Why do Robotaxi companies choose to develop L2 technology in today’s environment? What challenges will they face with hardware maturity, data reuse, and supply chain management? How do other industry players see this?
With the departure of Waymo’s former CEO John Krafcik at the beginning of April, and after Huawei’s first self-driving car debut and Su Qingtao’s statement that “companies focusing on RoboTaxi at this stage will go bankrupt,” the progressive route has gradually become the consensus in the self-driving car industry, and the technology roadmap of Robotaxi companies has been questioned like never before.
Critics of Robotaxi companies are not just practitioners and investors of the progressive route, but also some media commentators who have neither personally participated in self-driving technology development nor engaged in direct communication with self-driving car companies. In the eyes of these armchair analysts, Robotaxi companies are seen as “foolish.”
However, is this really the case?
After in-depth discussions with executives of many Robotaxi companies at the end of April, Jiuzhang Intelligence found that the leaders of these companies have a much more thorough understanding of the progressive route and the future prospects of their own companies than outsiders.
Previously, Robotaxi companies believed that developing L2 technology is actually more difficult than L4, but now these companies believe that the time is ripe for them to develop L2 technology.
Car manufacturers need L2 more than L4
On April 28th, during a media communication meeting at Pony.ai, Jiuzhang Intelligence asked Peng Jun a question: In the context of the increasing doubts about the direct development of L4 technology, has Pony.ai considered developing L2 technology and shifting towards the progressive route?
Peng Jun’s reply was that the classification of L2 and L4 is actually a concept from four or five years ago. Today, the gap between L2, L3, and L4 is getting smaller and smaller. We focus on the self-driving technology itself, rather than its classification.
Prior to this, Jiuzhang Intelligence also asked the same question to Mo Luyi, the general manager of Pony.ai Guangzhou, and his answer was similar to Peng Jun’s: “We are developing self-driving technology itself, and we are exploring all possible scenarios for its application in the future.”
At least we can understand that Pony.ai has not explicitly denied the possibility of developing L2 technology.
At this point, we have to consider such a question: What makes Pony.ai attractive to Toyota’s investment? Or, why does Toyota invest in Pony.ai? Is Toyota’s motivation just to sell a few hundred or thousand cars to Pony.ai for later modification into Robotaxi?
In Toyota’s values, full autonomous driving is not a must-have; their main purpose of developing autonomous driving is to assist rather than replace humans. As a result, Toyota’s focus is on Level 2 instead of Level 4 (which includes so-called Level 3 or Navigation Assisted Driving). Although Toyota isn’t short of money, they certainly hope that the 400 million US dollars investment in Pony.ai can be put to good use, and that good use is the Level 2 project. Before accepting Toyota’s investment, both sides must have reached a consensus on future cooperation in the Level 2 business, or at least have established implicit understanding.
In fact, not only Toyota, but also most rational traditional car companies (passenger car companies) do not consider Level 4 as a top priority, nor is it a particularly important matter. What they truly need for a considerable period of time is Level 2, which can help increase car sales through front-loaded technology, rather than only exist in a B2C mode and difficult to scale up and produce in volume like Level 4.
(Note: Level 2, which includes Navigation Assisted Driving and so-called Level 3, is what is referred to here. As according to the policy of the Ministry of Industry and Information Technology and the latest standard of the SAE, many car companies’ Level 2.9 to Level 3.9, in fact, are just Level 2.)
Therefore, applying the same logic, Nissan, Renault, Mitsubishi investing in WeRide, SAIC, Dongfeng investing in AutoX, and Daimler investing in Momenta, are also primarily targeting level 2 rather than level 4.
Opportunities for a “high-end hardware” level 2 due to low-cost hardware
The vice president of Robotaxi Company A said that they had previously rejected some OEM’s proposals for “downgrading to Level 2”. The reason was that their algorithm was written based on “Level 4 hardware”, and under current conditions, production cars of Level 2 could hardly afford chips and sensors that support Level 4, meaning that, if they were to do Level 2, they would also need to rewrite the algorithm; otherwise, there would be a problem of “small horse-drawn big car”.
Here, “small horse” refers to the hardware of Level 2, and “big car” refers to the algorithm of Level 4. In fact, the biggest challenge that Level 4 companies face when they “downgrade” to Level 2 is the problem of “small horse-drawn big car”. It is precisely for this reason that Level 4 companies have been hesitant about whether to “downgrade” and do Level 2 front-loaded projects for quite some time.
Before, these Robotaxi companies thought that they should focus on deepening Level 4 technology first, and then explore other business models. But now, they realize that perhaps the time has come to “downgrade” and do Level 2.The vice president of Company A stated, “The pricing of the models developed through the collaboration between Huawei and BAIC BJEV indicates that the hardware costs for the L4 level, including calculation platforms and sensors, have been controlled to around 200,000 yuan per set. Additionally, as the scale increases, the hardware prices will continue to decline. This allows us to see the possibility of ‘high-end hardware’ on L2 production cars.”
“High-end hardware” refers to applying the hardware originally only used on L4 test vehicles to L2 production cars.
For Robotaxi, this means that they do not have to modify the algorithm architecture when doing L2 production projects. As a result, technological transfer and adaptation efficiency can be significantly improved. “Now, asking me to do L2 work is too easy. With the computational power and sensors of L4, I can just put my L4 algorithm on top and do L2 work, couldn’t I?” said the vice president.
The vice president also admitted that their hardware costs are slightly higher than Huawei’s at present, but added, “Lowering hardware costs is not my main responsibility since as long as the OEM has batch orders, the supply chain costs will naturally come down quickly. Therefore, I do not believe that hardware cost is our main barrier to future development.”
In this sense, Huawei’s powerful appearance in autonomous driving and the “gradual upgrade route” becoming consensus is not a disaster for Robotaxi company, but rather an opportunity to hitch a ride on the industrial dividend.
The Prerequisite for Connecting L2 and L4 Data
For friends who are more concerned about data issues, they may have thought about this question: many car companies or Tier 1 suppliers claim that their autonomous driving is following the “gradual upgrade route,” iterating gradually from L2 to L4. So, does the road data collected during L2 have any use for training L4 algorithms?
This is a good question, but not rigorous enough – strictly speaking, the difference between L4 and L2 mainly lies in the decision-making algorithms, not in the sensor and computational power configuration, so the data may not be very different.
However, in practice, for cost considerations, most OEMs equip L2 production cars with sensors and computational power that are much lower than L4. A typical difference is that L2 cars usually do not have LiDAR and have fewer millimeter-wave radars. As a result, there is still a considerable gap between the data collected on L2 production cars and the requirements of L4 algorithms.
The chief development vice president of Robotaxi Company B said:
Usually, the data for L4 autonomous vehicles should include spatial information, texture information, and color information about the surrounding environment. The richness of this information is basically consistent with that obtained by humans, so even if the initial algorithm is poor, all of the data can help us create an algorithm platform in later stages.
In contrast, the data collected on L2 mass-produced vehicles, due to their relatively low sensor configuration (excluding LiDAR), only includes texture and color information about the surrounding environment, lacking spatial information. These data have little significance for training the L4 algorithm in the future.
Therefore, the completeness of information is a crucial indicator for measuring data quality, and it basically determines how high your ceiling is.
Over the past year, the author has consulted many L4 autonomous vehicle and automotive industry leaders as well as algorithm engineers on whether data collected from L2 vehicles can be used to train the L4 algorithm. The respondents have reached a consensus:
The sensor architectures of L4 and L2 vehicles must be consistent, which is a necessary condition for “connecting the data” of both vehicles. If L2 does not have a LiDAR while L4 requires one, the data is difficult to re-use for algorithm training without “translation,” which incurs high costs and is technically challenging.
There are two mainstream approaches to ensure “consistent sensor architecture”: sticking to the end, striving to achieve L4 without LiDAR (such as Tesla); or equipping L2 vehicles with LiDAR (as new Chinese automakers have done with “pre-installing hardware”).
For Momenta, which tries to develop both L2 and L4 lines and connect the data of both, the lack of LiDAR on L2 vehicles while LiDAR is equipped on L4 vehicles, as well as the addition of safety redundancy systems, make it difficult to completely connect the data. However, according to Cao Xudong’s previous introduction, “the difficulties caused by inconsistent sensors can be overcome, and currently about 80% of the data on L2 vehicles can be reused for L4 vehicles.”
However, as the price of LiDAR continues to fall, using higher-spec sensor configurations in the front-loading L2 solution may be a better choice.
Considering that the optional rate of LiDAR for the XPeng P5 is over 70%, and the ideal may make autonomous driving a standard feature, Robotaxi should choose a model with sufficient brand power and product strength that can make LiDAR a standard feature for L2 mass-produced projects. This way, the accumulated data will be valuable for L2 development in the future, and the data scale will also be more secure.Of course, in the ideal situation, the architecture of the sensors should be consistent across L2 to L4, and the number of various sensors should also be consistent. For example, if an L4 vehicle requires three LiDARs, but the L2 vehicle only has one, this will create a significant amount of back-end data processing work.
However, for most companies, the challenge of integrating data between the two levels is that the L2 solution mainly focuses on high-speed roads and city expressways, while the L4 solution mainly focuses on city roads, and the scenes and data are inconsistent. Therefore, to integrate the data, L2 needs to expand its coverage to city areas, while L4 needs to expand its coverage to high-speed scenarios (Momenta has been doing this already).
Can’t Manage the Supply Chain? Consider Collaboration with Tier 1
Of course, Robotaxi companies will not be able to produce L2 mass production projects overnight. When I talked to the CEO of a certain Robotaxi company about this issue in December last year, he pointed out that if they were to do L2, the supply chain management capability requirement would be higher than that of L4.
This is because, when doing L4, due to the relatively small size of the test fleet, they can purchase chips and sensors at a much higher cost. However, when doing L2, the role of an autonomous driving start-up company is usually as a Tier 1 or even Tier 2 supplier, and they need to provide a complete hardware and software solution to automakers or Tier 1 suppliers. At this point, not only performance but also cost and the ability to provide a stable supply are key considerations.
Typically, autonomous driving start-up companies do not have the bargaining power with suppliers. If your solution is good but has no cost advantage and delivery cannot be guaranteed, will automakers use it?
Also, at present, companies that directly do L4 are not required to go through the vehicle inspection process for key components. They only need to purchase a few vehicles and modify them for testing on the road. However, L2 solutions for mass production need to comply with a series of standards in the automotive industry, and it may take 2-3 years from securing orders to R&D and testing.
Perhaps, when Robotaxi companies are doing L2 mass production projects, they can consider “making the most of their strengths and avoiding their weaknesses,” that is to say, providing only the technical solutions themselves and leaving the supply chain integration and vehicle inspection to automakers or traditional Tier 1 suppliers?
Recently, when asked about this speculation about “how to best do mass production of L2 projects,” the COO’s response was, “Automakers and Tier 1 suppliers are both exploring this issue with us. Overall, our strategy will be more flexible rather than being limited to any one approach.”In the first half of this year, Bosch appeared on the investor list for Momenta’s Series C financing round. According to many industry insiders, Momenta is developing autonomous driving algorithms for Bosch, and Bosch is responsible for hardware. This confirms the speculation of “Nine Chapters Intelligent Driving” earlier.
Of course, apart from this, the extent to which OEMs or Tier 1 companies can open their chassis data and their degree of cooperation with Robotaxi companies is also a worthy question.
Xiao Ma Zhixing and Wenyuan’s Robotaxi ride experience is excellent, and one reason may be that Toyota and Nissan open their chassis data to Xiao Ma and Wenyuan to a greater extent, enabling them to fully optimize their vehicle chassis system.
Currently, many autonomous driving companies are struggling with perception algorithms, decision-making algorithms, and the importance of coordination with chassis manufacturers, which may not have received enough attention, and their value may even be underestimated. However, when perception and decision-making algorithms improve to a certain level, the quality of the chassis control system may become a key differentiating factor affecting the riding experience of various Robotaxi and L2 passenger cars.
Therefore, when Robotaxi chooses front-loading mass production partners, it may need to negotiate with the other party on the issue of opening up chassis data.
How do the startup Tier 1 companies view Robotaxi’s development in L2?
Of course, many startups have been developing L2 solutions for automakers long before Robotaxi decided to enter the L2 market. In addition to Momenta, there are also NIO Technologies, founded by founders with Tesla background; Zhixing Technology with traditional Tier 1 background; Foryou + Technology with OEM and Tier 1 background; and JiMu Intelligence, all of which are positioned as Tier 1 in the industrial chain.
NIO Technologies’ approach is actually similar to Momenta’s, which is “top-down.” Zhixing, Foryou +, and JiMu Intelligence are gradually moving up from L2, and even from L1. So how do these companies view the impact that Robotaxi’s development in L2 may bring?
A senior engineer from a certain company said:
“In the short term, their development of L2 will not affect us too much. Because our approaches are completely different. They are going down from L4, only in terms of functionality (transferring driving responsibility to the driver), but the algorithm won’t change much, and the hardware configuration won’t be reduced too much – their hardware architecture has been fixed. If the hardware is reduced, the workload of algorithm modification will be huge.””As a translator in the automotive industry, I am responsible for English translation, spelling proofreading, and wording improvement. To ensure the accuracy of the content, I will only provide corrections and improvements to the following Markdown text, without further explanation:
`”We cannot predict the long-term outlook, but currently, we are responding to customer demands by moving towards L3 and L4. However, our perspectives differ: they emphasize more on redundancy and safety, while we prioritize cost-effectiveness.”
The engineer stated that they have been researching for a long time, but still do not understand how to implement Momenta’s plan of “training L4 algorithms with L2 data.” He added, “L2 data used for L4 lacks many dimensions.”
During a casual conversation, the engineer mentioned an interesting point: traditional Tier 1 or OEM staff have weaker “data thinking” compared to those with AI backgrounds.
At the time, I asked a question: “If the route from L2 to L4 iteration is easy to achieve, then theoretically, shouldn’t traditional Tier 1 companies like Bosch who have the most ADAS shipments have the most data?”
The engineer responded, “These traditional Tier 1 companies lack data thinking and have not yet realized the value of data. Therefore, they have not equipped their ADAS systems with a T-box to transmit data. In contrast, companies like Momenta are inherently equipped with internet and data thinking. Our data thinking is, I can accept not making money, but I have to get the data.”
Furthermore, the engineer said, “Most of our company leaders come from traditional Tier 1 or OEMs and lack data thinking. Therefore, we have not spent much effort collecting data but rather focused on how to create products, scale quickly, and ensure reliability.”
From the comparison above, it is clear that these companies mentioned are not competitors of Robotaxi from their standpoint.
Of course, at this stage, it is difficult to evaluate who is truly stronger or weaker based merely on valuation or orders. Everyone is in the process of making up for their deficiencies (some are focusing on improving hardware, while others are focused on their software and data), and it all depends on who can fill in the gaps quicker.”`
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.