Per reporter: Sun Lei Per editor: Pei Jianru
A few days ago, Toyota Motor's subsidiary Woven Planet said that it will use a single visual solution to develop autonomous driving in assisted driving and higher-level autonomous driving projects, that is, to collect data through relatively low-cost cameras without using expensive sensors such as lidar, and promote autonomous driving technology. At this point, Toyota motor has also joined tesla's camp of promoting pure vision system autonomous driving technology.
It is worth mentioning that Toyota Motor is the earliest car company to propose the use of lidar, and Volkswagen, Daimler, Bosch, ZF, Waymo and other companies are also fans of the "lidar camp". But as Toyota joins Tesla's "visual camp", the industry has also begun to think about why Toyota has jumped out of the "lidar camp" proposed by itself, and what considerations it has for the transformation of autonomous driving technology research and development.
Image source: Daily Economic News Infographic
Visual schemes help reduce costs
At present, there are roughly two paths for automatic driving perception technology: one is a solution with machine vision as the core and the use of "millimeter wave radar + camera", typical representatives of enterprises such as Tesla, Baidu Apollo, etc.; the other is a sensor route with "high-precision map + lidar" as the core, representing huawei, Xiaoma Zhixing and so on.
In fact, Tesla also used a lidar solution, but due to its difficult cost to reduce, Tesla eventually turned to the visual technology solution. Tesla CEO Musk has also publicly shelled many times: "The radar solution on the vehicle is extremely stupid, and anyone or company that uses the lidar solution will eventually fail, and they will go to great lengths to use these expensive sensors as a problem in itself." ”
Musk said Tesla believes that automatic assisted driving can be achieved entirely through visual neural networks, because people are visual neural network drivers in the biological sense, and humans can do it, and AI drivers should also be able to do it. The excess sensors are actually "crutches", which increases the complexity of the system and the occupation of resources.
Image source: Visual China
According to semiconductor company ON Semiconductor, L2 autonomous driving costs $40 per vehicle camera, while L3 autonomous driving costs $185 per vehicle camera. In contrast, in the case of the continuous reduction in the cost of lidar, its current price is much higher than that of the camera. The data shows that a Luminar Iris lidar costs close to $1,000.
In an interview with Reuters, Woven Planet said it was able to use low-cost cameras to collect data and train autonomous driving systems effectively, just as effectively as using expensive sensors. The team at Woven Planet sees this as a "breakthrough" that could help reduce costs and better expand Toyota's self-driving technology.
It is reported that the cameras used by Woven Planet are 90% cheaper than the sensors previously used, and can be easily installed on vehicles. Woven Planet believes that using a large number of cars to collect different driving data is crucial to developing a robust self-driving car system, but testing self-driving cars using expensive sensors is too expensive and cannot be tested at scale.
"Requires a lot of data"
For self-driving vision technology solutions, their deep learning algorithms rely on a large amount of training data, which helps them get better over time. Currently, Tesla has been using cameras to collect data on more than 1 million cars driving on the road to develop its self-driving technology.
"We need a lot of data. The data collected from a small number of high-cost autonomous vehicles is completely insufficient. We want to unleash the advantage of big automakers like Toyota, which is to get a huge data corpus. Michael Benisch, vice president of engineering at Woven Planet, said in an interview that low-cost cameras will be installed on more vehicles, which in turn will enable the collection of more data.
In Michael Benisch's view, collecting data through cameras could also help give Woven Planet a better understanding of human driving patterns. Using visual positioning technology, the system can better track the true trajectory of human drivers when turning or driving in the lane, and improve accuracy.
However, the solution of autonomous driving vision technology is controversial in the industry. Cameras are susceptible to light, weather, and other factors, and detection accuracy may be compromised. Su Zhen, the former head of Huawei's autopilot business, previously publicly stated that the data quality and dimension of Tesla's pure visual perception scheme may not be enough, and the room for progress of the matching system and algorithm will be limited.
Michael Benisch acknowledges this as well. But he believes that with the development of technology, camera-based autonomous driving technology may surpass some self-driving technologies that use sensors such as lidar and millimeter-wave radar. "But when and how long it will take to reach a safe and reliable level remains unknown." He also stressed.
It is worth mentioning that Toyota will not give up exploring other sensor solutions. For commercial autonomous vehicles, such as the Robotaxi, which carries passengers, Toyota plans to use an integrated sensor suite on the vehicle. These vehicles will use common set of sensors, namely cameras, lidar and radar. Woven Planet believes that this is the best and safest way to develop Robotaxi at scale.
Daily economic news