his own XNGP is also working towards smoother and more comfortable acceleration and braking, and is expected to be upgraded in July.

In addition to these relatively simple scenarios, FSD is more decisive than human drivers when turning wide corners and encountering blind spots.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

The left turn into the road is also very accurate.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

Finally, He Xiaopeng summarized the performance of FSD:

Drive better than I do in California.

He Xiaopeng rated FSD's overall performance as “like an experienced human driver”, but some problems with FSD were also exposed during the test.

There are two more serious ones:

Driving blindly without navigation. Traffic light recognition problem.

The first problem was that after He Xiaopeng canceled the car navigation destination, the car was still driving aimlessly.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

I even decided on my own to turn at the T-junction.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

However, in reality, at this time, He Xiaopeng and the passengers in the car had not decided where to go.

FSD took them all the way, turning and cornering, until He Xiaopeng set the navigation destination again and got back on track.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

He Xiaopeng joked about this.Maybe FSD has gained wisdom and wants to invite him for coffee.

Another issue is the recognition of traffic lights at intersections.

At this time, FSD is inAggressive ModeBut for some reason, the car still stopped after the traffic lights at the intersection changed several times, and He Xiaopeng finally had no choice but to take over.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

It is worth mentioning that FSD had just been upgraded to version V12. The only time Musk took over during the live test was at a traffic light intersection.

The vehicle Musk was driving wanted tostraightwaited for a while for the red light, and whenLeft Turn SignalWhen it turned green, the car actually started and Musk immediately took over.

In addition to these two major issues, FSD still has some details that need to be improved.

For example, the SR depiction on the car screen does not show a roadblock ice cream cone, nor does it indicate the direction to turn.

There is also the problem of identifying animals. I saw a bird but was unable to identify it.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

The function of identifying animals is actually quite important. He Xiaopeng said that XNGP will do animal identification in the future.

When driving in China, you will indeed encounter some outrageous animals.

The video circulating online shows that the familiar “Xiao Yi” met on the road one night…An elephant.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

Of course, compared with kittens and puppies, it is easier to judge large animals like elephants by the idea of ​​occupying the network. All they need to do is stop them.

In addition to the above-mentioned issues, the evaluation also revealed the localization concerns of FSD after its introduction into China.

It can be divided into two scenarios: highway and urban area.

First, regarding highways, He Xiaopeng said that highways in the United States are more complicated than those in China because the speeds on highways are faster.

Therefore, after FSD is implemented, the high-speed driving performance may be good.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

However, the highways in China and the United States are slightly different. FSD runs very fast on the highway, including turning wide bends, but it may not be so aggressive in China.

Moreover, there is no ETC on highways in the United States, which also poses new challenges for FSD after it is introduced into China.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

As for urban areas, He Xiaopeng believes that domestic urban roads are 10 times more complicated than those in the United States.

Some sections of the road are indeed very empty, and it is impossible for things like electric scooters, Laotoules, and roadside starch stalls to exist there.

So some netizens who watched the video also said that He Xiaopeng let FSD off the hook and “did not push FSD to its limit.”

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

Therefore, it is still unknown how FSD will handle and optimize special scenarios after its introduction into China, and whether it can be invincible at its current level.

He Xiaopeng's ten-day trip to the United States, in addition to evaluating FSD, he also experienced the world's top three autonomous drivingWaymoof unmanned vehicles.

The new experience also made He Xiaopeng more resolute in his past judgment and revealed new thoughts.

He Xiaopeng: Future AI drivers should be like this

After He Xiaopeng took a ride on a Waymo driverless car and compared it with FSD, he thought that the Waymo driverless car performed better in the urban area of ​​San Francisco.

However, considering that Waymo's driverless cars use high-precision maps, this cannot directly prove that Waymo's capabilities are stronger than FSD.

In addition to comparing FSD, He Xiaopeng has another purpose for beating Waymo's driverless car:

Benchmarking.

He Xiaopeng previously stated in the 2023 annual earnings conference call that he had set a goal for the team that in key areas such as Beijing, Shanghai, Guangzhou and Shenzhen, the travel experience of XNGP should be comparable to that of Waymo driverless cars.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

After evaluating Waymo's driverless car and Tesla's FSD, He Xiaopeng expressed his feelings on Weibo and reiterated that he believes2025will beChatGPT moment of fully autonomous driving.

He Xiaopeng made a similar judgment in a financial report conference call last year. Considering the development of large models at the time, he believed that 2025 would be more likely to be the “iPhone 4 moment for fully autonomous driving.”

This is also related to Pony.ai’s co-founder and CTOMaster LouThe time of judgment coincides.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

So what will intelligent driving look like in the future?

fromTechnology ParadigmHe Xiaopeng believes that if intelligent driving is to be “usable all over the world” and smooth, it must be based on the “AI Model”, that is, end-to-end.

fromExperienceGenerally speaking, intelligent driving needs to be improved in two directions.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

It is necessary to learn the driver's “historical habits”, that is, AI should imitate the driver's driving style, and AI drivers should be different for each person.

It must also be able to “interact in real time” with human drivers. For example, if the AI ​​driver drives too fast or too aggressively on a certain section of road, the person sitting in the car can complain and the intelligent driving will drive slower.

One More Thing

Just as He Xiaopeng emphasized that end-to-end is the future and drew people's attention to FSD, a piece of news once again sparked heated discussion.

Long-term tracking and mining of Tesla software updates, people claim“Green God”HackergreenIt is said that Tesla added back the 300,000 lines of rule code that were deleted.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

So, is FSD end-to-end now?

Judging from various evaluations, FSD V12 is indeed a big improvement over V11. He Xiaopeng himself also lamented that he did not expect FSD to make such great progress in just a few months.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

Regarding this issue, the founder of a smart driving supplier previously speculated that Tesla may indeed still have rule codes that have not been completely eliminated.

But this part of the code may be used to run high-speed NOA. After all, there are fewer corner cases on the highway. It was developed based on rules in the past and the handling was also good.

The end-to-end architecture is used to handle more complex urban roads.

He Xiaopeng helps Musk debug: FSD drives randomly without navigation

There are also autonomous driving practitioners who agree with this, as they have just won a good ranking in an end-to-end international competition.

The practitioner believes that having an extra set of codes also means an extra layer of system redundancy.

Therefore, according to practitioners' speculation, adding back the rule code may be true, and moving to end-to-end is also true, and the two do not conflict.

If it is not a change in the technological paradigm, then what is the reason for the improvement in FSD capabilities?

What do you think?