News 147 Comments

Musk's Autonomous Driving: A Rehash of 2-Year-Old Radish Run?

"The autonomous driving technology of the two cars released by Musk did not achieve a leading position."

At 10:00 AM Beijing time on October 11th, Tesla's "Robotaxi Day" press conference was held in California, with the slogan "We, Robot," announcing to the world's technology media and enthusiasts that "the future has arrived."

Previously, Musk regarded this press conference as one that "will go down in history," believing it to be one of Tesla's most important moments since the Model 3.

As he anticipated, after the press conference, #TeslaPressConference# topped the hot lists of various content platforms, and the comment sections on social media platforms were filled with praise from netizens saying "the future has arrived" and "Musk is a time traveler."

What kind of press conference could cause such a big response?

What content was condensed in the short press conference?

At 10:00 AM on October 11th, countless content platform accounts were live streaming this press conference, with TikTok's live broadcast showing 100,000+ online viewers (later data showed a peak of 330,000 people online at the same time). However, under everyone's anticipation, Musk was a full hour late.

This delay seemed to be deliberately arranged. Because in 2016, Musk had said that it would be "only two years" until fully autonomous driving cars were realized; in 2018, he postponed it, saying "there is still one year left"; in 2019, he said "full self-driving will be achieved within the year"...

Advertisement

Until 10:54 AM, after the host's opening remarks, Musk walked into the futuristic silver self-driving taxi—"Cybercab"—with a brisk pace, not saying a word.

The streamlined body, stainless steel outer enclosure, and sharp minimalist sense... Walter Isaacson once recorded a sentence from Musk in his biography "Elon Musk"—"No side mirrors, no pedals, no steering wheel."Now that this statement has truly been fulfilled at the press conference, it predictably ignited the most enthusiastic applause and cheers from the live audience.

Previously, Musk had said, "When one of the vehicles appears on a street corner, people will think they have seen a creation from the future."

Judging from the on-site reaction, Musk deliberately chose the Warner Brothers film studio for the press conference, using the Hollywood backlot streets to showcase the futuristic landscape of his products. The silver silhouette of the vehicle moving through the street corner indeed lived up to expectations.

And before the shock from the Cybercab, which resembles the time-traveling sports car from the movie "Back to the Future," had subsided, an unmanned cargo truck resembling Iron Man's mask—Robovan—slowly drove in, with 20 people successively stepping down from Robovan, pushing the scene to a climax, as if bringing the bus from the movie "I, Robot" filled with robots into reality.

Subsequently, Musk's Optimus (Optimus Prime) robots marched out in neat formation, smoothly dancing on the scene and even mixing drinks for guests.

This scene suddenly explained why Musk set the theme slogan of this press conference as "We, Robot."

American writer Asimov named his short story collection "I, Robot" to highlight the various issues between humans and robots in the future. It seems Musk is determined to lead the future into the present.

The press conference ended earlier than many people expected—at 10:54, Musk officially appeared in a Robotaxi, and the press conference ended at 11:40, lasting only a little over half an hour.

However, the content displayed in this half-hour was enough to cause an impact.The press conference showcased a total of three Tesla products: the autonomous taxi "Cybercab," the autonomous freight vehicle "Robovan," and the humanoid robot "Optimus."

▶▷The autonomous taxi Cybercab, besides lacking a steering wheel or pedals, features doors that open upwards like butterfly wings. The cabin is quite small, only able to accommodate two passengers.

More notably, the vehicle has no plugs and charges wirelessly through induction charging. It still uses a pure vision FSD system and does not employ any lidar (which Musk refers to as expensive equipment).

In Musk's plan, the Cybercab is set to go into production in 2026, at the latest by 2027, but users can experience autonomous taxis ahead of time with Tesla's current models equipped with FSD.

This time, he promoted the potential for the cost of using Cybercab to be reduced to around $0.2 per mile in the future, significantly lower than the current transportation cost of about $1 per mile, and the vehicle cost will be below the $30,000 level.

The business model that the low-cost Cybercab can bring to mind, Musk said, is that once cars achieve full autonomy, they can increase usage by 5 to 10 times, with weekly usage time increasing from 10 to 100 hours, breaking through to 100 hours and even reaching over 160 hours, and even one person could manage a large-scale fleet.

▶▷The autonomous freight vehicle "Robovan" has an appearance that looks more like a streamlined large business vehicle or a single-section small train, capable of carrying 20 people or transporting goods, with travel costs as low as just 10 cents per mile.

Regarding the design of Robovan, Musk said, "The future should look like the future."

▶▷The humanoid robot "Optimus" is proclaimed by Musk to play many roles in the future, such as a tutor, pet litter scooper, home nanny, cyber bartender, etc., and even the price of Optimus will further decrease in the future, possibly ranging from $20,000 to $30,000 (approximately 140,000 to 200,000 RMB), equivalent to the price of a new energy vehicle.Two Paths for Autonomous Driving in China and the US

After the press conference, two starkly different market reactions emerged — ordinary people were elated, while industry professionals and investors remained skeptical.

The elation of the general public might stem from the fact that Musk, through Tesla's three products, seemed to bring the future within reach at the press conference.

On the other hand, the skepticism of industry professionals and investors might lie in the fact that the autonomous driving taxis showcased at the press conference were operated in highly controlled environments. They only demonstrated performance in specific scenarios but did not provide any technical details.

For example, the safety issue, which is of utmost concern to the public, was not mentioned at all during the press conference.

Professor Zhu Xichuan from Tongji University said: "If autonomous driving cannot solve the last long-tail challenge — safety, the era of autonomous driving will never truly arrive." Hou Cong, co-founder of Qingzhou Zhihang, also expressed a similar view: "What if the car gets stuck in the middle of the road? Do you ask the owner to take a taxi over and rescue the car?"

In the testing and operation of autonomous driving's Robotaxi and autonomous buses, there is an important indicator called "takeover rate": when a long-tail extreme situation occurs and autonomous driving cannot handle it safely, humans need to take over the vehicle for processing.

A lower takeover rate means a higher level of autonomous driving. If an autonomous vehicle encounters a 0.1% long-tail situation and does not have an appropriate processing mechanism, it could lead to a serious incident.

Autonomous driving is a safety issue and also a cost issue — how much cost users need to pay, including economic costs and safety and social costs.

Regarding solutions for cost and safety, two technical routes have emerged in this field — one is a vision perception plan that does not rely on high-precision maps, and the other is a fusion sensor plan based on high-precision maps and LiDAR.The levels of autonomous driving technology are divided into six levels: L0, L1, L2, L3, L4, and L5. Levels L1 to L3 are all considered assistive driving, requiring the driver to take over the system at any time, while L4 and L5 belong to the category of automated driving without the need for driver intervention.

As an autonomous taxi, Robotaxi naturally falls under the higher technical and safety requirements of Level L4. Two different approaches bring different cost and technical capability requirements for companies in the Robotaxi industry.

Companies that follow the "multi-sensor + high-precision map" technological route include Waymo internationally and Luobo Kuaibao, Didi Autonomous Driving, WeRide, and Pony.ai domestically. Companies that completely abandon lidar and high-precision maps in favor of a "pure vision" technological route include Tesla internationally and Xiaopeng Motors domestically, which is attempting to align with Tesla's 12th generation FSD (Full Self-Driving) system.

Currently, the "lidar + high-precision map" solution has higher hardware and computing power costs and is the mainstream choice for unmanned driving companies.

For pure vision solutions without high-precision maps, although the hardware cost per vehicle is generally two to three ten-thousand yuan lower than the former, the technical difficulty of solving long-tail problems encountered during vehicle operation is greater.

The large models of pure vision solutions require very large data samples to improve the growth rate of the model. Tesla's recently launched Full Self-Driving system FSD V12, which introduced end-to-end technology, has found a more spacious development direction for this route, transforming the autonomous driving industry from reliance on "genius engineers" to a competition of data and computing power.

*"End-to-end" refers to the input of information from the perception end, with the system directly outputting decisions, replacing previously man-made rules and modules with a large neural network model, and the model learns data in the form of high-quality human driving behavior video images.

The competition between these two routes has not yet seen a clear winner. However, although Tesla's press conference has boasted to the public a glimpse of the future, in the field of intelligent driving, Chinese consumers may be less easily surprised and may even become more demanding.

After all, there are already a group of leading players in the country, and many people have even experienced autonomous driving on the road.

So, for industry insiders and experts, what are the highlights of this press conference, and what are the levels of China and the United States in the field of autonomous driving? Let's take a look at the opinions of the big heads~At this press conference, the autonomous driving technology of the two cars released by Musk did not achieve a leading position.

In 2022, Baidu's sixth-generation Roborun had already possessed the same features - no steering wheel, no front and rear seats, and no pedals.

According to Musk, mass production is expected by 2026, striving to reduce the car cost to below $30,000. Comparing this with the price of the sixth-generation Roborun car, which is also around 200,000 yuan, Roborun has not only achieved mass production but has also launched trials in 13 cities across the country, accumulating more than 7 million orders, lagging behind domestic production.

From a commercialization perspective, Roborun officially registered its license plate in 2019, started trials around 2020, launched the sixth-generation car in 2022, and is now expanding the scope of the entire trial service. Currently, Roborun has achieved rolling profits and is gradually maturing into a closed-loop business model.

There are about hundreds of thousands of Roborun vehicles nationwide, with about two to three hundred backend control personnel. The vast majority of vehicles operate in autonomous driving mode, and emergency manual intervention is only required when an event that may threaten operational safety is judged to occur.

In contrast, Musk's two cars have not only failed to achieve commercial use but have not even achieved mass production.

In addition, domestic companies such as Pony.ai have also made good progress in the field of autonomous driving. For now, Musk's commercialization of autonomous driving technology has fallen behind Chinese companies.

Firstly, looking at the press conference, if judged solely by the videos released by Tesla, it is indeed very shocking. Eliminating the steering wheel and pedals means a truly driverless car.

However, in the past few years, consumers have seen too many industry cases where the promotion does not match the reality, so I believe that the judgment of this press conference needs to wait and let the market and consumers verify whether Musk's products can live up to the promotion as stated today.However, if the final outcome of its products truly lives up to the hype that Musk has been promoting, even if it only delivers 80% of what is promised, it would still be a milestone achievement. This kind of creation can truly liberate productivity and enhance production efficiency in related fields.

Secondly, analyzing the specific products, the autonomous taxis and cargo vehicles that Tesla showcased this time have similar counterparts in China. For instance, autonomous trucks are the simplest part of the process towards driverless technology because their application scenarios are relatively fixed. Heavy-duty trucks on highways, entering specific sections or parks, can achieve full-process automation, which is relatively simple from a technical standpoint.

As for autonomous taxis, whether it's Baidu's RoboTaxi in China, or WeRide, Pony.ai, or Didi's autonomous vehicles, from my personal experience, they have all reached the level of an experienced driver. I have experienced them multiple times without a single incident.

Regarding the humanoid robot "Optimus," if it can be combined with large models and understand people's real needs based on scene understanding, theoretically, Musk's claims of washing dishes, cooking, taking out the trash, and serving coffee can indeed be realized. It is possible that significant progress will be made within the next 5-10 years, and its forward-looking and practical aspects are worth looking forward to.

For Tesla's products to achieve true commercialization in the Chinese market, they must also face social and ethical issues.

At this press conference, I was hoping to see the technical direction of Tesla's FSD 13 version and the hardware configuration of the AI chip HW 5.0, but the conference did not introduce these aspects. Judging from Tesla's current FSD 12 version and AI chip HW 3.0, Tesla cannot achieve true autonomous driving.

The end-to-end model of Tesla's current FSD 12 version is still just a model with a parameter volume of 1 billion, using the Transformer architecture, and does not belong to a true large model. A so-called large model needs to be able to pre-train, generate content, and achieve analogical reasoning capabilities. Tesla's current end-to-end model is a small model, that is, a predictive AI model, which relies more on data and cannot solve the problem of edge scenarios that have not been seen before. It is equivalent to a car having only a cerebellum without a cerebrum, and thus it is impossible to truly solve safety issues.From this perspective, Tesla's recent press conference did not release any content of actual technological significance, and it is also difficult to establish investors' confidence in Tesla. The subsequent decline in the US stock market was the market's response to this.

Looking at the technical route, most domestic car companies and Tesla's technical routes are not consistent at present. Tesla's route does not represent the final result of the future.

For example, companies represented by Huawei have a technical route from perception to end-to-end. Domestic companies will not abandon lidar, nor will they abandon rule-based models. Just like in new energy vehicles, Tesla insists on pure electric, while domestic new energy car companies have chosen to "walk on two legs" - pure electric and hybrid.

The technical routes in the field of autonomous driving between Chinese and American companies, since no one has obtained the final technical results, will still present not completely consistent technical routes.

Of course, Tesla's data-driven, user data closed-loop, end-to-end approach is being learned by domestic companies, but they will not take the completely end-to-end model route. Domestic car companies do more in the perception end-to-end model, and still add rule-based models in motion planning and decision-making.

The advantage is that the end-to-end model can improve the upper limit and increase traffic efficiency; the disadvantage is that if all the rule-based models, especially the rule-based models of motion planning, are canceled, bad data will enter the model.

The danger lies in the fact that there are many illegal dangerous driving behaviors in human driver data, and moderate illegal behaviors can actually improve traffic efficiency. Then, under the reward function of traffic efficiency, the model will be "corrupted" by bad data.

Finally, looking at the commercialization and landing level, the commercialization and landing of L2+ assisted driving city NOA, China will be faster than the United States.

In terms of truly achieving unmanned driving, first, the original innovation ability of Chinese companies is not as good as Tesla; second, the Chinese government's ability to manage innovation in new industry fields is not as good as the US government; third, China's commercial demand is not as good as the United States. It is true that the United States cannot find ride-hailing drivers, and its labor costs for car drivers are much higher than those in China. It is false that no one in China is willing to drive ride-hailing cars and trucks, and unmanned driving will take away many people's employment opportunities.

Therefore, the commercialization of unmanned driving may ultimately land faster in the United States.In the best-case scenario for autonomous driving in China in the future, it is not about replacing drivers, but rather about assisting to reduce the labor intensity for drivers. For instance, Huawei's ADS 3.0 intelligent driving system, which is considered a well-executed urban NOV (Navigation on Vehicle), is currently only available in vehicles priced above 300,000 yuan. Even with the newly launched MDC 510 intelligent driving system this year, which can be equipped in vehicles priced at around 250,000 yuan, it is still too expensive for ride-hailing drivers.

If the cost of urban NOV in the future can be reduced to a level where it can be equipped in vehicles around 150,000 yuan, then assistive driving will be able to greatly help ride-hailing drivers in the future, rather than replacing them completely with autonomous driving technology.

Leave A Comment