2021 can be described as the first year of lidar mass production, in which many brands such as Weilai, Xiaopeng, Eian, and Great Wall have announced or listed products equipped with lidar.
In 2022, lidar has become a highly discussed topic in the field of new energy.
All kinds of signs show that in the rivers and lakes of autonomous driving today, there seems to be a laser faction emerging.
But there are other ideas about the future of autonomous driving. In April 2019, Tesla's head Musk said:
"Lidar is like a person with a bunch of appendixes, and the existence of the appendix itself is meaningless."
In April 2022, Woven Planet, a subsidiary of Toyota Motor Corporation, said:
"In future assisted driving and higher-level autonomous driving projects, we will use a single vision solution for development."
On the technical path of autonomous driving, Tesla, which represents the future, and Toyota, which symbolizes conservatism, have unexpectedly come together and gone farther and farther down the road of visualism.
What is the essential difference between laser and visual school? Where does the future of autonomous driving go? And the kings listen to me slowly.
The essential difference in basic logic
To understand the essential difference between laser and visual, we still have to return to the basic logic of automatic driving.
As a driver, our driving behavior can often be broken down into the following steps: eye observation, brain response, and finally to limb movements.
The process of autonomous driving is similar to that of us humans driving a car, and can also be divided into three modules:
Perception, decision-making, execution.
The focus of the debate between the laser school and the visual school is that in the construction of the automatic driving system, who is the real core of perception and decision-making?
Laser pie believes that the core module of autonomous driving should be perception, and their logic is:
Relying on the sensor matrix with lidar as the core to collect three-dimensional physical data, building a scene model and performing AI calculus on the basis of physical data, the real physical data is the core of the system, which is a set of bottom-up logic.
Unlike the laser school, the vision school believes that the core module of the automatic driving system should be decision-making. Their logic is:
Similar to human driving, relying on camera-based sensor groups to collect information, and then AI converts two-dimensional images into virtual three-dimensional models, AI decision-making and algorithm are the core of the system, which is a set of top-down logic.
After figuring out the essential difference between the laser school and the visual school, the advantages and disadvantages of the two in the actual application scenario are also very clear.
The technical advantages and disadvantages of visual and laser schools
The visual path has been used earlier in the field of autonomous driving, and the relevant technical characteristics have been agreed upon in the industry.
Since the core of the visual path is AI construction, whether it is converting a two-dimensional image into a three-dimensional model or generating behavioral decisions based on a virtual three-dimensional model, there must be a sufficiently intelligent AI.
With the current technical means, only a large amount of scene training can improve the intelligence of AI, but this is also the dead hole of the visual path:
After all, AI is not a human being, can only be based on the existing road experience to make decisions, once there is a scene that has not been learned, it is likely to be unable to judge the scene situation and lead to security risks.
Tesla famous scene
The biggest advantage of the visual path over the laser path is undoubtedly the cost.
First of all, at the sensor level, the cost of the sensor group based on the camera will be lower. On the issue of AI training, relying on the post-training of a large number of users can also achieve an effective transfer of costs, such as the "shadow" mode used on Tesla models, which can learn the driving habits of car owners independently.
The laser path, as a product of a later appearance, is more secure than the visual path.
We can think of lidar as a virtual bumper in an autonomous driving system, and the three-dimensional model built based on real physical data is more reliable; if we encounter unconventional situations that AI has not learned, the safety calibration of the system can also be biased towards physical data, so the laser path can theoretically provide better safety performance.
But the biggest problem with the laser path at present is the high price.
Early lidar products, such as Velodyne's related products released in 2017, have a unit price of up to $75,000; from 2008 to 2020, the price of related products is still stuck at about $1,000; in 2021, although Huawei has officially announced that the price of related products will be controlled at $500-200, it has not yet achieved mass production.
Where is the road to autonomous driving?
Regarding the path choice of automatic driving, in fact, it is not a proposition that fish and bear paws cannot be combined.
Based on the current downward trend of product prices from suppliers such as Huawei, we have reason to believe that the cost of lidar will gradually be accepted by OEMs in the future, and the listing of products such as Xiaopeng P5 also verifies this from the side.
Tesla and other visual factions have advantages in AI algorithms, and if the cost is reasonable enough, it is not impossible to "turn against the enemy" in order to strengthen system security.
In addition to the subjective will of the OEMs, the macro environment is also changing: the large transportation system built based on subsystems such as high-precision maps, intelligent roads, and V2X interconnection systems is the only way for future infrastructure construction, and the demand for underlying road data will surge under this trend.
For OEMs, the road information obtained by sensors such as lidar can not only generate greater commercial value in the future, but is more likely to be integrated into the big data system in the management of relevant state departments.
The so-called world will be divided for a long time, and it will be divided for a long time, and there may be no more disputes between laser school and visual school in the future autonomous driving jianghu.
Conclusion: In today's China, we not only have the world's largest new energy market, but also have many of the most technologically advanced and cost-effective parts suppliers.
And our independent brands, in recent years, they have adhered to the AI full-stack self-development and product intelligence improvement, making the current market acceptance of automatic driving technology more and more high.
Under the dual catalysis of the market and technology, autonomous driving technology in China may have a broader future.